C-ORAL-ROM
Studies in Corpus Linguistics Studies in Corpus Linguistics aims to provide insights into the way a corpus can be used, the type of findings that can be obtained, the possible applications of these findings as well as the theoretical changes that corpus work can bring into linguistics and language engineering. The main concern of SCL is to present findings based on, or related to, the cumulative effect of naturally occuring language and on the interpretation of frequency and distributional data. General Editor Elena Tognini-Bonelli Consulting Editor Wolfgang Teubert Advisory Board Michael Barlow
Graeme Kennedy
Rice University, Houston
Victoria University of Wellington
Robert de Beaugrande
Geoffrey Leech
Federal University of Minas Gerais
University of Lancaster
Douglas Biber
Anna Mauranen
North Arizona University
University of Tampere
Chris Butler
John Sinclair
University of Wales, Swansea
University of Birmingham
Sylviane Granger
Piet van Sterkenburg
University of Louvain
Institute for Dutch Lexicology, Leiden
M. A. K. Halliday
Michael Stubbs
University of Sydney
University of Trier
Stig Johansson
Jan Svartvik
Oslo University
University of Lund
Susan Hunston
H-Z. Yang
University of Birmingham
Jiao Tong University, Shanghai
Volume 15 C-ORAL-ROM: Integrated Reference Corpora for Spoken Romance Languages Edited by Emanuela Cresti and Massimo Moneglia
C-ORAL-ROM Integrated Reference Corpora for Spoken Romance Languages
Edited by
Emanuela Cresti Massimo Moneglia University of Florence
John Benjamins Publishing Company Amsterdam/Philadelphia
8
TM
The paper used in this publication meets the minimum requirements of American National Standard for Information Sciences – Permanence of Paper for Printed Library Materials, ansi z39.48-1984.
Cover illustration from original painting Random Order by Lorenzo Pezzatini, Florence, 1996.
Library of Congress Cataloging-in-Publication Data C-ORAL-ROM : integrated reference corpora for spoken Romance languages / edited by Emanuela Cresti and Massimo Moneglia. p. cm. (Studies in Corpus Linguistics, issn 1388–0373 ; v. 15) Includes bibliographical references and index. 1. Romance languages--Data processing. 2. Romance languages-Variation. 3. Romance languages--Spoken Romance languages. I. Cresti, E. (Emanuela) II. Moneglia, Massimo. III. Series. PC44.5.C2 2005 440’0285--dc22 isbn 90 272 2286 X (Eur.) / 1 58811 548 8 (US) (Hb; alk. paper)
2005041056
© 2005 of the book – John Benjamins B.V. © 2005 of the DVD – Università di Firenze; Universidad Autónoma de Madrid; Centro de Linguística da Universidade de Lisboa; L’Université de Provence. No part of the book and DVD may be reproduced in any form, by print, photoprint, microfilm, or any other means. John Benjamins Publishing Co. · P.O. Box 36224 · 1020 me Amsterdam · The Netherlands John Benjamins North America · P.O. Box 27519 · Philadelphia pa 19118-0519 · usa
Table of contents
Acknowledgements
xi
Preface
xv
Chapter 1 The C-ORAL-ROM resource Massimo Moneglia 1.1 Introduction 1 1.1.1 C-ORAL-ROM 1 1.1.2 Organisation of the volume 3 1.1.3 The issues of representation and comparability 4 1.1.4 C-ORAL-ROM sampling strategy 8 1.1.5 Comparing C-ORAL-ROM with other corpora 12 1.2 Prosodic tagging criteria 14 1.2.1 Prosodic breaks and utterance limits 15 1.2.2 Background of prosodic labelling 17 1.2.3 Utterance boundaries and labelling of discourse acts 19 1.2.4 Utterance limits in spontaneous speech 20 1.2.5 Prosodic labelling procedure and Alignment units 24 1.2.6 Conventions for prosodic tagging in the transcripts 25 1.2.6.1 Fragmentation phenomena 26 1.3 Textual format 27 1.3.1 Metadata 28 1.3.2 Dialogue representation 32 1.3.3 Conventions for transcription 33 1.3.4 Pauses 35 1.3.5 Restrictions on dialogue representation: The Intersection convention 36 1.3.6 Dependent lines 37 1.3.7 Alignment principle 38 1.3.8 Files, filename conventions and folder structure of the resource 1.3.8.1 Files 39 1.3.8.2 Filename conventions 39 1.3.8.3 Folder structure 40
1
39
Table of contents
1.4 WinPitch Corpus. A Text-to-Speech Analysis and Alignment Tool 40 Philippe Martin 1.4.1 Text-to-speech alignment 40 1.4.1.1 Automatic alignment 41 1.4.1.2 Alignment and transcription 41 1.4.1.3 Computer-assisted alignment 42 1.4.2 WinPitch Corpus features 43 1.4.3 Basic layout 46 1.5 C-ORAL-ROM PoS tagging 51 1.5.1 Minimal tagset requirements 52 1.5.2 Comparison between tagsets 52 1.5.2.1 PoS tagsets 52 1.5.2.2 Morpho-syntactic features of verbs 54 1.5.2.3 Non-standard tagsets 55 1.5.3 Tagging and frequency lists formats 56 1.6 Measurements of spoken language variability in the Romance languages 57 1.6.1 Mid-length of utterances (MLU) 58 1.6.2 Mid-length of the dialogic turn (MLTw) 59 1.6.3 Speed 60 1.6.4 Length of the tone unit (MLTone) 61 1.6.5 Fragmentation 62 1.6.6 Some conclusions 62 Chapter 2 The Italian corpus Emanuela Cresti, Alessandro Panunzi, and Antonietta Scarano 2.1 History of the corpus within the national framework 71 2.1.1 Historical overview 71 2.1.2 The LABLITA Corpus 75 2.2 Orthographic transcription 76 2.2.1 General criteria 76 2.2.2 Orthographic transcription of regional words 77 2.2.3 Diacritic marks 81 2.2.4 Interjections 84 2.3 Morpho-syntactic tagging 85 2.3.1 Tools and strategy adopted for automatic PoS tagging and lemmatisation 85 2.3.2 Tagset 86 2.3.2.1 Choices in the PoS tagset 88 2.3.3 Extended tagset for spoken language 92 2.3.3.1 Non-standard words tagset 93 2.3.3.2 Non-linguistic elements 93 2.3.4 Other choices 94
71
Table of contents
2.3.4.1 Regional and dialectal forms 94 2.3.4.2 Multiwords 94 2.3.4.3 Names 96 2.3.5 Evaluation 97 2.3.6 Specific problems with the morpho-syntactic tagging of spoken language 99 2.3.6.1 Words adjacent to utterance boundaries 101 2.3.6.2 Interruptions and retracting 101 2.3.6.3 PoS assignment in connection with secondary prosodic boundaries 102 2.4 Main data from lemmatisation 104 Chapter 3 The French corpus Estelle Campione, Jean Véronis, and José Deulofeu 3.1 History of the corpus within the national framework 111 3.2 Orthographic transcription 114 3.2.1 General criteria 114 3.2.2 Interjections 116 3.3 Morpho-syntactic tagging 116 3.3.1 Tagset 116 3.3.2 Multiword expressions 119 3.3.3 Tools and strategy adopted for automatic PoS tagging and lemmatisation 119 3.3.4 Evaluation 121 3.3.5 Main data from lemmatisation 123 Chapter 4 The Spanish corpus Antonio Moreno, Gillermo de la Madrid, Manuel Alcántara, Ana Gonzalez, José M. Guirao, and Raúl De la Torre 4.1 History of the corpus in the national framework 135 4.1.1 Historical overview 135 4.1.2 CORLEC features 137 4.1.3 C-ORAL-ROM features 138 4.1.4 Final remarks 140 4.2 Orthographic transcription 141 4.2.1 General criteria 141 4.2.2 Orthography for non-standard words 142 4.2.3 Interjections 142 4.3 Morpho-syntactic tagging 143 4.3.1 Tools and strategy adopted for automatic PoS tagging and lemmatisation 143
111
135
Table of contents
4.3.1.1 Electronic vocabulary 145 4.3.2 Tagset 146 4.3.2.1 Tagset adopted 146 4.3.3 The notion of “multiword” 147 4.3.4 Ambiguous clustering 148 4.3.5 Level of morpho-syntactic encoding of forms 149 4.3.6 Evaluation 151 4.3.7 Specific tagging problems with the Spanish spoken corpus 153 4.3.7.1 Retracting and interruption phenomena 153 4.3.7.2 Linguistic forms whose distribution is not consistent with the distributional characters of written language 154 4.3.7.3 Linguistic forms and non-standard forms used as discourse markers 156 4.3.8 Main data from lemmatisation 156 Chapter 5 The Portuguese corpus Maria Fernanda Bacelar do Nascimento, José Bettencourt Gonçalves, Rita Veloso, Sandra Antunes, Florbela Barreto, and Raquel Amaro 5.1 History of the corpus within the national framework 163 5.1.1 Historical overview 163 5.1.2 Reusing materials from existing databases at CLUL 168 5.1.3 New materials specifically collected and/or transcribed for the C-ORAL-ROM project 170 5.1.4 Final remarks 171 5.2 Orthographic transcription 172 5.2.1 Specific Portuguese conventions 172 5.2.1.1 General orthographic norms 172 5.2.1.2 Other specific conventions 173 5.2.2 Interjections 174 5.3 Morpho-syntactic tagging 175 5.3.1 Tools and strategy adopted for automatic PoS tagging and lemmatisation 175 5.3.2 Tagset 177 5.3.2.1 Categories and main options 177 5.3.2.2 Particular cases 182 5.3.2.3 Specific tags for the Portuguese spoken corpus 183 5.3.3 Lemmatisation of the spoken corpus 184 5.3.3.1 Specific lemmatisation choices 185 5.3.4 Evaluation 186 5.3.5 Specific problems with the morpho-syntactic tagging of spoken language 188 5.3.6 Some options of the Portuguese team 193
163
Table of contents
5.4 Main data from lemmatisation 195 5.4.1 The 100 most frequent verbs, nouns, adverbs, adjectives: A comparison between C-ORAL-ROM and CORLEX 195 5.4.2 Similarities and differences in the two corpora 200 5.4.3 Main data from word-forms 204 5.4.4 Lexical density 205 5.4.5 Multiword expressions 206 Chapter 6 Notes on lexical strategy, structural strategies and surface clause indexes in the C-ORAL-ROM spoken corpora Emanuela Cresti 6.1 Premises 209 6.1.1 The utterance 210 6.1.2 Comparison between speech and writing 211 6.1.3 Comparison among spoken Romance languages 213 6.1.4 Variation according to the corpus design 213 6.2 The noun vs. verb lexical strategy in speech 215 6.2.1 Lexical strategy and formality 217 6.3 Informational patterning 219 6.3.1 Informational patterning according to the corpus design 221 6.4 The verbal utterance 223 6.4.1 The verbal utterance according to the corpus design 224 6.5 The ‘non-structuring strategies’ in Italian 226 6.6 The structural types of utterances 228 6.6.1 General tendencies 230 6.6.2 Structural types according to the corpus design 232 6.7 Some remarks on Italian Media and Telephone 236 6.8 Surface clause indexes 237 6.8.1 General percentile data 238 6.8.2 Percentile data according to the corpus design 240 6.9 The informational positions of surface clause indexes 241 6.9.1 A general frame of correlation between syntactic functions and informational positions for Italian 242 6.9.2 Incidence of informational positions of surface indexes in Italian 250 6.10 Some remarks on coordination, subordination and negation in the four Romance languages (FRLs) 252
209
Table of contents
Appendix Evaluation of consensus on the annotation of terminal and non-terminal prosodic breaks in the C-ORAL-ROM Corpus Massimo Moneglia, Marco Fabbri, Silvia Quazza, Andrea Panizza, Morena Danieli, Juan Maríia Garrido, and Marc Swerts a.1 Goals of the evaluation 257 a.2 Evaluation background 258 a.3 Experimental setting 258 a.4 Selection of evaluators 260 a.5 Measurements and statistics 262 a.5.1 Evaluation data 262 a.5.2 First step: Binary comparison file 263 a.5.3 Second step: Ternary comparison file 264 a.5.4 Measurements 264 a.5.5 Kappa coefficient 266 a.6 Results 267 a.6.1 Binary Comparison 269 a.6.2 Ternary Comparisons 270 a.6.3 K-coefficient 271 a.7 Discussion 272 a.7.1 General results 273 a.7.2 Specific results 273 Bibliography
257
277
Acknowledgements
The C-ORAL-ROM Corpus for Spoken Romance Languages is the chief result of the C-ORAL-ROM project, which was funded within the Fifth Framework Program of the European Union, within the Information Society Technology Program (IST2000-26228). The official project web page and documentation is at http://lablita.dit.unifi.it/coralrom. The C-ORAL-ROM corpus is now available in two forms: 1. In this publication which presents the resource in compressed and encrypted format in one DVD. In this form, designed for research and wide distribution in the linguistic community, the resource can be accessed and studied only through the speech software and the text search software that are found on the DVD. 2. Through the ELDA Catalogue (http://www.elda.fr) in 9 DVDs where files are non-compressed and non-encrypted. This form is devoted to speech industry and speech laboratories which can exploit the potential of the resource for the purposes of research and speech technology. The project started in January 2001 and was completed in three and a half years by the following consortium, which was coordinated and managed by the LABLITA lab of the University of Florence:
Corpus and linguistic studies – – – –
Università di Firenze (Dipartimento di Italianistica, LABLITA); Direction: Emanuela Cresti Université de Provence (Description linguistique Informatizée sur Corpus, DELIC); Direction: Jean Véronis Centro de Linguística da Universidade de Lisboa (CLUL); Direction: Maria Fernanda Bacelar do Nascimento Universidad Autónoma de Madrid (Laboratorio de Lingüística Informática); Direction: Antonio Moreno Sandoval
Speech software –
Pitch Instruments France SARL; Direction: Philippe Martin
Acknowledgements
Distribution and dissemination – –
European Language Resource Distribution Agency (ELDA); Direction: Khalid Choukri Instituto Cervantes; Direction: Antonio Cid
Speech recognition –
Istituto Trentino di Cultura, Centro per la ricerca scientifica e tecnologica (ITCIRST); Direction: Daniele Falavigna
For the representation of spoken language in Media contexts in C-ORAL-ROM, various broadcasting companies allowed the publication of samples of their transmissions. On behalf of the consortium, we wish to acknowledge the following for their cooperation: – – – – – – – – – –
TECHE-RAI RTVE (Radio Televisión Española) Radio Televisión Madrid COPE (Cadena de Ondas Populares Españolas/Radio Popular) Onda Cero Radio RTP2: Rádio e Televisão de Portugal; SGPS, S.A. RDP (Radiodifusão Portuguesa), AS. (Antena 1 and Antena 2) SIC (Sociedade Independente de Comunicação), S.A. Rádio Notícias Produções e Publicidade
C-ORAL-ROM is in fact a result of the interaction of many different skills and interests in spoken language which characterise academia, the speech technology industry and the EU Commission in the present time. From the EU Commission representatives, we wish to thank project officer Bent Hauschildt who helped in structuring the project at its start, and project officer Philippe Gelin who took it over at its mid-term review and brought it to its conclusion: both gave constant support which has been essential for the coordination of this complex, collective work. We are also especially obliged to the project reviewers Johanna Moore and Louis ten Bosch: their suggestions have been important for enhancing the quality of the corpus and the reliability of its annotation. Besides the consortium, and the EU representatives, an extremely valuable advisory board acted within C-ORAL-ROM, representing the various technological and theoretical domains that are presently concerned with spoken language and with multilingualism: – – – –
Nuno Beires, PT-Inovação (Portugal Telecom, Lisboa) Claire Blanche-Benveniste, Ecole Pratique des Hautes Etudes (Paris, France) Bernard Cerquellini, INaLF (Institute National de la Langue Française, Paris, France) Morena Danieli, Loquendo (Torino, Italy)
Acknowledgements
– – –
Juan Maria Garrido, Telefonica I+D (Madrid, Spain) Marc Swerts, Tilburg University (Tilburg, Netherlands) Dominique Willems, Collate Research Network (University of Gent, Belgium)
We wish to thank all members of the board for their support. Their presence and patience in project meetings were crucial in the discussions and decisions involved in determining a corpus design strategy and format. Specific thanks must be extended to Morena Danieli, Juan Maria Garrido and Marc Swerts for their collaboration in the evaluation of the prosodic annotation, and Claire Blanche-Benveniste for her friendship, which provided support, ideas and contributions in a series of issues that are too many to be mentioned, from the conception of the project until its end. We also wish to express our gratitude to Anna Brita Stenström for reading the manuscript of this book and for requesting a number of interesting clarifications; and to Eugenio Picchi, who allowed the use of his technology and assisted us in the morpho-syntactic tagging of the Italian corpus. In addition, we wish to thank Silvia Quazza and her staff at Loquendo, and Kees Vaes at John Benjamins, for their work as sub-contractors in the C-ORAL-ROM project. It has been a pleasure working with them. We are also extremely grateful to Elena Tognini Bonelli for her support and advice on the form of both this book and the DVD. We are indebted too to Lisa Lim for taking on the language editing of this volume which has, after all, been written in English by authors with a Romance language competence. The Italian Department of the University of Florence has been in charge of all administrative matters, and we wish to express our thanks to the Director Anna Dolfi and to the secretarial staff Cinzia Baldi and Lucia Azalin for all their help. We have one regret: the extension of the C-ORAL-ROM project to Romanian and Catalan was not funded, and this collection of corpora thus does not represent two Romance languages whose importance is presently increasing in many respects. Finally, on behalf of the whole consortium, we wish to thank the numerous students and colleagues who have worked on the corpus recordings and transcriptions. Without their contribution, the C-ORAL-ROM collection would never have been possible. All responsibility for the work in this volume rests, of course, with the authors. Emanuela Cresti & Massimo Moneglia
Preface
Having been involved in the C-ORAL-ROM project during the preparation stage, I am glad to welcome this outcome of contrastive studies on four corpora of spontaneous spoken Romance Languages, Italian, French, Spanish and Portuguese. The editors, Emanuela Cresti and Massimo Moneglia, had to accept a double challenge: a scientific commitment in defining appropriate contrastive methods for spoken corpora, and a cooperative venture in integrating four types of material, methodological and theoretical resources, supported by different national traditions. Such an enterprise required a common assent. Italian, Spanish, Portuguese and French teams consented to a collective discipline, providing us with the first contrastive Romance study based on prosodic and pragmatic units (although it has to be admitted that French, according to a well-known national trend, proved to be less disciplined). Cresti and Moneglia deserve all the more credit: very fruitful results can be drawn from their report, with possible applications in various theoretical and practical domains, such as language perception, multi-layered linguistic analysis, early acquisition of prosodic patterns, multilingual inter-comprehension, cooperation with computational parsing on spoken languages, information-packaging according to different communicative situations, etc. Comparable data across the four languages were ensured by unified samplings, extracted from large corpora of adults recordings. Apart from traditional distinctions such as dialogue versus monologue and formal versus informal, analytical results imposed another classification involving the channel, by transmission or face-to-face, and an important subdivision between telephone and other media, the telephone being a dialogic event in its pure state, figuring the lowest step of speech structuring. Moreover, media and telephone productions did not seem to be directly related to such features as formality or informality, dialogic or monologic expressions. According to the Italian team’s hypothesis (based on twenty years of research on spoken Italian in LABLITA lab in Firenze), only prosody could supply the searchers with adequate units for intra- and inter-languages comparisons. Usual syntactic schemes were unsuccessful because of the high percentage of non-canonical patterns, and discourse units or speech acts were too elusive. The utterance, an expression marked by a prosodic terminal break, having tight correspondences with speech acts, was thus chosen as the fundamental analytical unit. As the investigators wanted to settle this unit at a psycholinguistic level, they even tested the perceptual recognition of utterances with experimental procedures. Philippe Martin’s WinPitch software offered accurate prosodic measurements for the four languages and an easy access to a general
Preface
text-to-speech alignment, a great asset of the present work. The whole constitutes a strong plea for considering the utterance to be the best unit for such a comparison. Even if some linguists will not wholly side with this choice (because they prefer more grammatical or more semantic grounds or other perspectives), one has to admit that this prosodically-based unit, acting in the way a useful algorithm would do, proved to be an efficient tool for comparing phenomena in the four languages. It could show, for instance, that tonal unit length varies in a language-dependent way, whereas correlations between utterance length, dialogic turns and speed are tightly correlated in all four languages. It also gave a nice point of reference for the distribution of grammatical elements such as coordination, relative-conjunction che/que or negation, by situating them as initial, medial or terminal, relative to prosodic delimitations. Among ensuing consequences, it can be stated, for instance, that, in the four languages, negations are sensitive to situations and are more frequent in informal than in formal speech, with the percentage of negations being the highest in Portuguese and Spanish. It could be stated too that, instead of assuming coordinative or subordinative functions inside the utterances, initial coordinatives, as well as initial subordinatives often open dialogic turns or connect speech acts, thus revealing an unexpected weight of non-subordinative and non-coordinative functions. Moreover, in formal monologues, initial pseudo-coordinatives and pseudo-subordinatives often aggregate together with negative expressions, creating patterns which could not be described within usual grammatical frames. In order to account for some kind of structural complexity inside utterances, two additional features were considered: a morphological one, the presence or absence of one or several finite verbs; and a prosodic one, the presence of one or several prosodic units. Here again, the choice might be contested by many specialists who would prefer more syntactic or more pragmatic approaches. But one has to agree that, combined with textual features, these characteristics allow some interesting generalisations as, for instance, the great proportion of verbless utterances in informal speech in the four languages, the regular absence of complexity in dialogic productions and some language-dependent regularities. A wide structural similarity among the four spoken Romance languages emerges from the comparisons, as shown in nice contrastive diagrams. Nevertheless, in many regards, French appears to be, as Raffaele Simone stated it (Simone 1997), “the less Romance of all Romance languages”. French required special processes for calculating word length because of its specific orthographic system, and special patterns for investigating negation (having three negatives ne, pas, non, where others have one or two) and relatives (two forms qui, que where the three other Romance languages have one). When compared to Italian, Spanish and Portuguese, French utterances showed a much greater number of simple verbal phrases (twice as many) and of simple prosodic units. And, on the contrary, French displayed fewer verbless utterances (even in telephone speech, where they tended to be more frequent) and fewer coordinatives in initial position, so significant for the other languages. The French drift, observed in many
Preface
previous comparative studies in Romance linguistics (Harris 1978), is given here new confirmation in new domains. In a recent textbook (Renzi & Andreose 2003), Lorenzo Renzi recalled the three main streams followed by usual studies in Romance Linguistics (i tre paradigmi degli studi romanzi): the classical period, the historical-comparative method and the structural-generative trend. A new one seems to be now emerging from linguistic comparisons based on spoken languages and operating with computational procedures. This Integrated Reference Corpora for Spoken Romance Languages moves along new paths in this famous domain. Already a successful outcome, it will certainly expand into important future developments. Claire BLANCHE-BENVENISTE Professeur émérite, Université de Provence, Ecole Pratique des Hautes Etudes
Chapter 1
The C-ORAL-ROM resource Massimo Moneglia
. Introduction .. C-ORAL-ROM The setting up of linguistic resources is a current issue in the European Union, and spoken language corpora play a specific role in this context. It is important to achieve knowledge on spoken language for the purpose of the development of linguistic engineering, which, in the era of digital communication, must address the primary role played by spontaneous speech within natural communication. This task is even more urgent in a multilingual context such as that of the EU, where citizens have the right to express themselves in their own native language, and it is the responsibility of the EU to enable the fulfillment of this principle. The study of statistically significant samples of everyday language use is also important for linguistic studies, whose long tradition has mostly been based on ideal models. Despite this general need, spontaneous spoken language is still under-represented, and audio information, which is essential to the development of a robust speech technology, is almost unavailable. The need of spoken corpora is even more strongly felt for Romance languages, as compared to English and other languages for which many resources are already available.1 The main goal of the C-ORAL-ROM project thus is to provide a comparable set of corpora of spontaneous spoken language of the main Romance languages, namely French, Italian, Portuguese and Spanish. The project was undertaken by a European consortium, and launched in 1999 by Emanuela Cresti, Massimo Moneglia, Claire Blanche-Benveniste, Fernanda Bacelar, Philippe Martin, Francisco Marcos Marín and Carlota Nicolás. Supported and co-ordinated by the University of Florence, it was then approved and funded within the IST program of the EU (C-ORAL-ROM IST 2000-26228). C-ORAL-ROM consists in total of 772 spoken texts and 121:43:07 hours of speech by 1,427 different speakers. Four comparable recorded resources of Italian, French, Portuguese and Spanish spontaneous speech sessions (roughly 300,000 words for each language) have been collected respectively by the following providers:2
Massimo Moneglia
i.
University of Florence (LABLITA, Laboratorio linguistico del Dipartimento di italianistica); ii. Université de Provence (DELIC, Description Linguistique Informatisée sur Corpus); iii. Centro de Linguística da Universidade de Lisboa (CLUL); iv. Universidad Autónoma de Madrid (Departamento de linguistica, Laboratorio de Lingüística Informática).3 Each recorded session in C-ORAL-ROM comes with the following main annotations: a.
Session metadata, containing essential information of speakers, recording situation, acoustic quality, source, and content of each session, ensuring clear identification of the various speech types documented; b. Orthographic transcription, in standard format, enriched by the tagging of terminal and non terminal prosodic breaks; c. Text-to-speech synchronisation, based on the alignment with the acoustic source of each transcribed utterance; d. Part of Speech (PoS) tag of each form in the transcribed texts and the corresponding frequency list of forms and lemmas.
This resource is stored in the DVD accompanying this book, and is integrated with two tools which allow the direct exploitation of the acoustic and textual information,4 respectively, WinPitch Corpus (© Pitch France) and Contextes (© Jean Véronis). The DVD presents C-ORAL-ROM in five sections: 1. Metadata of the recorded sessions of each language resource (see (a) above); 2. Multimedia corpora, which allow simultaneous access to textual and acoustic information through the speech software WinPitch Corpus; 3. Textual corpora, which allow searches on textual and PoS tagged files through the text search engine Contextes; 4. Frequency lists of forms and lemmas of each language collection; 5. Histograms and tables, derived from the main corpus annotation, which present general features and peculiar qualities (lexical, structural, syntactic) of spontaneous speech in the four Romance languages. The four corpora are orthographically transcribed in standard textual format (CHAT format, MacWhinney 1994), with representation of the main dialogue characteristics; that is, speaker’s turns; the main occurring non-linguistic and paralinguistic events; the prosodic breaks; and the segmentation of the speech flow into discrete speech events. The identification of speech events is one of the more particular features of CORAL-ROM. The speech flow and its transcription are divided into units which rank above the word level and belong to the domain of action, i.e. utterances. An utterance is the linguistic counterpart of a speech act (Austin 1962) and it is identified in C-ORAL-ROM through an heuristic method that is based on the perception of prosodic cues (Cresti 1994, 2000; Cresti & Firenzuoli 1999).5 Each utterance is
The C-ORAL-ROM resource
assumed to end with a perceptively relevant prosodic break having a terminal value. The textual string is divided into utterances following the annotation of prosodic breaks, discriminated in the speech flow through perceptive judgments.6 With WinPitch Corpus, which ensures a natural and meaningful text/sound correspondence, the annotated transcripts are aligned with their acoustic counterpart, with segments defined on independent layers, generating a database of all utterances in the resource (roughly 134,000 in the multilingual corpus).7 Besides text-to-speech and speech-to-text alignment, WinPitch Corpus allows an easy and efficient acoustic analysis of speech with real-time fundamental frequency tracking, spectrographic display, re-synthesis after editing of prosodic parameters, etc. The Contextes text search engine operates on the C-ORAL-ROM textual resource, allowing searches on textual items at different levels of annotation; that is, at word level, at lemma level, at PoS level, at metadata level, and at prosodic tagging level.
