compiling a spoken chinese corpus of situated discourse by luckboy


More Info
									Sampling situated discourse for spoken Chinese corpus
GU, Yueguo The Institute of Linguistics The Chinese Academy of Social Sciences Abstract Corpus sampling and representativeness constitutes the first and very important issue in corpus compilation. This paper first makes a brief assessment of three corpora of spoken English with regard to corpus sampling and representativeness. It then describes the way the problem was dealt with in compiling a spoken Chinese corpus of situated discourse. Key words: corpus sampling, representativeness, 1. Preamble Corpus sampling and representativeness constitutes the first and very important issue in corpus compilation. It is less difficult to be dealt with in written than in spoken language. Crowdy observes (1995:224): “With a corpus of spoken language there are no obvious objective measures that can be used to define the target population or construct a sampling frame.” This paper reports the way the problem was tackled in compiling a spoken Chinese corpus of situated discourse. On the face of it, there seems to be little connection between compiling a spoken Chinese corpus and globalization, the theme of this conference. It will become clear, as we proceed, that it is not. Nowadays the phenomenon of globalization is not confined to world economy or business. It has taken place in the field of language and linguistics as well. Natural language, in a way, is an innate endowment of human species. Surely languages differ from one another, but the belief that underneath the surface differences lies the fundamental sameness, or universals, is held more firmly than ever. Theories of language, or schools of linguistics, although constructed on the bases of a limited number of languages, tend to make a bold and ambitious claim of being applicable, with some revisions of course, to all the languages in the world. Studies of any individual language, therefore, contribute to the understanding of language as a global and human phenomenon. It is no exaggeration to say that linguistics as a scientific study of language, requires global efforts. For the last decade or so this has become particularly noticeable in corpus linguistics, one of the latest branches of linguistics that has become very quickly part of the mainstream. An international computerized corpus of English (ICE) was launched in 1988. According to Greenbaum (1991:85), at the

time of his writing, thirteen countries or regions (Hong Kong included) took part in the project, covering both English as the first and the second language. Also under compilation is a corpus of learners’ English with 14 countries including China as co-compilers (Granger, 1998:16). As English is nowadays regarded as a global language, Chinese, in terms of the number of speakers, is its close match, and an extensive investigation into it will be mostly welcome as part of the global efforts towards the understanding of language as species’ distinctive feature. In what follows we first make a brief assessment of three corpora of spoken English with regard to corpus sampling and representativeness. Then we share with readers the way we tackled it in compiling a spoken Chinese corpus of situated discourse under the auspices of the Chinese Academy of Social Sciences. Two appendixes are attached for the comparison of the present project with English corpora. 2. A brief appraisal of sampling methods and representativeness There are three major spoken English corpora currently available to public: the London-Lund Corpus (LLC), the Corpus of Spoken American English (CSAE), and the spoken part of the British National Corpus (BNC). Let us look at them one by one. LLC is indisputably the first influential spoken corpus. Historically it is derived from the spoken part of the Survey of English Usage Corpus (the SEU Corpus) . “The goal of the Survey of English Usage is to provide the resources for accurate descriptions of the grammar of adult educated speakers of English. For that purpose the major activity of the Survey has been the assembly and analysis of a corpus comprising samples of different types of spoken and written British English.” (Greenbaum and Svartvik 1990). It started in 1959, and was never intended to be machine-readable. It took some 25 years to complete the collection of the original 87 texts of transcribed speech totaling some 435,000 words. It was however computerized and supplemented with 13 more texts and becomes what is now known as the London-Lund corpus. Since the purpose was on the grammar of adult educated speakers of English, the sampling was confined to a particular type of population. It is justifiable on its own right. However, its limitations should not be forgotten. First, the language sampled is a sort of elite class. To what extent it can represent the state of language becomes questionable. This relates to the second limitation, namely adult educated speakers do not live in a specialized reserve land. They rub shoulders with non-educated speakers. So it is difficult to say that their way of speaking is not influenced by the latter. The London-Lund Corpus can be said to capture a variety of educated English. A

