Bilingually-Constrained (Monolingual) Shift-Reduce Parsing

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 1222–1231, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Bilingually-Constrained (Monolingual) Shift-Reduce Parsing Liang Huang Google Research 1350 Charleston Rd. Mountain View, CA 94043, USA lianghuang@google.com liang.huang.sh@gmail.com Wenbin Jiang and Qun Liu Key Lab. of Intelligent Information Processing Institute of Computing Technology Chinese Academy of Sciences P.O. Box 2704, Beijing 100190, China jiangwenbin@ict.ac.cn Abstract Jointly parsing two languages has been shown to improve accuracies on either or both sides. However, its search space is much bigger than the monolingual case, forcing existing approaches to employ complicated modeling and crude approxi- mations. Here we propose a much simpler alternative, bilingually-constrained mono- lingual parsing, where a source-language parser learns to exploit reorderings as ad- ditional observation, but not bothering to build the target-side tree as well. We show specifically how to enhance a shift-reduce dependency parser with alignment fea- tures to resolve shift-reduce conflicts. Ex- periments on the bilingual portion of Chi- nese Treebank show that, with just 3 bilin- gual features, we can improve parsing ac- curacies by 0.6% (absolute) for both En- glish and Chinese over a state-of-the-art baseline, with negligible (∼6%) efficiency overhead, thus much faster than biparsing. 1 Introduction Ambiguity resolution is a central task in Natu- ral Language Processing. Interestingly, not all lan- guages are ambiguous in the same way. For exam- ple, prepositional phrase (PP) attachment is (no- toriously) ambiguous in English (and related Eu- ropean languages), but is strictly unambiguous in Chinese and largely unambiguous Japanese; see (1a) I [ saw Bill ] [ with a telescope ]. wo [ yong wangyuanjin] [kandao le Bi’er]. “I used a telescope to see Bill.” (1b) I saw [ Bill [ with a telescope ] ]. wo kandao le [ [ na wangyuanjin ] de Bi’er]. “I saw Bill who had a telescope at hand.” Figure 1: PP-attachment is unambiguous in Chi- nese, which can help English parsing. Figure 1 for an example.1 It is thus intuitive to use two languages for better disambiguation, which has been applied not only to this PP-attachment problem (Fossum and Knight, 2008; Schwartz et al., 2003), but also to the more fundamental prob- lem of syntactic parsing which subsumes the for- mer as a subproblem. For example, Smith and Smith (2004) and Burkett and Klein (2008) show that joint parsing (or reranking) on a bitext im- proves accuracies on either or both sides by lever- aging bilingual constraints, which is very promis- ing for syntax-based machine translation which re- quires (good-quality) parse trees for rule extrac- tion (Galley et al., 2004; Mi and Huang, 2008). However, the search space of joint parsing is in- evitably much bigger than the monolingual case, 1Chinese uses word-order to disambiguate the attachment (see below). By contrast, Japanese resorts to case-markers and the unambiguity is limited: it works for the “V or N” attachment ambiguities like in Figure 1 (see (Schwartz et al., 2003)) but not for the “N1 or N2” case (Mitch Marcus, …

Content truncated to 500 words