Soft Cross-lingual Syntax Projection for Dependency Parsing

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 783–793, Dublin, Ireland, August 23-29 2014. Soft Cross-lingual Syntax Projection for Dependency Parsing Zhenghua Li , Min Zhang∗, Wenliang Chen Provincial Key Laboratory for Computer Information Processing Technology Soochow University {zhli13,minzhang,wlchen}@suda.edu.cn Abstract This paper proposes a simple yet effective framework of soft cross-lingual syntax projection to transfer syntactic structures from source language to target language using monolingual treebanks and large-scale bilingual parallel text. Here, soft means that we only project reliable dependencies to compose high-quality target structures. The projected instances are then used as additional training data to improve the performance of supervised parsers. The major issues for this idea are 1) errors from the source-language parser and unsupervised word aligner; 2) intrinsic syntactic non-isomorphism between languages; 3) incomplete parse trees after projection. To handle the first two issues, we propose to use a probabilistic dependency parser trained on the target-language treebank, and prune out unlikely projected dependencies that have low marginal probabilities. To make use of the incomplete projected syntactic structures, we adopt a new learning technique based on ambiguous labelings. For a word that has no head words after projection, we enrich the projected structure with all other words as its candidate heads as long as the newly-added dependency does not cross any projected dependencies. In this way, the syntactic structure of a sentence becomes a parse forest (ambiguous labels) instead of a single parse tree. During training, the objective is to maximize the mixed likelihood of manually labeled instances and projected instances with ambiguous labelings. Experimental results on benchmark data show that our method significantly outperforms a strong baseline supervised parser and previous syntax projection methods. 1 Introduction During the past decade, supervised dependency parsing has made great progress. However, due to the limitation of scale and genre coverage of labeled data, it is very difficult to further improve the performance of supervised parsers. On the other hand, it is very time-consuming and labor-intensive to manually construct treebanks. Therefore, lots of recent work has been devoted to get help from bilingual constraints. The motivation behind are two-fold. First, a difficult syntactic ambiguity in one language may be very easy to resolve in another language. Second, a more accurate parser on one language may help an inferior parser on another language, where the performance difference may be due to the intrinsic complexity of languages or the scale of accessible labeled resources. Following the above research line, much effort has been done recently to explore bilingual constraints for parsing. Burkett and Klein (2008) propose a reranking based method for joint constituent parsing of bitext, which can make use of structural correspondence features in both languages. Their method needs bilingual treebanks with manually labeled syntactic trees on both sides for training. Huang et al. (2009) compose useful parsing features based on word reordering information in source-language sentences. Chen et al. (2010a) derive bilingual subtree constraints with auto-parsed source-language sentences. During training, both Huang et al. (2009) …

Content truncated to 500 words