Improving Translation Quality by Discarding Most of the Phrasetable

Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pp. 967–975, Prague, June 2007. c⃝2007 Association for Computational Linguistics Improving Translation Quality by Discarding Most of the Phrasetable J Howard Johnson and Joel Martin Interactive Information Group National Research Council Canada Ottawa, Ontario, Canada firstname.lastname@nrc.gc.ca George Foster and Roland Kuhn Interactive Language Technologies Group National Research Council Canada Gatineau, Qu´ebec, Canada firstname.lastname@nrc.gc.ca Abstract It is possible to reduce the bulk of phrase- tables for Statistical Machine Translation us- ing a technique based on the significance testing of phrase pair co-occurrence in the parallel corpus. The savings can be quite substantial (up to 90%) and cause no reduc- tion in BLEU score. In some cases, an im- provement in BLEU is obtained at the same time although the effect is less pronounced if state-of-the-art phrasetable smoothing is employed. 1 Introduction An important part of the process of Statistical Ma- chine Translation (SMT) involves inferring a large table of phrase pairs that are translations of each other from a large corpus of aligned sentences. These phrase pairs together with estimates of con- ditional probabilities and useful feature weights, called collectively a phrasetable, are used to match a source sentence to produce candidate translations. The choice of the best translation is made based on the combination of the probabilities and feature weights, and much discussion has been made of how to make the estimates of probabilites, how to smooth these estimates, and what features are most useful for discriminating among the translations. However, a cursory glance at phrasetables pro- duced often suggests that many of the translations are wrong or will never be used in any translation. On the other hand, most obvious ways of reducing the bulk usually lead to a reduction in translation quality as measured by BLEU score. This has led to an impression that these pairs must contribute some- thing in the grand scheme of things and, certainly, more data is better than less. Nonetheless, this bulk comes at a cost. Large ta- bles lead to large data structures that require more resources and more time to process and, more im- portantly, effort directed in handling large tables could likely be more usefully employed in more fea- tures or more sophisticated search. In this paper, we show that it is possible to prune phrasetables using a straightforward approach based on significance testing, that this approach does not adversely affect the quality of translation as mea- sured by BLEU score, and that savings in terms of number of discarded phrase pairs can be quite sub- stantial. Even more surprising, pruning can actu- ally raise the BLEU score although this phenomenon is less prominent if state of the art smoothing of phrasetable probabilities is employed. Section 2 reviews the basic ideas of Statistical Machine Translation as well as those of testing sig- nificance of associations in two by two contingency tables departing from independence. From this, a filtering algorithm will be described …

Content truncated to 500 words