similar approach is adopted by the Corpus of Spoken American English. Chafe, Du Bois and Thompson (1991:69), compilers of the corpus, introduce their design thus: Spoken SAE [Standard American English] encompasses formal and informal styles, preaching, cursing, bureaucratese, slang, Southern accents, New York accents, and a host of regional, class, gender and ethnic accents. What unifies it is a shared set of grammatical rules and structures, which among spoken varieties of American English come closer than any other to the written variety as used in journalism, government publications, textbooks, and so on, while of course still differing from the written variety in the many ways that speaking may differ from writing. Because SAE encompasses a variety of accents, registers and styles, it is not the sole property of any particular group in American society. For every ethnic group of any size in the United States, there are members who speak SAE and members who do not. 30 half-hour conversations were planned. No attempt would be made at a rigorous sampling procedure. Instead, what was aimed at was a reasonable balance in the choice of speakers to achieve a wide representation of adult speakers of Standard American English (ibid, p. 71). This standard variety approach, with some advantages of its own way (e.g. it is convenient for contrastive study and educational purpose), is subject to criticisms relating to the reservation of the naturalness of language use. The standard variety, e.g. SAE, is given its identity before the corpus is compiled. The corpus cannot be used to represent its naturalness, nor be used to establish or demonstrate its identity. One cannot point to the corpus and say: Look, this is what SAE looks like. One can only say: This is what the compilers believe what SAE looks like. Another difficulty lies in sampling SAE speakers by filtering non-standard speakers out. When they are being audio-taped, the assumed SAE speakers may be interacting with non-standard speakers. Unless they are ‘commissioned’ to talk among themselves, the activities the standard and non-standard interactants are engaged in have to be properly filtered out as well. Demographic sampling method is adopted in compiling the spoken part of the British National Corpus (BNC). The sampling frame has been defined in terms of the language production of the population of British English speakers in the United Kingdom. Representativeness is achieved by sampling a spread of language producers in terms of age, gender, social group, and region, and recording their language output over a set period of time. (Crowdy, 1995:225) Sampling procedure is as follows. 124 adults (aged 15+) were recruited from

across the United Kingdom. Recruits were of both sexes and from all age groups and social classes. They were asked to record all of their conversations over a two to seven day period. The number of days varied depending on how many conversations each recruit was involved in and was prepared to record. They were also asked to fill in a conversation log with details of every conversation recorded, such as date, time and setting, and brief details of other participants. As Crowdy points out, a corpus constituted solely on the demographic model omits important spoken text types, for “many types of spoken text are produced only rarely by comparison with the total output of all ‘speech producers’: for example, broadcast interviews, lectures, legal proceedings, and other texts produced in situations where –broadly speaking—there are few producers and many receivers” (ibid). Consequently, the demographic component of the corpus is complemented with a context-governed component consisting of a range of text types divided into a priori linguistically motivated categories. At the top layer of the typology is a division into four equal-sized contextually based categories: educational, business, public/institutional, and leisure. Each is divided into the subcategories monologue (40 percent) and dialogue (60 percent). Within each subcategory a range of text types was defined. The sampling methodology was different for each text type but the overall aim was to achieve a balanced selection within each, talking into account such features as region, level, gender of speakers, and topic. Other features, such as purpose, were applied on the basis of post hoc judgements. (Burnard, 2000:4.2) The BNC effectively shows that a single method of sampling is undesirable for a spoken corpus. It calls for the combination of a variety of sampling techniques appropriate for particular types of spoken texts in question. 3. the SCCSD BJ-500 Project 3.1 the project A team was set up in 1999 to compile a five hundred hour spoken Chinese corpus of situated discourse in Beijing area (SCCSD BJ-500). The main purpose of the project is to capture the real-life dynamics of actual discourse. Beijing area is not treated as a regional parameter. Rather, it is a matter of convenience. Some recordings from other areas are also included. It is hoped that enough experience will be accumulated at the end of this project for the launching of a nation-wide programme of compiling a similar corpus bearing the name spoken Chinese. 3.2 research strategies and sampling methods Benefiting from the experiences of the spoken English corpora reviewed above, we have adopted two macro research strategies. First, we maintain the distinction

