Online Language Model adaptation via N-gram Mixtures for Statistical Machine Translation

Online Language Model adaptation via N-gram Mixtures for Statistical Machine Translation Germ´an Sanchis-Trilles Instituto Tecnol´ogico de Inform´atica Universidad Polit´ecnica de Valencia Valencia, Spain gsanchis@dsic.upv.es Mauro Cettolo FBK Fondazione Bruno Kessler Trento, Italy cettolo@fbk.eu Abstract The problem of language model adaptation in statistical machine translation is consid- ered. A mixture of language models is employed, which is obtained by clustering the bilingual training data. Unsupervised clustering is guided by either the devel- opment or the test set. Different mixture weight estimation schemes are proposed and compared, at the level of either sin- gle or all source sentences. Experimental results show that, by training different spe- cific language models weighted according to the actual input instead of using a single target language model, translation quality is improved, as measured by BLEU and TER. 1 Introduction The grounds of modern Statistical Machine Trans- lation (SMT), a pattern recognition approach to machine translation, were established in (Brown et al., 1993), where the problem of machine trans- lation was defined as follows: given a sentence f from a certain source language, an equivalent sen- tence ˆe in a given target language that maximizes the posterior probability is to be found. Such a statement can be formalized as: ˆe = argmax e Pr(e|f) = argmax e Pr(f|e) · Pr(e) (1) where Pr(f|e) stands for the translation probabil- ity and Pr(e) accounts for penalizing ill-formed sentences of the target language. c⃝ 2010 European Association for Machine Translation. More recently, a direct modeling of the posterior probability Pr(e|f) has been widely adopted, and, to this purpose, different authors (Papineni et al., 1998; Och and Ney, 2002) proposed the use of the so-called log-linear model, where Pr(e|f) = exp �K k=1 λkhk(f, e) � e′ exp �K k=1 λkhk(f, e′) (2) and the decision rule is given by the expression ˆe = argmax e K � k=1 λkhk(f, e) (3) where hk(f, e) is a score function representing an important feature for the translation of f into e, for example the target language model p(e), K is the number of models (or features) and λk are the weights of the log-linear combination. Typi- cally, the weights λk are optimized during the tun- ing stage with the use of a development set. In this paper, we deal with the problem of adap- tation of SMT models. Specifically, we focus on augmenting the Language Model (LM) compo- nent by introducing parameters that are adapted dynamically to the input text. With this purpose, the LM is implemented as a mixture of special- ized sub-LMs, which are conveniently estimated through some bilingual clustering of the train- ing data and then combined following different weighting schemes. The work described here rep- resents an important extension of what is presented in (Sanchis-Trilles et al., 2009): in fact, there the methods were tested on a small task like IWSLT; on the contrary, here the approach is assessed on the medium-sized Europarl task. Moreover, the clustering of training data does not exploit any su- pervised annotation of …

Content truncated to 500 words