Towards Discourse Classification for Chinese -- a Resource-poor Language.

摘要

Discourse raises issues about semantics, and especially the nature of coherence and cohesion of texts. Similar to part-of-speech tagging and syntactic parsing, discourse classification is fundamental in computational linguistics. But relatively, this issue is not well studied. The lack of annotated corpora brings limitations to research of discourse classification for most languages other than English (e.g., Chinese). Manual annotation for discourse classification is complex, time consuming and costly. To overcome this predicament, one alternative is to explore unsupervised learning methods. Nevertheless, previous work on English showed that unsupervised methods could only deal with coarse-grained discourse relations and suffered from low precision. Another possible way is to make use of discourse classification capabilities from other languages which have rich discourse corpora. But the problem of cross language discourse classification is still very much open for investigation. Using Chinese as the target, this thesis presents the first study on discourse classification for resource-poor language. Furthermore, we also annotate the first open discourse treebank for Chinese which includes 890 news articles.;At the beginning, we propose a novel bootstrapping unsupervised method based on semantic sequential representation (SSR) for discourse classification. SSR is a new representation for discourse instances which integrate basic bag-of-words information with lexical, semantic and word sequential information. Our method starts with a small set of cue-phrase-based patterns to collect large number of discourse instances which are later converted to SSRs. We then propose an unsupervised SSR learner to generate, weigh and filter new SSRs without cue phrases for recognizing discourse relations. Experimental results showed that our method outperformed previous unsupervised method by 7% in F-score. We also show that SSRs are effective features for supervised learning methods.;The SSR-based method (F-score = 0:63) ignores the ambiguities of discourse connectives. As a result, it suffers from low recall (Recall = 0:49). To discover and eliminate these ambiguities, we further propose a cross-language framework for discourse classification. In our framework, discourse classification for Chinese is achieved in two steps: (1) Discourse connective/trigger identification and (2) Sense classification. English Penn Discourse Treebank 2 (PDTB2) and Chinese-English parallel data are coupled to provide the training data for a co-training based framework. Experimental results showed that our method achieved significant improvement comparing to SSR based method. The proposed framework is practical and effective especially in coping with the intercommunity problem, which is common in cross-language discourse classification. Moreover, the proposed framework does not integrate any language specific features, making it theoretically applicable for other languages.;Every language has its unique characteristics, our cross-language framework which focuses on the common characteristics between languages is ineffective in detecting Chinese language specific characteristics. As a result, we package the corpus we used in this research to form the Discourse Treebank for Chinese (DTBC). DTBC adopts the principles of PDTB2, and at the same time, it incorporates the linguistic characteristics of Chinese. The ii annotation work adds a discourse layer to 890 articles from the Penn Chinese Tree Bank 5 (CTB5). DTBC is the first ever open Chinese discourse treebank, which will be an invaluable linguistic resource for future research in Chinese discourse.