between archive and corpus. An archive of situated discourse will include any recordings that fit the definition of situated discourse. A corpus of situated discourse will be a collection of spoken texts that are principally extracted from the archive. Second, two people, an academic and a community office clerk, were commissioned to audio-tape (using a digital MD recorder) all, if possible, their weekly activities. (The academic actually recorded 6 days, and the clerk overdid the task by recording two weeks instead of one.) This generates some access to non-idealized real-life data, and improves our understanding of what workplace discourse looks like in actual situations (see Gu 2001 for details). As to sampling methodology, we gave more weight to sociological considerations. Sinclair, a veteran corpus linguist, alludes to it when he writes (1991:13): The specification of a corpus --- the types and proportions of material in it --- is hardly a job for linguists at all, but more appropriate to the sociology of culture. The stance of the linguist should be a readiness to describe and analyse any instances of language placed before him or her. In the infancy of the discipline of corpus linguistics, the linguists have to do the text selection as well; when the impact of the work is felt more widely, it may be possible to hand over this responsibility to language-oriented social scientists. Sinclair’s job description of a linguist is surely quite attractive. However, he seems to overlook the fact what counts as linguistic data depends in part on how language is defined, and on from what perspective it is being looked at. These two issues cannot be left for language-oriented social scientists alone, whatever they might be, to solve for linguists. So the job is probably more profitably done jointly by linguists and language-oriented scientists. But anyway for the time being, with no language-oriented social scientists available, corpus-compiling linguists have to make serious and difficult decisions by themselves. In its formation discourse is both social and historical. This is particularly true of situated discourse, being situated in the sense that (1) It is situated to an actual social situation; (2) It is situated to actual users; (3) It is situated to an inter-subjective world of discourse; (4) It is situated to actual goals; (5) It is situated to spatial and temporal setting; (6) It is situated to the cognitive capacity of actual users; (7) It is situated to performance contingencies of actual users who are engaged in spontaneous talking with little pre-planning. So any sampling of such discourse is like chopping a slice and freezing it for later inspection. But we may do it like a careless butcher or like a professional anatomist.

The latter requires that we understand how the slice fits and functions in the overall architecture, and reserve its original texture to the best we can. In this connection Fairclough (1995), from a critical discourse analysis perspective, makes this inspiring observation: A principled basis for sampling requires minimally (a) a sociological account of the institution under study, its relationship between forces within it; (b) an account of the ‘order of discourse’ of the institution, of its IDFs [ideological-discursive formations] and the dominance relationships among them, with links between (a) and (b); (c) an ethnographic account of each IDF. Given this information, one could identify for collection and analysis interactions which are representative of the range of IDFs and speech events, interactional ‘cruxes’ which are particularly significant in terms of tensions between IDFs or between subjects, and so forth. In this way a systematic understanding of the functioning of discourse in institutions and institutional change could become a feasible target. (Fairclough, 1995:49) We may conclude at this point that the sampling methods for a corpus of situated discourse are the decisions to be made on the bases of sociological factors, discourse parameters and practical feasibilities. Sociological factors In the current state of affairs, life activities in Beijing are organized around two nexus: workplace and family. Here workplace includes both public workplace and self-employed workplace. According to the Beijing Yellow Book 1999, there are 67,783 social work units. Drawing on the original classifications we divide them into 6 major categories and 31 sub-categories, as shown in Table 2: Table 2 Categorization of Social Work Units
01 government, Parties and social organizations 4823 7.12% 0101 government depts 0102 depts of Parties 0103 social bodies 0104 foreign embassies 02 business and economy 53838 79.43% 0201 finance and insurance 0202 telecom 0203 energy 0204 mining and metallurgy 0205 manufacture 0206 construction 0207 transportation 0208 irrigation 0208 estate 0210 retail and wholesale 0211 husbandry and 3760 95 899 136 1340 840 1141 668 17659 4988 1173 2 919 15890 2748 5.55% 0.14% 1.33% 0.20% 1.98% 1.24% 1.68% 0.99% 26.05% 7.36% 1.73% 1.36% 23.44% 4.05%

