Phrase-Based Statistical Machine Translation: A Level of Detail Approach

Phrase-Based Statistical Machine Translation: A Level of Detail Approach Hendra Setiawan1,2, Haizhou Li1, Min Zhang1, and Beng Chin Ooi2 1 Institute for Infocomm Research, 21 Heng Mui Keng Terrace, Singapore 119613 {stuhs, hli, mzhang}@i2r.a-star.edu.sg 2 School of Computing, National University of Singapore, Singapore 117543 {hendrase, ooibc}@comp.nus.edu.sg Abstract. The merit of phrase-based statistical machine translation is often reduced by the complexity to construct it. In this paper, we ad- dress some issues in phrase-based statistical machine translation, namely: the size of the phrase translation table, the use of underlying transla- tion model probability and the length of the phrase unit. We present Level-Of-Detail (LOD) approach, an agglomerative approach for learn- ing phrase-level alignment. Our experiments show that LOD approach significantly improves the performance of the word-based approach. LOD demonstrates a clear advantage that the phrase translation table grows only sub-linearly over the maximum phrase length, while having a per- formance comparable to those of other phrase-based approaches. 1 Introduction Early approach to statistical machine translation relies on the word-based trans- lation model to describe the translation process [1]. However, the underlying as- sumption of word-to-word translation often fails to capture all properties of the language, i.e. the existence of the phrase where a group of words often function together as a unit. Many researchers have proposed to move from the word-based to the phrase-based translation model [2] [3] [4]. A phrase-based approach offers many advantages as a phrase translation captures word context and local re- ordering inherently [3]. It has become popular in statistical machine translation applications. There are typically two groups of approaches to constructing the phrase- based model. The first group learns phrase translation directly from the sen- tence pair. It learns both word and phrase units simultaneously. Although these approaches appear intuitive, it usually suffers from a prohibitive computational cost. It might have to consider all possible multi-word sequences as phrase can- didates and all possible pairings as phrase translations at the same time. R. Dale et al. (Eds.): IJCNLP 2005, LNAI 3651, pp. 576–587, 2005. c ⃝ Springer-Verlag Berlin Heidelberg 2005 Phrase-Based Statistical Machine Translation: A Level of Detail Approach 577 The second group of approaches learns phrase translations through word-level alignment: alignment template [2] and projection extension [6], just to name a few. In general, these approaches take the word-level alignment, a by-product of the word-based translation model, as their input and then utilize a heuristic mea- surement to learn the phrase translation. The heuristic measurement contains all possible configurations of word-level alignment on a phrase translation. It is noted that the underlying word-level alignment is just an approximation to the exact alignment. The approximation is reflected by a probability produced by the word-based translation model. The majority of approaches do not make use of this probability, whereas it may provide a valuable clue leading to a better phrase translation from a statistical point of view. Koehn, et. al [8] compared the representative of both groups and reported that learning phrase translation using a simple heuristic …

Content truncated to 500 words