.. Organisation of the volume The C-ORAL-ROM resource aims to represent the variety of speech acts performed in everyday language and to enable the induction of prosodic and syntactic structures in the four Romance languages, from a quantitative and qualitative point of view. This chapter will describe the set of choices adopted to this end. The criteria for the corpus design of the resource and for its prosodic annotation are presented respectively in Sections 1.1 and 1.2. In 1.3 the format for both metadata and dialogue representation is outlined, while in 1.4 the text-to-speech alignment method and tool are described. 1.5 will present the specifications of the C-ORALROM format for PoS tagging, and 1.6 will give some general measurements of the main language variations that can be found in the corpus structure. The next four chapters (2 to 5) will present the Italian, French, Spanish and Portuguese sub-corpora, respectively. Each chapter is similarly structured, providing descriptions along the following lines: (1) the history of the national corpus; (2) the choices made for the orthographic transcription of the oral material; (3) the PoS tagging tool and the strategy adopted, accompanied by the evaluation of PoS tagging correctness; and (4) the main data extracted from lemmatisation, for example, the most frequent lemmas in spoken language. The description of each sub-corpora, however, represents the peculiar point of view of each corpus provider on the resource itself. Other related arguments are also variously included as sub-sections in the four parallel chapters. More specifically, the state of the art of spoken resource collections in each nation, the problems encountered in PoS tagging of that particular language, and the peculiar characteristics of frequency lists derived from each spontaneous speech resource are treated in each chapter with more or less attention, depending on the research tradition of each team. A set of measurements highlighting the strategies of construction of spoken language utterances and their variation in the C-ORAL-ROM corpus design is also provided. These measurements are reported in a Diagram menu in the DVD, which
Massimo Moneglia
shows the recorded values in the four languages in Frames. The diagrams address the following main language data: i. lexical strategy: the proportion of nouns to verbs; ii. structural strategies: the level of structural complexity recorded in spoken texts through the main strategies adopted for the construction of the utterance, i.e. (a) verbal vs. non verbal; (b) prosodically simple vs. prosodically patterned; iii. structural types: the main structural types of utterances, i.e. compound verbal, simple verbless, simple verbal, and compound verbless; iv. clause indexes: the incidence of the main functional expressions, such as indexes of coordination, subordination and negation on the total utterances. Clearly, this work offers the reader some general results regarding the speech properties of the four Romance languages which can become a vital platform for comparison both with writing and with other spoken languages (Biber et al. 1999). It is also a first example of how research on speech which wishes to go beyond lexical frequencies and phonetic observations can be driven on the basis of utterances taken as reference units. A brief presentation of this is given in Chapter 6. Finally, the Appendix presents the results of the evaluation of the prosodic annotation in the C-ORAL-ROM corpus. The evaluation, executed by independent users of the speech industry sector, is aimed at guaranteeing the overall reliability of the data and computations presented.
.. The issues of representation and comparability The domain of spontaneous spoken language has become consolidated only in very recent times (see Biber 1988; Blanche-Benveniste et al. 1990; Cresti 2000; Givón 1979; Miller & Weinert 1998). This tradition agrees that spontaneous speech performances can be defined as every oral performance which does not execute a previous (written or scripted) text. From a positive point of view, speech events that may be considered spontaneous have a typical set of qualities, among which the following are the most relevant. Spontaneous speech events comprise: a. b. c. d.
face-to-face multi-modal interactions; inter-subjective reference to a deictic space; mental programming simultaneous with vocal execution (unscripted); contextually undetermined linguistic behaviour (unpredictable behaviour).
A long tradition of socio-linguistic studies (see Berruto 1987; Biber 2001; Biber et al. 1998; Gadet 1996a, b, 1997, 2000, 2003; Labov 1966) has frequently dealt with the significance of the sociological and contextual parameters in the definition of speech qualities, pointing out their variability. Many types of spontaneous speech exist, varying along the following parameters:
The C-ORAL-ROM resource
a. b. c. d. e. f. g. h. i.
a variety of possible structures of the communication event (monologue, dialogue, conversation between many participants, etc.); the channel of communication of the speech event; the sociological context, that is, the domain of society in which the speech event takes place (family life, private life, public life); programming conditions (totally unscripted vs. scripted or partially scripted performances); a variety of possible language registers and genres (diaphasic variation); socio-linguistic factors, such as sex, age, education, occupation of the speakers (diastratic variation); the geographical origin of speakers (diatopical variation); the task of the speech event; the topic of the speech event.
For instance, a tale told to a child and a row between husband and wife are different types of speech events which occur in family life; these events vary strongly in language register, dialogue structure, and of course in topic. Where public contexts are concerned, a lesson given by an university professor and a discussion in a talk show may present a set of features (viz. dialogical character, unscripted or partially unscripted programming, informal style) which allow them to be considered varieties of spontaneous speech as well, even if the speaker is performing a professional task. These two types vary in register, in communication structure, in channel, and in programming and control of the oral execution. The linguistic properties of the speech events may vary in connection with such non-linguistic variations The setting up of spontaneous speech databases is therefore a complex task which must ensure the representation of the most significant variations exploited by the different types of spontaneous speech performances (Berruto 1987; Biber 1988; De Mauro et al. 1993; Gadet 1996a, b, 2003). Spoken resources set up in controlled environments that have been collected to satisfy the need of speech technologies (such as telephone information, health dialogues, map tasking) constitute at present the majority of the available databases.8 Their acoustic/phonetic quality is excellent, but by definition they deal with restricted semantic domains, and in the given context the language behaviour is highly predictable. Should one wish to represent the spontaneous speech universe, the constitution criteria must ensure the widest possible variation in speech contexts, and the lowest control on the speech event, which is exactly the opposite of what dedicated resources do. On the basis of this need, the C-ORAL-ROM corpus is oriented towards the collection of corpora in natural environments; this, however, necessarily results in a lower acoustic quality of the resource (Labov 1966). Moreover, since C-ORAL-ROM exploits the rich archives created in previous years of research on spoken languages, the acoustic quality and the recording conditions of the resources are, inevitably, variable.
Massimo Moneglia
In any case, the acoustic quality of each recording and the most relevant information of the recording condition are always recorded in the metadata of each text. The following outline the requirements followed in the project for the acoustic format and the recording apparatus: i. Format: mono .wav files (Windows PCM); sampling frequency: 22,050 Hz, 16 bit. ii. Recording and storing process for old analogue recordings: directly derived in .wav files (20,050 Hz, 16 bit) from the original analogue tapes through a standard sound card (Sound Blaster live or compatible) with a professional sound editor. iii. Recording and storing process for new recording: a. dialogues: stereo DAT or minidisk recording (44,100 Hz) with unidirectional microphones, converted into mono .wav files (Windows PCM, 22,050 Hz, 16 bit) via SPDIF port of a standard sound card (Sound Blaster live or compatible) with a professional sound editor; b. conversations with more than two participants: mono DAT or minidisk recording with cardioids or omni-directional microphone, converted into mono .wav files via SPDIF port of a standard sound card (Sound Blaster live or compatible) with a professional sound editor. The sound files of the acoustic database are set on a quality scale (recording, volume, voice overlapping and noise) and are comparable with respect to it. The quality scale extends from the highest level of clarity of the voice signal to low levels of acoustic quality. The quality is gauged spectrographically (see the rules for marking the acoustic quality in 1.3). A further problem that a multilingual corpus of spontaneous speech must face is comparability between different language resources. Comparability of large written corpora has been tested in two forms: a. Parallel corpora (e.g. CRATER and EUROM4); b. Corpora of the same type or of the same specialised field in several languages. Clearly, with respect to the task of collecting multilingual spontaneous spoken (language) corpora, only the second alternative is, in principle, available, since, in the domain of speech, parallel corpora are possible only for reading and acting performances. In other words, it is impossible to achieve parallel spoken corpora without losing spontaneity. The prototype example for the second alternative in the domain of written resources is the relation between the Brown Corpus (Brown University, USA, early 1960s) and the LOB Corpus (Lancaster/Oslo/Bergen, 1970) which achieves a comparable sampling of American English and British English. In the spoken domain, comparability is quite easy to pursue with respect to resources based on the selection of a specific semantic domain (e.g. telephone information, health information, map tasking etc.), that is, with people in the same controlled situation doing the same things. However such resources are subject to elicitation parameters, within limited contexts, and therefore lack the main character of spontaneous speech.
The C-ORAL-ROM resource
If we assume that the representation of spontaneous speech must necessarily maximise spoken text variation, in a multilingual resource the more variability there is represented in each language resource, the more difficult it becomes to compare one language resource with the others. Therefore the comparability of the resource is a function of the application of specific variation parameters in a corpus design structure. The need to document many types of spontaneous speech in order to offer a significant representation of the spoken universe is confirmed by recent corpus linguistics studies, where all levels of language description vary when they are considered with respect to different types of spontaneous speech events. This is quite evident at the lexical level. The representation of a sufficient number of contexts, covering relevant types of speech events in the universe, is the only possible strategy to identify frequency lexicons. A high-frequency lexicon may be under-represented in specific pragmatic domains which, on the contrary, by definition, maximise the probability of occurrence of low-frequency lexical items (which are the real focal point of dedicated resources used for speech recognition). The frequency of Parts of Speech also varies according to the context of use. Nouns are known to be more frequent than verbs, but the relative frequency of nouns is lower in spoken language, and it is much lower in informal conversations with respect to formal contexts (1/1 vs. 1/3). Adjectives, on the other hand, are much more frequent in formal speech (Biber et al. 1999). More subtle lexical variations may also be noted. For instance, it is generally assumed that deverbal nouns (destruction) are characteristic of technical contexts and are much less frequent in everyday dialogues (Halliday 1989). Rouget (2000) shows that such nominalisation also occurs in spontaneous everyday dialogues, but typically in restricted syntactic positions, i.e. not depending on a verb. On the other hand, their syntactic position is unconstrained in technical contexts. Recent corpus-based grammatical descriptions of language have also highlighted the fact that the majority of complement clauses which depend on a verb depend on a putandi verb in spontaneous conversation, while complement clauses depend more frequently on a dicendi verb in broadcasting and media contexts (Biber et al. 1999). In corpus-based approaches, the induction of the main syntactic properties is strongly correlated to text variation parameters as well. For example, in English, both main types of dependent clauses (relative and complement clauses) vary in frequency according to socio-linguistic parameters. Generally speaking, in syntactic structures controlled by a noun, the frequency of both that-complement clauses and to-complement clauses is higher in formal language, while, in syntactic structures controlled by a verb, that-complement clauses are much more frequent in conversation (Biber 2000). Similar conclusions can be drawn with respect to relative clauses: relative constructions are much more frequent in formal speech, with their restrictive function generally being the most frequent (Biber et al. 1999). In other words, the pragmatic domain of corpora strongly influences the probability of occurrence of syntactic properties in the core area of grammar.
Massimo Moneglia
Spontaneous speech variation also displays variations in type and length of utterances. Recent work on spoken Italian (Cresti 2000) claims that the mid-length of the utterance (MLU) in dialogic speech sessions (family conversations, country wakes, conversations among work colleagues and conversations among university students) systematically diverges from the MLU of formal texts (university lectures and radio interviews). Moreover the number of speech act types in the previous Italian collections of spontaneous speech is much higher than those reported in semi-spontaneous corpora. Firenzuoli (2003) records over 80 speech act types distributed throughout the corpus, while in the map-task coding scheme (Anderson et al. 1991), the set of possible dialogue acts corresponds to roughly 16 possible moves (Stirling et al. 2001). It is clear from the preceding review that findings can differ drastically according to variation in parameters; C-ORAL-ROM aims to provide a sufficient representation of spontaneous speech variation.
.. C-ORAL-ROM sampling strategy C-ORAL-ROM sampling of the four Romance language resources is based on the following set of variation parameters which constitutes the semiological and sociological structure of the spontaneous speech corpus: a.
Dialogical structure: the recorded sessions are classified in monologues, dialogues, and conversations according to the interlocutor’s active participation in the exchange; b. Social context: the session recording of interactions belonging to family and private life are distinguished from those taking place in public; c. Channel: face-to-face interactions are distinguished from media productions and telephone recordings; d. Domain of use: domain of social environments, activities and professions, such as law, business, research, teaching, church, etc.; e. Register variation: recorded sessions where a formal language register is used are distinguished from those where informal uses are preferred; f. Speaker parameters: the main socio-linguistic qualities of speakers are registered; that is age, sex, education, occupation and geographical origin.
The C-ORAL-ROM corpora have been collected in Continental Portugal, Central Castilian Spain, Southern France, and Western Tuscany (the latter historically considered the source of the Italian language), and are intended to represent a possible standard usage that occurs in those areas.9 It must be pointed out that the language archive is not defined on the basis of the geographical origin of speakers as in dialectologicaloriented research (Contini et al. 2002). In other words, systematic diatopical variation is not represented in C-ORAL-ROM: in a multilingual collection of limited size such as this, diatopical restrictions must necessarily be established for each language, and the corpus cannot be concerned with the representation of geographical varieties. Such variation of pronunciation, lexical and morpho-syntactic features in a language,
The C-ORAL-ROM resource
Figure 1.1 Geographical origin of speakers in the four corpora
depending on geographical origin of speakers, can be represented only within wide intra-linguistic corpora, which takes into account the systematic representation of the geographical varieties of the language in question.10 C-ORAL-ROM thus represents the language actually spoken in a relevant national centre (namely Madrid, Lisbon, Aix-Marseille, and Florence) and in their neighbouring area (Cresti et al. 2002). Naturally, whatever other geographical varieties that may be employed in that reference centre may also occur, in varying proportions, and, in fact, instances of varieties of different geographical origin are recorded for each language. It should also be noted that for many participants in media emissions and other sessions recorded in public contexts, their geographical origin is not available. Figure 1.1 shows the number of speakers recorded in the four corpora and the distribution of their geographical origin. The goal of the collection is therefore limited to ensuring the representation of the sole diaphasic and diastratic varieties of the language that is spoken in the region in question (Berruto 1987; Gadet 1996b), and it is expected to present a sufficient variation for the study of communicative acts, lexicon, morphology, syntax and, as far as the main variety recorded may be considered significant, prosody. Nonetheless, since the speaker’s characteristics (sex, age, education, occupation and geographical origin) and those of the session (communication event, social context, register, topic) are recorded as part of the metadata, the user can always link any linguistic phenomenon to the specific situation in which it occurs and be aware of its value.
Massimo Moneglia
Given the above parameters, the main choices adopted in C-ORAL-ROM for the representation of the speech universe in the four 300,000-word corpora are the following: 1. 2. 3. 4.
Separating formal speech (50%) and informal speech (50%); Selecting distinct criteria for sampling the formal and informal parts of the corpus; Defining the dimension of samples in terms of number of words; Ensuring a sufficient representation of dialogical informal speech (which is the resource with higher added value); 5. Ensuring the representation of Formal language both in Broadcasting (Media) and in face-to-face contexts (which are labelled “Natural context” in the C-ORAL-ROM corpus design). Comparability across the four spoken corpora is ensured by means of common sampling criteria, and by the same weight in words in each domain type in the four corpora. Table 1.1 presents the matrix for all the languages in C-ORAL-ROM.
Table 1.1 Parameters for the description of spoken language in the corpus design matrix Language register Informal 150,000 words
Formal
Social context Family/private 124,500 words Public 25,500 words
Structure of the communication event Dialogue and Multi-dialogue 102,000 words Monologue 48,000 words
Channel Natural context 65,000 words
Typical domain of use Political speech Political debate Preaching Teaching Professional explanation Conference Business Law Talk shows Scientific press Reportage Interviews Sport News Weather forecast Private conversations 15,000 words Human-machine interactions 10,000 words
Formal
Media 60,000 words
Informal
Telephone 25,000 words
*At least 20,000 words from conversations with more than two participants in the Informal part.
The C-ORAL-ROM resource
Given the design scheme which offers a representation of the speech universe, the number of samples that are chosen for each domain in the universe is the main variable for the statistical relevance of the corpus. Although the overall dimension of the CORAL-ROM corpora determines a severe limit in this respect, it must be highlighted that this choice relies on two concurring needs that may have consequences on the comparability of the multilingual corpora and on their significance. The more samples are chosen, the more accurate the representation of the universe, but the amount of information in each sample must also be sufficient to represent each domain with respect to dialogue structure and syntax. In other words, while the samples must be large enough to offer a probability of occurrence of those properties that characterise each context, each corpus must also have a comparable number of samples. The C-ORAL-ROM corpus design strategy considers that the representation of spoken language must, in a specific language context, ensure both the possible appreciation of syntactic properties (micro-syntaxe) and the sufficient representation of different dialogue structures (macro-syntaxe and informational patterning; see BlancheBenveniste et al. 1990; Scarano 2003). Therefore the length of each sample must in principle be sufficient for both ends. The choices concerning sample length and number of samples consequently vary according to the corpus domains. Sample length is determined in terms of quantity of information (i.e. number of words) rather than in temporal units (minutes).11 Additionally, since formal texts, by definition, feature a more complex textual structure, two different strategies have been adopted for the formal and the informal categories, as follows: i.
Informal section: a. Short texts: at least 64 texts of approximately 1,500 words each.12 b. Long texts: from 8 to 10 texts of approximately 4,500 words each.
ii. Formal section: a. Formal in natural context: 2 or 3 samples for each domain of approximately 3,000 words each. b. Media: 2 or 3 samples of approximately 3,000 words for each domain, while only short samples for weather forecasts and news are provided. c. Telephone: text length is not defined (preferably 1,500 words as upper limit, no lower limit); the man-machine interactions domain should contain 10,000 words, comprising the complete interaction. In conclusion, given the above matrix, each corpus in the multilingual resource is not strictly comparable with the others with respect to the occurring semantic domains; however, as far as each corpus approximates the matrix, it demonstrates a comparable variation. This variation is significant with respect to the possible occurrence of spoken language structure/s at both syntactic and dialogue representation level.
Massimo Moneglia
.. Comparing C-ORAL-ROM with other corpora A corpus which has a very similar design structure to the C-ORAL-ROM matrix is the Spoken Dutch Corpus which, despite its different size, has been explicitly considered a model by the C-ORAL-ROM consortium. Both C-ORAL-ROM and the Dutch Corpus address roughly the same set of variation parameters; those of the latter are outlined in Table 1.2. The aim of the C-ORAL-ROM matrix is to define the crossing over between the communicative event’s structure parameters (Dialogue/Conversation vs. Monologue), the sociological context of use (Private vs. Public) and the channel (Broadcasting vs. Natural context), as the Dutch Corpus does. However some relevant differences in the corpus design strategy must be highlighted. The use of the formal/informal partition in the sampling strategy is absent in the Dutch corpus. This distinction of register is only partially equivalent to that of scripted/unscripted programming. While a formal register of speech is frequently partially scripted, it is also linked to other relevant conditions, such as: a. public use of speech; b. professional use of speech, in accordance with social functions and roles of the speaker in the community; c. the intentional task of accomplishing an oral text, that implies the treatment of a topic, an argumentation, conclusions, etc.13 In short, Formal speech may still be spontaneous, as long as it is not the execution of a previously prepared text.
Table 1.2 The Spoken Dutch Corpus design14 Dutch Corpus dialogue/ multilogue
monologue
private
unscripted
direct
conversations (‘face-to-face’) interviews distanced telephone conversations business transactions public broadcast more or less scripted interviews and discussions non-broadcast unscripted discussions, debates, meetings lectures private more or less scripted descriptions of pictures public broadcast unscripted spontaneous commentary more or less scripted news reports, current affairs programmes news commentary non-broadcast more or less scripted lectures, speeches read aloud text
The C-ORAL-ROM resource
For these reasons, as the C-ORAL-ROM matrix in Table 1.1 shows, genre variation is the main criterion applied in formal contexts, where differences in general professional functions (lawyers, teachers, priests), social occasions, and topics (health, business, art, science) are relevant. In contrast, genre variation is not strictly defined as a parameter in informal contexts; instead it is variations in social domain of use (family/private, public) and dialogue structure (monologue, dialogue, conversation) which represent the parameters systematically adopted. In other words, C-ORALROM adopts two different sampling strategies, one for the representation of spoken language in formal contexts and another for informal contexts. Only in formal contexts is the genre (or domain of application) of the recorded sessions strictly defined in a closed list of sub-types; this information is not provided for informal contexts. This choice has a theoretical value. While it is natural to assume the existence of a series of closed situations in which, in a certain social-historical context, the formal use of language is preferred, the same does not hold for the universe of informal speech. To feature typical contexts of use is a specific, marked trait of formal speech. For this reason, formal speech can be effectively identified by listing its most typical contexts of use. Conversely, the set of situations where informal language is used is open, and its domain cannot be represented by a list of typical contexts of use. No context is more typical than another. This feature is, in fact, far from obvious. For example, the Dutch corpus tries to define, as a comparability criterion, the contexts of its collection of monologues and dialogues, by determining the number of words for a closed set of domains of application (e.g. Business transaction, Picture description, Interview, Face-to-face conversation, Telephone). By observing a posteriori the sampling strategies used by these two collections, both aimed at the documentation of spontaneous speech and the comparability of data from various corpora, the concurrence of two criteria appears evident. It is a fact that, by strictly defining the genres, a higher degree of data comparability can be attained. However, the downfall of this practice is the a priori exclusion of significant spheres of informal speech, where genre characterisations still remain largely unexplored. This choice is significant for formal speech only. We can conclude that, by not defining explicitly the genres and domains of use of speech, C-ORAL-ROM’s corpus sampling guarantees, in theory, the possibility of occurrence in the collection of any significant genre. In other words, no genre has zero probability of occurrence. Considering the possible sampling strategy for a spontaneous speech corpus, it may be interesting to make explicit mention of another large project that was also taken into consideration when the C-ORAL-ROM corpus design was decided. The Corpus of Spoken Israeli Hebrew (CoSIH) does not feature a corpus design scheme.15 CoSIH was designed to integrate, in a random sampling, demographic and contextual criteria for both speakers and speech events. Day-long recordings of 950 informants (statistically representing the Israeli population) have been planned over a one-year period, along
Massimo Moneglia
with respective sociolinguistic data. These recordings must be evaluated, and a cell from each sample must be transcribed, to set up a five-million-word corpus.16 In principle the mapping of a continuous recording into a textual resource could also be a function of a random selection of one-hour recorded cells (consisting of 5,000-word units). However this method, appropriate from a statistical point of view, does not completely bypass the definition of criteria for the selection of relevant textual units. As a matter of fact, in order to select 5,000 words of a coherent continuous text, a huge amount of fragments, silences, and minor exchanges must be negatively selected. Therefore criteria like those specifying possible corpus variations in the corpus design scheme are also needed. The first results of this project however are expected shortly and will provide new evidence for the documentation of spontaneous speech variation.
. Prosodic tagging criteria The segmentation of the speech flow into discrete speech events is one of the most relevant questions for the analysis of speech resources. In the case of written language, the nature of the linguistic units ranking above word level is not controversial. Although the unit of analysis may be chosen at different levels (i.e. argument structures, sentences or clauses or head dependent structures, see Abeillé 2003), written language can be properly parsed into discrete objects that belong to the domain of syntax. In recent linguistic tradition this question is overtly raised for spoken language and specifically for spontaneous speech (see Blanche-Benveniste 1997a; Biber et al. 1999; Cresti 2000; Miller & Weinert 1998; Quirk et al. 1985). Although the reference units for spontaneous speech are commonly referred to as speech events or utterances, the definition of such entities is far from obvious (see the discussion in Scarano 2003). The definition of an utterance may be anchored to syntactic and/or semantic properties. One of the most common solutions is to identify it with a syntactic clause (Miller & Weinert 1998), clearly distinct from a sentence, the latter, on the other hand, being commonly assumed as the reference syntactic unit of writing. The Longman Grammar proposes a definition of C-Unit as an entity with or without a clause structure (Biber et al. 1999). Another recent definition is that proposed in the pronominal approach (Blanche-Benveniste 1997a), which identifies the nucleus of an utterance in a macro-syntactic domain based on a modalised noyau. Other solutions have been those based on a concept of predication (Voghera 1992), which in most cases has a verb as its core. However we must consider that figures show that in spontaneous speech, verbless contexts appear in around 30% of utterances (38% according to Longman Grammar). More specifically the statistical measurements made on the C-ORAL-ROM corpus show that verbless utterances constitute 38.1% in the Italian corpus, 24.1% in French, 37.23% in Spanish and 36.57% in Portuguese. Since on average 30% of utterances are verbless, all the definitions based on clause structure and verbal predication appear to be inadequate for spoken corpus analysis purposes.
The C-ORAL-ROM resource
A speech event may also be identified as a dialogue act, recorded in a dialogue representation scheme (see Sinclair & Coulthard 1975, and the literature cited below); it has been also interpreted as the minimal entity allowing a modification in the discursive memory (acte énocniatif in the framework of Berrendonner & Béguelin; see Berrendonner 1990, 2003). The background assumption for the C-ORAL-ROM project is that in spontaneous language the speech flow is divided into utterances, which are defined as the linguistic counterpart of speech acts, that is, entities belonging to the action domain. Speech acts are defined in the classic work of Austin (1962).17 Given that it is generally assumed that the identification of speech acts is quite a vague exercise, and that the notion lacks a formal definition (Fava 1995; François et al. 1990; Kempson 1977), the present literature is frequently sceptical of the possibility of identifying speech acts in the reality of spoken language, and, further, of finding linguistic regularities concerning this notion. In the rest of this section, we will first show how utterances are determined in CORAL-ROM from an operative point of view, and subsequently discuss the relation with other concurring criteria. The question of the replicability of the coding scheme will also be addressed. The first two issues will be dealt with in these sections, while the reader can refer to the Appendix of this book for the third question.
.. Prosodic breaks and utterance limits In C-ORAL-ROM, prosodic breaks are considered the most relevant cue for determining utterance boundaries. The relation between prosodic cues and the segmentation of the speech flow is a very complex argument (Simon 2004) that cannot be addressed in detail here. C-ORAL-ROM exploits some general properties of this relation to allow the segmentation of a large multilingual corpus in a clear and homogeneous manner. Tone units (or prosodic envelopes) are separated by prosodic breaks. A very general relation between tone units and information units has been recognised since the work of Halliday (Halliday 1976). From this point of view the division of speech into information units by mean of prosodic boundaries can be assumed. In fact, there is clear perceptual evidence for prosodic breaks, and they are also quite easily recovered in the speech flow (see 1.2.2 below). However this correlation is not sufficient for the detection of utterances in the speech flow, given that there is no necessary one-to-one correspondence between utterances and information units. An information unit may not be an utterance, but just a part of it. In other words, an utterance can be composed of one or more information units, and can therefore be prosodically performed by more than one tone unit, that is, through a complete intonation pattern (Cresti 2000; ’t Hart et al. 1990). A correlation between prosodic breaks and the segmentation of the speech flow into utterances can however be maintained, considering that classic studies on prosody have always highlighted the fact that utterances (or sentences) end with a terminal profile (Crystal 1975; Karcevsky 1931). From this point of view those prosodic breaks
Massimo Moneglia
that conclude an utterance can in principle be discriminated from those breaks that do not. Therefore those utterances that correspond to one tone unit may be distinguished from those featuring a complex pattern of tone units. In C-ORAL-ROM the perceptual detection of terminal breaks has been adopted as a heuristic method for determining utterance boundaries in the speech flow: each string ending with a terminal break is considered an utterance. The inter-annotator agreement reached in C-ORAL-ROM in the perceptual detection of this cue is reported in the Appendix. The rough equivalence between utterance and textual string ending with a terminal break is based on the idea that the performance of language actions is necessarily correlated to prosody, which constitutes the interface between the accomplishment of illocutionary and locutionary acts. An utterance can be performed by a simple or compound intonation pattern, that is, one that is made up of only one necessary tone unit (root), or by more optional tone units, with different prosodic features (for instance, prefix unit, suffix unit, in the framework of ’t Hart et al. 1990). Each tone unit (corresponding to an information unit) ends with a break, which is an object of perception (Hirst & Di Cristo 1998). Very roughly speaking, intonation patterns the utterance into information units where the necessary root unit specifies the illocutionary force (Cresti 2000; ’t Hart et al. 1990). For example, a simple noun phrase like the door can be considered a speech act in spontaneous speech (i.e. an assertion, a question, an order, a doubt, or other possible illocutionary act), only once it is performed with an appropriate root unit. The perception of the property “terminal” in a prosodic break seems to be at least in correlation with this: we perceive the break as terminal because an act has been accomplished. The accomplishment of an illocutionary act is the main property that a language event must has in order to be considered an utterance. The illocutionary force determines how the “propositional” content of the utterance must be interpreted in the world (e.g. the content describes the world if the action is declarative, or it specifies how to modify the world if the action is directive). An important consequence derives from the previous theoretical considerations. From an operational point of view the utterance can be defined as the minimal linguistic unit such that it allows a pragmatic interpretation in the world. Therefore, as far as the illocutionary values are conveyed by prosodic cues, each utterance must contain the prosodic cues that allow its pragmatic interpretation. In other words, prosodic cues are those that enable a competent hearer to interpret a linguistic string as a language activity. Competent speakers are extremely sensitive in their perceptual ability to detect even very slight prosodic variations, which mostly concern suprasegmental parameters, such as F0 movements, length, and intensity. More specifically, they are able to selectively perceive those variations that are voluntarily produced (’t Hart et al. 1990), as is the case when an intonation profile, conventionally codified in a language, is performed in order to express a specific illocutionary force.18 From the point of view of
The C-ORAL-ROM resource
perception, competent speakers are the only subjects that can judge, on the basis of its prosodic features, if a spontaneous speech string does or does not have the property to allow a pragmatic interpretation (illocutionary criterion).19 The strings of the speech flow that do not allow a pragmatic interpretation by a competent speaker, therefore, are not actual candidates for being an utterance.20 Those strings always correspond to language strings that do not end with a break perceived as “terminal”. Competent speakers, with their judgements, may not be able to assign to an utterance one specific speech act label from among the large number of possible speech acts that occur in spontaneous speech.21 For example, a competent speaker can be uncertain of whether a string like better to keep to the right or down the lane, uttered in a given context with the appropriate intonation, is to be considered “advice” or “prompting”. But every speaker will recognise if each information unit can or cannot receive a pragmatic interpretation. For example, a competent speaker will deny this property if the previous two strings appear within the same tone unit, as in better to keep the right down the lane, or in a patterned utterance with a topic-comment structure as down the lane / better to keep the right.22 In summary, with regard to its prosodic properties, an utterance can be simple, that is, featuring only one tone unit, or compound, that is patterned in various tone units. Competent speakers have the ability to detect prosodic breaks and to distinguish breaks that terminate a sequence of tone units from those which signal the flow of a same prosodic programme. This perceptual ability is enforced by the pragmatic correspondence of prosodic patterns ending with a terminal breaks with the accomplishment of illocutionary acts. The heuristic proposed in C-ORAL-ROM relies on these theoretical premises just outlined. Perceptively relevant terminal breaks are assumed to identify the utterance limits or, less compromisingly, to provide enough evidence that what follows a terminal break necessarily belongs to a different utterance, and therefore to a different pragmatic and linguistic domain. The various kinds of prosodic breaks are conceptualised and defined as follows: a.