03 education, academics, and arts



forestry 0212 technical service 0213 provincial representatives 0301 education 0302 research institutions 0303 arts 0304 mass media 0401 health 0402 sports 0403 social welfare 0501 daily life service 0502 public service 0760 environment protection 0770 horticulture 0780 funeral service 0601 military institutions

3656 845 3350 1457 507 1526 1365 ? ? 478 412 ? ? 41 27

5.39% 1.25% 4.94% 2.15% 0.75% 2.25% 2.01%

04 Health, sports and social welfare 05 public service





0.71% 0.61%

06 military




67783 Note: the statistics is provided by Zhang Hong, who wrote an automatic sorting program for the purpose. The figures should be taken as approximate numbers, not as accurate records. The inaccuracy is due to three sources. First the Yellow Book itself is inaccurate: over hundred errors were detected by us. Second, not all work units go to the Yellow Book. Third, the sorting engine is not powerful enough to differentiate some subtle categories, hence the question marks. In spite of these the figures serve as broadly valuable index.

Like workplace, family is also socially stratified. In terms of wealth some are now extremely rich, and some under poverty line, and the majority in between. In terms of social class, there are families of blue and white collar workers, academics, managers, CEOs, high-ranking officials, entrepreneurs, suburb farmers, etc. Surely situated discourse produced in workplace and family is open-ended. However, workplace and family can be in theory exhaustively listed. Thus the two were adopted as our sampling frame. Moreover, workplace and family are hierarchically structured and stratified. We thus have a stratified sampling within the sampling frame. Although it is, strictly speaking, not random sampling, it is “often more efficient to obtain a representative sample for predetermined social categories” (Wolfram and Fasold, 1997 [1974]:91). In this procedure, the social composition of the sample is first determined, then informants are chosen to represent these categories, which are sometimes referred to as cells of the sample. Informants can be chosen randomly until an adequate number is obtained to represent each cell. This procedure avoids the problem of over- and under- representation for particular social categories because the investigator stops selecting informants for given cells when a quota is reached.” (ibid)

Discourse parameters At the macro level, social organizations are structured primarily for particular social functions, which are in turn being fulfilled via discourse. However, this does not mean that social organizations, social functions and discourse types hold one-to-one relation to one another. First, social organizations are multi-functional, with some being prototypical to the organizations in questions, some less prototypical, and others overlapping with other organizations. Second, day-to-day activities in implementing the social functions vary and can be drastically different from one another. In our study of a week long discourse of an academic, we have found that it is very important to take the following 6 parameters into account (for details see Gu 2001): 1. The activities going on in work units occur in clusters of interactive discourse. It is the clusters that we actually capture by audio- or video- taping. These clusters constitute our everyday social and linguistic reality. 2. Talking is not only embedded in these clusters of interaction, but also part and parcels of the social activities at large. As our primary interest is in talking, it is necessary for us to draw a distinction between talking and non-talking (or simply called doing) activities. In real life situations, talking and doing can be interwoven in the following ways: a) Talking is the task with little doing, e.g., meeting, seminar, (it is an all-talking task. Note that turn-taking rules are based on such a type of talking-task relation); b) Talking is the main constitutive part of the task, some classroom discourse, doctor patient discourse (it is talking-dominated); c) Talking is a constitutive part of the task, e.g. giving instructions from time to time (doing is dominant, while talking tends to be fragmented); d) Talking and doing run in conflicting parallel, the achievement of the latter serves as a means to the goal of the former, e.g. business dinner (i.e. business table talk); e) Talking is an embedded social part of the task, e.g. talking over the meal (talking has no specific goal to reach); f) Talking is a decorative part of the task, e.g. talking accompanying tea-making; g) Talking is a hindrance to the task, e.g. talking over a written exam; h) Talking and doing are independent to each other. 3. The talking and doing activities among the interactants, on the other hand, are interwoven on another dimension: they can be conflictive, parallel, independent, relevant, etc. 4. The representativeness of situated discourse depends on the interwoven relations

