Transductive learning for statistical machine translation

Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 25–32, Prague, Czech Republic, June 2007. c⃝2007 Association for Computational Linguistics Transductive learning for statistical machine translation Nicola Ueffing National Research Council Canada Gatineau, QC, Canada nicola.ueffing@nrc.gc.ca Gholamreza Haffari and Anoop Sarkar Simon Fraser University Burnaby, BC, Canada {ghaffar1,anoop}@cs.sfu.ca Abstract Statistical machine translation systems are usually trained on large amounts of bilin- gual text and monolingual text in the tar- get language. In this paper we explore the use of transductive semi-supervised meth- ods for the effective use of monolingual data from the source language in order to im- prove translation quality. We propose sev- eral algorithms with this aim, and present the strengths and weaknesses of each one. We present detailed experimental evaluations on the French–English EuroParl data set and on data from the NIST Chinese–English large- data track. We show a significant improve- ment in translation quality on both tasks. 1 Introduction In statistical machine translation (SMT), translation is modeled as a decision process. The goal is to find the translation t of source sentence s which maxi- mizes the posterior probability: arg max t p(t | s) = arg max t p(s | t) · p(t) (1) This decomposition of the probability yields two dif- ferent statistical models which can be trained in- dependently of each other: the translation model p(s | t) and the target language model p(t). State-of-the-art SMT systems are trained on large collections of text which consist of bilingual corpora (to learn the parameters of p(s | t)), and of monolin- gual target language corpora (for p(t)). It has been shown that adding large amounts of target language text improves translation quality considerably. How- ever, the availability of monolingual corpora in the source language does not help improve the system’s performance. We will show how such corpora can be used to achieve higher translation quality. Even if large amounts of bilingual text are given, the training of the statistical models usually suffers from sparse data. The number of possible events, i.e. phrase pairs or pairs of subtrees in the two lan- guages, is too big to reliably estimate a probabil- ity distribution over such pairs. Another problem is that for many language pairs the amount of available bilingual text is very limited. In this work, we will address this problem and propose a general frame- work to solve it. Our hypothesis is that adding infor- mation from source language text can also provide improvements. Unlike adding target language text, this hypothesis is a natural semi-supervised learn- ing problem. To tackle this problem, we propose algorithms for transductive semi-supervised learn- ing. By transductive, we mean that we repeatedly translate sentences from the development set or test set and use the generated translations to improve the performance of the SMT system. Note that the eval- uation step is still done just once at the end of our learning process. In this paper, we show that such an approach can lead to better …

Content truncated to 500 words