A Framework for Discriminative Rule Selection in Hierarchical Moses

Proceedings of the First Conference on Machine Translation, Volume 1: Research Papers, pages 92–101, Berlin, Germany, August 11-12, 2016. c⃝2016 Association for Computational Linguistics A Framework for Discriminative Rule Selection in Hierarchical Moses Fabienne Braune1, Alexander Fraser1, Hal Daum´e III2, and Aleˇs Tamchyna3 1Center for Information and Language Processing, University of Munich, Germany 2Computer Science and UMIACS, University of Maryland 3Charles University in Prague, Prague Czech Republic Abstract Training discriminative rule selection models is usually expensive because of the very large size of the hierarchical gram- mar. Previous approaches reduced the training costs either by (i) using mod- els that are local to the source side of the rules or (ii) by heavily pruning out negative samples. Moreover, all previ- ous evaluations were performed on small scale translation tasks, containing at most 250,000 sentence pairs. We propose two contributions to discriminative rule selec- tion. First, we test previous approaches on two French-English translation tasks in domains for which only limited resources are available and show that they fail to improve translation quality. To improve on such tasks, we propose a rule selec- tion model that is (i) global with rich label-dependent features (ii) trained with all available negative samples. Our global model yields significant improvements, up to 1 BLEU point, over previously pro- posed rule selection models. Second, we successfully scale rule selection models to large translation tasks but have so far failed to produce significant improvements in BLEU on these tasks. 1 Introduction Hierarchical phrase-based machine translation (Chiang, 2005) performs non-local reordering in a formally syntax-based way. It allows flexible rule extraction and application by using a grammar without linguistic annotation. As a consequence, many hierarchical rules can be used to translate a given input segment even though only a subset of these yield a correct translation. For instance, rules r1 to r3 can be applied to translate the French sentence F1 below although only r1 yields the cor- rect translation E. (r1) X → ⟨ X1 pratique X2, practical X1 X2 ⟩ (r2) X → ⟨ X1 pratique X2, X1 X2 practice ⟩ (r3) X → ⟨ X1 pratique X2, X2 X1 process ⟩ F1 Une ´etude de l’ (int´erˆet)X1 pratique (de notre approche)X2. A study on the (interest)X1 practical (of our approach)X2. E A study on the practical (interest)X1 (of our approach)X2. The rule scoring heuristics defined by (Chiang, 2005) do not handle rule selection in a satisfac- tory way and many authors have come up with solutions. Models that use the syntactic structure of the source and target sentence have been pro- posed by (Marton and Resnik, 2008; Marton et al., 2012; Chiang et al., 2009; Chiang, 2010; Liu et al., 2011). These approaches exclusively take into account syntactic structure and do not model rule selection (see Section 6 for a detailed discus- sion). Following the work on phrase-sense disam- biguation by (Carpuat and Wu, 2007), other au- thors improve rule selection by defining features on the structure of hierarchical rules and combin- ing these with …

Content truncated to 500 words