Prosodic break: perceptively relevant prosodic variation in the speech continuum such as to cause the parsing of the continuum into discrete prosodic units. b. Terminal prosodic break: given a sequence of one or more prosodic units, a prosodic break is considered terminal if a competent speaker assigns to it, according to his perception, the quality of concluding the sequence. c. Non-terminal prosodic break: given a sequence of one or more prosodic units, a prosodic break is considered non-terminal if a competent speaker assigns to it, according to his perception, the quality of being non-conclusive.
.. Background of prosodic labelling The reliability of the perception of prosodic breaks in the speech continuum is based on their strong salience (’t Hart et al. 1990; Hirst & Di Cristo 1998). Therefore the C-
Massimo Moneglia
ORAL-ROM prosodic tagging procedure is considered accessible to labellers even after a brief training. This assumption is in line with other experiences of prosodic break labelling. Breaks have also been annotated (together with other cues), specifically in a sub-corpus of the Dutch corpus (Buhmann et al. 2002). In this, strong and weak breaks are discriminated and marked in the speech flow, and are reported in the transcripts (with the symbols | and ||, respectively).23 The prosodic annotation of prosodic breaks (both the distinction between “weak” and “strong”, and “terminal” and “non-terminal” ones) does not deal with modelling F0 movements, nor with marking lengthening or intensity peaks, as is the case with other different prosodic notations that have been widely used in the last two decades (e.g. the ToBI system). This type of prosodic annotation that does not constitute a “transcription of prosody” is significant as regards large corpora labelling strategies. The percentage of consensus on cues like “types of edge tone” and “types of pitch accent” was found to be quite low (Wightman 2002), while on the contrary a high degree of agreement on labelling prosodic breaks is found. For instance, in ToBI annotation, boundary tones score an agreement between 85% and 92%, and the prosodic annotation in the Dutch corpus by non-expert labellers has also been successfully verified in terms of K statistics (Buhmann et al. 2002).24 However the results obtained for the Dutch corpus cannot be automatically extended to C-ORAL-ROM. C-ORAL-ROM is a multilingual resource and many factors may influence the perception of prosodic cues in different languages. Moreover, although the annotation in the Dutch corpus may partially overlap that of C-ORALROM, it is not co-extensive. Strong breaks are defined as “severe interruptions of the normal flow of speech”, while weak breaks are defined as “weak but still clearly audible interruptions of the speech flow”. Now, it is very likely for all terminal breaks to be perceived as severe interruptions of the speech flow, but a remarkable number of nonterminal breaks also share this property. For instance, the prosodic break connected to prefix intonation (see ’t Hart et al. 1990 and the example in Figure 1.4 below) is a severe interruption of the speech flow, but under no circumstance may it be perceived as a terminal break. In other words, a strong break may not have the functional value of terminal breaks (that is, of marking the end of the utterance), and therefore it has lower linguistic value. Given these premises, the prosodic tagging accomplished in C-ORAL-ROM has been evaluated by an external institution in order to confirm the reliability of this annotation. This evaluation was performed by mother-tongue non-expert evaluators hired by one of the main individuals in speech technology (Loquendo, Turin), and it is found that the annotation of prosodic breaks, especially terminal ones, has a very high consensus; reliability is therefore high. An extensive report on this is found in the Appendix.
The C-ORAL-ROM resource
.. Utterance boundaries and labelling of discourse acts Traditional speech act theory has been considered insufficient for managing dialogue flow. Language activities for discourse management have been introduced, comprising devices such as turn-taking, repair, reference/information, and attention, which have been proposed for analysing concrete dialogues of everyday use. The first scheme for dialogue acts, based on the study of real conversations, was worked out by Sinclair and Coulthard (1975), and developed by Stenström (1994). In this tradition the dialogue exchange is described as consisting of dialogue transactions, dialogue turns, dialogue moves, and dialogue acts, which are defined in a coding scheme. Dialogue coding has been very useful in a range of research fields, especially in computational linguistics and language technology; consequently, some best practices have emerged.25 Some coding schemes have been developed to cover specific applications such as LINDA answer systems, as well as translation systems, such as VERBMOBIL (Alexandersson et al. 1997; Siegel 1997). Others are more general and are similar to that conceived by Sinclair and Coulthard (1975) in their analysis of classroom conversations. The main general schemes have been developed for Map Tasking (Carletta et al. 1997) and DAMSL (Discourse Representation Initiative). Given the relevance of this recent tradition for the annotation of speech corpora, the relation with the C-ORAL-ROM annotation of utterance boundaries must be explicitly considered. It is important to note that, once the relation between prosodic cues and speech act performances is recognised, the parsing of the speech flow into discrete speech events does not rely on the recognition by the labeller of a specific performed action. Instead, recognition of utterance limits is independently motivated. This is not irrelevant from the point of view of the annotation of dialogue acts. Although the coding scheme for dialogue acts provides a closed list of possible moves, a competent speaker may find it difficult to identify and define the performed act. The replicability of the coding scheme is, as a matter of fact, one of the main problems of the annotation of dialogue acts, even in quite restricted domains. Once the utterance limits are identified, the language string corresponding to an utterance is the linguistic entity which is suitable for receiving a certain tag. In other words, the definition of utterance limits is a matter of direct perception, while the assignment of a specific value to a dialogue act is a categorisation issue involving our knowledge of linguistic values. The annotation of dialogue acts parallels current labelling practices for the transcription of prosody in ToBI (Avesani 1996; Grice et al. 1996; Nolan & Grabe 1997): both these practices add qualitative labels to a language domain according to an annotation scheme. While the dialogue labeller selects a specific tag for discourse acts performed by the textual string in the dialogue structure, in ToBI, the labeller interprets a prosodic movement and tags it in a certain way.26
Massimo Moneglia
Based on these premises, it can be seen that the annotation of terminal and nonterminal breaks does not describe the prosodic movement that actually occurs in correspondence with a specific speech segment, but rather it selects the specific segment where, according to perception, a significant movement occurs. At the same time the annotation does not specify which proper speech act is performed by a sequence of words, but rather, specifies which sequence of words performs an act, for prosodic reasons. Clearly, the hypothesis underlying this scheme is that the previous annotation processes are quite different from one another. In short, while the segmentation of speech into utterances through terminal breaks is a matter of direct perception, the interpretation of intonation movements and the identification of dialogue acts mainly involves categorisation and understanding. For this reason, the three levels should be separated. The annotation of terminal and non-terminal breaks, however, specifies a common interface for both levels of annotation. Once the relevant domain for prosodic movements and speech acts is determined, this will probably allow a better interpretation of both the relevant prosodic movements and the functional, dialogical value of the speech event. The same consideration can hold for syntactic features. Utterances cannot be identified and defined on the basis of syntactic properties as clauses can, for instance, but once an utterance is identified on the basis of a terminal break, any kind of morpho-syntactic and lexical evaluation can be driven on it.27 In other words, the annotation of terminal and non-terminal breaks links the domain of prosody with the domain of speech act analysis and may be significant for improving the general state of the art on spoken text representation and analysis.
.. Utterance limits in spontaneous speech The pragmatic definition of speech events is not the only theoretical alternative that has been proposed. Besides the syntactic definition cited above, in current work on spoken language, a further alternative has often been used, namely, the practical definition of an utterance as the linguistic sequence between two silences (see TEI guidelines28 ). Such a criterion may sound preferable to the prosodic criterion adopted in C-ORAL-ROM, as it appears to be more objective, whereas the prosodic criterion may be considered arbitrary, as it is based on perception. In what follows, we will point out two empirical arguments strictly concerned with the analysis of spontaneous speech: (a) that syntactic structures appear frequently underdetermined, and that their definition is rather a function of prosodic cues; and (b) that the timing of the wave signal is not linguistically significant, as it is, at the same time, a criterion which is both too weak and too strong to determine the utterance boundaries in spontaneous speech corpora. The following is a dialogic exchange between a beautician and a client who is about to undergo a depilation of her legs. This text, taken from the Italian sub-corpus (file ifamdl15 of the multimedia corpus in the DVD), allows an empirical evaluation of the
The C-ORAL-ROM resource
difficulties encountered by both criteria. Let us consider the bare transcription (i.e. words only), accompanied by the essential contextual information: *EST: o vieni dai [come on then] %act: the beautician invites her client to begin the hair removal *CLA: a patire [to suffer] *EST: no ascolta qui sopra sì [no listen up here yes] %act: the beautician gets closer to the leg to be depilated *CLA: qui sì [here yes] The first thing to note is that the third and the fourth turns are verbless. For this reason their syntactic structure is underdetermined by the actual syntactic data and may be compatible with many possible interpretations. The following punctuations highlight eight possible dependency structures for the first case and three dependency structures for the second: No. Ascolta. Qui? Sopra? Sì. No. Ascolta qui sopra. Sì. No. Ascolta. Qui sopra, sì. No, ascolta. Qui. Sopra sì. No, ascolta. Qui sopra, sì. No. Ascolta. Qui. Sopra sì. No, ascolta. Qui. Sopra sì. No. Ascolta, qui sopra. Sì.
[No. Listen. Here ? At the top ? Yes.] [No. Listen here. Yes.] [No. Listen. Up here, yes.] [No, listen, here. Yes, on top.] [No, listen. Up here, yes.] [No. Listen. Here. Yes, on top.] [No, listen. Here. Yes, on top.] [No. Listen, (what about) up here? Yes.]
Qui sì. Qui, sì. Qui. Sì.
[Here, ok.] [Here, yes.] [Here. Yes.]
As the translations suggest, all the word groupings, which all correspond to distinct possible utterance boundaries, are consistent with the contextual domain. Therefore, neither syntactic nor contextual information are sufficient to determine the actual structure of the previous turns. In speech, however, these turns are in fact not ambiguous. Once the information borne by terminal and non-terminal breaks is perceptively recovered, the reference units can be determined with precision. The speech act labels in the dependent tier may help the reader to achieve the proper interpretation, as illustrated in the following: *EST: o vieni / dai // [come on then] %ill: invitation *CLA: a patire // [to suffer] %ill: ironical assertion *EST: no // ascolta / qui sopra ? sì // [no // listen / (what about) up here ? yes //] %ill: reassurance (1) question, introduced by a conative (2) self answer (3)
Massimo Moneglia
*CLA: qui ? sì // [here ? yes] %ill: question (1) answer(2) In other words, the index which determines the choice of possible structure for both turns under investigation is the prosodic structure, and not the reverse. This is a relevant property of the utterance that follows from the fact that each utterance accomplishes a distinct illocutionary act. This act defines the relevant domain for linguistic relations in the speech flow. Once two strings belong to two distinct utterances, then the linguistic expressions belonging to those strings do not share the same linguistic domain. They belong to two autonomous speech acts, they depend on two illocutionary forces, and they are the function of two distinct locutionary and prosodic programmes. In present approaches to multimedia spoken language archives, a “from silence to silence” criterion is the most common, probably because the automatic recognition of pauses in the speech flow is quite an easy task to be pursued, given the actual technology. This, in fact, at the outset, seems very reasonable, because utterance boundaries frequently occur together with significant wave interruptions. Figure 1.2 illustrates the third turn from the dialogue examined, where, after the utterance no, there is an interruption of approximately 600 ms (indicated in light grey) that marks the break before the beginning of the next utterance. This is not, however, as straightforward as it seems. To illustrate a different – and common – case, we examine the fourth turn of the dialogue, shown in Figure 1.3. The speech flow of this turn corresponds to two distinct speech acts (marked in black and light grey, respectively), but actually there is no pause separating the two speech events.
Figure 1.2 Pause before a new utterance
The C-ORAL-ROM resource
Figure 1.3 Two utterances not separated by pause
Figure 1.4 Pause within an utterance, after a Topic unit
The distinction is still possible because prosody perception appears to be sensitive even to the 20 Hz discontinuity occurring at the start of the second utterance. A converse case may also occur: a perceptively relevant prosodic break may be accompanied by a pause even if the break does not mark the end of the utterance. Figure 1.4 illustrates a typical utterance with Topic-Comment structure, taken from the same dialogue.29 The first element has prefix intonation and is perceived as nonconclusive, while the second string is conclusive (’t Hart et al. 1990): *EST: . . . lei / prima veniva tutte le settimane // [she / used to come every week once //]
Massimo Moneglia
The Topic unit is separated from the Comment unit by a pause (in light grey). If we assumed the silence to silence criterion, the sequence would be wrongly considered a sequence of two distinct utterances. In summary, the concept of utterance as a sequence between two silences does not match the identification of an utterance determined on a prosodic basis. Utterances may occur with no need for pauses between them, and therefore in such instances, it would be a case of too strong a criterion; conversely, the occurrence of a pause is not a sufficient cue to infer the conclusion of an utterance, hence making it too weak a criterion. A quantitative measurement of the incidence of both over-extension and underextension of the silence to silence criterion can be proposed on the basis of C-ORALROM data. The French corpus is tagged using both temporal and perceptual criteria. Pauses of more than 200 ms were detected automatically in the speech flow and annotated in the transcripts. At the same time the corpus was also tagged with respect to terminal and non-terminal prosodic breaks, as perceived by the expert operators who transcribed and tagged the corpus. On the basis of the results of this double tagging, we recorded that approximately 63% of sequences ending with a terminal break are accompanied by a pause, while 37% of sequences ending with a terminal break do not bear a pause. Conversely, approximately 42% of breaks that were considered non-terminal are also accompanied by a pause.
.. Prosodic labelling procedure and Alignment units Based on the empirical results above, the identification of utterance boundaries by means of prosody appears to be significant from a linguistic point of view; however it must also be stressed that it is in fact an easy annotation to be added to spoken corpora. Once the concept of prosodic break is clear to the transcribers, they are able to listen directly to the speech flow and use direct perception to annotate the terminal and non-terminal breaks in real time, together with transcription of the segmental information. The segmentation of speech into utterances is therefore simultaneous to both the transcription and the annotation of the main prosodic boundaries. Tagging consists of the marking of prosodic breaks in the orthographic transcription of speech. In C-ORAL-ROM each word boundary is considered a possible position for a prosodic break, but not within-word breaks. Each word boundary in C-ORAL-ROM transcripts thus necessarily has one of the following values: 1. no break 2. terminal break 3. non-terminal break The labelling is based only on perceptual judgments and in principle does not require any specific linguistic knowledge, although the notion of speech act is always familiar to the expert transcribers (comprising PhDs and PhD students) who annotated the corpus.
The C-ORAL-ROM resource
The process of prosodically tagging utterance boundaries in C-ORAL-ROM involves linking the annotation of the speech signal with the relevant linguistic information; in other words, alignment. The definition of the text-to-speech interface in C-ORAL-ROM is based on the idea that the access to acoustic information in a multimedia corpus must go hand in hand with the representation of prosody. Given that in C-ORAL-ROM all the textual information is tagged simultaneously with respect to prosodic parsing and utterance limits, each textual unit corresponding to an utterance can be easily and directly aligned to its acoustic counterpart, thus ensuring a natural and meaningful text-to-sound correspondence. Such a method can be proposed as a possible standard for storing oral language in multimedia and multi-modal language resources. In Section 1.4 the technical process and methods for text-to-speech aliment will be discussed. The following outlines the standard procedure adopted in C-ORAL-ROM: a. Tagging of prosodic breaks simultaneous with the transcription by a first labeller. b. Revision of tagging by a different labeller in connection with the revision of transcripts. c. Revision of tagging and specific challenge of terminal breaks by a third labeller in connection with the alignment. This process ensures control over inter-annotator relevance of tags and a maximum accuracy in the detection of terminal breaks. The accuracy with respect to non-terminal breaks is by definition lower.
.. Conventions for prosodic tagging in the transcripts Discriminating between terminal and non-terminal breaks is mandatory in all the C-ORAL-ROM transcripts. When a prosodic break occurs, the prosodic tagging is specified with a dedicated symbol taken from the CHAT tradition, as shown in Table 1.3. Each symbol represents a type of break which is defined in the annotation schema which permits a very restricted set of alternatives. C-ORAL-ROM allows the prosodic tagging to be displayed at two hierarchical levels, with greater or lesser attention to the annotation of the types of terminal breaks and to disfluency in the speech performance (Shriberg 1994). For terminal breaks, a // tag (double slash) is inserted in the transcription each time a prosodic break is perceived as terminal. At the richer level of transcription, three more types of terminal breaks, that appear to be extremely evident perceptually, can be specified, namely, interrogative utterances, intentionally suspended utterances, and interrupted utterances. Each of these is given a dedicated symbol, as listed in Table 1.3. At less detailed levels of transcription, interrogatives and intentionally suspended utterances are marked with the generic terminal break // tag, with no added specification. For non-terminal breaks, the / symbol (single slash) is inserted in the transcription to mark the internal prosodic parsing of a textual string which ends with a terminal
Massimo Moneglia
Table 1.3 Summary of prosodic break types in C-ORAL-ROM30 Symbol
Type description
// ? (optional)
Conclusive prosodic break Conclusive prosodic break such that the utterance has an interrogative value . . . (optional) Conclusive prosodic break such that the utterance is left intentionally suspended by the speaker + Conclusive prosodic break such that the utterance is interrupted by the listener or by the speaker himself / Non conclusive prosodic break [/] (optional) Non conclusive prosodic break caused by a false start [//] (optional) Non conclusive prosodic break caused by a false start (retracting) such that the linguistic material is only partially repeated [///] (optional) Non conclusive prosodic break caused by a false start (retracting) such that the linguistic material is not repeated
Break value Terminal Terminal Terminal Terminal Non-terminal Non-terminal Non-terminal
Non-terminal
break. This sign is inserted in the position where a prosodic break that is not perceived as terminal is detected in the speech flow.
... Fragmentation phenomena This annotation scheme embodies the generalisation that a prosodic break always occurs when a disfluency arises in the speech performance. This is true in the case of the interrupted utterances mentioned above, but it is also true in the case of false starts or retracting, which are probably the most frequent and interesting fragmentation phenomena of spontaneous speech (Blanche-Benveniste 2003). When fragmentation occurs, this is perceived as a clear break in the prosodic programme. Therefore a fragmentation never takes place without a simultaneous break in prosodic fluency. Two types of fragmentation can be noted: 1. Interruptions. An interruption (non-completion) of the utterance may be due to many reasons: a change of the linguistic programming by the speaker, an interruption caused by the listener, or by other events in the environment. Interruptions may be accompanied by word fragmentation, i.e. interruption before the end of the last word of the utterance, although the absence of fragmentation is the more frequent case. The interruption mark “+”31 is counted as a kind of terminal break, and is inserted in the transcription in the position of the interruption. The following is an example of interruptions within a dialogic turn: *MAR: è perché forse / non c’ è stata + dico / tu / che perdi il capo / che + perché / a lui / gli piacciono i soldi / Ida // [it depends perhaps / there has not been + I say / you / that lose your head / that + because / he / he likes money / Ida // ]
The C-ORAL-ROM resource
2. Retracting and/or restart and/or false start(s). Retracting (or false start) is the most frequent fragmentation phenomenon in spontaneous speech. The speaker hesitates while trying to find the best way to express himself and retracts his speech before choosing between two alternatives. This phenomenon is, generally, clearly distinguishable from interruptions or changes in programming, reported above, which do not feature speaker’s hesitations. Contrary to interruptions, retracting is almost always accompanied by the repetition (complete or partial) of the linguistic material, and clearly causes a loss of the informational value of the retracted material, which is abandoned by the speaker in favour of the chosen alternative. As in the case of interruption, in retracting phenomena, a change of prosodic envelope is again necessary. In other words, the retracting between two elements cannot be accomplished in the same prosodic envelope. Therefore, retracting is always accompanied by a prosodic break marked with the symbol “[/]”,32 illustrated in the following example: *MAX: mio &cug [/] mio fratello non nuota // [my &cous [/] my brother doesn’t swim //] Retracting breaks are considered a type of non-terminal breaks and are highlighted only at richer levels of transcription. The symbol is inserted in the transcription after each set of fragments, in the position where a restart begins. At poorer levels of transcription, retracting phenomena are not treated as a special kind of prosodic break caused by fragmentation, and only the generic non-terminal break sign is used after each set of fragments where a restart begins, as in the following Portuguese example: *FER: [2] / o / o touro / realmente / estava a / estava a dar cabo do cavalo // [the / the bull / actually / was to / was attacking the horse //]
3. Retracting/interruption ambiguity. In some cases it is hard to decide whether a fragmentation phenomenon fits with the definition of “restart” or “interruption”. This can be the case when an alternative to the locution in object is chosen, but no repetition is involved. In this case a supplementary sign, “[///]”, can optionally be used marking the fact that a retracting phenomenon probably occurs with neither partial nor complete repetition of the linguistic material. The ambiguous mark is counted as a non-terminal break.
. Textual format C-ORAL-ROM’s format can be defined as a relation between two independent levels: a textual format, in which each recorded session is transcribed, and a text-to-speech alignment format, which allows access to the acoustic data from the textual information. The principles which guide the transcription of spontaneous speech will be highlighted in this section, while the alignment method will be explained in Section 1.4.
Massimo Moneglia
C-ORAL-ROM’s textual format is an implementation of the CHAT architecture (MacWhinney 1994). Its structure comprises two main levels:
1. Metadata. Metadata contain all possible information on the recorded session, allowing its identification and localisation within the resource. Metadata are recorded as an ordered set of Headers lines placed before the session transcription and are compiled following explicit rules and terminology. 2. Dialogue representation. Sets of relevant linguistic and non-linguistic information embodied in the recorded session are presented in textual form. The dialogue representation principles allow the written representation of the spoken events and the relevant circumstances which enable their interpretation. Dialogue representation features two main levels of information: i.
Text lines: orthographic transcription of the speech events occurring in the recorded session, divided as follows: a. Vertically, in dialogic turns (introduced by a speaker label); b. Horizontally, by prosodic parsing and utterance limit, representing terminal and non-terminal prosodic breaks of the speech continuum.
ii. Dependent tiers: contextual information.
.. Metadata Metadata are listed in a closed set of types, comprising: a. b. c. d. e. f.
the context of the recorded session; its size; the speakers; the acoustic quality; the source; the person who can provide information on the session.
Each type of metadata is introduced by “@” immediately followed by a label, a colon “:”, and an empty space, and then followed with PCdata ending with “enter”. In order to allow better retrieval, PCdata refer as much as possible to a previously defined standard terminology (closed vocabulary). Table 1.4 summarises the various metadata types, and Tables 1.5 to 1.8 provide further details regarding the vocabulary and the formation rules. The Situation field describes the set of contextual information regarding the session and how it was recorded, as detailed in Table 1.6, and is represented as an ordered set of sub-fields separated by a comma.33 The Class field is an ordered set of sub-fields which mirrors the corpus design classification of each recorded session, as illustrated in Table 1.7.
The C-ORAL-ROM resource
Texts in the collections are labelled with respect to the acoustic quality of the sound source according to the criteria outlined in Table 1.8.34 Table 1.4 Metadata types Label type @Title: @File:
Description
One or two words in the object language which help to identify the text. Filename (note that the names of the audio and text files differ only in the extension which is not included in the Filename). @Participants: Three capital letters identifying each speaker, followed by the corresponding first name, plus a sub-field with an ordered set of information about the speaker. @Date: Date of the recording, in the form day/month/year; e.g. 20/06/2001. @Place: Name of the city where the recording session took place. @Situation: Ordered set of information: genre, role of the participants, surroundings, main actions performed, recording conditions; e.g. gossip between friends at home during dinner, not hidden, research participant. @Topic: The main topic dealt with in the speech event (max. 50 characters); e.g. traffic problems. @Source: Name of the collection leading to a copyright holder; e.g. LABLITA_CORPUS; CORPAIX. @Class: Set of fields identifying the session in accordance with the C-ORAL-ROM corpus structure, separated by commas. @Length: Length of the transcribed audio file in minutes (’) and seconds (”); e.g. 12’ 15”. @Words: Number of words in the text file. @Acoustic_quality: Acoustic quality of the recording labelled according to specific criteria (A, B or C). @Transcriber: Name of the person(s) in charge of the text, who can provide further information. @Revisors Names of the revisors. @Comments: Transcriber’s comments on the text. *
Massimo Moneglia
Table 1.5 Ordered set of sub-fields for each Participant Type
Description
Vocabulary
Sex
Sex of the speaker
Closed vocabulary. One of the following conventional capital letters, according to the description in brackets: M (Male) W (Female)35 O (non-defined; e.g. artificial devices)
Age
Age of the speaker
Closed vocabulary. One of the following conventional capital letters for each range of age between brackets: A (18–25) B (25–40) C (40–50) D (over 60)
Education
Level of education Closed vocabulary. One number according to the according to the degree degree of schooling between brackets: of schooling 1 (primary school or illiteracy) 2 (high school) 3 (graduates or university students)
Profession
Speaker’s occupation
Role
Role in the recorded Open vocabulary. Name of the role (e.g. father; event (even if it is the professor) same as the profession)
Geographical origin/linguistic influence
Name of the place of origin of the speaker
Open vocabulary. Name of the profession (e.g. professor; secretary; student)
Open vocabulary (e.g. Ile de France; Castilla)
The C-ORAL-ROM resource
Table 1.6 Ordered set of sub-fields for Situation Type
Description
Vocabulary
Genre
Information helping to define the genre of activity to which the linguistic event belongs.
Open vocabulary (e.g. gossip; chat; quarrel; discussion; narration; claim; etc.). The neutral case is “talk”. The information of the Class field (dialogue; conversation; etc.) is not repeated.
Reciprocal role
Reciprocal position of the participants.
Open vocabulary (e.g. friends, colleagues, relatives, citizens).
Ambience
Surroundings in which the recordings Open vocabulary (e.g. in a silent studio; took place. on the street; at home; in a shop; at school; in the office; etc.).
Action
Main action performed during the speech event (if any).
Recording conditions
Status of recording with respect to the Closed vocabulary. Choice between the observer’s paradox in spontaneous alternatives in each of the following two speech resources. sets: a. hidden vs. not hidden; b. participant researcher vs. observing researcher vs. researcher not present.
Open vocabulary (e.g. while ironing; during depilation; etc.).
Table 1.7 Ordered set of sub-fields for Class Informal Type
Type
Type
Formal Type
family/private public
formal in natural context
media
telephone
Sub-type monologue dialogue conversation
Sub-type political speech political debate preaching teaching professional explanation conference business law
Sub-type news sport interview weather forecast scientific press documentary talk show
Sub-type private conversation human-machine interaction
sub-sub-type monologue dialogue conversation
Massimo Moneglia
Table 1.8 Labels for acoustic quality Type
Properties
Label
Digital recordings
Digital recordings with DAT or minidisk apparatus and unidirectional microphones. a. Analogue recordings with good microphone response; b. Low background noise; c. Low percentage of overlapped utterances; d. F0 computing possible for most of the files. a. Low-quality analogue recordings with poor microphone response; b. Background noise; c. Medium percentage of overlapped utterances; d. F0 computing possible in many parts of the files.
