Elastic-substitution decoding for Hierarchical SMT: efficiency, richer search and double labels

Elastic-substitution decoding for Hierarchical SMT: efficiency, richer search and double labels Gideon Maillette de Buy Wenniger1 gemdbw AT gmail.com Khalil Sima’an2 k.simaan AT uva.nl Andy Way1 andy.way AT adaptcentre.ie 1ADAPT Centre, School of Computing, Dublin City University, Dublin, Ireland 2Institute for Logic Language and Computation (ILLC), Faculty of Science, University of Amsterdam, Amsterdam, The Netherlands Abstract Elastic-substitution decoding (ESD), first introduced by Chiang (2010), can be important for obtaining good results when applying labels to enrich hierarchical statistical machine translation (SMT). However, an efficient implementation is essential for scalable application. We describe how to achieve this, contributing essential details that were missing in the original exposition. We compare ESD to strict matching and show its superiority for both reordering and syntactic labels. To overcome the sub-optimal performance due to the late evaluation of features marking label substitution types, we increase the diversity of the rules explored during cube pruning initialization with respect to labels their labels. This approach gives significant improvements over basic ESD and performs favorably compared to extending the search by increasing the cube pruning pop-limit. Finally, we look at combining multiple labels. The combination of reordering labels and target-side boundary-tags yields a significant improvement in terms of the word-order sensitive metrics Kendall reordering score and METEOR. This confirms our intuition that the combination of reordering labels and syntactic labels can yield improvements over either label by itself, despite increased sparsity. 1 Introduction Elastic-substitution decoding (ESD) – also known as soft label matching or soft-constraint decoding – is an effective method to gain maximal benefit from the use of labels to enrich hierarchical phrase-based statistical machine translation (SMT), and was first introduced by Chiang (2010). This method removes many of the disadvantages of working with labeled grammars when labels are strictly enforced. We discuss the requirements and details of an efficient implementation in the first part of this paper, to benefit other researchers that want to apply ESD. In the second part of the paper we further strengthen the empirical evidence for the success of ESD. This is done by comparing strict and soft-labeled (ESD) systems for Chinese– English translation, using four different types of labels. Next, we describe how the results of ESD can be further improved for small label sets by diversifying the search, exploring all alternatively labeled versions of each rule source-side type during cube pruning initialization instead of only the single best one. This is compared against the more crude approach of just increasing the search space by increasing the cube pruning pop-limit. Finally, we explore the effect of combining multiple labels, either the two types of reordering labels or a reordering Proceedings of MT Summit XVI, vol.1: Research Track Nagoya, Sep. 18-22, 2017 | p. 201 label with a syntactic label. All source code for both ESD and labeled grammar extraction is made publicly available with this publication.1 2 Background and Related Work Hierarchical phrase-based SMT (or hierarchical SMT for short) (Chiang, 2005) is the hierarchical generalization of phrase-based SMT (Koehn et al., …

Content truncated to 500 words