Alignment templates: the RWTH SMT system

Alignment Templates: the RWTH SMT System Oliver Bender, Richard Zens, Evgeny Matusov, Hermann Ney Chair of Computer Science VI RWTH Aachen University D-52056 Aachen, Germany {bender,zens,matusov,ney}@cs.rwth-aachen.de Abstract In this paper, we describe the RWTH statistical machine translation (SMT) system which is based on log-linear model combination. All knowledge sources are treated as feature functions which depend on the source language sentence, the target language sentence and possible hidden variables. The main feature of our approach are the alignment templates which take shallow phrase structures into ac- count: a phrase level alignment between phrases and a word level alignment between single words within the phrases. Thereby, we directly consider word contexts and local reorderings. In order to incorporate additional models (the IBM-1 statistical lexicon model, a word deletion model, and higher order language models), we per- form n-best list rescoring. Participating in the International Work- shop on Spoken Language Translation (IWSLT 2004), we eval- uate our system on the Basic Travel Expression Corpus (BTEC) Chinese-to-English and Japanese-to-English tasks. 1. Introduction The goal of machine translation is the translation of a text given in some source language into a target language. We are given a source string f J 1 = f1...fj...fJ, which is to be translated into a target string eI 1 = e1...ei...eI. Among all possible target strings, we will choose the string with the highest probability: ˆeI 1 = argmax eI 1 n Pr(eI 1|f J 1 ) o = argmax eI 1 n Pr(eI 1) · Pr(f J 1 |eI 1) o (1) The decomposition into two knowledge sources in Equation 1 is known as the source-channel approach to statistical machine trans- lation [1]. It allows an independent modeling of target language model Pr(eI 1) and translation model Pr(f J 1 |eI 1). The target lan- guage model describes the well-formedness of the target language sentence. The translation model links the source language sentence to the target language sentence. It can be further decomposed into alignment and lexicon model. An alternative to the classical source-channel approach is the direct modeling of the posterior probability Pr(eI 1|f J 1 ). Using a log-linear model [2], we obtain: Pr(eI 1|f J 1 ) = pλM 1 (eI 1|f J 1 ) = exp ů M P m=1 λmhm(eI 1, f J 1 ) ÿ P e′I 1 exp ů M P m=1 λmhm(e′I 1, f J 1 ) ÿ The hm denote the feature functions. As a decision rule, we obtain: ˆeI 1 = argmax eI 1 ( M X m=1 λmhm(eI 1, f J 1 ) ) (2) This approach is a generalization of the source-channel approach. It has the advantage that additional models or feature functions can be easily integrated into the overall system. The overall architecture of the log-linear model combination is summarized in Figure 1. The argmax operation denotes the search problem, i.e. the gen- eration of the output sentence in the target language. We have to maximize over all possible target language sentences. In a …

Content truncated to 500 words