A Phrase-Based Hidden Semi-Markov Approach to Machine Translation

Proceedings of the 13th Annual Conference of the EAMT, pages 168–175, Barcelona, May 2009 A Phrase-Based Hidden Semi-Markov Approach to Machine Translation Jes´us Andr´es-Ferrer Instituto Tecnol´ogico de Inform´atica Universidad Polit´ecnica de Valencia jandres@iti.upv.es Alfons Juan Dpto. de sist. inform´aticos y computaci´on Universidad Polit´ecnica de Valencia ajuan@dsic.upv.es Abstract Statistically estimated phrase-based mod- els promised to further the state-of-the-art, however, several works reported a perfor- mance decrease with respect to heuristi- cally estimated phrase-based models. In this work we present a latent variable phrase-based translation model inspired by the hidden semi-Markov models, that does not degrade the system. Experimental re- sults report an improvement over the base- line. Additionally, it is observed that both Baum-Welch and Viterbi trainings obtain the very same result, suggesting that most of the probability mass is gathered into one single bilingual segmentation. 1 Introduction The machine translation problem is stated as the problem of translating a source sentence, xJ 1 , into a target sentence, yI 1. In accordance with the sta- tistical approach to machine translation, the opti- mal translation ˆy of a source sentence x is given by the fundamental equation of statistical machine translation (Brown and others, 1993) ˆy = arg max y∈Y⋆ p(x | y) p(y) (1) where p(x | y) is approximated by an inverse translation model and p(y) is modelled with a lan- guage model; which is usually instanced by a n- gram language model (Chen and Goodman, 1996). The first approaches to model the translation probability in Eq. (1), were based on word dic- tionaries. These word-based models, the so-called IBM translation models (Brown and others, 1993), c⃝ 2009 European Association for Machine Translation. tackled the problem with word-level dictionaries plus alignments between words. However, current systems model the inverse conditional probability in Eq. (1) using phrase dictionaries. A phrase is understood here as any sequence of source or tar- get words. This phrase-based methodology stores specific sequences of target words (target phrase) into which a sequence of source words (source phrase) is translated. However, a key concept of this approach is the procedure through which these phrase pairs are inferred. The common approach consists in us- ing the IBM alignment models (Brown and others, 1993) to obtain a symmetrised alignment matrix from which coherent phrases are extracted (Och and Ney, 2004). Then, a simple count normalisa- tion is carried out in order to obtain a conditional phrase dictionary. Alternatively, some approaches infer the phrase dictionaries statistically. For instance, a joint prob- ability model for phrase-based estimation is pro- posed in (Marcu and Wong, 2002). In that work, all possible segmentations were extracted using the EM algorithm (Dempster et al., 1977), without any matrix alignment constraint, in contrast to the ap- proach followed in (Och and Ney, 2004). Based on this work, another work (Alexandra Birch and Koehn, 2006), constrained the EM to only con- sider phrases which agree with the alignment ma- trix, thus reducing the size of the phrase dictionar- ies (or tables). A possible drawback of the …

Content truncated to 500 words