A
Analogue recordings
Analogue recordings
B
C
.. Dialogue representation As introduced in the beginning of Section 1.3, for a given recorded session which comprises a continuous set of acoustic data, dialogue representation includes: (1) the notion of text, that is, the representation of a session in discrete written form; and (2) the notion of dialogic turn, which defines the maximal structural components of such a discrete unit. The textual representation of both sessions and dialogic turns must be consistent with the representation of the speech events that make up the dialogue itself. The definitions outlined in Table 1.9 provide a clear distinction between the relevant entities of dialogue structure and speech event representation, allowing a possible textual representation of dialogue. Table 1.9 Entities for dialogue representation Concept
Definition
Session Text
Set of dialogic turns corresponding to one metadata set. Each recorded session as an ordered collection of transcribed dialogic turns referring to a given metadata set. Dialogic Turn Continuous set of speech events performed by a single speaker’s voice. The dialogic turn changes if, and only if, a speech event by another speaker occurs. Utterance The minimal speech event by a single speaker such that it can be pragmatically interpreted as a speech act. Word A speech event perceived as a phonetic unit, such that it conveys a meaning.
Turn representation. A dialogic turn by one speaker is expressed by “*”, immediately followed by three capital letters identifying the speakers in the metadata, then followed by “:” and one space before the transcription of the speech event. Each dialogic turn ends with an “enter” (Table 1.10).
The C-ORAL-ROM resource
Table 1.10 Turn representation Convention*
Description
^\*[A-Z]{3}:\s{1}
Dialogic turn of a given speaker
* Expressed with regular expression notation.
Utterance representation. Each utterance is represented by a series of transcribed words, ending with the termination symbol “//” or other symbols having a terminal value such as “?” “+”; “. . . ” (see Table 1.3), as in the following example: *MOR: I’m going home // Are you tired too? *PIL: bye bye // A dialogic turn can also be filled with non-linguistic or paralinguistic material according to the transcription convention (see Table 1.12),36 as follows: *MAX: are you sure? *PIL: hhh //
Word representation. Each word is transcribed as a continuous sequence of characters between two empty spaces, in accordance with the orthographic conventions of a language. .. Conventions for transcription General criteria. The transcription of a dialogic turn expresses, horizontally, the sequence of speech events (i.e. utterances) that occur in each dialogic turn. In principle, no “enter” can occur within the sequence of utterances.37 Transcription follows the standard orthography of every language in the C-ORAL-ROM resource and is integrated with special signs devoted to handling spoken language phenomena, as detailed in Section 1.3.3.38 Overlapping. The speech of one speaker is, in spontaneous dialogues, frequently overlapped by the speech of another speaker, who may insert his dialogic turn in the first speaker’s turn. The overlapping therefore determines a relation of temporal equivalence between two or more speech portions in different dialogic turns. In C-ORALROM, overlapping is represented by the conventions summarised in Table 1.10 and elaborated on below. The overlapped text in both dialogic turns is placed between a pair of angled brackets, i.e.
. Moreover, the sign “[<]” may optionally appear at the beginning of the second turn, immediately before the overlapped text, to mark that an overlapping relation holds between the two pieces of text between brackets in the two turns, as illustrated in the following examples:
Massimo Moneglia
*ABA: María mi ha detto che [/] // [Maria told me that / //] *BAB: [<] <non viene> // [she is not coming //] In C-ORAL-ROM, overlapping is marked only when it occurs with at least two words in two different turns; this means that the overlapping of single syllables is left unmarked or reported to word boundaries. When, due to the simultaneous occurrence of more than one dialogic turn, it is impossible to attribute the speech to a speaker, a fixed variable is used to mark a mixedturn (see Table 1.11), e.g.: *XYZ: chi è va bene //39 [is who is it all right]
Spoken language noise phenomena. As mentioned in Section 1.3.2, transcription follows standard orthography of each language concerned, and is integrated with special signs for handling spoken language phenomena, as summarised in Table 1.12. Table 1.11 Symbols for overlapping Symbol
Description
<>
Brackets which mark the beginning and the end of the overlapped text of a given speaker. Symbol which specifies the overlapping relation between two bracketed textual strings belonging to two speakers. Turn of overlapped speech by non-identified speakers.
[<] *XYZ:
Table 1.12 Symbols for spoken language phenomena Symbol
Description
& hhh xxx yyy yyyy
Speech fragments Paralinguistic or non-linguistic elements Incomprehensible word Non-transcribed word Non-transcribed audio signal
Fragments. All incomplete words and/or phonetic fragments are immediately preceded by the ampersand symbol “&”, as in the following example of an incomplete utterance:40 *MAX: mio &cug [my &cous] or in the following lengthening of the programming time:41
The C-ORAL-ROM resource
*MOR: credo che si chiami / &eh / Giovanni // [I think they call him/ &eh / Giovanni //]
Paralinguistic elements. All paralinguistic elements (laughing, crying, etc.) are not counted as a word occurrence in a frequency list and are indicated as “hhh”. Further information of the element is explained in a dependent tier (see Section 1.3.6). Incomprehensible words. All words that are not properly understood are reported (and counted as word occurrences in a frequency list) as “xxx”. Non-transcribed words. When a word must be deleted for reasons concerning privacy or decency, it is substituted by a “yyy” variable, to be counted as a word, as in the following example:42 *MOR: il dottor yyy è un cretino // è proprio un bello yyy // [doctor yyy is an idiot // he’s a right yyy //]
Non-transcribed audio signals. When, for whatever reason, part of the audio cannot be transcribed, a single variable “yyyy” is inserted in the transcripts, which is not dependent on the length of the signal. The said variable may be subject to alignment, but will not be counted as a word.43 Interjections. Interjections are conventional phonetic elements with dialogic functions and low lexical or grammatical meaning. Interjections are not fragments and are transcribed in accordance with the lexicographical tradition of each Romance language. New interjections discovered in the corpus are transcribed tentatively and their presence is reported in an added glossary. Non-standard words. In the absence of any previous orthographic tradition, nonstandard regional expressions found in the corpus are transcribed tentatively, normalised according to the graphical conventions in use in each language’s tradition, and reported in an added glossary. .. Pauses Pauses in the speech flow are indicated with “#” and are included in the transcription only if clearly perceived as a significant interruption of speech fluency (Table 1.13). Other pauses can be reported optionally. No distinction is made with respect to the length of the pause. In the French corpus pauses are detected automatically (see Chapter 3). The “#” symbol is not a sign of prosodic parsing and never substitutes for the marks dedicated to prosodic breaks.
Massimo Moneglia
Table 1.13 Symbol for pauses Symbol Description #
Definition
Pause in the speech flow A perceptively relevant silence in the speech continuum, or, in any case, a silence longer than 250 ms.
.. Restrictions on dialogue representation: The Intersection convention In spontaneous spoken language the event of an intersection of dialogic turns by different speakers occurs frequently; this means that a dialogic turn may arise before the end of the turn that immediately precedes it. In those cases, the representation of a dialogue as a vertical ordered collection of dialogic turns cannot be maintained. Because the CORAL-ROM format forces to represent the sequence of turns in a temporal order, an Intersection convention has been adopted.
Intersection convention. In the C-ORAL-ROM transcripts, a slash placed at the beginning of a turn, i.e. immediately after the turn label and before the transcribed text, is a convention expressing that the turn in question is only virtual. Although the transcribed language strings follow the dialogic turn of another speaker, they still belong to the preceding turn of the same speaker.44 Three major cases of intersection have been detected in C-ORAL-ROM: 1. Simple intersection of turns; 2. Intersection of turns due to interruption by the listener; 3. Intersection of turns due to overlapping.
Intersection of turns. In some cases, a complete intersection of dialogic turns may occur. In the following example, ELA, without actually interrupting the speaker, starts a brief dialogic turn, while MAX goes on with his turn: *MAX: in linea di principio / voi dovreste / [in principle / you had /] *ELA: ah // *MAX: / seguire le regole di trascrizione // [ to follow the transcription rules //]
Interruption and intersection. The speaker frequently interrupts his utterance in connection with the intervention of the listener. However, despite this interruption, he still continues the same turn, but with another utterance. For example, in the following situation, a speaker inserts himself in a dialogic turn, interrupting it, but the original speaker goes on with his turn despite the interruption: *MAX: in linea di principio / voi dovreste + [in principle / you had +] *ELA: ah // *MAX: / dovete seguire le regole di trascrizione // [you had to follow the transcription rules //]
The C-ORAL-ROM resource
Overlapping and turn intersection. Overlapping, which is very frequent in spontaneous speech, is a relation between texts belonging to different turns. In most cases, overlapping also causes interruption and if the speaker goes on after the interruption, this also causes intersection of turns. In order to capture this generalisation, overlapping is always annotated following the Intersection convention.45 In C-ORAL-ROM, when the overlapped text in the upper turn continues after the overlapping, the intersection convention has been applied in two alternative ways: a.
by transcribing over the overlapped turn until the first terminal break, as follows: *ABA: Maria mi ha detto che [/] più / al concerto // *BAB: [<] <non viene> // *ABA: / perché non si sente //
b. by interrupting the transcription of the overlapped turn when the overlapping ends, as follows: *ABA: Maria mi ha detto che [/] *BAB: [<] <non viene> // *ABA: / più / al concerto // perché non si sente // Only the first alternative compels the system to assume the generalisation that each formal turn ends with a terminal break, and allows the alignment of each utterance. Table 1.14 Intersection convention Configuration of symbols Operational definition :/
Relation that converts the turn in which it appears into a linear sequence which includes the preceding turn of the same speaker.
.. Dependent lines Information of three kinds, regarding the text reported in a dialogic turn, is optionally given in lines following a dialogic turn. Those lines, that are not computed as linguistic information are usually referred as dependent lines in the CHAT tradition: a. Alternatives proposed for the transcription of the text; b. Comments regarding the pragmatic context and the visual modality of communication; c. Other comments. These options are summarised in Table 1.15. The link between the information in the dependent line and a word, or series of words, in the dialogic turn can be specified using a serial number which indicates the position in the dialogic turn of the word referred to.46 In the following example, the reported alternative concerns the third word of the utterance
Massimo Moneglia
*MAX: voglio mangiare // pasta // [I want to eat // pasta //] %alt: (3) basta [stop] Table 1.15 Dependent lines types Sign
Definition of the type of information
%act: %sit: %add: %par: %exp: %amb: % sce: %alt: %com:
Actions of a participant while speaking Events or state of affairs occurring in the speech situation Participant to whom the speech is addressed Speaker’s gestures or paralinguistic aspects Explanations necessary for the understanding of the turn, or signs in the text (hhh) Description of the Setting in a media emission Description of the Scene in a media emission Alternative transcription Transcribers’ comments
.. Alignment principle The Alignment of C-ORAL-ROM is based on two choices: a. Specification of the Alignment unit at utterance level (as previously defined). b. Rough equivalence between terminal breaks and utterance limits. Each text is aligned with respect to perceptively relevant terminal breaks annotated in the original transcripts, in order to deliver a database of utterances.47 The alignment of transcribed texts is achieved through the assistance of the alignment tool by means of insertion in the text of a ($) tag after a terminal break, while the audio is played (at a reduced speed). The system assigns two temporal units to each alignment unit: a.
end of the alignment unit: the temporal unit of the sound file in the instant in which the tag is inserted; b. beginning of the alignment unit: the temporal unit which marks the end of the previous segment. The expert operator in charge of the alignment (a postdoctoral researcher or doctoral student) always considers whether the accomplished alignment unit truly corresponds, in his perception, to a speech segment ending with a terminal break. The operator may add or delete the terminal breaks annotated in the original according to his personal perceptual perspective of the speech signal, thus improving the quality of the annotation.
The C-ORAL-ROM resource
.. Files, filename conventions and folder structure of the resource ... Files 1. Audio Files: encrypted MP3 files; sampling frequency: 22,050 Hz; 16 bits 2. Text Files: encrypted .TXT files 3. Alignment files: encrypted .XML files
... Filename conventions The filenames of the C-ORAL-ROM resource bear information of three types, in the following order: a. The represented language; b. The text type; that is, the field and sub-field to which each text belongs in the corpus structure; c. The serial number identifying each text in its sub-field. The following are the conventions adopted in each C-ORAL-ROM language collection: a.
Language Country code: F (French), I (Italian), P (Portuguese), E (Spanish) b. Text type This is detailed in Table 1.16. Table 1.16 Codes for text type Field
Sub-field
informal
fam (family-private) pub (public)
mn dl cv
(monologue) (dialogue) (conversation)
formal
nat (natural context)
ps pd pr te pe bu co la nw mt in rp sc sp ts pv mm
(political speech) (political debate) (preaching) (teaching) (professional explanation) (business) (conference) (law) (news) (weather forecast) (interview) (documentary) (scientific press) (sport) (talk show) (private) (man machine)
med (media)
tel (telephone)
Philippe Martin
c.
Text identification number A two-digit number is used to identify the text within the sub-field; e.g.: efamdl01 (Spanish, Family-private, Dialogue, 01) efamdl02 (Spanish, Family-private, Dialogue, 02) efammn01 (Spanish, Family-private, Monologue, 01)
... Folder structure In the multimedia corpus, the files of each language collection are delivered within the same folder structure, which directly mirrors the C-ORAL-ROM corpus design. For each session, the following are delivered into folders: a. the encrypted MP3 file; b. the encrypted .TXT file of the transcripts; c. the encrypted .XML file defining the text-to-speech alignment in WinPitch Corpus format. The encrypted .TXT files of the transcripts and the encrypted .TXT files of the PoS tagged transcripts delivered two non-structured directories for each language collection.
. WinPitch Corpus. A Text-to-Speech Analysis and Alignment Tool Philippe Martin Although putting a speech corpus together seems to be a simple task involving speechto-text transcription and alignment, this ceases to be the case when the volume of data becomes large. Problems inherent in transcription, even for well-recorded speech data, are not trivial (Blanche-Benveniste 2002), and, even with the use of modern signal editing software, the manual transcription and alignment of just one hour of data quickly becomes cost-prohibitive. The development of adequate and user-friendly tools is thus essential for the elaboration of large speech corpora. WinPitch Corpus, which has been used in the C-ORAL-ROM project and is available in the DVD accompanying this book, is one of these tools, developed to facilitate the alignment process, and to render the analysis of very large corpora an easier and more pleasant task.
.. Text-to-speech alignment Text-to-speech alignment establishes a bi-univocal relationship between units of speech and units of text. In its simplest implementation, each unit of text (be it a syllable, word, phrase or sentence) receives a temporal index corresponding to the time position of its equivalent in the sound file. Once this process is done, an operator can select an aligned unit of text and listen to the corresponding speech segment. The acoustic analysis of the speech sound, such as melodic curve and spectrogram, can also be displayed at the same time. Conversely, the selecting of a speech seg-
The C-ORAL-ROM resource
ment will highlight the corresponding segment of text in its orthographic or phonetic transcription. Text-to-speech alignment is frequently used in multimedia language learning software, where the user can easily listen to the sound corresponding to a specific word or sentence merely by clicking on the appropriate text segments. Other important applications are found in fundamental research in phonetics, in synthesis-by-rule developments, or in speech recognition validation tests. In prosodic research, for example, the user should be able to quickly get an analysis display of the acoustic prosodic parameters simply by highlighting a part of the text of interest.
... Automatic alignment Emerging automatic text-to-speech alignment processes present recurrent limitations: 1. Their performance depends on speakers’ voice characteristics, which cannot be too different from the models adopted in these methods. 2. The recording signal-to-noise ratio must be high enough to reduce the error rate to an acceptable level. Radio and television broadcasts generally meet this requirement, but spontaneous speech recordings made in various public environments (street, public transportation, etc.) rarely fulfil this criterion. Echo in the speech signal is another aggravating factor. 3. The overlapping of speakers’ voices, frequently found in spontaneous dialogues, constitutes an additional aggravating factor. Furthermore, use of recordings for phonetic and general linguistic research requires an acceptable quality of the recording itself, such as a good frequency response curve and a low phase distortion. All these considerations seem to indicate that a human operator is often required to obtain a reliable text-to-speech alignment. The responsibility for all the problems mentioned above is then transferred to the operator, who, with appropriate and ergonomically well-designed tools, should perform better than by manually correcting, one by one, the errors made by an automatic system of alignment.
... Alignment and transcription Text-to-speech alignment can be executed in two modes, depending on whether the text pre-exists or not. If the text must be created, the operator proceeds by selecting segments of speech in sequences (which can be played back at reduced speed to enhance intelligibility) and typing the corresponding text perceived, either orthographically or in phonetic transcription (WinPitch Corpus allows the use of any font defined in Unicode). This is illustrated in Figure 1.5. During this process, a database is automatically updated, and will be later saved directly in Excel® or XML format.
Philippe Martin
Figure 1.5 Simultaneous transcription and alignment. The user sequentially defines segments of speech and enters the corresponding text
... Computer-assisted alignment Experimental studies have shown that improved coordination between visual spotting of words and positioning of a mouse on a computer screen can be obtained by slowing down speech playback by a suitable factor, depending on the size of the text object to spot. Larger chunks of text require less processing time, and thus allow a faster speech playback rate in the process. WinPitch Corpus-assisted text-to-speech alignment is based on this principle. The pre-existing text is displayed dynamically in a window while the corresponding speech sound is played back at a slower speed. This can be adjusted continuously on the fly by a factor ranging from 1/7 to 2. At each identification of a speech unit to segment and align (whether a syllable, word, phrase or sentence), the operator clicks with the computer mouse on the text segment perceived. The programme records the position of the cursor on the text window, which defines the end of the text segment to align, and the time of the click, remapped on the real time scale of the speech wave. This process continuously generates a database of pointers linking segments of text and segments of speech. This is shown in Figure 1.6. Various tools are provided to backtrack, fine-tune speech segment limits (with the help of a displayed spectrogram), dynamically modify limits of overlapping voices, etc.
The C-ORAL-ROM resource
Figure 1.6 Assisted alignment by slowing down speech playback. At each mouse click on a unit of text perceived at slower speed (top right window), bidirectional pointers are generated automatically between the corresponding speech segment (bottom right window) and a database (left window)
.. WinPitch Corpus features a. Slowing down speech. Variable-rate speech playback is the central engine of the assisted text-to-speech aligner. It uses a modified version of the PSOLA algorithm (Moulines & Charpentier 1990) and allows high quality re-synthesis of natural speech. This quality is strongly dependent on precise pitch marking, and therefore on reliable fundamental frequency analysis. Errors in pitch marking (e.g. missing markers, double markers) induce an echo effect due to the misalignment of pitch chunks added in the PSOLA algorithm. F0 estimation is obtained by the spectral comb method (Martin 1981), and the speed playback factor can vary from 1/7 to 2 (the latter meaning that speech is played back at double speed). This rate is dynamically adjustable by the user while the alignment is processed, allowing operations on very large files, and continuous speed adjustment as required by the operator. b. Automatic layer assignment. Pre-existing text can be organised, following a simple convention for naming speaker turns, so that segments are automatically assigned to their corresponding layers. The user does not have to worry about speakers’ turns while executing the aligning, as the programme will put segmented text in the appro-
Philippe Martin
Figure 1.7 Fine-tuning of speech segment limits with the help of a simultaneously displayed spectrogram (which allows precise segmentation in case of speakers’ overlapping)
priate layer assigned to each speaker. Eight layers are presently available, but future extension will provide for an unlimited number of speakers. c. Fine-tuning and speaker overlap. Once the assisted alignment session ends, the programme automatically displays the text under the corresponding speech segments, represented by their acoustic analyses (spectrogram, intensity and melodic curves, waveform). The user can then more precisely adjust the segments by dragging their limits using the mouse, with the help of visual inspection of the corresponding spectrogram or other available acoustic information. Overlapping segments can then be easily and precisely defined. This is illustrated in Figure 1.7. d. Unicode font. WinPitch Corpus is Unicode compliant, which means that any Unicode-defined font can be used for transcription, alignment and labelling. Furthermore, a right-to-left real-time display is provided for languages such as Arabic or Hebrew. e. Input formats. The programme uses most multimedia file formats, such as wav, mp3, aiff, au, wma, snd for audio, and mpeg, mpg, m1v, mp2, mpa, asf, avi, wmv for video. The user can select a time window (limited only by the available RAM size) to store the audio information from a selected time reference of the video or audio file. When loaded, audio and video channels are separated, and only the audio channel stays resident in memory to be processed and analysed as a simple wav file. Very large
The C-ORAL-ROM resource
files (limited only by the hard disk capacity) can thus be handled, transcribed, aligned and analysed. f. Hierarchical transcriptions. The available transcription layers can be assigned by the user to different transcription entities. In the C-ORAL-ROM project, for instance, each layer is assigned to a particular speaker in the dialogue, whereas in other applications, some layers can be assigned to syntactic tagging, syllabic segments, etc. In such cases, the segmentation and alignment are hierarchical, and the hierarchy is automatically maintained in fine-tuning operations. This means that the corresponding limits of segments of various levels (phones, syllables, words, syntactic groups, etc.) are adjusted together automatically if any of the limits is moved. This ensures a proper synchronisation of all levels of the transcription. In addition, this feature can be turned off when required, for instance, when layers are assigned to different overlapping speakers. g. Highlighting and marking. Speech segments can by highlighted individually in any colour, and can also be marked for sampling and statistical analysis. Marked segments are directly sampled, with a constant time interval or a constant number of samples, and their values exported to Excel® without further user intervention. h. Documents produced after alignment. The documents produced after alignment are defined in XML format (see Figure 1.8). i. Multimodal alignment with WinPitch Corpus. The multimodal alignment with WinPitch Corpus allows the synchronisation of acoustic and video information (F0 , waveform, time) (see Figure 1.9).48
Figure 1.8 The resulting text-to-speech alignment defined in XML format
Philippe Martin
Figure 1.9 Simultaneous display of acoustic and video information of a multimedia file
.. Basic layout WinPitch Corpus has three main windows, as shown in Figure 1.10: a.
A command window, displaying main command dialogue boxes, accessed from the toolbar, the operation menu or the dialogue box tabs; b. A navigation window, for selection of speech segments in the sound file; c. An analysis window, displaying the graphic representation of the selected speech part acoustical analysis. WinPitch Corpus operates with an XML file to define an alignment between a speech file (extension *.wav or *.mp3) and a text file (extension *.txt or *.rtf), the *.rtf file supporting Unicode text. Loading the alignment file (*.xml) will automatically load the corresponding sound and text files, whereas a new alignment session requires loading the sound file first, followed by the text to be aligned. To load an existing file, get the Transfer dialogue and click on Load in the Align (*.xml / *.alg) section, as shown in Figure 1.11. Loading a file will overwrite any existing data on the current analysis window. A new window does not have to be created when a file is loaded. Another, and often more convenient, way to load a file (*.wav, *.mp3, or *.xml) is by dragging its file name from the File dialogue to the analysis or navigation window, as illustrated in Figure 1.12. To drag a file, position the cursor on its name, press the mouse left button, move the cursor with the mouse keeping the left
The C-ORAL-ROM resource
Figure 1.10 WinPitch Corpus basic layout
Figure 1.11 Set of buttons for alignment
Figure 1.12 Dragging file names
Philippe Martin
button down, and release the button when the cursor is anywhere inside the analysis or navigation window. The File dialogue displays the hierarchy of the storage devices (diskette, hard disk, CD-ROM, etc.) and operates like Windows Explorer, but shows only directories and files that can be processed by WinPitch Corpus. Each type of file is displayed with a specific icon, as follows: for sound files (*.wav): for alignment files (*.alg or *.xml): and for WinPitch files (*.wp2): This scheme allows for quick loading by dragging, and the analysis of files belonging to the same directory, as the file already loaded in WinPitch is overwritten by the newly loaded file. Once a sound file has been loaded from memory, it can be navigated to retrieve sound sections to listen to and analyse. Sound files which are simply recorded and not saved can also be navigated and analysed (but it is of course wiser if the data are saved). Any of the following processes can be used, alone or in combination, in navigation:
1. Block selection. A “block” is defined in the navigation section as follows: Position the cursor on the starting edge of your block, press the mouse left button, move the cursor with the left button down, and release the button when the cursor is positioned on the ending edge of the block. When the mouse button is released, the analysis window is automatically updated to reflect the analysis of the time frame defined by the block. This is illustrated in Figure 1.13.
Figure 1.13 Navigation block analysis
The C-ORAL-ROM resource
The left and right edges of a block can be dragged to change the block position and size. To move a block edge, position the cursor near the chosen edge. The cursor will take a double arrow shape, indicating that it can be dragged to its new position. To do this, press the mouse left cursor and keep it down until the desired position has been reached. The block can also be moved without changing its size, by positioning the cursor inside the block until it takes a quadruple arrow shape. The whole block can then be dragged to its new position. The navigation window can also be panned by reaching either the left or right edge by moving the block. To remove a block, click with the left mouse anywhere outside it in the navigation window, or press the right mouse button.
2. Navigation window zoom and pan. Zooming in and out is possible using the left slider in the navigation window, and panning left and right is done with the top horizontal slider of the navigation window. The vertical slider on the right side adjusts the amplitude of the waveform displayed, but it has no effect on the actual wave amplitude in the sound buffer. The check box on the bottom right side is used to reverse the phase of the waveform; this also does change the phase of the sound wave in the sound buffer. These are all shown in Figure 1.14. 3. Analysis window zoom. The analysis window can be enlarged virtually up to 10 times by adjusting the click on the
value in the Play dialogue. To get the Play dialogue,
icon on the toolbar or click on the Play
Figure 1.14 Zooming, panning, amplifying waveform, and phase checkboxing
Philippe Martin
dialogue tab. This mode generates a virtual enlarged analysis window, which can be panned with the Analysis horizontal top slider. This is illustrated in Figure 1.15.
4. As shown in Figure 1.16, a block in the analysis window can also be defined, in order to select a sound segment, or for various other functions (e.g. to define highlighted
Figure 1.15 Zooming pan slider and pan level
Figure 1.16 Blocking in the analysis window and in the navigation window
The C-ORAL-ROM resource
sections, to save an aligned segment, etc.). Each time an analysis block is defined or edited (i.e. when the edges of an entire block are moved), the corresponding sound is played back, with variable speed if the Enable slow block is checked.
. C-ORAL-ROM PoS tagging C-ORAL-ROM also provides a textual entry of the transcripts where each form is tagged with respect to the part of speech (PoS) and lemma. It should be noted at the outset that PoS tagging of a multilingual spontaneous spoken resource is still a new task and that few antecedents can be found in the literature concerning the automatic PoS tagging of spontaneous speech.49 The development of multilingual tagsets for Romance languages, the integration of traditional tagsets with PoS for spoken language phenomena, the development of multilingual language technologies for automatic PoS tagging, and the level of encoding and the manual revision standards are, consequently, still issues which need to be explored further, and resolving them completely cannot, unfortunately, be properly addressed in the C-ORAL-ROM project because of the constraints of both time and economy. A specifically dedicated project would have been necessary to this end. Nonetheless, the consortium felt that the state of the art of automatic PoS tagging reached by the present HLT (Human Language Technology) for each Romance language would allow significant annotation of the C-ORAL-ROM transcripts with PoS with limited financial effort. This would ensure the exploitation of the resource for syntactic purposes, as well as enhance the level of awareness regarding PoS tagging of spontaneous speech. Moreover the current practices of PoS tagging in each team were already strongly based on common EU standard tagsets already established in previous projects (LINGUA-EAGLES-MULTEXT-PAROLE) and recently revised or adapted for current research. Thus, although the full integration of the four resources under a common PoS tagging scheme is not possible, the automatic PoS tagging of each resource ensures qualities that are sufficient for the exploitation of the resource. More specifically, C-ORAL-ROM ensures: (1) the use of a common format; (2) comparability of the adopted tagsets; (3) a minimal threshold of syntactic information sufficient for linguistic research; (4) an evaluation of the performance of the automatic PoS in spoken resources with respect to written resources; and (5) a better understanding of the specific difficulties that spontaneous speech presents for PoS tagging, from both a technological and theoretical point of view. In the remainder of this section, the tagsets used in each PoS tagged resource are compared and the C-ORAL-ROM tagging format presented. The specific choices adopted and the results obtained for each language are explained in the description of each resource.
