Bilingually-Guided Monolingual Dependency Grammar Induction

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 1063–1072, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics Bilingually-Guided Monolingual Dependency Grammar Induction Kai Liu†§, Yajuan L¨u†, Wenbin Jiang†, Qun Liu‡† †Key Laboratory of Intelligent Information Processing Institute of Computing Technology, Chinese Academy of Sciences P.O. Box 2704, Beijing 100190, China {liukai,lvyajuan,jiangwenbin,liuqun}@ict.ac.cn ‡Centre for Next Generation Localisation Faculty of Engineering and Computing, Dublin City University qliu@computing.dcu.ie §University of Chinese Academy of Sciences Abstract This paper describes a novel strategy for automatic induction of a monolingual de- pendency grammar under the guidance of bilingually-projected dependency. By moderately leveraging the dependency in- formation projected from the parsed coun- terpart language, and simultaneously min- ing the underlying syntactic structure of the language considered, it effectively in- tegrates the advantages of bilingual pro- jection and unsupervised induction, so as to induce a monolingual grammar much better than previous models only using bilingual projection or unsupervised in- duction. We induced dependency gram- mar for five different languages under the guidance of dependency information pro- jected from the parsed English translation, experiments show that the bilingually- guided method achieves a significant improvement of 28.5% over the unsuper- vised baseline and 3.0% over the best pro- jection baseline on average. 1 Introduction In past decades supervised methods achieved the state-of-the-art in constituency parsing (Collins, 2003; Charniak and Johnson, 2005; Petrov et al., 2006) and dependency parsing (McDonald et al., 2005a; McDonald et al., 2006; Nivre et al., 2006; Nivre et al., 2007; Koo and Collins, 2010). For supervised models, the human-annotated corpora on which models are trained, however, are expen- sive and difficult to build. As alternative strate- gies, methods which utilize raw texts have been in- vestigated recently, including unsupervised meth- ods which use only raw texts (Klein and Man- ning, 2004; Smith and Eisner, 2005; William et al., 2009), and semi-supervised methods (Koo et al., 2008) which use both raw texts and annotat- ed corpus. And there are a lot of efforts have also been devoted to bilingual projection (Chen et al., 2010), which resorts to bilingual text with one lan- guage parsed, and projects the syntactic informa- tion from the parsed language to the unparsed one (Hwa et al., 2005; Ganchev et al., 2009). In dependency grammar induction, unsuper- vised methods achieve continuous improvements in recent years (Klein and Manning, 2004; Smith and Eisner, 2005; Bod, 2006; William et al., 2009; Spitkovsky et al., 2010). Relying on a predefined distributional assumption and iteratively maximiz- ing an approximate indicator (entropy, likelihood, etc.), an unsupervised model usually suffers from two drawbacks, i.e., lower performance and high- er computational cost. On the contrary, bilin- gual projection (Hwa et al., 2005; Smith and Eis- ner, 2009; Jiang and Liu, 2010) seems a promis- ing substitute for languages with large amount of bilingual sentences and an exist-a ing parser of the counterpart language. By project- ing syntactic structures directly (Hwa et al., 2005; Smith and Eisner, 2009; Jiang and Liu, 2010) across bilingual …

Content truncated to 500 words