Adaptation of the translation model for statistical machine translation based on information retrieval

EAMT 2005 Conference Proceedings 133 Adaptation of the Translation Model for Statistical Machine Translation based on Information Retrieval Almut Silja Hildebrand, Matthias Eck, Stephan Vogel and Alex Waibel Interactive Systems Laboratory Carnegie Mellon University 5000 Forbes Avenue Pittsburgh, PA, 15213, USA silja@cs.cmu.edu, matteck@cs.cmu.edu, stephan.vogel@cs.cmu.edu, waibel@cs.cmu.edu Abstract. In this paper we present experiments concerning translation model adaptation for statistical machine translation. We develop a method to adapt translation models using in- formation retrieval. The approach selects sentences similar to the test set to form an adapted training corpus. The method allows a better use of additionally available out-of-domain training data or finds in-domain data in a mixed corpus. The adapted translation models significantly improve the translation performance compared to competitive baseline sys- tems. 1. Introduction The goal of this research is to improve the trans- lation performance for a Statistical Machine Translation system. The basic approach is to adapt the translation models. Statistical machine translation can be de- scribed in a more formal way as follows: () argmax(|) argmax(|) * Pt Pst Pts t t t ⋅ = = Here t is the target sentence, and s is the source sentence. P(t) is the target language model and P(s|t) is the translation model used in the de- coder. Statistical machine translation searches for the best target sentence from the space de- fined by the target language model (LM) and the translation model (TM). Statistical translation models are usually ei- ther phrase- or word-based and include most notably IBM1 to IBM4 and HMM (Brown et al., 1993; Vogel et al., 1996). All models use available bilingual training data in the source and target language to estimate their parameters and approximate the translation probabilities. Typically, the more data is used to estimate the parameters of the translation model, the bet- ter it can approximate the “true” translation probabilities. This will obviously lead to a higher translation performance. However if a signifi- cant amount of out-of-domain data is added to the training data, translation quality can drop. One reason for this is that a general translation model P(s|t), that was trained on in-domain and out-of-domain data, does not fit the topic or style of individual texts. Unfortunately the mean- ing of quite a number of words and phrases is ambiguous; this results in the fact that their translation highly depends on the topic and con- text they are used in. For example the word ‘leg’ is usually thought of as a body part (‘He broke his leg’). In sports, especially bicycling, the word ‘leg’ can also have the meaning of ‘stage’ (‘US Postal wins fourth leg’). Similar to this meaning is the use in aviation with the phrase ‘single leg airline’. This fact would not be a problem if the trans- lations for ‘leg’ were the same in every case. But this is rarely true. German for example uses different words for the upper three meanings of ‘leg’. So a translation that might be totally ac- ceptable for one specific topic, applied to test …

Content truncated to 500 words