The ORD Speech Corpus of Russian Everyday Communication One Speaker s Day : Creation Principles and Annotation

Transcription

1 The ORD Speech Corpus of Russian Everyday Communication One Speaker s Day : Creation Principles and Annotation Alexander Asinovsky, Natalia Bogdanova, Marina Rusakova, Anastassia Ryko, Svetlana Stepanova, and Tatiana Sherstinova St. Petersburg State University, St. Petersburg, Universitetsakaya nab. 11, , Russia {a.s.asinovsky,mvrusakova,sherstinova}@gmail.com, {nvbogdanova 2005,aryko,stsvet 2002}@mail.ru Abstract. The main aim of the ORD speech corpus is to fix Russian spontaneous speech in natural communicative situations. The corpus presents the unique linguistic material, allowing to perform fundamental research in many scientific aspects and to solve different practical tasks, especially in speech technologies. The paper concerns methodology and description of the ORD corpus creating and presents the system of annotations. 1 Introduction Beginning from 1990s national corpora of spontaneous speech are created in many countries of the world. The first audio corpus based on spontaneous oral speech of subjects organized in a demographically balanced sample is a part of the Spoken Component of the British National corpus [12]. So far, there is no really representative corpus of Russian everyday speech, though some databases of oral speech have been elaborated by various scientific groups for at least recent 40 years [1]. Although there is some progress achieved in the exploration of informal speech, this domain certainly calls for further investigation. For example, it is not known how many wordforms (morpheme, sentence) tokens are produced and perceived by a speaker per hour, day, month; how many different language units he (she) processes per various periods of time or what is the average duration of sounding speech within these periods. The lack of linguistic information of this kind is well comprehensible: carrying out investigations needed to achieve such results is extremely labour- and time-consuming, and technical tools for such explorations became available comparatively not long ago. Now both technical means and theoretical background for the exploration of spontaneous oral speech do exist. Nevertheless, these investigations should be treated as highly innovative, because they can be held only by professional teams including many researchers. During the last decades a great amount of Russian natural oral speech data was collected. Unfortunately, when gathering speech material researchers use different methods of collecting and pursue their individual scientific goals. As a result, the available resources are not uniform, uncoordinated and it is rather difficult to integrate the data V. Matoušek and P. Mautner (Eds.): TSD 2009, LNAI 5729, pp , c Springer-Verlag Berlin Heidelberg 2009

2 The ORD Speech Corpus of Russian Everyday Communication One Speaker s Day 251 into a single representative corpus open for general use. Therefore, shaping of the fullfledged descriptive database of the Russian oral informal speech is the prerequisite for the progress in those linguistic spheres which are concerned with the native speaker, natural speech and, in a larger scale, communicative behaviour in general. 2 The Main Concepts of the ORD Corpus 2.1 Methodological Background The main goal of the present investigation is to fix Russian spontaneous speech in natural communicative situations. Firstly, it means that nothing should interfere with the usual habits of speakers communicative behaviour in particular speech situations during recording. For example, speech communication at breakfast should be realized in those circumstances that are habitual to every individual: in the same setting and with the same interlocutors as usual, with the same level of accompanying noise (with an open window, refrigerator functioning, etc.). Secondly, every subject should realize his speech ability in standard situations, not changing the usual length of utterances, as well as speech topics and repertoire during recording. For example, if a subject got used to read a newspaper, he should not give up this habit and should not organize an unusual for this moment communication (e.g., intentionally talking with other members of the family or inviting unusual guests) in order to enlarge the amount of speech material within the recorded time. At the same time at the first stage of our investigation the participants were asked to make a deliberate choice of a particular day for recordings. For example, it is preferable to make the recording on an ordinary day, rather than on a day when the subject is about to do something unusual (e.g., to go for an excursion, to be absent from home for the whole day because of a seasonal rush at his office, etc.) [2]. 2.2 Technical Equipment Recording is made with the help of a dictaphone. At first the subject makes all the necessary settings and fasten the dictaphone to his clothes (into a pocket or a special bag). Such a mode of recording inevitably causes non-uniformity of the quality of the obtained speech data. In our research we have been using Olympus WS-320M dictaphone, which ensures more that 35 hours of high-quality recordings. A relatively low level of quality in the recording obtained in this way (if compared to studio recordings) is an unavoidable consequence of undertaking a fieldwork aimed at the natural experiment with human speech behaviour. 2.3 Speakers Selection and Training At this stage of the research the study is not directly aimed at the description of the Russian language in the whole richness of its manifestations. There is only one form of its functioning that is being studied, viz. the speech of naïve native speakers of Russian who live in a city where the Standard language is dominating, where there is no significant impact on the part of dialects or other languages, and where the population is professionally heterogeneous and unbiased in terms of age and gender distribution.

