Degrees of Orality in Speech-like Corpora: Comparative Annotation of Chat and E-mail Corpora

Degrees of Orality in Speech-like Corpora: Comparative Annotation of Chat and E-mail Corpora Eckhard Bick Institute of Language and Communication, University of Southern Denmark Campusvej 55, DK 5230 Odense M eckhard.bick@mail.dk Abstract. This paper describes and evaluates the automatic grammatical annotation of a chat and an e-mail corpus of together 117 million words, using a modular Constraint Grammar system. We discuss a number of genre-specific issues, such as emoticons and personal pronouns, and offer a linguistic comparison of the two corpora with corresponding annotations of the Europarl corpus and the spoken and written subsections of the BNC corpus, with a focus on orality markers such as linguistic complexity and word class distribution. Keywords: chat corpus, e-mail corpus, orality, parsing, Constraint Grammar, NLP 1 Introduction Traditional speech corpora with phonetic transcription are very labour-intensive to create, involving a huge effort in data collection, sound file management and manual transcription. Automatic transcription is today an alternative, at least for languages with a mature language technology base such as English, but the method is not error free and commercial tools will produce standard orthography, not phonetic transcription. Thus, Luz et al. (2008) report transcription speeds of 22-30 words per minute, for an ASR-assisted post editing method, with a final error rate of 3.3% - 7.83%., translating into 20 man-years of work for the one-pass one- annotator transcription of a 25 million-word corpus. A third alternative - and the position taken in this paper - is to use data where people write in a speech-like fashion, without the constraints of ordinary written production, thus in fact providing their own transcriptions. This is the case for both chat- and sms-data, and to a certain degree e-mail text. The paper describes and evaluates the annotation of two such corpora, the Enron e-mail corpus and our own Fantasy chat corpus, comparing these to the written and oral sections of the BNC and the English section of the Europarl corpus which could be described as "listener-transcribed" rather than "speaker- transcribed" and also differs from our other data in representing fairly formal, parliamentary speech. 2 The corpora The Enron corpus is a corpus of corporate e-mails, called the Enron Email Dataset, and made available for research by William Cohen on his website (http://www.cs.cmu.edu/~wcohen/, http://www.cs.cmu.edu/~enron/). The data was originally made public, and posted to the web, by the (US) Federal Energy Regulatory Commission during its investigation, and later prepared by the CALO Project (http://www.ai.sri.com/project/CALO). Our chat corpus was compiled from 4 different fantasy chat logs from Project JJ (http://www.projectjj.com), administrated and made available by Tino Didriksen. The logs were collected between August 2002 and August 2004, and cover the topics (a) Harry Potter, (b) Goth Chat, (c) X Underground and (d) Amarantus: War in New York. The Europarl corpus used here is the English part (both original and translated) of the European Parliament Proceedings Parallel Corpus 1996-2003, prepared by Philipp Koehn. The corpus was retrieved from his website at http://www.isi.edu/ ~koehn/europarl/. PACLIC 24 Proceedings 721 The BNC (British National Corpus, …

Content truncated to 500 words