Domain Adaptation in Statistical Machine Translation with Mixture Modelling

Proceedings of the Second Workshop on Statistical Machine Translation, pages 177–180, Prague, June 2007. c⃝2007 Association for Computational Linguistics Domain Adaptation in Statistical Machine Translation with Mixture Modelling ∗ Jorge Civera and Alfons Juan Universidad Polit´ecnica de Valencia Camino de Vera s/n 46022 Valencia, Spain {jorcisai,ajuan}@iti.upv.es Abstract Mixture modelling is a standard technique for density estimation, but its use in sta- tistical machine translation (SMT) has just started to be explored. One of the main advantages of this technique is its capabil- ity to learn specific probability distributions that better fit subsets of the training dataset. This feature is even more important in SMT given the difficulties to translate polysemic terms whose semantic depends on the con- text in which that term appears. In this pa- per, we describe a mixture extension of the HMM alignment model and the derivation of Viterbi alignments to feed a state-of-the-art phrase-based system. Experiments carried out on the Europarl and News Commentary corpora show the potential interest and limi- tations of mixture modelling. 1 Introduction Mixture modelling is a popular approach for density estimation in many scientific areas (G. J. McLach- lan and D. Peel, 2000). One of the most interest- ing properties of mixture modelling is its capability to model multimodal datasets by defining soft parti- tions on these datasets, and learning specific proba- bility distributions for each partition, that better ex- plains the general data generation process. ∗Work supported by the EC (FEDER) and the Spanish MEC under grant TIN2006-15694-CO2-01, the Conseller´ıa d’Empresa, Universitat i Ci`encia - Generalitat Valenciana un- der contract GV06/252, the Universidad Polit´ecnica de Valen- cia with ILETA project and Ministerio de Educaci´on y Ciencia. In Machine Translation (MT), it is common to encounter large parallel corpora devoted to hetero- geneous topics. These topics usually define sets of topic-specific lexicons that need to be translated taking into the semantic context in which they are found. This semantic dependency problem could be overcome by learning topic-dependent translation models that capture together the semantic context and the translation process. However, there have not been until very recently that the application of mixture modelling in SMT has received increasing attention. In (Zhao and Xing, 2006), three fairly sophisticated bayesian top- ical translation models, taking IBM Model 1 as a baseline model, were presented under the bilingual topic admixture model formalism. These models capture latent topics at the document level in order to reduce semantic ambiguity and improve translation coherence. The models proposed provide in some cases better word alignment and translation quality than HMM and IBM models on an English-Chinese task. In (Civera and Juan, 2006), a mixture exten- sion of IBM model 2 along with a specific dynamic- programming decoding algorithm were proposed. This IBM-2 mixture model offers a significant gain in translation quality over the conventional IBM model 2 on a semi-synthetic task. In this work, we present a mixture extension of the well-known HMM alignment model first proposed in (Vogel and others, 1996) and refined in (Och and Ney, …

Content truncated to 500 words