Ngram-Based Statistical Machine Translation Enhanced with Multiple Weighted Reordering Hypotheses

Proceedings of the Second Workshop on Statistical Machine Translation, pages 167–170, Prague, June 2007. c⃝2007 Association for Computational Linguistics Ngram-based statistical machine translation enhanced with multiple weighted reordering hypotheses Marta R. Costa-juss`a, Josep M. Crego, Patrik Lambert, Maxim Khalilov Jos´e A. R. Fonollosa, Jos´e B. Mari˜no and Rafael E. Banchs Department of Signal Theory and Communications TALP Research Center (UPC) Barcelona 08034, Spain (mruiz,jmcrego,lambert,khalilov,adrian,canton,rbanchs)@gps.tsc.upc.edu Abstract This paper describes the 2007 Ngram-based sta- tistical machine translation system developed at the TALP Research Center of the UPC (Uni- versitat Polit`ecnica de Catalunya) in Barcelona. Emphasis is put on improvements and extensions of the previous years system, being highlighted and empirically compared. Mainly, these include a novel word ordering strategy based on: (1) sta- tistically monotonizing the training source cor- pus and (2) a novel reordering approach based on weighted reordering graphs. In addition, this system introduces a target language model based on statistical classes, a feature for out-of-domain units and an improved optimization procedure. The paper provides details of this system par- ticipation in the ACL 2007 SECOND WORK- SHOP ON STATISTICAL MACHINE TRANSLA- TION. Results on three pairs of languages are reported, namely from Spanish, French and Ger- man into English (and the other way round) for both the in-domain and out-of-domain tasks. 1 Introduction Based on estimating a joint-probability model between the source and the target languages, Ngram-based SMT has proved to be a very competitive alternatively to phrase-based and other state-of-the-art systems in previ- ous evaluation campaigns, as shown in (Koehn and Monz, 2005; Koehn and Monz, 2006). Given the challenge of domain adaptation, efforts have been focused on improving strategies for Ngram-based SMT which could generalize better. Specifically, a novel reordering strategy is explored. It is based on extending the search by using precomputed statistical information. Results are promising while keeping computational ex- penses at a similar level as monotonic search. Addition- ally, a bonus for tuples from the out-of-domain corpus is introduced, as well as a target language model based on statistical classes. One of the advantages of working with statistical classes is that they can easily be used for any pair of languages. This paper is organized as follows. Section 2 briefly reviews last year’s system, including tuple definition and extraction, translation model and feature functions, de- coding tool and optimization criterion. Section 3 delves into the word ordering problem, by contrasting last year strategy with the novel weighted reordering input graph. Section 4 focuses on new features: both tuple-domain bonus and target language model based on classes. Later on, Section 5 reports on all experiments carried out for WMT 2007. Finally, Section 6 sums up the main conclu- sions from the paper and discusses future research lines. 2 Baseline N-gram-based SMT System The translation model is based on bilingual n-grams. It actually constitutes a language model of bilingual units, referred to as tuples, which approximates the joint proba- bility between source and target languages by using bilin- gual n-grams. Tuples are extracted from a word-to-word …

Content truncated to 500 words