between the prototypical doing-activities of the unit and the prototypical talking-activities of the unit. Given a work unit, two levels of analysis are required: macro sociological analysis, that is, the unit being treated as a whole, and micro performance analysis, that is, the actual doing and talking by individual members of the unit. 5. Activities that are actually being carried out and are potentially recordable fall into a continuum: officially prototypical, officially non-prototypical, officially tolerable, and officially unacceptable. 6. In workplace or even within a family, accents vary. For example, Father talks with a strong Jiangsu accent, Mother with a North China accent, and children with local accents. Practical feasibilities It is a formidable task to collect authentic and spontaneous talks in actual social situations. It is even more so to record those that are prototypical of the category in question. For one thing we have to watch for the opportunities. For another, when the opportunities arise, permissions have to be obtained for the recording. For still another, the person who does the actual recording must be a “within”, or be at least made welcome. These set the limits to practical feasibilities. 3.3 Preliminary categories The deliberation of sociological factors, discourse parameters and practical feasibilities leads to the final choice of a stratified sampling with a predetermined typology of spoken text types, as shown in Figure 1. Note that the predetermined typology plays two roles in the project. (1) It acts as a blueprint for data collection; (2) it represents an ideal the project is aimed at. The objective may be eventually reached or fallen short of, or even overtaken. Of course it can also be amended in the process of the project being implemented. 3.4 Sampling procedure The sampling procedure goes as follows. The project team members recruit “insiders” for each sub-category. The recruits are then given a brief training on how to operate the digital MD recorder, and how to fill in the recording information log. The official documents about the project such as authorization certificate, commitment to privacy, commitment to academic research only, etc. are made available on demand to the recruits. They are asked to record, if possible, 10 hour spontaneous talks, ideally talks which they think are typical in their work units. It is up to them to choose to do the recording surreptitiously or openly. The only requirement from the project is that permission should be obtained from the person in charge of the work unit. The recruits can recruit their own “insiders” if they wish, which actually did happen in some cases. So the number of recruits is

not fixed, and keeps increasing as the project ploughs ahead. Figure 1 predetermined typology of spoken Chinese corpus of situated discourse
discourse of government, Parties and social organizations business discourse educational and academic discourse legal and mediatory discourse Major activities of organization mass media discourse discourse of medicine and health discourse of sports political discourse public service discourse public welfare discourse religious and superstitious discourse Societal discourse activities common to organization administrative discourse banquet discourse discourse of celebration and ceremony discourse of entertainment and leisure office discourse political study discourse telephone discourse pathological discourse special discourse
criminal discourse

Spoken Chinese Corpus of Situated Discourse

military discourse Miscellaneous family of high-ranking officials family of entrepreneurs family of businessmen Family of academics Family of white collar Family of blue collar Family of suburb farmers Family of immigrant labour

familial discourse

family discourse in a metropolis

4. Retrospection As the first phase of the project is completed, in retrospection the ensuing points are worth noticing. 1. the sampling method and sampling procedure adopted are sound and effective. However, the representativeness of the recorded data is less satisfactory. It is largely due to the unavailability of the targeted data, either for short of access

or permission or for short of opportunity. 2. it is difficult to find recruits for some sub-categories such as military, police, non-CPC Parties, criminals, and family of certain class. 3. As we make it paramount to obtain and keep all the important information of the actual situation, we realize that it is too much of a task for audio-taping. So video-taping (by using a digital handycam) was added whenever it was possible. 4. Since it is our research objective to capture real-life discourse, no attempt was made to filter out any accents, Putonghua or otherwise. Finally for those readers who may be interested in the preliminary results of the project, Appendix 1 is attached which shows the recordings already collected for the archive mentioned in 3.1 above. Appendix 2 shows the categories of the spoken part of SEU corpus and the BNC spoken corpus for the interest of comparison. References Aijmer, K, Altenberg, B. eds. 1991. English Corpus Linguistics: Studies in Honour of Jan Svartvik. Longman, London Biber, D. 1993. Representativeness in corpus design. Literary and Linguistic Computing 8(4): 243-57 Burnard, Lou, 2000. Reference Guide for the British National Corpus (World Edition) Edwards, J. A., Lainpert, M. D. eds. Talking Data: Transcription and Coding in Discourse Research. Lawrence Erlbaum, Hillsdale, NJ pp 33-43 Chafe, W. L., 1994. Discourse, Consciousness, and Time: The Flow and Displacement of Conscious Experience in Speaking and Writing. University of Chicago Press, Chicago Chafe, Wallace L., John W. Dubois and Sandra A. Thompson (1991), “Towards a new corpus of spoken American English”. In Aijmer, K, Altenberg, B., eds. pp. 64-82 Crowdy, S., 1993. Spoken corpus design and transcription. Literary and Linguistic Computing 8(4): 259-65 Crowdy, S., 1995. The BNC spoken corpus. In Leech, Myers, and Thomas, 1995:224-234 Fairclough, N. 1995. Critical Discourse Analysis. Longman

