Triplet Lexicon Models for Statistical Machine Translation

Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 372–381, Honolulu, October 2008. c⃝2008 Association for Computational Linguistics Triplet Lexicon Models for Statistical Machine Translation Saˇsa Hasan, Juri Ganitkevitch, Hermann Ney, Jes´us Andr´es-Ferrer†∗ Human Language Technology and Pattern Recognition, RWTH Aachen University, Germany †Universidad Polit´ecnica de Valencia, Dept. Sist. Inform´aticos y Computaci´on {hasan,ganitkevitch,ney}@cs.rwth-aachen.de jandres@dsic.upv.es Abstract This paper describes a lexical trigger model for statistical machine translation. We present various methods using triplets incorporating long-distance dependencies that can go be- yond the local context of phrases or n-gram based language models. We evaluate the pre- sented methods on two translation tasks in a reranking framework and compare it to the re- lated IBM model 1. We show slightly im- proved translation quality in terms of BLEU and TER and address various constraints to speed up the training based on Expectation- Maximization and to lower the overall num- ber of triplets without loss in translation per- formance. 1 Introduction Data-driven methods have been applied very suc- cessfully within the machine translation domain since the early 90s. Starting from single-word- based translation approaches, significant improve- ments have been made through advances in mod- eling, availability of larger corpora and more pow- erful computers. Thus, substantial progress made in the past enables today’s MT systems to achieve acceptable results in terms of translation quality for specific language pairs such as Arabic-English. If sufficient amounts of parallel data are available, sta- tistical MT systems can be trained on millions of ∗The work was carried out while the author was at the Hu- man Language Technology and Pattern Recognition group at RWTH Aachen University and partly supported by the Valen- cian Conselleria d’Empresa, Universitat i Ci`encia under grants CTBPRA/2005/ and BEFPI/2007/014. target source e e′ f Figure 1: Triplet example: a source word f is triggered by two target words e and e′, where one of the words is within and the other outside the considered phrase pair (indicated by the dashed line). sentence pairs and use an extended level of context based on bilingual groups of words which denote the building blocks of state-of-the-art phrase-based SMT systems. Due to data sparseness, statistical models are of- ten trained on local context only. Language mod- els are derived from n-grams with n ≤ 5 and bilin- gual phrase pairs are extracted with lengths up to 10 words on the target side. This captures the local dependencies of the data in detail and is responsi- ble for the success of data-driven phrase-based ap- proaches. In this work, we will introduce a new statistical model based on lexicalized triplets (f, e, e′) which we will also refer to as cross-lingual triggers of the form (e, e′ → f). This can be understood as two words in one language triggering one word in another language. These triplets, modeled by p(f|e, e′), are closely related to lexical translation probabilities based on the IBM model 1, i.e. p(f|e). Several constraints and setups will be described later on in …

Content truncated to 500 words