3 252 A. Asinovsky et al. St. Petersburg is a model city of this type. This is the reason why it was St. Petersburg where the subjects were being selected. The objective of the present research supposes that subjects should never be chosen among those speakers whose occupation requires any specific level of speech control or self-consciousness in this respect; indeed, it is known that these professional skills might have a strong impact on speech production by individuals. The preparatory work with the subject consists of two stages: 1) providing the necessary degree of naturalness of the subject s behaviour in communicative situations; 2) providing proper technical quality of the recordings necessary for further speech analysis. The first of these two stages is indeed very complicated. During a pilot experiment, in which the very members of the research team (i.e., linguists themselves) recorded their days-of-speech, it was detected that despite being highly motivated in terms of striving for adequate results, the subjects were hardly able to forget that they were bearing a dictaphone: they could not abandon the idea that their communication with surrounding people, especially with their close relatives, was going to become available to their acquaintances/colleagues. Finally, it was decided to gather recordings under conditions of full and utter anonymity for the sake of obtaining their maximal naturalness. The following procedure has been elaborated. A non-member of the research team, a psycholinguist by specialization, joins the procedure. He addresses potential subjects, for example, a group of people working in a particular enterprise. When instructing potential subjects he guarantees that under no circumstances he would personally come in touch with the future materials obtained from those subjects. When the recordings are made, this intermediate collaborator hands them over to the research team. The subjects do not indicate in any form their actual names, but they have to fill in a questionnaire concerning their age, profession, place of birth and other sociological data. As a result the research team obtains recordings of speakers who are absolutely unknown to them; moreover they have never met one another in person. The shortcoming of this approach is that no control on the part of professional linguists is possible at recording stage, hence there is a relatively high ratio of spoilage and technical flaws in recorded files, which make impossible phonetic analysis of correspondent fragments. The second stage is implemented by way of instructing the subjects-to-be how to use dictaphones for the best results. 2.4 Creating the ORD Corpus The abbreviation ORD stems from Russian Odin Rechevoj Den, literally translated as one day of speech. A demographically balanced group of subjects was formed for the first series of recordings, including 30 participants (the base speakers or informants) representing various social and age strata in the population of St. Petersburg. After detailed instructions the subjects made recordings of all speech communications in which they took part during one day. Besides, all of them filled in the questionnaires and passed through psychological testing.

4 The ORD Speech Corpus of Russian Everyday Communication One Speaker s Day 253 The overall length of the recorded material is 240 hours, from which 170 hours contain speech data quite suitable for further analysis. Beside the subjects speech, that of their 520 interlocutors was also recorded. Interlocutors were people of different ages (from 3 to 68 years), professions and occupations (e.g., salesmen, conductors, managers, lecturers, doctors, librarians, IT-professionals, students, and many others), being in friendly, family, professional or other relations with the subjects. Materials contain diverse genres and styles of speech: professional conversations with colleagues, telephone talks, lectures, practical lessons of foreign languages, communications with friends and relatives during airings, parties, breakfasts, dinners, etc. The topics in these conversations also range from discussions of teeth problems with a dentist to conversations about religion, life and death. Recordings were made at home, while traveling by the public transportation, while walking outside, at the university, in the military college, in coffee bars, in the shops, in the amusement park, etc. The corpus was divided into 2202 communication episodes. 134 episodes are already transcribed in detail [2]. The recordings are processed by professional linguists. The initial stage of this process consists in preliminary description of the material and its orthographic transcription. Generally, when studying oral speech, the burden of deciphering is often put onto the shoulders of the very participants of the dialogues; these are the persons who are often able to interpret the least hearable fragments and to provide the fullest extralinguistic information relevant for the production of the texts in question. In our research, the deciphers could not witness the communicative behaviour of speakers for the sake of speech naturalness. The fact that the recordings are transcribed by non-participants results in partial loss of information. This loss should be viewed as an inevitable toll for the naturalness of the material gathered. 3 System of Annotations 3.1 Annotation Software Two professional annotation tools ELAN and Praat are used to annotate the ORD corpus. The first one is ELAN (EUDICO Linguistic Annotator) developed at the Max Planck Institute for Psycholinguistics (Nijmegen, The Netherlands), with the aim to provide a sound technological basis for annotation and exploitation of multi-media recordings. This professional tool allows to create, edit, visualize and search annotations for video and audio data. Being specifically designed for the analysis of language, ELAN is a convenient tool for multi-level annotations of communication and speech [5]. Figure 1 shows a fragment of multi-level annotation of the ORD corpus for a number of tiers made in ELAN. ELAN is used for annotation of most tiers (Episodes, Frases, Events, Words, etc.) apart from phonetic ones (Sounds, Syllables, and others presented in phonetic transcription), which are made in Praat. Praat is a professional phonetic annotator which allows to analyze and manipulate speech. The Praat program was created by Paul Boersma and David Weenink (the Institute of Phonetics Sciences, University of Amsterdam, The Netherlands) [8]. The structures of ELAN and Praat annotations are fully compatible.

