Best Practices for Speech Corpora in Linguistic Research Workshop ...

More documents

Recommendations

Info

codes marked through special Unicode characters entered through combinations of the F1 and F2 function keys with other characters. This system is described at http://talkbank.org/CABank/codes.html and in MacWhinney and Wagner (2010) 2. Morphological and syntactic lines. The MOR and GRASP programs compute these two annotation lines automatically. The forms on these lines stand in a one-to-one relation with main line forms, excluding retraces and nonwords. This alignment, which is maintained in the XML, permits a wide variety of detailed morphosyntactic analyses. We also hope to use this alignment to provide methods for writing from the XML to a formatted display of interlinear aligned morphological analysis. 3. Phonological line. The %pho line stands in a one-to-one relation with all words on the main line, including retraces and nonwords. This line uses standard IPA coding to represent the phonological forms of words on the main line. To represent elision processes, main line forms may be grouped for correspondence to the %pho line. The Phon program developed by Yvan Rose and colleagues (Rose, Hedlund, Byrne, Wareham, & MacWhinney, 2007; Rose & MacWhinney, in press) is able to directly import and export valid TalkBank XML. 4. Error analysis. In earlier versions of the system, errors were coded on a separate line. However, we have found that it is more effective to word-level code errors directly on the main line, using a system specifically elaborated for aphasic speech at http://talkbank.org/AphasiaBank/errors.doc. 5. Gesture coding. Although programs such as ELAN and Anvil provide powerful methods for gesture coding, we have found that it is often difficult to use these programs to obtain an intuitive understanding of gesture sequences. Simply linking a series of gesture codes to the main line in TalkBank XML is similarly inadequate. To address this need, we have developed a new method of coding through nested coding files linked to particular stretches of the main line. These coding files can be nested indefinitely, but we have found that two levels of embedding are enough for current analysis needs. Examples of these gesture coding methods can be found at http://talkbank.org/CABank/gesture.zip. 6. Special coding lines. CLAN and TalkBank XML also support a wide variety of additional coding lines for speech act coding, analysis of written texts, situational background, and commentary. These coding tiers are not aligned only to utterances and not to individual words. 8. Dissemination Platforms The fundamental idea underlying the construction of TalkBank is the notion of data sharing. By pooling their hard-won data together, researchers can generate increasingly accurate and powerful answers to fundamental research questions. The CHILDES and TalkBank web sites are designed to maximize the 60 dissemination of the data, programs, and related methods. Transcript data can be downloaded in .zip format. Media can be downloaded or played back over the web through QuickTime reference movie files. The TalkBank browser allows users to view any TalkBank transcript in the browser and listen to the corresponding audio or see the corresponding video in continuous playback mode, linked on the utterance level. We also provide methods for running CLAN analyses over the web, which we are now supplementing with analyses that use the XML database as served through the Mark Logic interface.To teach the use of the system, we have produced manuals, instructional videos and powerpoint demonstrations which we use in a wide variety of workshops internationally 9. Conclusion Together these various TalkBank facilities provide a comprehensive, interoperable set of best practices for the coding of spoken language corpora for research in linguistics, psycholinguistics, speech technology, and related disciplines. New methods and improvements to these practices are continually in development, as we expand the database to include a fuller representation of the many forms of spoken communication. 10. References Jefferson, G. (1984). Transcript notation. In J. Atkinson & J. Heritage (Eds.), Structures of social interaction: Studies in conversation analysis (pp. 134-162). Cambridge: Cambridge University Press. MacWhinney, B. (2008). Enriching CHILDES for morphosyntactic analysis. In H. Behrens (Ed.), Trends in corpus research: Finding structure in data (pp. 165-198). Amsterdam: John Benjamins. MacWhinney, B., Fromm, D., Forbes, M., & Holland, A. (2011). AphasiaBank: Methods for studying discourse. Aphasiology, 25, 1286-1307. MacWhinney, B., & Wagner, J. (2010). Transcribing, searching and data sharing: The CLAN software and the TalkBank data repository. Gesprächsforschung, 2, 1-20. Malvern, D. D., Richards, B. J., Chipere, N., & Purán, P. (2004). Lexical diversity and language development. New York: Palgrave Macmillan. Rose, Y., & MacWhinney, B. (in press). The Phon and PhonBank initiatives. Sagae, K., Davis, E., Lavie, A., MacWhinney, B., & Wintner, S. (2010). Morphosyntactic annotation of CHILDES transcripts. Journal of Child Language, 37, 705-729. Sagae, K., Lavie, A., & MacWhinney, B. (2005). Automatic measurement of syntactic development in child language Proceedings of the 43rd Meeting of the Association for Computational Linguistics (pp. 197-204). Ann Arbor: ACL.
