Search-Aware Tuning for Machine Translation

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1942–1952, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Search-Aware Tuning for Machine Translation Lemao Liu Queens College City University of New York lemaoliu@gmail.com Liang Huang Queens College and Graduate Center City University of New York liang.huang.sh@gmail.com Abstract Parameter tuning is an important problem in statistical machine translation, but surpris- ingly, most existing methods such as MERT, MIRA and PRO are agnostic about search, while search errors could severely degrade translation quality. We propose a search- aware framework to promote promising par- tial translations, preventing them from be- ing pruned. To do so we develop two met- rics to evaluate partial derivations. Our tech- nique can be applied to all of the three above-mentioned tuning methods, and ex- tensive experiments on Chinese-to-English and English-to-Chinese translation show up to +2.6 BLEU gains over search-agnostic baselines. 1 Introduction Parameter tuning has been a key problem for ma- chine translation since the statistical revolution. However, most existing tuning algorithms treat the decoder as a black box (Och, 2003; Hopkins and May, 2011; Chiang, 2012), ignoring the fact that many potentially promising partial translations are pruned by the decoder due to the prohibitively large search space. For example, the popular beam-search decoding algorithm for phrase-based MT (Koehn, 2004) only explores O(nb) items for a sentence of n words (with a beam width of b), while the full search space is O(2nn2) or worse (Knight, 1999). As one of the very few exceptions to the “search-agnostic” majority, Yu et al. (2013) and Zhao et al. (2014) propose a variant of the per- ceptron algorithm that learns to keep the refer- ence translations in the beam or chart. How- ever, there are several obstacles that prevent their method from becoming popular: First of all, they rely on “forced decoding” to track gold derivations that lead to the reference translation, but in practice only a small portion of (mostly very short) sen- (a) 0 1 2 3 4 (b) Figure 1: (a) Some potentially promising partial trans- lations (in red) fall out of the beam (bin 2); (b) We identify such partial translations and assign them higher model scores so that they are more likely to survive the search. tence pairs have at least one such derivation. Sec- ondly, they learn the model on the training set, and while this does enable a sparse feature set, it is or- ders of magnitude slower compared to MERT and PRO. We instead propose a very simple framework, search-aware tuning, which does not depend on forced decoding, and thus can be trained on all sen- tence pairs of any dataset. The key idea is that, besides caring about the rankings of the complete translations, we also promote potentially promis- ing partial translations so that they are more likely to survive throughout the search, see Figure 1 for illustration. We make the following contributions: • Our idea of search-aware tuning can be ap- plied (as …

Content truncated to 500 words