5 254 A. Asinovsky et al. Fig. 1. A fragment of multi-level annotation of the ORD corpus made in ELAN. Tiers: Frase (orthographic transcripts), SynIdeal (syntagmas), Words (orthographic), WordIdeal and WordReal (words in phonetic transcription), POS (part of speech), GramForm (grammatic form), SyntRoles (syntactical role of the word), Syllable (real phonetic transcription), MorphemeReal (morpheme in real phonetic transcription), SoundReal (sounds in real phonetic transcription). 3.2 Annotation Principles The main principles of multi-level annotation for a spoken corpus were given in [3]. Primary annotation of speech in the ORD corpus is made in ELAN and implies annotation of the followings data types: Frase - orthographic transcripts of phrases, which are the main units of description. Besides transcripts it contains references to pauses and noisy fragments. Independent type. Event - nonlanguage audio events (dog barking, squeak of a door, phone ring, radio program, etc.), including as well some voiced events (e.g., cough, yawn, laugh, moan, etc.) not connected with speech. Independent type. Speaker is a person who pronouncedcorrespondentphrase (Frase) either a base speaker (subject) or one of his/her communicators. This type depends on Frase (Symbolic Association). Voice is the special characteristic of speech for the given phrase (Frase) or its segment (e.g., hoarse, whisper, scanning, irritated, imitating, ironical, dramatic, etc.). Depends on Frase (Included In). FraseComment contains all kinds of comments on phrase realisation and researcher s remarks. Depends on Frase (Symbolic Association). Notes may contain other useful information. Independent type. Episode (Communication episode) refers to general communicational situation. Independent type. These main 7 data types are being annotated in ELAN on 7 correspondent tiers: Frase, Events, Speaker, Voice, FraseComments, Notes, Episodes.

6 The ORD Speech Corpus of Russian Everyday Communication One Speaker s Day 255 Some spelling rules for transcripts (tier Frase): Speech is written in standard orthography. Speech is divided into conventional sentences [10]. Transcripts may include the following symbols: Symbol Meaning / (//) marks of syntagmatic (phrasal) division H fragment of unintelligible speech P barely intelligible speech Π pause () short hesitation pause (...) long hesitation pause, which may be filled (m-m), (a-a), hesitation pause filled by sounds (e-e), (a-m)!,? marks indicating exclamatory and interrogative utterances ca... interrupted word... unfinished phrases (?) questionable or ambiguous transcript # change of a speaker in overlapping remark inserted by another speaker in overlapping speech fragments Speaker s code is attached to each phrase (e.g., M2). If a person-interlocutor is unknown or unidentified, his/her code is marked by symbol X (e.g., MX1). In case interlocutors speak simultaneously we use symbol # as a delimiter for individual codes (e.g., S01#F5) or (e.g., S07@M2 means that while S07 is speaking his second male interlocutor pronounces an insert remark). Figure 2 shows a sample of primary annotation made for three tiers. Fig. 2. Sample of primary annotation for three tiers (Frase, Speaker, Events) The second stage of corpus annotation is made on lexical level and includes tagging for the following main tiers: Words (spelling), POS, GramForm (grammatical form), and SyntRole (syntactic role). Description of annotations made on phonetic levels in Praat may be found in [9].