Massimo Moneglia
.. Minimal tagset requirements C-ORAL-ROM’s four sub-corpora are tagged using different tagsets because of the differences in both the languages and the tools adopted for natural language processing systems. Nonetheless, in order to ensure comparability within the whole corpus, a compulsory minimal threshold of information has been established in the tag codes; such a minimal tag set is defined as follows: 1. Part of Speech (PoS) tag is compulsory for each class of word and, if applicable, for Locutions. 2. Specifications on Mood and Tense features are compulsory for Verbs, or, alternatively, the Finite/Non-finite character must be specified [± person] (necessary for the detection of verbless utterances). 3. Specification of the Coordinative/Subordinative quality is compulsory for Conjunctions. 4. Specification of the Common/Proper trait is compulsory for Nouns. The common level of morpho-syntactic tagging features a distinction, within the set, of elements which are not strictly linguistic but which are peculiar to spontaneous speech texts,50 providing two special tags to this aim (with the actual code of the tag depending on the specific tagset) for: a. paralinguistic elements: mh; he; ehm (used for filled pauses); b. extralinguistic elements: hhh (used for laughs and coughs).
.. Comparison between tagsets ... PoS tagsets The main PoS categories related to the tagsets used for each language are listed in Table 1.17. The first column comprises the maximal projection of the word categories present in the four tagsets; in the remaining four columns, the projection is mapped on the tagset of each language, showing the labels used. As far as the main PoS tags are concerned, as seen in Table 1.17a, the four tagsets are completely coincident with respect to the individuation of the main lexical categories (Nouns, Verbs, Adjectives, Adverbs), the functional particles (Prepositions, Conjunctions) and the Interjection class. This ensures a common baseline of annotation that makes the comparison of the high-frequency words of the corpora and the main lexical indexes extracted from the tagged resource reliable. Looking at the different C-ORAL-ROM tagsets in both parts of Tables 1.17, it can be seen that the maximum projection of categories is almost completely represented by the French tagset, which follows the EAGLES specifications on morpho-syntactic annotation of corpora.51 The different choices adopted by the other tagsets are justified by either the different language-specific tools for the automatic tagging procedure, designed with different perspectives on PoS classes and macro-categorisation, or by the differences between the languages. More specifically, the Portuguese and Spanish
The C-ORAL-ROM resource
tagsets include the speech-specific categories of discourse markers, used for the tagging of words with a pragmatic value at discourse level (e.g. mira, vamos, Sp.; pá, portanto, Por.). The Portuguese tagset also uses a specific label for the emphatic elements (e.g. lá, cá) that occur in the speech flow. The second part of the comparison table, Table 1.17b, shows the more complex relations among the different tagsets. The treatment of articles in Romance languages exhibits only a minor difference in categorisation. From a distributional point of view, articles in Romance languages are determiners; this is represented both in the French and in the Spanish tagsets, which have a “definite determiner” class for the articles, while the Italian and the Portuguese tagsets opted for an independent class, in accordance with grammatical tradition. This difference does involve the structure of the tagset: where articles are univocally identified as definite determiners, they constitute a sub-category; where they are identified Table 17a. Synopsis of PoS tag sets (1) Tagset projection
French
Italian
Portuguese
Spanish
nouns verbs adjectives adverbs prepositions conjunctions interjections discourse markers emphatic
NOM VER ADJ:QUA ADV PRE CON INT
S V A B E C I
N V ADJ ADV PREP CONJ INT MD ENF
N V ADJ ADV PREP C INT MD
Table 17b. Synopsis of PoS tag sets (2) Tagset projection
French
Italian
Portuguese Spanish
articles
DET:DEF R ART (definite determiner) (articles) (articles)
DETd (definite determiner)
demonstrative determiners DET:DEM demonstrative pronouns PRO:DEM
DIM
DEM
DETdem
possessive determiners possessive pronouns
DET:POS PRO:POS
POS
POS
DETposs
personal pronouns clitic
PRO:PER
PER
PES CL
PPER
rel-int-excl determiners rel-int-excl pronouns
DET:INT PRO:RIN
REL
REL
PR
indefinite determiners indefinite pronouns numbers (cardinals) numerals (ordinals)
DET:IND PRO:IND NUM ADJ:ORD
IND
IND
Q (quantifiers)
N NA
NUMc NUMo
Massimo Moneglia
as proper articles, they constitute a main category. However, this does not decrease the comparability among the tagsets. The main differences between the tagsets involve pronouns, determiners and pronominal adjectives.52 The treatment of such a lexical section was faced through different annotation schemes in the various initiatives which dealt with PoS tagging in a cross-linguistic perspective.53 In this respect, the different languages in C-ORAL-ROM cannot be described in the same way. For example, possessives in Italian are pronominal adjectives, but they cannot be determiners, in complementary distribution with articles or demonstratives; in French, however, possessives can be both pronouns and determiner (they also can be adjectival elements).54 In this case, differences among the tagset structures reflect different linguistic behaviours among languages. In C-ORAL-ROM, the class of personal pronouns is treated in the same way in all the tagsets, with the exception of the Portuguese tagset where clitic pronouns are separated with a specific tag. The French tagset bundles all the semantic categories of demonstratives, possessives, indefinites and relatives into two distributional macro-classes, determiners and pronouns, which are then divided into semantic sub-categories. The Italian and Portuguese tagsets do not distinguish these four types of lexical items by their functional value (pronominal or adjectival), but rather have each of them in single categories. The Spanish tagset is mostly coincident with the Italian and Portuguese ones for demonstratives, possessives and relatives. Where it differs is in the merging of indefinites, cardinals and ordinals into a single quantifiers class. This choice underlines the fact that, from a functional perspective, indefinites and numerals are quantifiers. From a distributional point of view, however, while cardinal numbers and indefinites have similar characteristics, ordinals do not occur in the same position, and instead present a distribution similar to that of adjectives. This behaviour is represented in the French tagset through an adjectives macro-class, which is split into qualifiers and numerals, while numbers and indefinites are treated separtately. The Italian tagset has separate classes for cardinals (Number) and ordinals (Numeral Adjective), while the Portuguese tagset has a single numerals class which is split into two sub-classes of cardinals and ordinals
... Morpho-syntactic features of verbs As can be seen in Table 1.18, the different tools used for the four languages encode the morphological traits of verbs at different levels. The Italian and Spanish tagsets provide a complete description of the verbal forms, giving specifications of mood, tense, person, number and, for the inflected participles forms, gender. The French and Portuguese tagsets offer information only on mood and tense. The last feature in Table 1.18 has to do with the annotation of auxiliary forms of verbs. The French tagset is the only one that does not contain the distinction between main and non-main use of verbs; conversely, information that the verb is auxiliary is specifically indicated in the Spanish and the Portuguese tagsets. Moreover, in the Portuguese tagset the adjectival past participles receive a special label (PPA), to distinguish
The C-ORAL-ROM resource
Table 1.18 Synopsis of the morpho-syntactic encodings for verbs French
Italian
Portuguese
Spanish
MOOD
indicative subjunctive conditional imperative infinitive participle
indicative subjunctive conditional imperative infinitive participle gerund
indicative subjunctive conditional imperative infinitive participle gerund
TENSE
present past imperfect future
present past imperfect future
indicative subjunctive conditional imperative infinitive participle gerundive adjectival past participle present past imperfect plusperfect future
PERSON
NUMBER GENDER (only for participles) VERB TYPE
first second third singular plural masculine feminine common main non-main
present past imperfect future first second third singular plural masculine feminine
main auxiliary
main auxiliary
these uses from the ones within the compound tenses (VPP). As for the Italian corpus, it distinguishes between the uses of verbs essere (‘to be’) and avere (‘to have’) as main and non-main ones (the latter includes auxiliaries and copulas).
... Non-standard tagsets The non-standard phenomena that occur in spoken language need specific labels in the automatic tagging procedure. Table 1.19 shows a comparison between the different phenomena encoded in the four corpora. The Italian, Portuguese and Spanish corpora mostly agree in the encoding of the main phenomena regarding paralinguistic elements (mostly pause fillers, phonetic supports or word fragments) and extralinguistic ones (laughs and coughs or other non-linguistic sounds), as shown in Table 1.19a. The French resource, however, does not provide a complete account of such phenomena, and does not mark the extralinguistic elements found in the transcripts in the tagset. Table 1.19b, which presents further non-standard phenomena, shows that the Italian non-standard tagset is the broader one, encoding facts of different nature (see Section 2.3.3 for the description of such phenomena).
Massimo Moneglia
Table 19a. Synopsis of the non-standard tagsets (1)
extralinguistic paralinguistic
support & fillers fragments
French
Italian
Portuguese
Spanish
–
XLG
EL
– –
PLG PLG
PL FRAG
<sup> –
Table 19b. Synopsis of the non-standard tagsets (2)
foreign words new formations acquisition forms onomatopoeia meaningless forms euphonic particles non understandable words
French
Italian
Portuguese
Spanish
XXX:ETR – – – – XXX:EUP –
(POS) + K (POS) + Z ACQ ONO
ESTR – – – – – Pimp
– – – – – –
X
.. Tagging and frequency lists formats The output of the morpho-syntactically tagged text represents both the speaker codes and the prosodic breaks, in order to allow context-bound grammatical studies (i.e. within utterances or tone units contexts). The text is given horizontally, with respect to the legibility of dialogic turns. The format of the tag for each lexical entry features three elements divided by backslash separators (\), as follows: a. the first element is the word-form; b. the second element is the lemma (in capital letters); c. the third element is the code which includes the PoS category (in capital letters) and, optionally, the morpho-syntactic description (msd) of the word form. The result is a pattern with the following structure:55 wordform\LEMMA\PoSmsd e.g. *CAR: come\COME\B andò\ANDARE\Vs3ir / a\A\E casa\CASA\S vostra\VOSTRO\POS ? (Italian text) The output of the frequency lists is encoded in a standard form, to ensure the highest comparability among the four Romance languages. The frequency lists are presented in two formats: a.
by lemmas, containing 4 columns for rank, lemma, PoS, and frequency, as illustrated in Figure 1.17; and b. by word-forms, containing 5 columns for rank, form, lemma, PoSmsd, and frequency, as illustrated in Figure 1.18.
The C-ORAL-ROM resource
rank 1 2 ... 32 33 ...
LEMMA IL ESSERE ... BELLO CASA ...
PoS R V ... A S ...
frequency 4213 3840 ... 352 289 ...
Figure 1.17 Sample of lemmas frequency list
rank 1 2 ... 39 ... 46 ... 150 ... 589.
form il è ... bello ... bella ... era ... era
LEMMA IL ESSERE ... BELLO ... BELLO ... ESSERE ... ERA
PoS(msd) R V(s3ip) ... A ... A ... V(s3ii) ... S
frequency 2214 1587 ... 125 ... 109 ... 56 ... 22
Figure 1.18 Sample of forms frequency list
. Measurements of spoken language variability in the Romance languages One of the main characteristics of spoken language as compared to written language is the huge variability of the speech events according to individual characteristics, context of use, and semantic domain. Each Romance corpus in C-ORAL-ROM is designed to offer a representation of the main contexts of use of the spoken domain. In parallel, corpora are annotated with a set of linguistic information which is relevant for the highlighting of spoken language variability. The main parameters used for the description of spoken language domain (Dialogue structure, Sociological domain of use, Genre, Semantic domain of application, Channel) are represented in the corpus design schema, as reported in Table 1.1 above. As outlined in Section 1.5, corpora are tagged with respect to: a. Dialogue structure: Dialogic turns; b. Prosodic breaks: Terminal, non-terminal breaks and fragmentation events; c. Significant discrete units: utterances in the speech flow. Once relevant information regarding the spoken events are marked in the speech resource, corpus-based contrastive studies allow the investigation of correlations be-
Massimo Moneglia
tween the context in which a given spoken event occurs and its main linguistic qualities. Spoken language variability can thus be investigated by measuring the average and the variation coefficient of the linguistic units. The following are the standard variation parameters that reflect the overall structure of the texts recorded in the corpus: a. b. c. d. e.
Mid-length of utterances (in words56 ) (MLU); Mid-length of the dialogic turn (in words) (MLTw); Speed (words per second); Mid-length of the tone unit (in words) (MLTone); Fragmentation.
Cross-linguistic studies of standard measurements in the domain of Romance languages have two main implications: (i) tendencies that are consistent at a crosslinguistic level testify to their linguistic significance; and (ii) if the range of variation among Romance languages turns out to be well defined, then language-specific characteristics can emerge. The line diagrams reported below reproduce the mid-value of the above variation parameters in the various contexts of the corpus design structure and are also stored in the Standard Measure Menu of the Diagram section in DVD. The reader can refer to the DVD for the Variation coefficient registered by each Measure. Figures are identified below with the numbering they have in DVD. The line diagrams indicate strong correlations with the fields represented in the corpus design structure and show predictable characteristics at cross-linguistic level. In brief, the following general tendencies, which will be elaborated on in the following sub-sections, are observed. i.
MLU has a positive correlation with MLTw and exhibits highly predictable values especially in informal dialogic structures. ii. Both MLU and MLTw have an inverse correlation with Speed. iii. MLTone and Speed are however not predictable on the basis of the variation parameters, but instead vary according to language-specific features. iv. Fragmentation varies mainly in accordance with speaker characteristics.
.. Mid-length of utterances (MLU) MLU has strong cross-language correlations with respect to all the variation parameters represented in the corpus structure. In general, it is: a. b. c. d. e.
much higher in formal language; higher in public contexts; significantly higher in monologues; lower in transmission texts than in formal natural contexts; the lowest in the telephone domain, which is lower even than informal dialogues.
The C-ORAL-ROM resource
MLU 25,0 20,0 15,0 10,0 5,0 tel priv fam d/c pub d/c nat d/c fam m
Italian
French
media
Spanish
pub m
nat m
Portuguese
Figure DVD 4.1 Mid-length of utterances in words (MLU) across domains
As can be seen in Figure DVD 4.1, MLU is highly predictable in all four languages in informal dialogic structures: Spanish, Italian and Portuguese vary around the same average (5–7) while French records a higher average (9–11). MLU is however more variable in media and monologues. MLU may also vary consistently from text to text in both formal and informal monologues (with variation coefficient being around 40%), probably in accordance with different strategies adopted for text organisation. The variation coefficient for MLU in the four collections (shown in Figure DVD 4.1), is constantly low in telephone (>20%) and informal dialogues (20–25%), while it is high in media emissions, in accordance with the emission format (over 50%).
.. Mid-length of the dialogic turn (MLTw) MLTw shows a regular increasing trend according to corpus design, a pattern which is shared by the four languages (see Figure DVD 4.2). The variation coefficient of MLTw (see the diagram menu DVD 4.2) also shows a smooth increasing trend, with the only exception being the Italian Formal in Natural Context domain where a significantly higher value is recorded. MLTw has a positive correlation with MLU. Monologues, which have only one long turn, always have a higher MLU value, while in dialogue structures (telephone, family private, public informal and formal), MLU and MLTw co-vary in the four collections. In other words, in spontaneous speech the longer the turns of a text are, the longer each utterance is.
Massimo Moneglia
MLTw 45,0 35,0 25,0 15,0 5,0 tel priv Italian
fam d/c French
pub d/c Spanish
nat d/c Portuguese
Figure DVD 4.2 Mid-length of dialogic turn in words (MLTw) across domains
.. Speed Speed is a relevant predictable quality for discriminating between the four languages in the various spoken styles. As can be seen in Figure DVD 4.3, in the informal dialogic sub-corpus (telephone, family, public), each Romance language records a different speed: speed is lowest in Italian at less than 2.7 w/s, constantly over 3.5 w/s in French, and between 3 and 3.5 in Portuguese and Spanish. The variation coefficient (shown in the DVD in Diagram menu 4.3) is low in telephone conversation and informal dialogues. Conversely, the average value of formal monologues discriminates among the four languages to a lesser degree, while still presenting a low variation coefficient (<20%). Speed, which is limited by articulatory factors, therefore appears to be bound by language-specific phonetic structures. This difference is more evident in informal
Speed 4,0 3,5 3,0 2,5 2,0 1,5 tel priv fam d/c pub d/c nat d/c fam m
Italian
French
media
Spanish
Figure DVD 4.3 Speed, in words per second, across domains
pub m
nat m
Portuguese
The C-ORAL-ROM resource
dialogues rather than in formal corpora, where speed is reduced in connection with the speech performance’s task. Speed turns out to have an inverse correlation both with MLTw and with MLU. A constantly higher value for speed in telephone conversations and informal dialogues and a lower one in formal contexts can be observed cross-linguistically. Together with the findings for MLU and MLTw in 1.6.1 and 1.6.2, this indicates that the longer the turn, the slower the flow of speech.
.. Length of the tone unit (MLTone) MLTone shows strong upper limits, due to breath constraints, and, in theory, may vary in accordance with the prosodic, rhythmic properties of languages and their syllabic structure. Unlike MLU, MLTone cannot be predicted along sociological and structural variation parameters. Breath constraints force a low, predictable, average value. It is constant and similar in Italian, Spanish and Portuguese (2.5–3.5 words), despite the fact that these languages have different prosodic structures (variation coefficient from 10– 20%). The marked difference from the other three languages which is exhibited in French (4–6 words), which can clearly be seen in Figure DVD 4.4, is of course the main phenomenon to be explained. The word weight in terms of number of written words and written syllables is probably the origin of the great part of this difference of French: in speech, the reduction of syllables with respect to their orthographic counterpart is significant, and systematic. This makes possible to produce a higher number of words within the breath unit.
MLTone 6,0 5,0 4,0 3,0 2,0 tel priv fam d/c pub d/c nat d/c Italian
French
fam m media Spanish
pub m
nat m
Portuguese
Figure DVD 4.4 Mid-length of tone unit in words (MLTone) across domains
Massimo Moneglia
Fragmentation Phenomena
100% 50% 0% tel priv fam d/c pub d/c nat d/c fam m Italian
French
media pub m nat m Spanish
Figure DVD 4.5 Incidence of fragmentation, in %, for three corpora, across domains
.. Fragmentation The incidence of fragmentation phenomena (i.e. interrupted utterances and false starts) when compared to the total utterance has been noted for Italian, Spanish and French corpora only, and is found to be high in all collections. This testifies to its relevance as a peculiar feature of spoken language. As can be seen in Figure DVD 4.5, the incidence of fragmentation varies from 20% to 60% in utterances. Quite surprisingly, in all collections it has lower incidence (around 20%) in the informal dialogues, in telephone conversations and media emissions, while it ranges from 40% to 60% in the formal contexts. Fragmentation turns out to be higher when the spoken performance is taskoriented, while it is relatively lower in everyday use. However, although this parameter is predictable on average, it is also strongly determined by individual factors, as is shown by the lower incidence of fragmentation in media emissions (where professional speakers are involved), and confirmed by a high variation coefficient, all through the corpus nodes (40–100%) (see Diagram 4.5 in the DVD).
.. Some conclusions In spoken language, the utterance limits specify the domain of the main linguistic relations. For instance, argument structure, constituency, head dependency, and chunking relations hold among elements of a same utterance. MLU values allow Human Language Technologies to make hypotheses about the probable occurrence of the utterances and therefore to better determine the domain of actual linguistic relations. This may be significant for many purposes, such as adequacy of speech synthesis performances, selection of relevant domain for speech recognition, and automatic indexing technologies.
The C-ORAL-ROM resource
The spoken language domain exhibits a strong variability with respect to MLU, but certain variation tendencies can be clearly noted. The data presented show that, in all four languages under investigation, the mid-length of the utterance varies according to the structure of the communicative event, the sociological domain of use, and the channel; this suggests that the length of the linguistic object which is the outcome of the speech act depends on non-linguistic factors. In other words, non-linguistic conditions determine the goals of the linguistic performance. Accordingly, the linguistic means may be limited to a few words or may need complex structures with many words. The range of variation turns out to be quite predictable in informal dialogues, which are the prototypical domain of application of spontaneous speech. In informal dialogues, values are cross-linguistically recorded within two ranges (5–7 words per utterance in Italian, Spanish and Portuguese, and around 10 in French), with a consistent variation between French and the other Romance languages, specifically in the informal dialogic domains. For all four languages, such values strongly diverge from the formal-monologic domains, which consistently have higher values. The variation coefficient with respect to MLU in informal contexts is also much lower, testifying to the significance of the correlation. The severe restriction of the number of words comprising one utterance in natural dialogic contexts must thus be seriously considered when dealing with speech. The tendency to have much longer utterances is cross-linguistically verified in two domains: (a) monologic texts, and (b) contexts requiring a formal use of language. However while the length of the utterance is predictable as having higher values in those domains (always over 10 words per utterance), the range of variation is also much higher. Such variation is found to be high specifically in media speech. No language-specific tendency can however be highlighted on the basis of the data at our disposal. In order to make a more detailed analysis of MLU in such non-prototypical domains, the corpus sampling must deal with more subtle distinctions such as genre and semantic domain. The C-ORAL-ROM corpus does not however account for a sufficient amount of samples to document those variations. MLU shows significant correlation with the length of the dialogic turn but, at least in general, not with the length of the tone unit. In all four languages, the length of the utterance and the length of the turn co-vary according to socio-structural parameters. Specifically, the more formal the context, the more the linguistic task requires a long turn, and the more each utterance turns out long and structured. This tendency may seem natural, but it must be pointed out that this is not in principle necessary. The reverse possibility could also be considered, that is, the more complex the linguistic performance, the more each piece of information might be made simple. However this theoretical alternative does not apparently apply to natural languages in their steady state. Turn alternation in a given spoken dialogue is therefore a cue for predicting utterance length. On the contrary, in all four languages, the length of the tone unit appears independent from the contexts of use, and is fairly constant within each language. The
Massimo Moneglia
variation in values of MLTone is however relevant at the cross-linguistic level. With the Romance languages, the similarity between Italian, Spanish and Portuguese is evident, while French strongly diverges, probably due to a different syllabic weight of words. The relation between MLTone and MLU must be studied carefully. While MLTone is constant throughout the various corpus domains, MLU shows much variation. In general, the two measurements do not co-vary, The absence of a strict correlation between MLU and MLTone must be highlighted for its linguistic significance. While the latter is a function of the number of syllables and breath constraints, the former is independent of both factors. In other words, the functional values of the utterance that vary in accordance with the tasks requested by the context of use determine its length, that therefore is not strictly bound by the phonological structure of a language. The constant values that the tone unit length has in the various context of uses shows, on the contrary, that its length is not strictly correlated to the context of use. The parameter of speed is found to vary according to both corpus design and language-specific factors. Within the Romance family, speed turns out to be a characteristic particular to each language, and is confirmed as a genuine language-specific datum by the low variation coefficient. Differences are however mainly noted in the informal dialogic domain, where French and Spanish have clearly higher values. Crosslinguistic differences are lower in the formal contexts, where speed decreases in all languages for reasons presumably linked to the task of the speech performance. Given this general tendency, the variation in speed between informal and formal contexts is more marked for those languages with an overall higher speed. The results regarding speed are additionally interesting when compared with variations parameters as MLTw and MLU. Speed has an inverse correlation with both MLTw and MLU: the longer the turn and the utterance, the slower the flow of speech. The consistent frequency of fragmentation events with respect to the number of utterances in spoken language is confirmed by the available data. This is extremely relevant for speech analysis at both syntactic and prosodic levels. All fragmentation events cause the onset of a PoS sequence that, by definition, is inconsistent with the grammar. Moreover all fragmentation events cause a prosodic rupture of speech fluency and the onset of prosodic patterns that are also inconsistent with prosodic rules. The frequency of fragmentation events determines sequences that have little or no meaning from an informational point of view. Given their percentile incidence, their recognition is therefore vital for the understanding of spontaneous speech. Measurements verify that there is a significant similarity of incidence of fragmentation at cross-linguistic level; this suggests that fragmentation is a stable characteristic of spoken language. Despite this meaningful result, no significant correlation can be pointed out with respect to the corpus structure. Cross-linguistically, it is verified that nonprofessional speakers have the tendency for more fragmentation in task-oriented speech performance, rather than in everyday language. However the high variation coefficient indicates that the tendency for fragmentation is more a speaker’s character-
The C-ORAL-ROM resource
istic whose probability of occurrence cannot reasonably be predicted on the basis of the data analysed.
Notes . The status of a spoken resource for each of the Romance languages represented in C-ORALROM is presented in Chapters 2 to 5. Other non-English corpora collections must also be cited for their relevance in the state of the art. Apart from the Dutch and the Israeli corpora extensively described in this chapter, the following are at present the main references: – DGD (Datenbank Gesprochen Deutsch) of Mannheim: 27 corpora (spoken standard, spontaneous speech, relevant German dialects, foreign speakers’ language) collected from 1955. http://dsav-wiss.ids-mannheim.de/DSAvINFO.HTM – CALLHOME German Transcripts corpus: Constituting 100 unscripted telephone conversations between native speakers of German, the CallHome German lexicon consists of 318,807 words. http://www.ldc.upenn.edu/Catalog/CatalogEntry.jsp?catalogId=LDC97T15 – The Kiel corpus of read and spontaneous German, which has been collected and labelled segmentally since 1990. http://www.ipds.uni-kiel.de/forschung/kielcorpus.en.html – The Swedish Spoken Language Corpus: Comprising 1,230,663 tokens, it is an incrementally growing corpus of spoken language from different social activities (Allwood 1999); Göteborg University. http://www.ling.gu.se/projekt/old_tal/SLcorpus.html – The Corpus of Spontaneous Japanese: This project aims at the building of a large-scale spontaneous speech corpus (approximately 7 million words, with a total speech length of 800 hours) and its exploitation to improve the speech recognition technologies (see Furui et al. 2000; Maekawa et al. 2000). For English corpora the following are at present the main available collections: – London-Lund Corpus: 100 texts (435,000 words from conversations, telephone conversations, talks, lessons, radio) with prosodic annotations; collected by the Survey of English Usage, University College, London (published). – The British National Corpus (BNC): 100,000,000 words (10% spoken language); available online at http://thetis.bl.uk/. – ICE BG – Written and Spoken English: 1 million words parsed and tagged; Department of English Language and Literature, University College, London. – Corpus of London Teenage Language: 500,000 words within the BNC. http://www.hd.uib.no/ colt/ – COBUILD Bank of English: 320 million words, with the spoken word represented by transcriptions of everyday conversation, radio broadcastings, etc. http://titania.cobuilt.collins.co. uk/boe_info.html – University of Lancaster annotated corpora (Lancaster/IBM Treebank): Six parsed corpora, comprising 1 million words from debates at the houses of parliament; 250,000 words from the American printing house for the blind; 800,000 words from IBM manuals. http://www.comp.lancsac.uk/computing/research/ucrel/corpora.html – CANCODE, Cambridge and Nottingham Corpus of Discourse in English: 5 million words of spontaneous spoken English; built up by CUP and the University of Nottingham as a part of the Cambridge International Corpus. http://uk.cambridge.org/elt/corpus/cancode.htm
Massimo Moneglia
– ANC, American National Corpus (technical director: Nancy Ide, Vassar College); The first release includes 10 million words, not balanced; the final release will comprise 100 million words of written and spoken American English. http://americannationalcorpus.org/ – TRAINS: Corpus includes 98 task-oriented dialogues, amounting to 6.5 hours of speech, approximately 5,900 speaker turns, and 55,000 transcribed words. http://www.cs.rochester.edu/ research/trains/ – ATHELSAN, Corpus of Spoken, Professional American-English: Contains two main subcorpora of 1 million words each, the first consisting mainly of academic discussions, the second containing transcripts of White House press conferences. http://www.athel.com/cspatg. html – CSAE, Santa Barbara Corpus of Spoken American English; collected by University of California; scientific director: John W. Du Bois. http://www.ldc.upenn.edu/Projects/SBCSAE/ – The CHRISTINE Project: This extends the SUSANNE Corpus to cover spontaneous, informal spoken English; sponsored by the Economic & Social Research Council (UK) and directed by Geoffrey Sampson. http://www.grsampson.net/RChristine.html – WSC, The Wellington Corpus of Spoken New Zealand English: Comprises one million words of spoken New Zealand English collected in the years 1988 to 1994. http://helmer.aksis.uib.no/ icame/wsc/ . The C-ORAL-ROM databases are anonymous. All speech segments that might have offended the user for decency reasons have been erased and substituted with a beep in the audio signal. Speakers authorised each provider to use the recorded data for all ends in the C-ORAL-ROM project, including publication and language technology applications. The authorisation models are available at http://lablita.dit.unifi.it/coralrom/authorization_model. The use of radio and TV emissions has been authorised by the broadcasting companies which also provided the raw data for the database. Acknowledgement of all companies is given in the Copyright section of the DVD. The authorisation databases have been checked by ELDA (European Language Resource Distribution Agency) and delivered to the Commission. . Other partners in the C-ORAL-ROM Consortium: European Language Distribution Agency (ELDA), France, will distribute the resource in the industrial market sector; Instituto Cervantes (IC), Spain, is in charge of dissemination and exploitation for language acquisition purposes; Istituto Trentino di Cultura (ITC-Irst), Italy, tested the resource in current multilingual speech recognition technologies. . Files are encrypted and can only be accessed through the programmes included in the DVD. Minimal configuration required: Pentium III, 1 GHz, 252-megabytes Ram, S-blaster or compatible sound card, running under Windows 2000 or XP only. See Read-me file in the DVD for details. . See the discussion in 1.2.1. . The level of inter-annotator agreement on prosodic tag assignment has been evaluated by an external institution (LOQUENDO) and is reported in the appendix of this book. . In the French resource, terminal breaks are annotated but the alignment mainly follows pauses in the speech flow. . See the reference to those resources for each language in Chapters 2 to 5, and more in general in the ELRA catalogue. . This limit is quite severe for Italian, where local varieties may strongly diverge from the standard (De Mauro et al. 1993).
