Dependency Parsing for Weibo: An Efficient Probabilistic Logic Programming Approach

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1152–1158, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Dependency Parsing for Weibo: An Efficient Probabilistic Logic Programming Approach William Yang Wang, Lingpeng Kong, Kathryn Mazaitis, William W. Cohen Language Technologies Institute & Machine Learning Department Carnegie Mellon University Pittsburgh, PA 15213, USA. {yww,lingpenk,krivard,wcohen}@cs.cmu.edu Abstract Dependency parsing is a core task in NLP, and it is widely used by many applica- tions such as information extraction, ques- tion answering, and machine translation. In the era of social media, a big chal- lenge is that parsers trained on traditional newswire corpora typically suffer from the domain mismatch issue, and thus perform poorly on social media data. We present a new GFL/FUDG-annotated Chinese tree- bank with more than 18K tokens from Sina Weibo (the Chinese equivalent of Twit- ter). We formulate the dependency pars- ing problem as many small and paralleliz- able arc prediction tasks: for each task, we use a programmable probabilistic first- order logic to infer the dependency arc of a token in the sentence. In experiments, we show that the proposed model outperforms an off-the-shelf Stanford Chinese parser, as well as a strong MaltParser baseline that is trained on the same in-domain data. 1 Introduction Weibo, in particular Sina Weibo1, has attracted more than 30% of Internet users (Yang et al., 2012), making it one of the most popular social media services in the world. While Weibo posts are abundantly available, NLP techniques for ana- lyzing Weibo posts have not been well-studied in the past. Syntactic analysis of Weibo is made difficult for three reasons: first, in the last few decades, Computational Linguistics researchers have pri- marily focused on building resources and tools us- ing standard English newswire corpora2, and thus, 1http://en.wikipedia.org/wiki/Sina Weibo 2For example, Wall Street Journal articles are used for building the Penn Treebank (Marcus et al., 1993). there are fewer resources in other languages in general. Second, microblog posts are typically short, noisy (Gimpel et al., 2011), and can be considered as a “dialect”, which is very differ- ent from news data. Due to the differences in genre, part-of-speech taggers and parsers trained on newswire corpora typically fail on social media texts. Third, most existing parsers use language- independent standard features (McDonald et al., 2005), and these features may not be optimal for Chinese (Martins, 2012). To most of the applica- tion developers, the parser is more like a blackbox, which is not directly programmable. Therefore, it is non-trivial to adapt these generic parsers to language-specific social media text. In this paper, we present a new probabilistic de- pendency parsing approach for Weibo, with the following contributions: • We present a freely available Chinese Weibo dependency treebank3, manually annotated with more than 18,000 tokens; • We introduce a novel probabilistic logic programming approach for dependency arc prediction, making the parser directly pro- grammable for theory engineering; • We show that the proposed approach outper- forms an off-the-shelf dependency parser, as …

Content truncated to 500 words