Toward the Harmonization of Metadata Practice for Spoken Languages Resources Christopher Cieri ❄ , Malcah Yaeger-Dror ❄● ❄ Linguistic Data Consortium, University of Pennsylvania, ● University of Arizona ❄ 3600 Market Street, Suite 810, Philadelphia, PA 19104, USA E-mail: ccieri@ldc.upenn.edu, malcah@email.arizona.edu Abstract This paper addresses issues related to the elicitation and encoding of demographic, situational and attitudinal metadata for sociolinguistic research with an eye toward standardization to facilitate data sharing. The discussion results from a series of workshops that have recently taken place at the NWAV and LSA conferences. These discussions have focused principally on the granularity of the metadata and the subset of categories that could be considered required for sociolinguistic fieldwork generally. Although a great deal of research on quantitative sociolinguists has taken place in the Unites Stated, the workshops participants actually represent research conducted in North and South America, Europe, Asian, the Middle East, Africa and Oceania. Although the paper does not attempt to consider the metadata necessary to characterize every possible speaker population, we present evidence that the methodological issues and findings apply generally to speech collections concerned with the demographics and attitudes or the speaker pools and the situations under which speech is elicited. Keywords: metadata, sociolinguistics, standards 1. Introduction The brief history of building digital, shareable language resources (LRs) to support language related education research and technology development is marked by numerous attempts to create and enforce standards. The motivations behind the standards are numerous. For example, standards offer the possibility of making explicit the process by which LRs are created, establishing minimum quality levels and facilitating sharing. Nevertheless, there have been instances in which the pre-mature or inappropriate promulgation or adoption of standards has lead to its own set of problems (Osborn 2010, p. 74ff, Mah, et. al. 1997) as researchers struggle to apply to their use cases standard that were not truly representative and perhaps not intended to be. To reduce the potential effort expended in developing, promoting and using proposed standards that may subsequently be found difficult to sustain, we propose that standardization is a late step in a multipart process that begins with understanding, progresses to documentation that may itself encourage consistency in practice within small groups at which point the question of standardization begins to ripen. 2. Background The present workshop seeks to survey current initiatives in speech corpus creation with an eye toward standardization across sub-disciplines. Such standardization could permit resource sharing among researchers working in conversation and discourse analysis, sociolinguistics and dialectology among others and between those fields and others who depend upon similar kinds of data including language engineers (Popescu-Belis, Zufferey 2007). Coincidentally, the authors have been involved in a number of workshops on related themes including a series taking place at the annual NWAV (New Ways of Analyzing Variation) meetings on speech data collection, annotation and distribution including documentation and metadata 61 description. More recently they lead a workshop funded by the U.S. National Science Foundation at the 2012 winter meeting of the Linguistics Society of America 1 . The principal topics of the latter were metadata description and related legal issues in the creation of spoken language corpora for sociolinguistics. This paper constitutes a summary of efforts within that community to begin understanding metadata encoding practice as a first step toward consistency, sharing and standardization. 3. Towards Standardization Before metadata practice can be standardized, individual researchers must first understand their practices, the variations among them, the causes for variation, the tradeoffs of different approaches and their potential uses. In particular, researchers need to know if they can apply their metadata categories consistently, a question that is not frequently asked but must be if the goal is to adopt a standard that will be used by many independent groups with the intent of sharing corpora. Once the practice is understood it must be documented so that potential users can evaluate it and competing practices can be harmonized to permit appropriate comparisons. With adequate documentation independent researchers can decide if they want to adopt consistent practices. 4. Metadata Within sociolinguistics, some researchers’ position is that each study requires its own set of demographics. However, the ultimate consensus at the workshops was that cross community comparative corpus-based studies are only possible if there is a shared set of specific coding choices. Some of the demographic information is generally accepted within the larger sociolinguistic community: sex, birth year, years of education, and some designation of job description are fairly common 1 http://projects.ldc.upenn.edu/NSF_Coding_Workshop_L SA/index.html
Page 1 and 2:
Best Practices for Speech Corpora i
Page 3 and 4:
Editors Michael Haugh Griffith Univ
Page 5 and 6:
Author Index Broeder, Daan ........
Page 7 and 8:
A linguistics-based speech corpus J
Page 9 and 10:
Figure 2: Grammatical tags are visi
Page 11:
In Jokinen, Kristiina and Eckhard B
Page 14 and 15:
2.2 Parameters of the Corpus Design
Page 16 and 17: switching and code mixing, we have
Page 18 and 19: � 12
Page 20 and 21: French and Russian screen versions
Page 22 and 23: singularity, expressiveness, semant
Page 24 and 25: is often fluid in terms of communic
Page 26 and 27: annotation. Researchers have often
Page 28 and 29: � 22
Page 30 and 31: in various research and/or student
Page 32 and 33: As previously mentioned, the degree
Page 34 and 35: 9. References Corpus of Spoken Gree
Page 36 and 37: In addition to part-of-speech tags
Page 38 and 39: POS description example translitera
Page 40 and 41: Figure 4: Different syntactic analy
Page 42 and 43: Herbert H. Clark and Thomas Wasow.
Page 44 and 45: The ‘externality’ of DA arises
Page 46 and 47: 6. Data analysis Each turn is annot
Page 48 and 49: � 42
Page 50 and 51: 3. manual phonetic transcription (t
Page 52 and 53: • comparative linguistic research
Page 54 and 55: ��
Page 56 and 57: ��
Page 58 and 59: The global corpus data model is a s
Page 60 and 61: a) b) c) Figure 6: Web experiment o
Page 62 and 63: � 56
Page 64 and 65: conversation across 26 languages; t
Page 68 and 69: demographic fields, as are designat
Page 70 and 71: methods for eliciting metadata. Che
Page 72 and 73: � 66
Page 74 and 75: difference between corpora that can
Page 76 and 77: database can be reconstructed at an
show all

Best Practices for Speech Corpora in Linguistic Research Workshop ...

Create successful ePaper yourself

Delete template?

Save as template?