The C-ORAL-ROM resource . For spoken corpora representative of diathopic variation see: for Italian, LIP, Lessico di Frequenza dell’Italiano Parlato (De Mauro et al. 1993); CLIPS, Corpora e Lessici di Italiano Parlato e Scritto, coordinatore F. Albano Leoni http://www.cirass.unina.it); for French, CRFP, Corpus de référence du français parlé, collected in 1999–2000 at the Université de Provence; for Portuguese, Portugues falado (Bacelar do Nascimento 2000 and 2001a); for British English, SED, Survey of English Dialects: The Spoken Corpus, recorded in England 1948–1973 (Orton et al. 1978; Upton et al. 1994) and CANCODE, the Cambridge and Nottingham Corpus of Discourse in English http://uk.cambridge.org/elt/corpus/cancode.htm; for American English, CSAE, Santa Barbara Corpus of Spoken American English http://www.ldc.upenn.edu/Projects/SBCSAE/ . Around 5% variation on the corpus design figures is always allowed. . Up to 20% of this part may be constituted by texts of different length. . In Chapter 6 it is possible to appreciate all the most relevant lexical, structural and surface syntactic characteristics of an oral formal text. . From the project web site http://lands.let.kun.nl/cgn/doc_Dutch/topics/design/design.htm# intro . A detailed description of the design can be found in Izre’el et al. (2001: 171–197). . A cell was defined as a recorded segment designated to include 5,000 words of coherent continuous text. . In the original Austin definition, a speech act is the simultaneous performance of a locutionary act (“act of saying”), an illocutionary act (language action conventionally performed through the locutionary act) and a perlocutionary act (both the goals and the consequences of the act, in a broad sense). Language actions are a finite set and are defined by language conventions. In the Searle vulgata of speech act theory (Searle 1969), which is even more well known in the linguistic community, an utterance (u) is the result of the application of an illocutionary force (F) to a propositional content (p). According with Searle, F express a possible performative verb. Mirroring the relation between a performative verb and its embedded clause, the utterance is linearly represented as F(p) = u. The present framework, however, directly refers to Austin’s definition, which is more coherent with our assumption. The main formal indexes of illocutionary force in spontaneous speech belong to the prosodic domain, rather then to lexicon, therefore we must consider that prosodic programmes are not applied to the language string, but simultaneously performed while the string is uttered. Crucially the illocutionary force specifies which language action is accomplished by the utterance and therefore it constitutes a necessary feature of the utterance itself. . According to the theoretical framework of the IPO tradition, prosody features two types of F0 movements: “programmed, voluntary F0 changes and physiologically determined, involuntary fluctuation” (’t Hart et al. 1990: 39). Only the first type of movements is a selective object of perception for competent speakers. . The illocutionary criterion has been extensively applied to Italian spontaneous speech corpora for both adult and child language corpora (see references in Cresti 2000 and references to corpora in http://lablita.dit.unifi.it). . The properties of “completeness” and/or “autonomy” that are frequently referred as a necessary condition for the definition of utterance (Simon 2004) can be better understood in connection with the illocutionary criterion. A language entity is judged complete (or autonomous) if it can be pragmatically interpreted, and vice versa.
Massimo Moneglia . Some experimental research in this domain has been conducted in LABLITA Lab on large Italian spoken corpora. Roughly 80 different types of illocutionary acts, pragmatically recognisable by native speakers, have been found. More specifically, those which are most commonly employed (about 30) have specific intonation profiles, dedicated to the expression of a well identified speech act (see Cresti 2000; Firenzuoli 2003). . In other words the annotation of utterance boundaries in accordance with the illocutionary criterion is a much weaker practice than the annotation of speech act labels or discourse act labels (see also 1.2.3 below). . The prosodic annotation of breaks in the Dutch corpus also considered in-word breaks, which are not marked in C-ORAL-ROM, and other cues (e.g. lengthening, etc.). . In the Dutch corpus, a “substantial consistency” (Landis & Koch 1977) of the annotation for strong and weak prosodic breaks has been quantified by means of K-coefficient (Cohen 1960), recorded in the conventional range between 0.61 and 0.80 points. . For an overview, see the result of the IST project Mate. . For an exhaustive discussion of the question, see Wightman (2002) and Martin (2001). . Percentile data, such as those presented in Chapter 6, on lexical, morphologic and surface syntactic characters of the Romance spoken corpora can be derived on the base of the annotation of utterance boundaries. . The following is the definition given in the guidelines for the transcription of speech in the TEI P4: “Each distinct utterance in a spoken text is represented by a element, described as follows: a stretch of speech usually preceded and followed by silence or by a change of speaker”. See http://www.tei-c.org/P4X/TS.html#TSOV . See Cresti (2000) for the definition of Topic-Comment relation, and Firenzuoli and Signorini (2003) for morpho-syntactic and prosodic characteristics. . Typical examples for each break type in the four Romance languages are provided in the DVD in the folder “Types of prosodic breaks” of the Multimedia corpus. . No distinction connected to possible causes of interruption is considered in this frame (e.g. the CHAT format marks when the interruption is caused by the listener). However, the format explicitly marks the distinction between interruption and intentional suspension, which frequently occurs at the end of utterances. Intentional suspension must be marked with a generic utterance limit “//”, or specified, as “. . . ”. . Retracting phenomena with complete or partial repetition can both be expressed by this symbol, therefore, in principle, all traditional CHAT [//] should be simplified to [/] in this system, although the use of both traditional CHAT symbols is tolerated in the C-ORAL-ROM format. . For media corpora, the situation field is filled with the name of the programme. . Sessions in which F0 analysis is not significant are labelled D and excluded from sampling. . In human-machine interaction, in correspondence with the machine’s artificial voice, the value “Woman” or the value “x” is reported. . In this case the dialogic turn is filled by a communication event instead of a speech event. The relation between the two concepts is left undefined in C-ORAL-ROM; this resource marks the main communication events in a dialogue, but is not specifically devoted to the study of such events, which must be considered in a multimodal framework.
The C-ORAL-ROM resource . This may not be the case in overlapping and other intersection phenomena. See below. . No phonetic transcription. . In these cases the transcription may not be reliable for perceptual reasons. . Incomplete words are never subject to rebuilding, except, of course, for systematic phonetic phenomena (e.g. elision, breaking off of the last syllable, etc.). Those phenomena may or may not be mirrored in the transcription, following the orthography of each language and the traditions in editing oral text. Such choices must be detailed in the notes to the corpus edition. . Note that the lengthening of syllables, which is quite a common and perceptively relevant phenomenon in spoken language, is not marked in this system. However, following the philosophy of marking prosodic breaks, the system automatic assumes the generalisation that lengthening necessarily causes a prosodic break. . When a word is not transcribed in the text, it is substituted with a beep of a similar length in the acoustic signal. . Music or advertising fragments in media may not be transcribed, and instead substituted by a variable (that may or may not be aligned). In the case of very long music or advertising fragments in a media corpus, the fragment can be cut and the cut noted in a dependent line. . Each slash immediately following the speaker mark is of course not counted as a prosodic break. . When a dialogic turn continues despite the insertion of another turn in it that partially overlaps it, the traditional CHAT format allows the following linear representation (which has been abandoned in C-ORA-ROM for practical reasons): *ABA: Maria mi ha detto che [/] più / al concerto // perché non si sente // [Maria told me that [/] no more // to the concert // because she doesn’t feel well //] *BAB: [<] <non viene> // [<she doesn’t come> //] This representation is not consistent with the Intersection convention . In long monologues, where the serial number of a word is high, a “: /” convention can be used for the insertion of a dependent line immediately following the commented item. . The alignment of the French resource is by pauses; see Chapter 3. . The multimodal interface is not exploited in C-ORAL-ROM. . The best-known corpora of spoken English – Switchboard Telephone Corpus at Penn Treebank website http://www.cis.upenn.edu/∼treebank/home.html; London-Lund Corpus (Svartvik 1990); Lancaster-IBM Corpus (Knowles, Williams, & Taylor 1996) – received PoS tagging following different tagsets and methods. The recent national project of Spoken Dutch Corpus (CGN, Corpus Gesproken Nederlands, see Boves & Oostdijk 2003) outlined a procedure for exploiting existing taggers (TnT tagger, Brill tagger, maximum entropy memory based taggers) and lexical resources for the annotation of corpora with a new tagset (van Eynde et al. 2000; Zavrel & Daelemans 2000). The Japanese National Project on Spontaneous Speech Corpus (see Maekawa et al. 2000) tried to develop a wider series of specific tags for annotation of spoken language. . See Maekawa et al. (2000). . See Calzolari and Monachini (1996).
Massimo Moneglia . “One of the major problems presented by this category is that the Romance tradition makes use of the label Pronominal Adjective, whereas other languages (e.g. English) use Determiner. The two classes are not mappable one to the other. This is a crucial problem, involving the different behaviour of the Determiners and Pronominal Adjectives” (Calzolari & Monachini 1996). . In the NERC scheme, the problem of the use of different tags, i.e. Determiner and Pronominal Adjective, in different traditions has been raised and the proposal made to opt for the label Determiner, without any further functional distinction. The TEI work-group on annotation (see TEI 1991) has made the choice of inserting Determiners in the Adjective category, with the value pronominal. . See the GENELEX model which distinguishes the different cases of Pronouns, Determiner and Adjectives (which includes Pronominal Adjectives), exemplified by the following set: le nôtre (Pron), nôtre chien (Det), le chien est nôtre (Adj). . Regarding multiwords, there are two possibilities for their tagging, with respect to the different ways for identifying locutions in each tagset: e.g. a_priori\A_PRIORI\B a\A_PRIORI\B1 priori\A_PRIORI\B2 (Portuguese resource) . All measurements regarding words strictly refer to graphic words.
Chapter 2
The Italian corpus* Emanuela Cresti, Alessandro Panunzi, and Antonietta Scarano
. History of the corpus within the national framework .. Historical overview Before the 1990s when the first structured corpora of spoken Italian were published, it was possible to refer to many databases collected by certain researchers for their studies on spoken Italian, which have been published in the last twenty years (Bazzanella 1994; Berretta 1994; Berruto 1990; Bianconi 1980; Cresti 1987; Giannelli 1994; Orletti 1994; Poggi Salani 1977; Pontecorvo & Duranti 1996; Sornicola 1981; Stammerjohann 1970; Voghera 1992). However, though information about these corpora are available, neither their sound counterpart nor their transcripts are published (apart from a few excerpts of single texts), and their size and running time can only be deduced by references made to them in the mentioned works (Berretta 1994; Koch & Osterreicher 1990). There are also some Ph.D. theses which include their corpora in appendices, like those of Caputo (1996), Rossi (1998) and Voghera (1990). Proper speech corpora made their first appearance in the Nineties, the first one among them being the Lessico di frequenza dell’italiano parlato (LIP, Frequency Lexicon of Spoken Italian). A historical and chronological overview of the most important spoken Italian corpora, research projects and websites dedicated to them follows: 1. The LIP corpus (see De Mauro et al. 1993) is not only the first, but also one of the most important collections of spoken Italian, and doubtless the most used in linguistic research. LIP was accomplished by Università La Sapienza of Rome, in collaboration with the IBM Italia foundation between 1990 and 1992, and was directed by T. De Mauro. Its texts, which account for 57 hours’ recording time, and over 500,000 words, were collected in four cities (Milan, Florence, Rome and Naples, each accounting for 125,000 words) in order to represent the diatopic variety of Italian language. These texts also represent the diaphasic and diamesic variety, because they are organised over five different types of texts: (a) face-to-face conversations; (b) telephone conversations; (c) non-free dialogical interactions; (d) monologues (lectures, sermons, etc.); and (e) radio and TV programmes. Group (c) includes texts in which the dialogical interaction does not proceed with-
Emanuela Cresti, Alessandro Panunzi, and Antonietta Scarano
out restraint, but is guided by one of the speakers: typical examples are interviews, classroom interactions and debates. Each register is represented by samples of texts amounting approximately to 100,000 word tokens. The frequency lexicon was collected in book form as a dictionary, accompanied by two floppy disks containing the 500,000 words of which the corpus consists; this is the broad orthographic transcription of the recordings, whose criteria include the indication of pauses, of overlapping of different speakers and of vowel lengthening, but not of punctuation. The recordings, on magnetic band, are kept by the Universita˛ La Sapienza in Rome, but are not publicly available. The only parts of the LIP corpus accessible to the public are the orthographic transcripts of the recordings. 2. Another activity which employs the LIP corpus is the Banca dati dell’italiano parlato (BADIP, Spoken Italian Databank). Part of Karl-Franzens-Universitat of Graz’s Language Server, this is an initiative which intends to make data of spoken Italian available electronically, i.e. freely accessible on the internet, and to give researchers unified access to it. At present, its efforts are concentrated on an online edition of the entire LIP Corpus. BADIP also contains extra-linguistic data, PoS tags, lemmas, and word frequencies. Future activities will be dedicated to the publication of the audio part of LIP, to accommodate further data from this corpus as well as from other corpora of spoken Italian. The corpus can be accessed at http://languageserver.uni-graz.at/badip/. 3. The Archivio delle Varietà di Italiano Parlato (AVIP, Varieties of Spoken Italian Archive) project was co-funded by MURST (Ministero dell’Universita˛ e della Ricerca Scientifica e Tecnologica) in 1997 and completed in 1999. The project was coordinated by P. M. Bertinetto of the Scuola Normale Superiore di Pisa with collaboration from the Linguistics Laboratory of the Scuola Normale Superiore di Pisa (Bertinetto 2001), CIRASS (Università degli Studi di Napoli “Federico II”), the Phonetics Laboratory of the Istituto Universitario Orientale, Naples, the Department of Electrotechnics and Electronics of Politecnico di Bari, and the Linguistics Department of Università del Piemonte Orientale, Vercelli. The corpus contains speech from different regional Italian varieties, collected in Pisa, Naples and Bari. Totalling 37,600 words, the corpus is aligned text-to-speech following dialogic turns. It includes 39 semi-spontaneous map-task dialogues, produced by young adult speakers, and 5 video-recorded dialogues, produced by children with normal and impaired hearing, for a total of about 14 hours. Only 15 adult speakers’ dialogues, plus the 5 produced by children, giving a total of about 350 minutes, are orthographically transcribed. 75 minutes of speech are phonetically segmented and labeled, with a smaller subset also prosodically labeled, and four dialogues are annotated at textual level. The entire documentation of the project is available by ftp, as well as in the CD version1 (which can be requested at [email protected]), and is freely accessible at ftp://ftp.cirass.unina.it/cirass/pub/avip/. The recordings of all AVIP texts are downloadable in WAWE format. 4. The Archivio di Parlato Italiano (API, Spoken Italian Archive) project, co-funded by MURST in 1999, began in November 1999 and was completed in November
The Italian corpus
2001. It was coordinated by F. Albano Leoni (CIRASS), with collaboration from Scuola Normale Superiore di Pisa, Università di Pisa, Università di Venezia, Università del Piemonte Orientale, Politecnico di Bari, and the Department of Neurosciences and Interhuman Communication of Università degli Studi di Napoli “Federico II”. API added a set of labelling levels to those of the AVIP corpus. Electronic tools which integrate those fine-tuned in the AVI project have also been developed. 5. The consortium coordinated by CIRASS also worked towards the creation of Corpora Linguistici per l’Italiano Parlato e Scritto (CLIPS, Linguistic Corpora for Spoken and Written Italian). This corpus was financed by MIUR, and began in February 2000 and concluded in February 2003. The idea for the project came from F.Albano Leoni at the Federico II University of Naples, and involved the participation of other institutions: Scuola Normale Superiore di Pisa, FUB, ISCTI, Università di Lecce. CLIPS accounts for about 100 hours’ worth of recordings, and is articulated, from a diaphasic point of view, in semi-spontaneous maptask dialogues, in radio and TV broadcastings (news, cultural and entertainment programmes, advertisements), in telephone conversations, and read texts. CLIPS represents regional Italian varieties well, as its recordings were collected in 15 different places, chosen according to linguistic, demographic and social-economic criteria. The corpus is dedicated to technological applications (training of recognition systems, fine-tuning of automatic segmentation tools and labelling), as well as linguistic analysis (systematic analysis and linguistic variability). It was partly orthographically transcribed (about 300,000 words, comprising 30% of the total) and aligned text-to-speech following dialogic turns. About 10% of the corpus was also labelled phonetically. The transcription and the labelling were achieved following international standards, adopting EAGLES recommendations. The documents published by the AVIP consortium have been transferred onto CD, distributed by the consortium, and made available on the FTP site arranged by the Federico II University of Naples’ CDS. 6. The Italiano Parlato (IPAR, Spoken Italian) project was conceived as a conclusion to the AVIP and CLIPS projects, and was co-funded by MURST in 2001. It was coordinated by F. Albano Leoni (CIRASS), and also involved Università degli Studi di Napoli ‘Federico II’ (CIRASS and Department of Neurosciences and Interhuman Communication), Seconda Università di Napoli, Università di Perugia, Università di Torino, Università “La Sapienza” di Roma, Scuola Normale Superiore di Pisa, Università di Pisa, Università del Piemonte Orientale, Università di Venezia, Università di Siena, Università di Salerno, and Istituto Universitario Orientale di Napoli. Its aim was a more in-depth linguistic analysis of spoken Italian which could represent the main regional Italian varieties collected in the AVIP and CLIPS corpora .The project was also aimed towards the fine-tuning of other tools for the automatic analysis of speech, particularly the prosodic component.
Emanuela Cresti, Alessandro Panunzi, and Antonietta Scarano
The next two corpora, LIR and CiT, are examples of corpora specifically collected to investigate the diamesic features of radio and television speech and their specific lexical characters. 7. The Lessici di frequenza dell’italiano radiofonico (LIR, Frequency Lexicons of Radiophonic Italian) project was co-funded by MURST in 1995. It was coordinated by N. Maraschio and executed by the Italian Grammar Centre of the Accademia della Crusca; Dipartimento di italianistica dell’Università di Firenze and Scuola Normale Superiore di Pisa also took part in it. The research work was made possible by the support given by Giovanni Nencioni, a pioneer in Italian speech studies, at that time President of Accademia della Crusca. LIR accounts for around 68 hours of radiophonic speech and over 500,000 transcribed words. The recordings were made in 1995, and taken from the broadcastings of nine nationwide-diffused radio stations: the three RAI stations, Radio DJ, RTL 102.5, Rete 105, Radio Italia, Radio Radicale, and Radio Vaticana. The CD-ROM, still a prototype for the time being, contains the transcripts of the texts aligned with the recordings. The texts are accompanied by metadata information on the station, speaker and communication type. The transcription is orthographic, and integrated by the highlighting of some of the most typical speech phenomena, such as auto-corrections, clippings, overlapping, and hesitations. The corpus was automatically lemmatised using a version of E. Picchi’s DBT program, which also enabled the recovering of multiwords, idiomatic expressions, and foreign words. Separate frequency lexicons for each of the radio stations are also scheduled to be conducted. 8. Work on the Corpus di Italiano Televisivo (CiT, Televised Italian Speech) corpus started in 1998 and is still in progress. CiT is kept by the Department of Language Sciences of the University for Foreigners in Perugia, and its coordinator is S. Spina. The corpus collects texts from television broadcastings and consists of 250,000 words, orthographically transcribed; its aim is the analysis of lexical and grammatical features of televised Italian. CiT was manually annotated, following the standards set by the Text Encoding Initiative (TEI). The labels concern the following different levels: (a) structure of television programmes; (b) grammatical categories; and (c) lexicon (foreign words, technical words and multiwords). The corpus is available for consultation in demo version at http://www.sspina.it/cit/cit.htm. Other corpora with different purposes are: 1. The Acoustic Phonetic And Spontaneous Speech Corpus (APASCI) (Angelini et al. 1994). It was collected by IRST and is distributed by ELRA.2 APASCI’s audio counterpart consists of continuous speech sentences, designed to cover a large number of phonetic contexts. The corpus consists of two portions. The first part is devoted to the training and validation of a speaker-independent speech recognition system, comprising 3,900 utterances from 188 speakers. The second part is designed for the training and validation of speaker-dependent recognition sys-
The Italian corpus
tems using 6 speakers: 3 males and 3 females. The APASCI corpus was collected in a quiet room (SNR * 40 dB) by means of a high quality close-talk microphone. 2. The Italian Fixed Network Speech SpeechDat(M) Corpus (FIXED0IT) was recorded within the SpeechDat(M) project (LRE-63314), funded by the European Commission. The corpus contains the speech of about 1,000 speakers, in approximately equal numbers of males and females, and was designed to support the creation of voice-driven teleservices. Most items are read, while some are spontaneously spoken. All speech is transcribed at orthographic level; moreover, age and regional background of the speakers are provided. A pronunciation dictionary is added, containing all the words which the corpus features, with a corresponding SAMPA broad-class phonemic transcription. Validation and premastering of the CD-ROMs were performed by the Speech Processing Expertise Centre (SPEX) of Leidschendam, The Netherlands. 3. The PIXI Corpus, created during the Eighties within an Anglo-Italian environment, published by Gavioli in 1990, and also available in electronic format, is also worthy of notice. The PIXI Corpus (Gavioli & Mansfield 1990) consists of 450 naturally occurring conversations recorded in bookshops in England and Italy, for the purpose of cross-cultural comparison of discourse structure and pragmatics. The Italian part of the corpus consists of around 30,000 words in 142 texts, some of which are very brief. The recordings were made in different villages of Central Umbria, although the actual locations of the recordings have been anonymized. They are available in electronic form from the Oxford Text Archive, and in book form from Gavioli and Mansfield (1991), together with in-depth details of the data collection, discourse contexts, analytic approach, and bibliography of related publications. However, very few metadata are given. A written request has to be sent to OTA; see http://ota.ahds.ac.uk/ for details. No recordings are available. 4. The CHILDES ITALIA (Bortolini & Pizzuto 1997) project, executed by CNR of Rome, the Psychology Department of Università La Sapienza in Rome, the “Stella Maris” Institute of Care and Research in Pisa and the Infantile Neuropsychiatry Department of Università di Pisa, LABLITA of Università di Firenze, and CNR of Padova, allows the creation of longitudinal collections regarding normal and impaired acquisition of Italian; part of this was included in the international CHILDES database (see Bortolini & Pizzuto 1997). The CHILDES system provides free tools for studying language in infancy data and conversational interactions (MacWhinney 2000). The CHILDES project provides a large database of first and second language acquisition data from over 30 languages in CHAT format.
.. The LABLITA Corpus The LABLITA Corpus, from which the Italian C-ORAL-ROM is taken, is kept by the Linguistics Laboratory of the Italian Department in Florence, which carries out research on corpus linguistics and experimental research on the intonation of spoken Italian. The collection of the corpus began in the 1970s, and is continually updated,
Emanuela Cresti, Alessandro Panunzi, and Antonietta Scarano
enabling it to be acknowledged as an “open” diachronic corpus. It includes the first spontaneous Italian speech corpus, which was collected in 1965 in Florence by Stammerjohann, who gave it to the laboratory as a gift (Stammerjohann 1970; Scarano & Signorini forthcoming; Signorini & Tucci forthcoming). The LABLITA corpus is made up of three corpora: (a) the Spontaneous Adult Language corpus, comprising approximately 152 hours, 62 of which are transcribed, for a total of about 640,000 words; (b) the Early Acquisition corpus, comprising longitudinal and transversal collections of recordings of infants from 13 to 36 months of age, which account for around 72 hours and 210,554 transcribed words. As in the CORAL-ROM corpus, the orthographic transcripts of the LABLITA corpus recordings are in CHAT format, and are integrated by prosodic tagging. The audio is digitally mastered in wav files (22.050 hz – 16 bits). The LABLITA corpus is in the process of being aligned using the utterance-based system developed in C-ORAL-ROM. All of LABLITA’s texts are accompanied by complete metadata, according to the IMDI format at http://corpus1.mpi.nl/IMDI/metadata/IMDI.imdi. The Spontaneous Adult Speech corpus, which comprises texts mostly collected in Florence in a natural environment, was conceived to fully represent the spoken diaphasic and diastratic varieties. It is entirely composed of spontaneous texts, which are classified according to the following parameters, which roughly match those of the C-ORAL-ROM corpus design: text formality or informality, social projection of the communication context (family, private, public), number of speakers involved in each communicative event (monologue, dialogue, conversation), and free or restrained modality of dialogical turning. The entire collection of LABLITA corpora is not freely accessible, but can be consulted for research activities within research programs. Before the C-ORAL-ROM Project, a sample (approximately 7 hours and 60,000 transcribed words) was published in the volume Corpus di italiano parlato (Cresti 2000). A large sampling of LABLITA texts with text/audio/acoustic parameters alignment is scheduled to be freely accessible on the web.
. Orthographic transcription .. General criteria In the transcription of the texts included in the C-ORAL-ROM Italian corpus, the use of initial capital letters is reserved for: a. b. c. d.
Christian names (Calvino; Paola); Toponyms (Parigi, Prato); Odonyms (via Burchiello, Porta Romana, Via Pian de’ Giullari, Ponte alla Vittoria); Television-programme names (Fantastico);
The Italian corpus
e. Film titles (Piccolo grande uomo, Il secondo tragico Fantozzi); f. Book titles (Memoriale, Vangelo, Divina Commedia);3 g. Band names (Depeche Mode; Neganeura). Abbreviations are written entirely in block capitals, without full stops between the letters, e.g. BTP, MPS. Capitals are used in accordance with their role in traditional grammar, and with the swings that grammar lets them perform: e.g., in expressions which contain toponyms and odonyms preceded by a common name, only the proper name always requires an initial capital letter, whereas the common name can be written beginning with an initial capital as well as with a small letter; as for titles of works, the use of capitals is only compulsory for the first letter (see Serianni 1988: 55).
.. Orthographic transcription of regional words This section complements the glossary which accompanies the transcripts of the oral texts included in the C-ORAL-ROM Italian corpus. It is reported in the DVD, in the Italian section of the corpus Metadata Menu. The glossary features: a. regional or local words, which need an explanation; b. idiolectic words for which a spelling form was chosen in the transcripts; c. regional variety forms, characterised by phonetic phenomena within the word or by phonosyntactic phenomena in the linear sequence. Words belonging to category (a) have been defined by other terms which are more common and understandable in the rest of Italy; we shall not attend any further to these words, as no spelling choices are necessary for them. All the other types of words recorded in the glossary will be addressed in this section, because of the choices in spelling required as to their written form. A written form has been attributed to idiolectic words, and to words which are frequent and common in usage, but have not undergone orthographic normalisation in standard Italian, thus not appearing in dictionaries. These include interjections, onomatopoeia, baby-talk words, foreign words, hapax legomena, and true idiosyncratic forms; for their treatment, see Sections 2.2.4 and 2.3.3.1. Among the words to which a written form has been assigned, there are also very common expressions, which, we felt, needed a spelling more in keeping with their pronunciation, such as: vabbè = va bene [all right] vabbò = va buono [ironical form, meaning ‘all right’] These include not only the result of actual lexicalisation of multiwords, but also Tuscan variations of standard Italian forms, like: quande = quando [when] quante = quanto [how much] ’un = non [not and other negatives]
Emanuela Cresti, Alessandro Panunzi, and Antonietta Scarano
dumila = duemila [two thousand] vu’ = voi [you (plural)] e’ = [third person clitic pronoun, plural (him/them) and first person clitic pronoun, singular (I)] to’ = tuo [yours (singular)] For those verbal forms which differ from their paradigm in standard Italian, and which have not been brought back to it, the spelling form adopted reproduces their pronunciation as correctly as possible, with a phoneme/grapheme correspondence which follows the standard Italian system:4 andesti avessin potette rompé rompiede rompò fo vo facci
instead of instead of instead of instead of instead of instead of instead of instead of instead of
andasti avessero potè ruppe ruppe ruppe faccio vado faccia
[you went] [had they] [was able to] [broke] [broke] [broke] [I’m doing] [I’m going] [do (imperative, courtesy form)]
Tuscan forms characterised by an inflectional allomorph, which makes them homophonous with other standard verbal forms, have not been brought back to their paradigm, but loyally transcribed according to their pronunciation; therefore, in our transcription, they appear as homographs of other standard Italian verbal forms, but in the glossary, they are recorded with their dual value, as illustrated in Table 2.1. Verbal polysemic forms of the Tuscan dialect have been transposed, through a graphemic cluster (-gn cluster and -ne as clitic), as in: dagnene = dagliene, daglielo, dargliene, darglielo [give (some) to him/her/it, give it to. . . , to give (some) to. . . , to give it to. . . ] portagnene = portaglielo, portagliene, portaglielo, etc. [take it to him/her, take (some) to. . . ] The above forms are strongly characteristic from a diastratic point of view: they belong to a lowly variety of Tuscan. Orthographic solutions which are loyal to the actual pronunciation of the words were brought about by the adoption of a general criterion of sociolinguistic significance of a phenomenon. The spirantisation of voiceless occlusive consonants in intervocalic position, which is common to all Tuscan speakers, is only highlighted in the transcript when it is marked, and thus significant from diastratic point of view; other phenomena (like rhotacism, for example) are highlighted because they are significant in their own right from a sociolinguistical point of view.