Granger, Sylviane, 1998. Learner English on Computer. Longman Greenbaum, Sidney and Svartvik, Jan, 1990. London-Lund corpus of spoken English. In Jan Svartvik, ed. Gu, Yueguo, 1999. Towards a model of situated discourse analysis. In Ken Turner, ed. The Semantics/Pragmatics Interface from Different Points of View. Elsevier Science B.V. Gu, Yueguo, 2001. Towards an understanding of workplace discourse. In Christopher Candlin, ed. Theory and Practice in Professional Discourse. City University of HK Press


Leech, G. N. 1992. Corpora and theories of linguistic performance. In Jan Svartvik ed. Directions in Corpus Linguistics. Proceedings of Nobel Symposium 82, Stockholm, 4-8 August 1991. Mouton de Gruyter, Berlin/New York pp 105-22 Leech, G. N., Greg Myers, and Jenny Thomas, eds. 1995. Spoken English on Computer. N.Y.: Longman Sinclair, J. 1991. Corpus, Concordance, Collocation. Oxford: Oxford University Press Svartvik, Jan, ed. 1990. The London-Lund Corpus of Spoken English: Description and Research. Lund University Press, Lund Wolfram, Walt and Ralph W. Fasold, 1974. The Study of Social Dialects in American English. Englewood Cliffs, NJ: Prentice-Hall


Appendix 1 The First Phase Work of SCCSD
discourse of government, Parties and social organizations business discourse educational and academic discourse legal and mediatory discourse Major activities of organization mass media discourse discourse of medicine and health discourse of sports political discourse public service discourse public welfare discourse religious and superstitious discourse Societal discourse activities common to organization administrative discourse banquet discourse discourse of celebration and ceremony discourse of entertainment and leisure office discourse political study discourse telephone discourse pathological discourse special discourse
* criminal discourse

Spoken Chinese Corpus of Situated Discourse

* military discourse miscellaneous * family of high-ranking officials * family of entrepreneurs * family of businessmen family of academics family of white collar * family of blue collar family of suburb farmers * family of immigrant labour

family discourse in a metropolis familial discourse

family discourse in a family of academics small town family of white collar Categories marked with * are still empty. Categories marked with & are not in the original typology.


Appendix 2 The categories of the spoken part of SEU corpus: surreptitious Face-to-face conversation dialogue Spoken part of SEU corpus monologue prepared to be written The BNC spoken corpus 1. Demographic component --- everyday conversation recorded by 124 (15+) recruits, plus COLT contribution (around 700 hours) Age group 0-14 15-24 25-34 35-44 45-59 60+ texts 26 36 29 22 20 20 w-units 265382 660847 848162 839622 957382 634663 % 6.30 15.71 20.16 19.96 22.76 15.08 s-units 41035 97994 121750 126688 136540 86556 % 6.72 16.04 19.94 20.74 22.36 14.17 Public discussion telephone Non-surreptitiou

spontaneous to be spoken

2. context-governed component --- 4 major categories (40% monologue, 60% dialogue) Educational and informative Lectures, talks, educational demonstrations (a university and a school) News commentaries Classroom interaction (home tutorials included) Business Company talks and interviews Trade union talks Sales demonstrations Business meetings

Consultations (medical, legal, business and professional) Public/ or institutional Political speeches Sermons Public/government talks Council meetings Religious meetings Parliamentary proceedings Legal proceedings Leisure Speeches Sports commentaries Talks to clubs Broadcast chat shows and phone-ins Club meetings


To top