7 256 A. Asinovsky et al. 4 Conclusion Though creating of the ORD corpus is still in progress, its diverse speech material and annotations have given birth to a number of linguistic and interdisciplinary researches. Let us mention just some of the papers. In [4] the most frequent reduced wordforms of spontaneous Russian speech are described. The work [10] discusses general problems of separation of various linguistic units from a real speech stream, whereas the paper [11] describes dynamics of communication episodes in everyday life. Article [6] relates to psycholinguistic studies and concerns the dependence of speech characteristics on speaker s mental state and personality. In [13] one may find sociological description of the speakers recorded for the ORD, and the paper [7] presents study on the everyday rhetoric and its techniques. The new project which have been started on the material of the ORD corpus is creating an audio dictionary of Russian morphemes. Acknowledgements The first recordings and database creating of the ORD corpus were supported by the Russian Foundation for Humanities within the framework of the project Speech Corpus of Russian Everyday Communication One Speaker s Day (project # e/Ya). Nowadays creating of the corpus is supported by the program of the Russian Ministry of Education titled Sound Form of Russian Grammar System in Communicative and Informational Approach and by the grant of the Russian Foundation for Humanities Development of an Information System for Monitoring of Russian Spoken Language (project # v). References 1. Asinovsky, A.S., Arkhipova, E.A., Bogdanova, N.V., Rusakova, M.V., Ryko, A.I., Stepanova, S.B., Sherstinova, T.Y.: Polevaya lingvisticheskaya praktika. Uchebno-metodicheskij kompleks slozhnoj struktury. Chast 1. Teoreticheskie osnovy i metodika sbora lingvisticheskikh dannykh dl a predstavlenia ikh v linguisticheskom korpuse russkogo yazyka. St. Petersburg (2007) 2. Asinovsky, A.S., Bogdanova, N.V., Rusakova, M.V., Stepanova, S.B., Sherstinova, T.Y.: Zvukovoj korpus russkogo yazyka povsednevnogo obschenia Odin rechevoj den : koncepcia i sosytoyanie formirovania. In: Kompjuternaya lingvistika i intellektualnye tekhnologii. Vypusk, Moscow. Po materialam mezhd. konferencii Dialog, vol. 7 (14), pp (2008) 3. Asinovsky, A.S., Koroleva, I.V., Rusakova, M.V., Ryko, A.I., Philippova, N.S., Stepanova, S.B.: On Integral Multilevel Annotation of a Spoken Russian Corpus. In: Proc. the XIIth International Conference Speech and Computer SPECOM 2007, Moscow (2007) 4. Bogdanova, N.V.: Allegrovye formy russkoj rechi: ot proiznositel noj redukcii k pis mennoj fiksacii i leksikalizacii v yazyke. Mat-ly XXXVII mezhd. filologicheskoj konferencii. Vypusk 18. Fonetika. St. Petersburg (2008) 5. ELAN - Linguistic Annotator. Version 3.6,

8 The ORD Speech Corpus of Russian Everyday Communication One Speaker s Day Koroleva, I.V.: Individual nye sostoyania i svoistva yazykovoj lichnosti: vliyanie na lingvisticheskuju strukturu vyskazyvanij. Mat-ly XXXVII mezhd. filologicheskoj konferencii. Vypusk 21. St. Petersburg. pp (2008) 7. Markasova, E.V.: Ritoricheskaya enantiosemia v korpuse russkogo yazyka povsednevnogo obschenia Odin rechevoj den. In: Kompjuternaya lingvistika i intellektualnye tekhnologii. Vypusk Dialog, Moscow, vol. 7(14), pp (2008) 8. Praat: Doing Phonetics by computer, 9. Ryko, A.I., Stepanova, S.B.: Mnogourovnevaya lingvisticheskaya razmetka zvukovogo korpusa russkogo yazyka. In: Kompjuternaya lingvistika i intellektualnye tekhnologii. Vypusk. Po materialam mezhd. konferencii Dialog, Moscow, vol. 7 (14), pp (2008) 10. Ryko, A.I., Stepanova, S.B.: Problemy vychlenenia jedinic analiza spontannogo ustnogo teksta. In: Mat-ly XXXVII mezhd. filologicheskoj konferencii. Vypusk, St. Petersburg, vol. 21, pp (2008) 11. Sherstinova, T.Y.: Odin rechevoj den na vremennoj shkale: o perspektivakh issledovania dinamicheskikh processov na materiale zvukovogo korpusa. In: Vestnik Sankt-Peterburgskogo universiteta, Seria 9: Filologia, Vostokovedenie, Zhurnalistika, Chast 2, St. Petersburg, vol. 4, pp (2008) 12. The British National Corpus Zobnina, E.A.: Social nye characteristiki govoriaschego: objektivnye dannye i ekspertnaya ocenka rechi (po materialam zvukovogo korpusa Odin rechevoj den. In: Mat-ly XXXVII mezhd. filologicheskoj konferencii. Vypusk, St. Petersburg, vol. 21, pp (2008)