The Italian corpus
Table 2.1 Homographic verbal forms Tuscan form
Mood and person
Corresponding Italian form
Homophonous & homographic Italian form
abbi [to have]
3a singular person subjunctive present
abbia
abbi: 2a singular person imperative present
dividano [to divide]
3a plural person indicative present
dividono
dividano: 3a plural person subjunctive present
intendano [to mean]
3a plural person indicative present
intendono
intendano: 3a plural person subjunctive present
governavi [to govern]
2a plural person indicative past
governavate
governavi: 2a plural person indicative past
possano [can]
3a plural person indicative present
possono
possano: 3a plural person subjunctive present
rendano [to give back] tenevi [to keep]
3a plural person indicative present 2a plural person indicative past
rendono
rendano: 3a plural person subjunctive present tenevi: 2a singular person indicative past
devan [must]
3a plural person indicative present
devono
devano: 3a plural person subjunctive present
mettan [to put]
3a plural person indicative present
mettono
mettano: 3a plural person subjunctive present
vedano, vedan [to see]
3a plural person indicative present
vedono
vedan: 3a plural person present
vengan [to come]
3a plural person indicative present
vedono
vengano: 3a pers. pl. subjunctive present
tenevate
Depending on the degree of evidence of the spirantisation, the spirantised consonant has been omitted, or an h has been inserted, in keeping with a traditional vernacular spelling form: poho passai staa sentìo
instead of instead of instead of instead of
poco passati stata sentito
[a small quantity] [gone, gone by (pl.)] [been (f.)] [heard]
An approach towards the actual pronunciation has also been achieved in the rendering of the intervocalic /v/. This not only applies to well-known forms – which are consolidated in literary tradition, like avea, dicea – but also to forms like guardaa = guardava [was looking at]. Below is a short list; the reader can refer to the DVD for the full set. aere = avere [to have] aerlo = averlo [to have it] andaa = andava [he/she/it was going] andao = andavo [I was going]
Emanuela Cresti, Alessandro Panunzi, and Antonietta Scarano
aveo, avei, avea, aveano, avean = avevo, avevi, aveva, avevano, avevano [I had, you had, he/she/it had, they had] credeo = credevo [I thought] arriò = arrivò [he/she/it arrived] dèo, dèi, dèe, dèano, dèan = devo, devi, deve, devono, devono [I must, you must, he/she/it must, they must] dao = davo [I was giving] doveo, dovea, dovean = dovevo, doveva, dovevano [I had to, he/she/it had to, they had to] A phenomenon considered significant, and therefore highlighted by the graphic transposition, is the loss of the semi-consonant /w/ in the /kw/ cluster (corresponding to the graphematic cluster qu in standard Italian spelling). In this case, we have resorted to the cluster corresponding to occlusive velar consonant (ch), a solution also generally adopted in traditional vernacular spelling: chello, chelli instead of quello, quelli [that, those] chesto, chesta instead of questo, questa [this (m., f.)] cherce instead of querce [oak tree] As mentioned before, cases of rhotacism (l → r) are orthographically highlighted; in order to avoid cases of homography, the tonic vowel is marked with an accent: mórto = molto scarza = scalza (verb scalzare) sòrdi = soldi sòrdo = soldo rinvortata = rinvoltata vorte = volte quarcuna = qualcuna
[very, a lot] [undermines] [money] [penny] [wrapped (f.)] [times] [a few (f.)]
Other speech phenomena whose effects were loyally transposed in writing are various assimilation, epenthesis, epithesis and metathesis phenomena. Various assimilation phenomena: a.
al + consonant > geminated consonant: ammeno < almeno (con ulteriore possibile evoluzione aimmeno) [at least] attro, attri < altro, altri (con possibile evoluzione antro, antra) [other, others] infizzaa < infilzava [was piercing]
b. nl > ll: domall’ altro < doman l’altro < domani l’altro [the day after tomorrow] c.
r + l > ll: prendello < prenderlo [to take it]
The Italian corpus
Another possible result is the degemination of double consonants and the closing of the post tonic which precedes -ll: prendilo = prenderlo [to take it] leggile = leggerle [to read them (f.)] perdile = perderle [to lose them (f.)] riprendila = riprenderla [to retrieve it (f.)] d. -cn- > nn: tennicamente < tecnicamente [technically] Other phenomena: a.
epenthesis phenomena: anderà < andrà [he/she/it will go]
b. epithesis phenomena: barre < bar [bar, pub] chie? < chi? [who?] c.
metathesis phenomena: drento < dentro [inside] presempio < per esempio [for example]
Some phonetic peculiarities of southern Italian varieties have also been rendered graphemically, using the graphic solutions adopted for the transcription of southern dialects, for example: Palatalisation of cluster with /p/: chiù < più < plus (lat.) [more] Loss of /l/ before consonant: vota < volta [time, as in ‘one time’]
.. Diacritic marks Two graphic symbols have been used to make word reductions evident and to avoid cases of homography: the accent and the apex. Accents have been used to distinguish non-standard Italian forms from their homographic or homophonous standard forms: méssi = misi [I put] different from messi (noun, f. pl.) [crops] mésse = mise [he put] past different from messe (noun, f. pl. of messa) [mass]; participle [put (f. pl.)] Tonic vowels resulting from the contraction of diphthongs have been transcribed with a grave accent, for example:
Emanuela Cresti, Alessandro Panunzi, and Antonietta Scarano
bòna, bòni, bòno = buona, buoni, buono [good (f. sing, m. pl., m. sing.)] còcere = cuocere [to cook] The contraction of diphthongs has not been made evident in words where the accent does not fall on the diphthong: bongiorno = buongiorno An apex (apostrophe) has been used where clipping and aphaeresis occur.
Clipping phenomena Cases of vowel apocope, which come after a liquid or nasal consonant – which are also described in Italian grammar books – are not highlighted, but rendered graphically. The following examples are a result of this: buon ben voglian vòl
[good] [well] [they want] [wants]
devan [they must] mettan [they are putting] vengan [they are coming]
The various kinds of vocalic and syllabic clipping, or falling of segments at the end of a word, have instead been signalled with an apex: allo’ < allora [then] aspe’ < aspetta (2a ) [wait] avra’ < avrai [you should have] bra’ < bravo [good boy] capi’ < capito [udestood] du’ < due [two] ecce’ < eccetera [etxetera] fa’ < fai, fare [you do; to do] gua’ < guarda [look] le’ < lei [she] lu’ < lui [he] ma’ < mai; mamma [momy] mi’ < mia, mio [my; mine]
pa’ < padre [dady] po’ < poi [then] sa’ < sai [you know] se’ < sei [you are] so’ < sono [I’m] su’ < suo, sua [his; her] t’ < tu [you] telefonera’ < telefonerai [you will call] tu’ < tuo-a-e, tuoi [your; yours] ve’ < vedi [look] vie’ < vieni (imperative) [come] vò’ < vuoi, vuole [you want; he want] vorre’ < vorrei [I would like]
In vernacular texts, the third person masculine subject clitic gli is sometimes used, for which the chosen transcription is always gl’, without vowels, so as to highlight its non-syllabic character. With regard to the pre-consonantic determinative article il, which in Florentine results in [i] followed by the strengthening of the consonant that comes after it – as in iccane, iggatto, etc. – an apex was used, in keeping with a traditional vernacular spelling form, thus writing i’ cane instead of il cane, etc.5 The same written form of the article il was adopted for “analytical” pronunciations of articulated prepositions – that is, a preposition resulting from the fusion between
The Italian corpus
a simple preposition and an article – each transcribed as “preposition + article” in order to avoid confusing them with the articulate preposition of the corresponding plural form: a i’ for al [at the] da i’ for dal [from the] su i’ for sul [on the] The use of an apex after proper articulated prepositions must be explained; in these cases, the apex marks the loss of the vowel, caused by a process similar to clipping: a’ da’ co’ de’ ne’ su’
for for for for for for
ai dai coi del, dei nei sui
[at the, pl. to (m. pl.)] [from (them), pl.] [with the, pl.] [of the, m., sing. and pl.] [in the, m.pl.] [on the, m.pl]
Finally, synthetic forms of articulate prepositions, which can at times result as ambiguous, have also been marked by an apex. In practice, once transcribed, these could be indistinguishable from the correspondent simple preposition. This would never happen in pronunciation, because only articulate prepositions produce the doubling of the following consonant: di’ for del [of the, m. sing.] ni’ for nel [in the, m. sing.] An unusual case is represented by the demonstratives quel/quei, which behave like articulate prepositions: que’ for quei [those, m.] qui’ for quel [that, m.] Apexes have again been used for prepositions which feature consonant apocope: pe’ < per [for] co’ < con [with] Cases of apocope of final vowels and of final syllables of verbs with infinite inflection have received a spelling common to both. With regard to vocalic apocope, which occurs in Tuscany in front of a consonant followed by an accented vowel, the result is an progressive assimilation of the ‘r + consonant’ group (geminated consonant: andare male > andar male > andammàle). As for syllabic apocope, cases like andà’ and mètte’ occur (and the accent does not change its position). A common spelling form was adopted for these phenomena, which appear together in the same texts with a considerable frequency: the apocope is signalled, whatever its nature, with an apex at the end of the word, but not the vowel on which the accent occurs, which stays the same as in the verbs with infinite inflection without apocope: anda’, mette’.
Emanuela Cresti, Alessandro Panunzi, and Antonietta Scarano
anda’ < andare [to go] analizza’ < analizzare [to analize] apri’ < aprire [to open] ave’ < avere [to have] capi’ < capire [to undestand] compra’ < comprare [to buy]
rifa’ < rifare [to redo] ritorna’ < ritornare [to come back] sape’ < sapere [to know] sede’ < sedere [to seet down] smette’ < smettere [to give up] spende’ < spendere [to spent]
Aphaeresis phenomena The vocalic or syllabic aphaeresis and the loss of initial word segments are marked with an apex at the beginning of the word: cidenti < accidenti [damn] ’nnaggia < mannaggia [damn] ’fatti < ’nfatti [in facts] ’petta < aspetta [wait] ’giorno < ’ngiorno [good morning] ’spetta < aspetta [wait] ’gna < bisogna [we need] ’st’, ’sta, ’ste < quest’, questa, queste [this; these] nfatti < infatti [in facts] ’sto, ’sti < questo, questi [this; these] Cases in which the vocalic or consonantal aphaeresis generates forms which have become common use as variations, and therefore appear in dictionaries, have not been signalled by an apex at the beginning of the word, e.g.: briaca, briache = ‘ubriaca’, ‘ubriache’ [drunk, f., sing. and pl.] ché = ‘perché’ [because]
.. Interjections The following is a list of the most frequent interjections and discourse particles; greetings, wishes and curses are also included (Table 2.2). Table 2.2 Interjections and discourse particles Rank
LEMMA
Occ.
Rank
LEMMA
Occ.
1 2 3 4 5 6 7 8 9 10 11 12 13 14
eh mh ah vabbe’ oh mah beh ciao boh magari buonasera bah he mamma_mia
1944 893 682 215 160 151 133 122 49 47 41 38 28 24
15 16 17 18 19 20 21 22 23 24 25 26 27 28
ohi via grazie uh han eee’ mamma oddio accidenti basta madonna per_carita’ ueh ehm
24 22 20 15 14 13 11 10 8 8 7 6 6 5
The Italian corpus
Table 2.2 (continued) Rank
LEMMA
Occ.
Rank
LEMMA
Occ.
29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50
ieh perdio ahi bu ich uhm arrivederci buongiorno chi_se_ne_frega hei macche’ mannaggia sh vaffanculo aho’ buonanotte che_palle cristo dagli ehi male meno_male
5 5 4 4 4 4 3 3 3 3 3 3 3 3 2 2 2 2 2 2 2 2
51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72
ooh ahm amen caspita grazie_al_cielo hi hum ia in_bocca_al_lupo ma_va’ maramao mhm moh ohe’ ohiohi perdinci prego toh uffa uhi umh vabbuo’
2 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
. Morpho-syntactic tagging .. Tools and strategy adopted for automatic PoS tagging and lemmatisation The main antecedent in automatic tagging of spoken Italian is the LIP corpus (De Mauro et al. 1993). However, neither the tagger nor the tagged resource are available.6 The automatic tagging procedure of lemmatisation and morpho-syntactic annotation of the Italian C-ORAL-ROM corpus is based on the PiTagger tool, created and developed by Eugenio Picchi within the ILC-CNR of Pisa. PiTagger was chosen on the basis of three main reasons: (a) the availability of the tool on the market; (b) the EAGLES-like tagset; and (c) PoS annotation, and lemmatisation joined in a coherent system. PiSystem is an integrated procedure for textual and lexical analysis, which consists of the following main components: 1. DBT text encoding and analysis modules; 2. a morpho-syntactic analyser (PiMorpho); 3. a Parts of Speech (PoS) tagger and lemmatiser (PiTagger). The DBT encoding provides a first parsing of the text, known as tokenisation, and represents the pre-analysis level of the morphological analyser. The encoded text is
Emanuela Cresti, Alessandro Panunzi, and Antonietta Scarano
then given to the PiMorpho, which assigns all the possible alternatives of MorphoSyntactic-Description (MSD) to each lexical item. For the PoS disambiguation, PiTagger uses two other input-resources: an electronic dictionary and a training corpus. In detail, these resources consist of: a.
the DMI, a morphologic dictionary of Italian language, developed within the ILC at CNR, Pisa; it collects 106,090 lemmas encoded with PoS specifications and inflectional tags (Zampolli & Ferrari 1979; Calzolari, Ceccotti, & Roventini 1983); b. a Training Corpus of 50,000 words, manually tagged; c. a statistical database extracted from the Training Corpus (BDR). These resources are built on a coherent tagset for PoS and morpho-syntactic annotation. The disambiguation phase is processed by statistic measurements (on tri-grams) extracted from the Training Corpus and stored in the BDR. The main programme of this procedure estimates the maximum likelihood pattern among the possible alternatives given by the morphological component, with a transitional probabilistic method (Picchi 1994). In this environment, the level and the precision of the analysis depends on the information archived in the BDR. Statistics must be defined on a specific level of linguistic information, i.e. lexical patterns or PoS sequences. The BDR uses a hybrid set of specifications called Disambiguation Tags, which defines the threshold of the analysis relevant for statistic measurements, i.e.: a. PoS codes; b. the lemmas ESSERE and AVERE; c. the MSD tags about non-finite moods of verbs. This information is the same used by the PiTagger in the disambiguation phase.
.. Tagset The PoS tagset used is, for the most part, in agreement with the EAGLES recommendations for the morpho-syntactic annotation of Italian language (Monachini 1996). These guidelines are built on the basis of previous experiences on multilingual corpora tagging (NERC project)7 and lexical encoding (MULTILEX and GENELEX models),8 and adopted within the LRE-MULTEXT and MLAP-PAROLE projects,9 within the framework of the European standardisation for multilingual resources.10 The EAGLES tagset for Italian is quite a traditional one, given that it follows the PoS given by descriptive grammars. The aim of this tagset is for it to be adequate both for morpho-syntactic tagging and for the building of lexical resources (Calzolari & Monachini 1996; Leech & Wilson 1993), in order to create a coherent basis on which to operate within various computational linguistics tasks.11 The C-ORAL-ROM tagset, entirely oriented towards corpus tagging, features some adjustments in relation to specific category spaces:
The Italian corpus
1. with regard to verbs, in order to operate a distinction between main verbal instances and non main ones (auxiliaries and copulas); 2. with regard to pronouns and determiners, in order to achieve a semantic-oriented classification of such lexical objects (independent from their functional value). Table 2.3 presents the tagset used for the PoS tagging of Italian C-ORAL-ROM, each with a short definition and an example. The rows in grey refer to the tagset choices Table 2.3 The Italian C-ORAL-ROM PoS tagset
Emanuela Cresti, Alessandro Panunzi, and Antonietta Scarano
which do not follow the EAGLES guidelines; these are detailed in the paragraph following. Tables 2.9 and 2.10 show the tagset extension used to deal with non-standard forms and other special items in the transcript.
... Choices in the PoS tagset Morphological annotation of verbs. A detailed morphological analysis is provided for verbs, to enable the extraction of data which are necessary for grammatical studies. For example, the morpho-syntactic description of verb forms is necessary to determine whether an utterance is a verbal or nominal one. To this end, the trait of finiteness is crucial to distinguish nominal uses from proper verbal uses. The former are found in utterances constituted by one verb in non-finite form, which has neither a subject nor an argument structure, and the resulting utterance has the characteristics of a nominal one in Romance languages: e.g. volare! cantare! ‘to fly!’ ‘to sing!’ Tables 2.4 and 2.5 below show the level of annotation for verbs, divided between finite and non-finite forms. In the first row, the features considered by the tagset are given; optional features are bracketed. If a given feature can assume different values, these are listed in the rows below. The tags are presented in the positional order that they assume in the tagged text. Table 2.4 Verbs: Finite forms Verb
(non-main)
Number singular
V
Person s
W plural
Mood
first
1
second
2
p third
3
Tense
(encl.)
indicative
i
present
p
subjunctive
c
past
r
conditional
d
imperfect
i
imperative
m
future
f
_E
Examples: amo\AMARE\Vs1ip amalo\AMARE\Vs2mp_E
Table 2.5 Verbs: Non-finite forms Verb
V
(non-main)
W
Gender
Number
masculine*
M
feminine*
F
common*
N
singular*
plural* * only for participles Examples: amarlo\AMARE\Vfp_E amata\AMARE\Vfspr
Mood s
Tense
infinite
f
participle
p
gerund
g
p
present
(encl.) p _E
past
r
The Italian corpus
Main and non-main verbs. The morpho-syntactic description of verbal forms represents a crucial problem for automatic PoS tagging. At the current state of the art, it is not easy to achieve a good level of automatic recognition for both auxiliaries and copulas.12 Furthermore, the Pi-Tagger procedure for Italian corpora does not mark the distinction between auxiliaries [ESSERE (‘to be’) and AVERE (‘to have’)] and copulas [ESSERE] and the predicative uses of those verbs.13 This leads to a lack of information in the tagged resource which is significant for the linguistic studies on C-ORAL-ROM corpora (in verbal vs. non verbal utterances; percentage of name vs. verbs), and in general for the linguistic exploitation of the resource. The Italian C-ORAL-ROM finds a way out of this problem, encoding both auxiliaries, verbs [ESSERE and AVERE], and copulas [ESSERE] with the [non-main] feature. An automatic post-edit procedure provides a distinction between the different uses of these verbs: MAIN ones (fully semantic, predicative verbs) and NON-MAIN ones (auxiliaries and copulas). The MAIN versus NON-MAIN distinction represents the first level of annotation for the verbal function (predicative and of “grammatical”) and allows the identification of predicative verbs (marked MAIN) with positive consequences for the corpus-based linguistics studies. The negative consequence of such a treatment is to have the auxiliary and copula instances of ESSERE falling into the same category, with no distinction between them. The tag dedicated to the NON-MAIN verbs is a capital W following V, the main tag. Table 2.6 shows the strategy adopted to tag the verbs ESSERE and AVERE in the most common contexts and uses. The automatic annotation of the MAIN/NON-MAIN value of the occurrence of verbs ESSERE and AVERE was achieved through a post-edit procedure that provided to operate the distinction, on the basis of a set of contextual rules, on the already tagged text. Pronouns and pronominal adjectives. With regard to the categorical space outlined by the traditional categories of “pronouns” and “pronominal adjectives”,14 the Italian C-ORAL-ROM tagset diverges from the EAGLES standards. Specifically, no distinction is made between pronominal and adjectival uses of the possessives, indefinites and determinatives, which are all merged into the same category. This choice follows the strategy used in the tagset for the Portuguese C-ORAL-ROM. Table 2.7 shows the relations between the Italian categories in EAGLES and C-ORAL-ROM. As shown in Table 2.7, while it is possible to map the EAGLES categorisation on the C-ORAL-ROM one, the reverse procedure is not applicable. The EAGLES tagset is structured on a sub-categorisation system within Pronouns and Determiners (macrocategories which represent the general PoS tags); in the Italian C-ORAL-ROM tagset, however, following extensive criteria, each category is treated as a proper PoS, merging the pronominal and the adjectival uses of the same lemmas. The EAGLES choices, according to a most accurate level of linguistic description, are motivated by the perspective of the achievement of a level of analysis that would be
Emanuela Cresti, Alessandro Panunzi, and Antonietta Scarano
Table 2.6 Essere and avere tagging Value
Examples
Classification Tag
predicate era\ESSERE\V il due settembre [it was September the second] mio nonno aveva\AVERE\V solo la bicicletta [my grandfather only had a bicycle] c’ è\ESSERE\V un cane in mezzo alla strada [there’s a dog in the middle of the road] quello che hai visto è\ESSERE\V il padre di Luca [the man you saw is Luca’s father]
MAIN
V
copula
NON-MAIN
VW
NON-MAIN
VW
Marco è\ESSERE\VW un ingegnere (NP) [Marco is an engineer] sono\ESSERE\VW troppo stanco (A) per uscire di nuovo [I’m too tired to go out again] il vaso è\ESSERE\VW rotto (PP) [the vase is broken]
auxiliary mio figlio se ne è\ESSERE\VW andato di casa a diciotto anni [my son went away from home at eighteen] il ladro è\ESSERE\VW stato\ESSERE\VW catturato [the thief has been caught] il gatto ha\AVERE\VW mangiato tutto [the cat has eaten everything] ho\AVERE\VW corso un’ora di seguito [I have run for a whole hour]
Table 2.7 Categories in EAGLES and C-ORAL-ROM EAGLES Pronouns subcat. Personals Reflexives Relatives Interrogatives Exclamatives Demonstratives Indefinites Possessives
Determiners subcat.
C-ORAL-ROM Merged categories Personals
Relatives Interrogatives Exclamatives Demonstratives Indefinites Possessives
Relatives
Demonstratives Indefinites Possessives
consistent with the lexicon encoding standards.15 However, such a choice encounters problems from several points of view: a.
The traditional label of “pronominal adjectives”, widely used in the description of Romance languages, cannot be entirely mapped to the notion of determiners, mostly used within the English and American traditions.16
The Italian corpus
b. From a theoretical point of view, Italian articles are determiners (for example, they are in complementary distribution with respect to demonstratives). As Romance tagsets normally feature articles separated from the other PoS (in accordance with the TEI proposal), the determiner word class would be inadequate according to these choices.17 c. Neither the syntactic nor the semantic values of the elements which would be tagged as determiners are uniform; on the other hand, the pronominal class would be a hybrid one, including both determiners in pronominal functions and strictly pronominal elements. The Italian system can be summarised as follows: a.
possessives in Italian are mostly adjectival elements (Renzi et al. 1988; Serianni 1988) but they can never be determiners (occurring with articles or other determiners, both as pronouns, as in (1), and as adjectives, as in (2)): (1) ogni cappella c’ aveva la\IL\R sua\SUO\POS [every chapel had its own] (2) me lo vende un\UN\R mio\MIO\POS amico [a friend of mine is selling it to me]
b. indefinites are quantifiers: (3) tremilacinquecento caffè / non son tanti\TANTO\IND [3500 coffees aren’t many] (4) si potrebbero risolvere / tanti\TANTO\IND problemi [many problems could be solved] c.
determinatives are true determiners (in complementary distribution with articles), mostly with deictic value: (5) un ferro / uguale a quello\QUELLO\DET + [an iron the same as that one] (6) era allucinante / quell’\QUELLO\DET uomo // [he was terrifying, that man]
d. personals are pronouns (they cannot appear with determiners), whose referents are unequivocally determined though a deixis on the basis of the participants to the dialogue and their common knowledge; e. relatives are mostly pronouns (also used in adjectival functions) which have an anaphoric value (with respect to an NP in the speech context) or act as interrogative markers (within a wh- question). The main definition underlying the above classification is based on lexical-semantic considerations. The definition criteria adopted avoids the merging of the above semantic categories into heterogeneous macro-classes which do not represent a valid
Emanuela Cresti, Alessandro Panunzi, and Antonietta Scarano
generalisation with respect to their distributional characters. Moreover, this reduction in the tagset allows a lower error rate of the automatic PoS tagging results. For these reasons this classification has been preferred for the Italian tagset to the distributional one proposed in EAGLES.
.. Extended tagset for spoken language The spoken language transcripts of C-ORAL-ROM corpora contain elements of different types.18 More specifically, besides the linguistic elements which belong to the dictionary, there is also a wide variety of non-standard linguistic forms in the corpora. The following cases have been distinguished: a. b. c. d.
foreign words (they); new formations (bruttozzi, torniante); onomatopoeia (fffs, zun); language acquisition forms (aua, cutta).
Moreover, a wide series of non-linguistic phenomena may also be involved in the speech flow. Such phenomena are identified in the transcripts following the C-ORALROM format and are encoded in the tagged resource as special elements: a. word fragments (&pa, &costrui); b. phonetic support elements and pause fillings (&he, &mh); c. laughs and coughs. Table 2.8 shows the underlying structure of the tagsets, with respect to the classification of the elements occurring in the transcripts. In accordance with this general scheme, the tagsets for non-standard linguistic elements and for non-linguistic elements, and the specifications determining these choices, are described in the paragraphs following.
Table 2.8 General structure of the Italian C-ORAL-ROM tagset ROOT classification linguistic elements
non-linguistic elements (NL tagset)
Secondary classification Standard (PoS tagset)
compositional non-compositional
non-standard (NS tagset)
compositional non-compositional
paralinguistic extralinguistic
Elements classified all parts of speech interjection (according to tradition, within PoS) foreign and new forms onomatopoeia, language acquisition forms fragm. words, phonetic supports, pause fillings coughs and laughs
The Italian corpus
Table 2.9 Non-standard (NS) tagset Non-standard element Compositional Foreign forms New formations Non-compositional Acquisition Onomatopeic
Tag
Example
(PoS+)K (PoS+)Z
they\PERK torniante\SZ
ACQ ONO
cutta\ACQ zun\ONO
... Non-standard words tagset Although foreign words, occasional neologisms, new formations, onomatopoeia and language acquisition forms, as listed in Table 2.9, are not widely present in the corpus, their treatment is important when the quality level of the tagging results is taken into consideration. The main feature which marks the distinction between these forms is the syntactic value of these elements within the linguistic structures: while foreign and new forms are compositional elements (which follows the syntactic criterion), onomatopoeia and language acquisition forms are non-compositional. For example, the foreign word in example (7) and the new formation in example (8) produce complex Noun Phrases (underlined in the text) and preserve the agreement features. Conversely, the onomatopoeic element in example (9) is not compositional; i.e. it lacks both syntactic and argumental relationships with respect to the other linguistic elements of the utterance: (7) *ROS: le &mac [/] raccomandazioni del gruppo per la prevenzione / gruppo spread\AK / [the &mac [/] recommendations of the prevention group / spread group] (8) *ANG: [. . . ] con quelli babbussi\SZ brutti che arrivavono . . . [with those ugly bad-guys which were coming] (9) *MON: i’ tronco / dentro / quando t’ arrivi a segarlo / fffs\ONO / s’ allarga // [the trunk / inside / when you start to saw / fffs / it widens] This generalisation on the non-compositional value of certain lexical elements has the consequence of establishing a macro-class encompassing onomatopoeia and acquisition forms (as non-standard elements) and interjections (standard elements).19 As a rule, all non-standard forms should not be considered tokens for lemmatisation. Since their occasional use does not conform to the definition of lemma, they are reported in the frequency lists as simple forms without lemma specification.
... Non-linguistic elements Non-linguistic elements present in the transcript (see the C-ORAL-ROM format requirements in the introduction) and all words (or word chains) that the transcribers have not understood are tagged with special codes,20 illustrated in Table 2.10.
