Improving Phrase-Based Translation with Prototypes of Short Phrases

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 301–304, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Improving Phrase-Based Translation with Prototypes of Short Phrases Frank Liberato†, Behrang Mohit‡, Rebecca Hwa†‡ †Department of Computer Science ‡Intelligent Systems Program University of Pittsburgh {frank,behrang,hwa@cs.pitt.edu} Abstract We investigate methods of generating addi- tional bilingual phrase pairs for a phrase- based decoder by translating short sequences of source text. Because our translation task is more constrained, we can use a model that employs more linguistically rich features than a traditional decoder. We have implemented an example of this approach. Experimental re- sults suggest that the phrase pairs produced by our method are useful to the decoder, and lead to improved sentence translations. 1 Introduction Recently, there have been a number of successful attempts at improving phrase-based statistical ma- chine translation by exploiting linguistic knowledge such as morphology, part-of-speech tags, and syn- tax. Many translation models use such knowledge before decoding (Xia and McCord, 2004) and dur- ing decoding (Birch et al., 2007; Gimpel and Smith, 2009; Koehn and Hoang, 2007; Chiang et al., 2009), but they are limited to simpler features for practi- cal reasons, often restricted to conditioning left-to- right on the target sentence. Traditionally, n-best rerankers (Shen et al., 2004) have applied expen- sive analysis after the translation process, on both the source and target side, though they suffer from being limited to whatever is on the n-best list (Hasan et al., 2007). We argue that it can be desirable to pre-translate parts of the source text before sentence-level decod- ing begins, using a richer model that would typically be out of reach during sentence-level decoding. In this paper, we describe a particular method of gen- erating additional bilingual phrase pairs for a new source text, using what we call phrase prototypes, which are are learned from bilingual training data. Our goal is to generate improved translations of rel- atively short phrase pairs to provide the SMT de- coder with better phrasal choices. We validate the idea through experiments on Arabic-English trans- lation. Our method produces a 1.3 BLEU score in- crease (3.3% relative) on a test set. 2 Approach Re-ranking tends to use expensive features of the en- tire source and target sentences, s and t, and align- ments, a, to produce a score for the translation. We will call this scoring function φ(s, t, a). While φ(·) might capture quite a bit of linguistic information, it can be problematic to use this function for decoding directly. This is due to both the expense of com- puting it, and the difficulty in using it to guide the decoder’s search. For example, a choice of φ(·) that relies on a top-down parser is difficult to integrate into a left-to-right decoder (Charniak et al., 2003). Our idea is to use an expensive scoring function to guide the search for potential translations for part of a source sentence, S, even if translating all of …

Content truncated to 500 words