Learning Translation Boundaries for Phrase-Based Decoding

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 136–144, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Learning Translation Boundaries for Phrase-Based Decoding Deyi Xiong, Min Zhang, Haizhou Li Human Language Technology Institute for Infocomm Research 1 Fusionopolis Way, #21-01 Connexis, Singapore 138632. {dyxiong, mzhang, hli}@i2r.a-star.edu.sg Abstract Constrained decoding is of great importance not only for speed but also for translation qual- ity. Previous efforts explore soft syntactic con- straints which are based on constituent bound- aries deduced from parse trees of the source language. We present a new framework to es- tablish soft constraints based on a more nat- ural alternative: translation boundary rather than constituent boundary. We propose sim- ple classifiers to learn translation boundaries for any source sentences. The classifiers are trained directly on word-aligned corpus with- out using any additional resources. We report the accuracy of our translation boundary clas- sifiers. We show that using constraints based on translation boundaries predicted by our classifiers achieves significant improvements over the baseline on large-scale Chinese-to- English translation experiments. The new constraints also significantly outperform con- stituent boundary based syntactic constrains. 1 Introduction It has been known that phrase-based decoding (phrase segmentation/translation/reordering (Chi- ang, 2005)) should be constrained to some extent not only for transferring the NP-hard problem (Knight, 1999) into a tractable one in practice but also for im- proving translation quality. For example, Xiong et al. (2008) find that translation quality can be signif- icantly improved by either prohibiting reorderings around punctuation or restricting reorderings within a 15-word window. Recently, more linguistically motivated con- straints are introduced to improve phrase-based de- coding. (Cherry, 2008) and (Marton and Resnik, 2008) introduce syntactic constraints into the stan- dard phrase-based decoding (Koehn et al., 2003) and hierarchical phrase-based decoding (Chiang, 2005) respectively by using a counting feature which ac- cumulates whenever hypotheses violate syntactic boundaries of source-side parse trees. (Xiong et al., 2009) further presents a bracketing model to include thousands of context-sensitive syntactic constraints. All of these approaches achieve their improvements by guiding the phrase-based decoder to prefer trans- lations which respect source-side parse trees. One major problem with such constituent bound- ary based constraints is that syntactic structures of the source language do not necessarily reflect trans- lation structures where the source and target lan- guage correspond to each other. In this paper, we investigate building classifiers that directly ad- dress the problem of translation boundary, rather than extracting constituent boundary from source- side parsers built for a different purpose. A trans- lation boundary is a position in the source sequence which begins or ends a translation zone 1 spanning multiple source words. In a translation zone, the source phrase is translated as a unit. Reorderings which cross translation zones are not desirable. Inspired by (Roark and Hollingshead, 2008) which introduces classifiers to decide if a word can begin/end a multi-word constituent, we build two discriminative classifiers to tag each word in the source sequence with a binary …

Content truncated to 500 words