Emanuela Cresti, Alessandro Panunzi, and Antonietta Scarano
Table 2.10 Non-linguistic (NL) tagset Non-linguistic element
Tag
Example
Paralinguistic Extralinguistic Non-understandable words
PLG XLG X
&he|PLG hhh\XLG xxx\X
.. Other choices ... Regional and dialectal forms Spoken language, Italian in particular, is characterised by a wide presence of regional and dialectal forms.21 This problem is particularly relevant in the verb class, because of the complexity of the Italian verbal inflection and the relevance of dialects within the inflectional paradigms. The regional inflected forms are inserted in a pre-dictionary already analysed and tagged with respect to the full set of morpho-syntactic features and standard lemma information. ’apito\CAPIRE\Vmspr [understood] ’cchiappavono\ACCHIAPPARE\Vp3ii [they were catching]
... Multiwords Definition. A multiword is a complex lexical unit, made up of a group of words (two or more) which have a single linguistic value. The meaning of a multiword expression does not strictly depend on the composition of the single meanings of each word (even if in many cases it is possible to find a head constituent); moreover, the word class of a multiword can be completely independent from the PoS values of its components (e.g. un sacco di is a sequence of an article, a noun and a preposition, but it assumes the value of an indefinite quantifier). In general, multiwords assume a holistic value as a whole, creating an independent lexical entry with specific semantic value and syntactic function. An explicit list of criteria is needed to distinguish the objects we consider as real multiwords from other expressions whose meaning is made up of syntactic and semantic relationships. The criteria followed to test whether a compound lexical expression is a multiword may follow different perspectives (see Voghera 1994); for example: 1. insertions within the word-chain are not possible (morpho-syntactic cohesion); 2. it is not possible to interchange the elements which compose the multiword expression (morpho-syntactic fixity); 3. in the speech flow, it is impossible to insert a prosodic break between two elements of a multiword (prosodic cohesion). These criteria are used to identify different degrees of cohesion with respect to the compound lexical expressions, rather than to be applied at the same time to all multi-
The Italian corpus
words. The presence of at least one of the characters described in the list is necessary to claim that a generic “compound expression” is a real multiword. Verbal multiwords were excluded from the list, because of the difficulties in identifying the borderline between the verbal composition (syntactic and thematic structures) and the different types of real verbal multiwords (phrasal verbs, analytic forms, locutions). The criteria used to identify other kinds of multiwords do not predict the behaviour of verbal multiwords, whose status frequently remains ambiguous. For example, the following expression can be used in Italian as a phrasal verb,22 as in example (10): (10) sono molti anni che [corre dietro] a quel suo sogno (verbal multiword); [he’s been [running after] that dream of his for many years] but in other contexts, such as (11), the same word patterns do not create phrasal verbs: (11) per tutta la gara [ha corso] dietro al vincitore (verb + preposition);23 [he [has run] behind the winner for the whole time of the race] Multiwords in Italian can belong to several word classes, not only to the verbal one. The following list shows some examples of compound lexemes categorised as adjective, adverb, conjunction, preposition and interjection:24 di bassa lega\A a viso aperto\B in modo da\C a causa di\E meno male\I
[lowly] [openly] [so as to] [because of] [it’s just as well]
As for nouns, in the Italian C-ORAL-ROM a complex nominal lexeme is considered a multiword when: 1. the replacement of the compound expression with its head is not possible; e.g. acqua minerale ‘mineral water’ is replaceable with acqua ‘water’, so it is not a multiword; aria condizionata ‘air conditioning’ is not replaceable with aria ‘air’, so this is a multiword; 2. even if the replacement of the head is possible, the compound expression is a non-grammatical pattern (two nouns in linear sequence); e.g. linguaggio macchina ‘machine language’ is replaceable with linguaggio ‘language’ but the word pattern is non-grammatical, therefore it is a multiword. The possibility of alternative categorisations for a multiword is similar to the same possibility for simple words: e.g. pagamento in_contanti\IN CONTANTI\A ‘cashpayment’; pagare in_contanti\IN CONTANTI\B ‘to pay cash’.
Ambiguous clustering of multiwords. The linear sequence of words which compose a multiword expression can be read, in some cases, as a syntactic cluster, thus creat-
Emanuela Cresti, Alessandro Panunzi, and Antonietta Scarano
ing ambiguity in the tokenisation. As shown in the following examples (12) and (13), we cannot automatically tag an expression like un sacco di as a whole lexical item, because in (13) each element of the expression, which in (12) is a multiword, has its own syntactic value within the constituent: (12) per fare il purè ci vogliono UN_SACCO_DI\IND patate [it takes A LOT OF potatoes to make purée] (13) UN\R SACCO\N DI\E patate di solito pesa due chili [A SACK OF potatoes usually weighs two kilos] At the current state of the art, it is still impossible to obtain an automatic lemmatisation for these objects, which need a semantic interpretation in the context to be disambiguated.
... Names Proper names are identified in transcriptions as the only words starting with a capital letter, thus avoiding any problems of identification of such items. In the tagset, these elements are characterised by the SP code, both for single names and for multiword proper nouns. Multiword names are joined with underscores in the lemmatised output. The expressions considered as proper nouns (with examples for each type) are listed below: a.
Anthroponyms: names, surnames, nicknames Massimo\MASSIMO\SP, Fausto_Bertinotti\FAUSTO BERTINOTTI\SP
b.
Toponyms: Africa\AFRICA\SP, San_Gottardo\SAN GOTTARDO\SP
c.
Works of art, books, movies, TV, music bands, games: Alice_nel_Paese_delle_Meraviglie\ALICE NEL PAESE DELLE MERAVIGLIE\SP, Pearl_Jam\PEARL JAM\SP
d.
Names of institutions, associations and organisations: UNESCO\UNESCO\SP, Federcalcio\FEDERCALCIO\SP
e.
Names of companies, brands, products: Sony\SONY\SP, Calvè\CALVE’\SP
f.
Acronyms: DC\DC\SP, IVA\IVA\SP
g.
Religious events Natale\NATALE\SP
The Italian corpus
.. Evaluation The C-ORAL-ROM Italian resource comprises 306,638 tagged tokens.25 Since the nonstandard and regional forms were inserted in a special pre-dictionary, the PiTagger system reached a 100% recall of the number of tokens: total number of tokens: 306,638 tagged tokens: 306,638 total recall = 100% The evaluation of the precision of the automatic PoS-tagging procedure is based on a random sampling of 1/100 tokens picked out of the whole C-ORAL-ROM Italian resource, and evaluated in their utterance contexts. Each token is extracted from a different utterance, also randomly selected. The random sampling obtained sufficiently represents the whole. The size of the sampling has been considered sufficient from a statistical point of view, as it ensures a 95% confidence interval lower than 1% (see the evaluation of the French PoS tagging in this volume).26 The manual revision of the tagged samples evaluates the automatic procedure with respect to different degrees of accuracy: errors in sub-categorisation, errors in morpho-syntactic description of verbs, errors in main tag category and in lemma assignation. The statistic precision of these levels of annotation is shown in Table 2.11, where it can be seen that the more significant errors, that is, those of PoS tag and lemma errors, only involve around 10% of tokens.27 The results of the evaluation of PoS tag errors (in the first level of errors in Table 2.11) are shown in a confusion matrix in Table 2.12, listing the errors in PoS assignment with respect to each word class: in each row the errors are recorded by category, while in the columns the correct PoS to which the token belongs is reported.28 The last column records the number of cases in which the PoS is overextended, while the bottom row represents cases of under-extension.
Table 2.11 Precision of automatic tagging procedure Total Sampling Not Decidable Total Evaluated All Correct Correct PoS tag 1. PoS tag errors 2. Lemma errors 3. Sub-category errors 4. Morpho-syntactic description errors Total Errors
3,100 31 3,069 2,726 2,773 296 2 38 7 343
1.00% 88.82% 90.36% 9.64% 0.07% 1.24% 0.23% 11.18%
Emanuela Cresti, Alessandro Panunzi, and Antonietta Scarano
Table 2.12 Confusion matrix
The data show that most mistakes occur in the Nouns category, numbering 103 out of 296 total errors.29 Considering the frequency of the category in the evaluation corpus, the projected overextension of this word class would be around 15%. This probably depends on the statistical normalisation applied by the PiTagger system, which, very roughly, assigns to Nouns the highest probability of occurrence (see Picchi 1994). Another interesting result is that the Adverbs and Interjections category is consistently under-extended with respect to its actual weight: over 10% under-extension for Adverbs and 8.5% under-extension for Interjections. Verbs show a lower incidence of errors, of 3.95% over-extension and 3.44% under-extension, with a roughly correct projection of the frequency of the category on the total. From the confusion matrix in Table 2.12, it is possible to obtain data on precision, recall and f-measure for each category. Table 2.13 details these measurements, which give an overall estimate of the automatic tagging procedure.30
The Italian corpus
Table 2.13 Precision recall and f-measure for each PoS PoS
tp
fp
fn
precision
recall
f-measure
DIM E V POS* R I B PER S N C IND A REL* NA*
65 253 559 10 181 195 502 170 440 40 200 50 90 24 2
0 7 23 1 11 6 25 27 103 1 32 21 27 12 0
0 11 20 0 14 23 85 11 19 11 39 2 33 23 5
100.00% 97.31% 96.05% 90.91% 94.27% 97.01% 95.26% 86.29% 81.03% 97.56% 86.21% 70.42% 76.92% 66.67% 100.00%
100.00% 95.83% 96.55% 100.00% 92.82% 89.45% 85.52% 93.92% 95.86% 78.43% 83.68% 96.15% 73.17% 51.06% 28.57%
1.0000 0.9656 0.9630 0.9524 0.9354 0.9308 0.9013 0.8995 0.8782 0.8696 0.8493 0.8130 0.7500 0.5783 0.4444
.. Specific problems with the morpho-syntactic tagging of spoken language Lemmatisation and PoS tagging need a detailed and coherent definition of the relevant context within which the statistics on disambiguation have to operate. Therefore, resources must include annotations regarding context boundaries. Written language is defined by period boundaries, which are marked by punctuation signs; on the contrary, the minimal relevant unit in spoken language is the utterance, which must be detected and identified within the speech flow. In the following examples, two possible transcripts of the same dialogic turn are presented: the first one without prosodic boundaries (followed by a second row bearing the PoS tags related to each word), the second one according to the Italian C-ORAL-ROM corpus (words in italics are ambiguous). (14) *NIC: sì dice che taglia porta io lì per lì un’ elle penso Adv V Conj/REL S/V S/V PER Adv Prep Adv Art S V (15) *NIC: sì // dice / che taglia porta ? io / lì per lì / un’ elle / penso // [yes, he/she says, which size do you take? There and then, I (say), a large, I think] The PoS of ambiguous elements (underlined in the second row) can be decided only in connection with the context boundaries (represented by utterance boundaries). If the disambiguation process could work on the simple bare transcripts, it would also operate on a non-coherent word pattern. The long PoS sequence reported in the example must be parsed through the context boundaries of the prosodic annotation, and then disambiguated by a PoS tagger working within the relevant contexts. Without prosodic boundaries, the disambiguation of the PoS tags belonging to the words in the dialogic turns would be arbitrary.
Emanuela Cresti, Alessandro Panunzi, and Antonietta Scarano
The prosodic annotation provided by the C-ORAL-ROM resource thus gives the automatic lemmatiser the information on the relevant linguistic unit which constitutes the context of the disambiguation process. PiTagger considers the utterance limits reported in the transcripts as the proper domain of application for the statistics. Although the utterance boundaries have been defined, in previous experiences, by the automatic detection of pauses (longer than 200 ms; see Uchimoto et al. 2002), in the C-ORAL-ROM corpora such boundaries are provided by the marking of terminal prosodic breaks (//, ?, . . . ). Besides the speech-specific feature of language constituted by utterance boundaries, during the evaluation of automatic tagging the operators marked three kinds of contexts relevant to the spoken domain. 1. Under-specification of the PoS in a given context (it is always impossible to express a disambiguation judgement). These cases (1%) have not been counted as part of the evaluation samples: (16) che\CHE\CS?REL? mi\MI\PER piace\PIACERE\Vs3ip tanto\TANTO\B // [which I like so much] 2. Cases in which the PoS disambiguation is possible only if the evaluator is provided with information on the audio; that is, with information not computable by the System. In these cases (2.7%) the PoS assignment by the system has been evaluated. The system failed to assign the right Tag in 83% of cases: (17) e\E\CC allora\ALLORA\CC?B? / decise\DECIDERE\Vs3ir di\DI\E partire\PARTIRE\Vf // [and so (he) decided to leave] 3. Cases of uncertainty, found in connection with the secondary boundaries, in which the operator, in accordance with the tagset, can in principle assign a tag, but also points out a lack of adequacy of this categorisation with respect to the spoken domain. In these cases too (3.1%), the Tags provided by the system have been evaluated positively or negatively, according to the tagset. (18) dai\DARE\V? / prendiamola\PRENDERE\V // [come on / let’s take it] The underspecification not connected to the utterance boundaries occurred in roughly 7% of the selected sample (3,100 words). A second level of evaluation provides us with an estimate of the incidence of these phenomena on the number of errors. As a result, it becomes possible to identify the contexts in which the system encounters problems and to select which of them are caused by specific features of spoken language. As will be detailed in the next sections, the automatic PoS tagging procedure is made complex in spontaneous speech at three main levels:
The Italian corpus
a. words adjacent to utterance boundaries; b. fragmentation phenomena (retracting and interruptions); c. secondary prosodic boundaries.
... Words adjacent to utterance boundaries One word utterances. In the set of total errors in PoS tag (296), 26% are one-word utterances where an ambiguous word occurs, shown in examples (19) and (20). In this case, which is typical of spontaneous speech, the statistics of the disambiguation system, based on PoS order, are radically under-determined: (19) *PRO: esatto\ESATTO\A?B? // [exactly] (20) *ELA: cosa\COSA\S?REL? + [what]
Words in peripheral position. Apart from single words which make up an utterance, in the collected data it can be observed that around 22% of the tagged tokens with a PoS error occur in the first position of the utterance, and that roughly 13% occur in the last position. In other words, this is when the system is faced with little contextual information to make a decision. (21) che\CHE\CS* devo\DOVERE\Vs1ip dire\DIRE\Vfp / io\IO\PER sono\ESSERE\V nelle\IN\E_R mani\MANO\S del\DI\E_R mio\MIO\POS compagno\COMPAGNO\S // [what can I say, I’m in my partner’s hands] (22) e\E\CC vai\ANDARE\Vs2ip sempre\SEMPRE\B diritto\DIRITTO\S* // [and keep going straight on] These data, together with the previous results on single word utterances, tell us that in 61% of cases the errors in PoS tagging occur with a peripheral position of the tagged word. To improve the results on words adjacent to utterance boundaries, the tagging system should be able to take into account the main prosodic breaks as relevant positions for the disambiguation procedure, and to be trained on a training corpus comprising such information. Given the low average utterance length which characterises spontaneous speech, the percentage of such contexts appears to be quite relevant (roughly 30% of words).
... Interruptions and retracting Interruption and retracting phenomena are a trait peculiar to spoken language. They introduce irregular and unpredictable PoS chains. An interruption involves the ending of the utterance, and it may create both an anomalous syntactic configuration (which is impossible to find in written language) and a lack of information necessary to decide on disambiguation, as illustrated in (23):
Emanuela Cresti, Alessandro Panunzi, and Antonietta Scarano
(23) *LIA: [. . . ] per farsi perdonare / che + Prep V V Conj/Rel? [to be forgiven / for] The lemmatisers have neither rules or statistics about these kinds of endings which, by definition, are not regular. The retracting phenomenon is often considered under the generic label of “disfluency phenomena” which occur in the speech flow. Retracting phenomena produce non-informative prosodic units within the utterances, with an irregular PoS sequence. Taking the utterance as the relevant context in which to apply the disambiguation rules, the retracting phenomenon (marked with “[/]” in the transcripts) generates unpredictable linear patterns, as in (24): (24) *AAA: non vedevo bene la [/] la [/] la strada // Adv V Adv Art Art Art Noun [I couldn’t see the road well] In the total number of utterances within which the selected word was wrongly tagged (296), interruptions and retracting appear in around 10% of cases.31
... PoS assignment in connection with secondary prosodic boundaries Some relevant correlations between the prosodic boundaries in the transcripts and the distribution of particular forms emerged during the lemmatisation and PoS tagging processes. In various cases, the secondary prosodic marks isolate lexical or support elements which assume peculiar values within the speech flow. Non-standard support elements. The non-standard forms he and mh constitute the phonetic supports (of the vowel and the consonant type) used in the spoken Italian language. The positions in which they occur are systematically marked by prosodic boundaries, which point out the relation between the prosodic structure of an utterance and the informational values of the elements which occur in different tone units. In utterances like the following ones, these forms may be considered (with regard to the PoS tagging) alternatively as interjections or as paralinguistic elements: (25) *MIC: no / comunque / &he / vorrei vedere il film di [/] di Troisi . . . [no / however / er / I’d like to see Troisi’s film] (26) *LIA: mh / facile // [mh / easy] These are typical and highly frequent phenomena in spoken language: the first one, as in (25), is used to take time; the second one, as in (26), to mark the taking of a dialogic turn. Such non-standard forms have a functional value within the dialogic flow, but their PoS is uncertain and their distributional character cannot be foreseen on the basis of written language data or standard linguistic rules.
The Italian corpus
Standard forms with PoS assignment problems. Even the distribution of standard forms is not consistent with the definition of the categories that are based on syntactic relationships of the words within the clause structure. The following examples show two peculiar values that conjunctions and adverbs may assume in connection with secondary prosodic boundaries, which isolate these words at the extremities of utterances (first/last position): (27) *GNA: allora / stavamo discutendo // Conj/Adv? [so / we were discussing] (28) *EEE: passiamo ad un altro argomento / via // Adv/Int? [let’s change the subject / come on] According to the traditional categorisation, allora, as in (27), may be an adverb or a conjunction, while via, as in (28), may be an interjection or an adverb. In these cases, we are not able to disambiguate the proper PoS for these ambiguous expressions. We can only highlight that these words assume special functional values in such contexts. Moreover, the tagging system used does not take the secondary boundaries into account, and it treats the words in question within a single utterance context, as if belonging to a linear pattern. The disambiguation is estimated, by default, on the basis of the maximum likelihood of a tag sequence, but in this case this process is based on an improper word pattern (e.g. allora/stavamo and argomento/via), given that the two elements do not show a proper syntactic relation. With via in (28), we could decide, on the basis of the secondary boundaries, that the ambiguous element is an interjection at the end of the utterance. On the contrary, with allora in (27), the disambiguation process is not only underdetermined, but also arbitrary. As a matter of fact, the syntactic categories of conjunction and adverb are both unsatisfactory in defining the function of this kind of forms, which is not syntactic. As these elements have such a trait, it would be correct to treat them as non-syntactic. The label of “Discourse Marker”, well-known in the literature (see references in the Chapter 6) and mostly used to denote elements that have a prominent value within the conversational relations in the speech flow, would be useful to tag elements of this kind as mainly informative ones, without further remarks on their syntactic value.
Allocutive emphatic. Similar problems emerge with fixed verb-forms isolated in the speech flow within non-terminal prosodic breaks. The following examples show this speech-specific phenomenon, in which both the PoS category and the lemma of the words senti and guarda (‘listen’ and ‘look’, imperative forms) are uncertain: (29) *SER: senti\? / e poi [/] e poi chi c’ era? V? Conj Adv Conj Adv REL Adv V [listen / and then [/] and then who was there?]
Emanuela Cresti, Alessandro Panunzi, and Antonietta Scarano
(30) *SRE: ti stavamo aspettando / guarda\? // PER V V V? [we were waiting for you / look] The words senti and guarda can never change inflection in this position, and they are more like fixed expressions with a pragmatic function (allocutive or phatic) rather than properly semantic instances of the verbs “SENTIRE” or “GUARDARE”.
Reported speech contexts. Other cases in which the distribution in spoken language is not consistent with the distributional characters found in written language may also arise when reported speech comes about. In this case, the expression which introduces the reported speech creates a linear sequence of categories in which the expression is not separated by proper punctuation, as it would be in written language, as in (31) or an odd sequence, for example, three verbs in sequence, as in (32): (31) *MAR: [. . . ] dice / allora vo a vedere // V Adv V Prep V [(he/she)says / I’ll go and see then] (32) *NOR: [. . . ] Vincenzo / lasciala fare / dice + Noun V V V [Vincenzo / leave her alone /(he/she) says] It can be maintained that, in connection to secondary prosodic boundaries, the statistics extracted from the written training corpora may hardly be recognised and properly applied. In such positions, a word-form may have pragmatic values which match neither its typical PoS, nor its lemma. Other non-standard forms with pragmatic functions can appear in similar isolated positions. The data collected in the evaluation phase of the automatic tagging show that 13.2% of the errors are represented by words in a single tone unit (i.e. isolated by secondary prosodic breaks), mostly standard words with uncertain classification. A tagging system able to assume information on both primary and secondary prosodic boundaries as contextual inputs is needed to achieve an adequate treatment of these forms in the prosodic context. The training of tools on corpora which comprehend such an annotation level will be highly relevant for the improvement of the results on the automatic tagging for spoken corpora. The state of the art of automatic taggers for Italian does not feature such an annotation level.
. Main data from lemmatisation The high frequency lists extracted from the Italian C-ORAL-ROM corpora were compared with a previous frequency lexicon of spoken Italian (LIP) and with a written resource (Linguistic Miner).32
The Italian corpus
Table 2.14 High frequency verbs, excluding auxiliaries and modal verbs: comparison of C-ORAL-ROM and other corpora Rank
C-ORALROM
LIP
Written resource
Rank
C-ORALROM
LIP
Written resource
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50
fare dire andare vedere sapere stare venire capire mettere dare partire sentire arrivare prendere pensare parlare guardare trovare portare credere sembrare bisognare chiamare lavorare cercare entrare chiedere conoscere aspettare passare rimanere riuscire piacere ricordare cominciare lasciare comprare cambiare vivere mangiare tornare tenere leggere rispondere diventare riguardare selezionare usare aprire finire
fare dire andare vedere sapere stare dare parlare venire capire mettere sentire pensare trovare guardare chiamare prendere portare credere arrivare ricordare chiedere scrivere scusare tenere sembrare bisognare passare conoscere riuscire cercare lasciare aspettare lavorare finire rimanere leggere bastare entrare mangiare mandare cominciare partire riguardare parere diventare pagare uscire cambiare succedere
fare dire venire andare vedere stare sembrare mettere chiedere parlare trovare prendere diventare entrare continuare prevedere tenere decidere spiegare pubblicare pensare arrivare dare perdere attraversare rimanere capire considerare cercare tornare lasciare sapere ottenere riguardare ricordare raggiungere rispondere costituire rendere portare apparire cominciare uscire seguire sentire credere consentire presentare vivere definire
51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100
perdere capitare nascere esprimere seguire buttare versare uscire chiudere costruire pagare parere servire sperare creare mandare pigliare scegliere toccare morire vendere rendere ritornare spiegare raccontare scusare decidere provare continuare iniziare esistere incontrare presentare mancare succedere scrivere discutere camminare fermare interessare funzionare imparare ripetere giocare proporre costare durare riconoscere intendere produrre
presentare servire perdere tornare piacere continuare ringraziare sperare spiegare telefonare provare usare morire aprire ascoltare rendere rispondere significare trattare interessare vivere vendere mancare seguire studiare considerare esistere ripetere scegliere fermare ritornare porre nascere pigliare chiudere occupare decidere aiutare raccontare tirare iniziare riconoscere intendere valere proporre organizzare comprare togliere dimenticare dipendere
rappresentare conoscere riuscire realizzare utilizzare esistere cambiare scrivere guardare chiamare restare pagare effettuare lavorare intendere raccontare dichiarare affermare avvenire risultare sostenere precisare bisognare partire ritenere finire dimostrare annunciare passare mancare aprire indicare aggiungere giocare evitare superare scegliere destinare prestare fornire parere svolgere aspettare controllare riconoscere studiare registrare convincere offrire proporre
Emanuela Cresti, Alessandro Panunzi, and Antonietta Scarano
Table 2.15 High frequency adverbs: comparison of C-ORAL-ROM and other corpora Rank
C-ORALROM
LIP
Written resource
Rank C-ORALROM
LIP
Written resource
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50
non si’ no ci poi come piu’ cioe’ qui ecco insomma li’ cosi’ bene sempre allora proprio quindi po’ quando molto gia’ ancora prima perche’ forse dove oggi mai niente su pratico* ora adesso invece qua meglio via subito comunque per_esempio tanto senno’ nulla eccetera giu’ vero* almeno soprattutto mica
non si’ no poi più’ bene ecco poco qui così’ sempre molto proprio solo già ancora adesso prima invece qua li’ certo oggi appunto ora meno forse mai pure veramente la’ tanto eccetera meglio soltanto dopo fuori magari avanti via praticamente ormai niente insieme subito abbastanza quasi almeno soprattutto ieri
non come ci su dopo prima ancora quando poi sempre ieri dove molto oggi fino invece mai oltre quasi forse soprattutto quindi poco qui pure comunque soltanto allora almeno no po’ ormai subito fuori adesso inoltre nulla appena niente infine ecco tuttavia dietro intanto sopra insomma domani davvero addirittura piuttosto
51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100
purtroppo fa sicuramente tutto domani probabilmente male dentro mo’ troppo naturalmente giù’ addirittura intanto certamente ovviamente neanche assolutamente mica completamente sotto direttamente nemmeno spesso chiaramente davanti stasera altrimenti vicino appena sopra effettivamente evidentemente davvero giustamente piuttosto solamente su presto eventualmente piano stamattina dietro esatto oltre semplicemente perlomeno nulla indietro accanto
neppure probabile* dentro diretto* accanto finora naturale* nemmeno magari certo* finale* attorno attuale* particolare* persino neanche perfino abbastanza completo* ovvio indietro assoluto* sicuro* immediato* peraltro purtroppo vero* rispettivo* semplice* altrettanto altrimenti affatto successivo* esclusivo* evidente* recente* pratico* esatto* talvolta contemporaneo* stasera facile* deciso* definitivo* sostanziale* bene stavolta qua ufficiale* perfetto*
abbastanza la’ solo sicuro* meno naturale* poco assoluto* ok davvero qualcosa magari dopo domani ormai purtroppo ovvio troppo ieri dentro intanto chiaro* più_o_meno fuori male probabile* stasera addirittura rispetto_a certo* nemmeno mo pure piuttosto dietro avanti completo* mezzo quasi diretto* evidente* fino indietro semplice* soltanto insieme sopra in_effetti a_posto effettivo*
* The asterisk indicates adverbs in “-mente” reported to lemma without the suffix by the PiTagger tool.
The Italian corpus
Table 2.16 High frequency adjectives: comparison of C-ORAL-ROM and other corpora Rank
C-ORALROM
LIP
Written resource
Rank
C-ORALROM
LIP
Written resource
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50
bello grande certo vero diverso solo giusto bellino bravo importante possibile nuovo esatto chiaro piccolo mezzo politico italiano pronto unico complesso buono difficile prossimo automatico grosso ferroviario pubblico scorso ultimo interessante locale maggiore particolare forte strano alto tranquillo giovane facile preciso sociale libero lungo mondiale piano umano famoso tedesco attento
grande bello certo vero importante diverso buono nuovo piccolo ultimo giusto prossimo solo vario politico possibile grosso chiaro bravo particolare italiano unico difficile generale lungo vecchio libero caro presente alto mezzo scorso economico uguale facile forte pubblico strano nazionale solito semplice migliore normale rosso esatto povero sociale attento bianco nero
grande certo diverso possibile vero scientifico economico unico pronto bello attuale fiscale regionale mondiale relativo finanziario elettorale comunale successivo culturale amministrativo commerciale recente facile professionale intero speciale numeroso costretto prossimo operativo aperto ulteriore monetario capace produttivo tradizionale necessario vario complessivo televisivo famoso definitivo disponibile urbano giusto agricolo tedesco antico significativo
51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100
fondamentale inutile amico centrale generale normale bene fine incredibile migliore professionale civile comunale continuo da_solo aperto brutto elettorale azionario felice grave nero regionale semplice vicino attuale basso lontano pieno cosiddetto leggero nazionale operaio personale a_parte generico internazionale militare necessario pericoloso positivo scusato usato vecchio comune fotografico intero sanitario scientifico sicuro
storico aperto brutto famoso preciso attuale sicuro maggiore contento reale interessante positivo basso incredibile pronto verde minimo precedente caldo necessario breve centrale personale civile giovane comune inutile perfetto serio tranquillo vicino 