Automatic Cross-Lingual Similarization of Dependency Grammars for Tree-based Machine Translation

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 501–510, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics Automatic Cross-Lingual Similarization of Dependency Grammars for Tree-based Machine Translation Wenbin Jiang 1 and Wen Zhang 1 and Jinan Xu 2 and Rangjia Cai 3 1Key Laboratory of Intelligent Information Processing Institute of Computing Technology, Chinese Academy of Sciences, China 2School of Computer and Information Technology, Beijing Jiaotong University, China 3Research Center of Tibetan Information, Qinghai Normal University, China jiangwenbin@ict.ac.cn Abstract Structural isomorphism between languages benefits the performance of cross-lingual ap- plications. We propose an automatic al- gorithm for cross-lingual similarization of dependency grammars, which automatically learns grammars with high cross-lingual sim- ilarity. The algorithm similarizes the an- notation styles of the dependency grammars for two languages in the level of classifica- tion decisions, and gradually improves the cross-lingual similarity without losing linguis- tic knowledge resorting to iterative cross- lingual cooperative learning. The dependency grammars given by cross-lingual similariza- tion have much higher cross-lingual similar- ity while maintaining non-triviality. As appli- cations, the cross-lingually similarized gram- mars significantly improve the performance of dependency tree-based machine translation. 1 Introduction Due to the inherent syntactic regularity of each language and the discrepancy between annotation guidelines of linguists, there is not necessarily struc- tural isomorphism between grammars of different languages. For many cross-lingual scenarios such as information retrieval and machine translation, re- lationships between linguistic units are expected to be (at least roughly) consistent across languages (Hwa et al., 2002; Smith and Eisner, 2009). For cross-lingual applications, syntactic structures with high cross-lingual similarity facilitates knowledge extraction, feature representation and classification decision. The structural isomorphism between lan- guages, therefore, is an important aspect for the per- formance of cross-lingual applications such as ma- chine translation. To achieve effective cross-lingual similarization for two grammars in different languages, an ad- equate algorithm should both improve the cross- lingual similarity between two grammars and main- tain the non-triviality of each grammar, where non- triviality indicates that the resulted grammars should not give flat or single-branched outputs. Differ- ent from constituency structures, dependency struc- tures are lexicalized without specialized hierarchical structures. Such concise structures depict the syn- tactic or semantic relationships between words, and thus have advantage on many cross-lingual scenar- ios. It is worth to perform cross-lingual similariza- tion for dependency grammars, but the special prop- erty of dependency grammars makes it hard to di- rectly adopt the conventional structure transforma- tion methods resorting to hand-crafted rules or tem- plates. Both graph-based models (McDonald et al., 2005) and transition-based models (Nivre et al., 2006) factorize dependency parsing into fundamen- tal classification decisions, that is, the relation- ships between words or the actions applied to cur- rent states. We assume that cross-lingual simi- larization can also be factorized into fundamen- tal classification decisions, and propose an au- tomatic cross-lingual similarization algorithm for dependency grammars according to this assump- tion. The algorithm conducts cross-lingual sim- ilarization on the level of classification decisions 501 with simple …

Content truncated to 500 words