Improving Syntax-Augmented Machine Translation by Coarsening the Label Set

Proceedings of NAACL-HLT 2013, pages 288–297, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Improving Syntax-Augmented Machine Translation by Coarsening the Label Set Greg Hanneman and Alon Lavie Language Technologies Institute Carnegie Mellon University Pittsburgh, PA 15213 USA {ghannema, alavie}@cs.cmu.edu Abstract We present a new variant of the Syntax- Augmented Machine Translation (SAMT) for- malism with a category-coarsening algorithm originally developed for tree-to-tree gram- mars. We induce bilingual labels into the SAMT grammar, use them for category coars- ening, then project back to monolingual la- beling as in standard SAMT. The result is a “collapsed” grammar with the same expres- sive power and format as the original, but many fewer nonterminal labels. We show that the smaller label set provides improved trans- lation scores by 1.14 BLEU on two Chinese– English test sets while reducing the occur- rence of sparsity and ambiguity problems common to large label sets. 1 Introduction The formulation of statistical machine translation in terms of synchronous parsing has become both the- oretically and practically successful. In a parsing- based MT formalism, synchronous context-free grammar rules that match a source-language input can be hierarchically composed to produce a corre- sponding target-language output. SCFG translation grammars can be extracted automatically from data. While formally syntactic approaches with a single grammar nonterminal have often worked well (Chi- ang, 2007), the desire to exploit linguistic knowl- edge has motivated the use of translation grammars with richer, linguistically syntactic nonterminal in- ventories (Galley et al., 2004; Liu et al., 2006; Lavie et al., 2008; Liu et al., 2009). Linguistically syntactic MT systems can derive their label sets, either monolingually or bilingually, from parallel corpora that have been annotated with source- and/or target-side parse trees provided by a statistical parser. The MT system may exactly adopt the parser’s label set or modify it in some way. Larger label sets are able to represent more precise, fine-grained categories. On the other hand, they also exacerbate a number of computational and modeling problems by increasing grammar size, derivational ambiguity, and data sparsity. In this paper, we focus on the Syntax-Augmented MT formalism (Zollmann and Venugopal, 2006), a monolingually labeled version of Hiero that can cre- ate up to 4000 “extended” category labels based on pairs of parse nodes. We take a standard SAMT grammar with target-side labels and extend its label- ing to a bilingual format (Zollmann, 2011). We then coarsen the bilingual labels following the “label col- lapsing” algorithm of Hanneman and Lavie (2011). This represents a novel extension of the tree-to-tree collapsing algorithm to the SAMT formalism. Af- ter removing the source-side labels, we obtain a new SAMT grammar with coarser target-side labels than the original. Coarsened grammars provide improvement of up to 1.14 BLEU points over the baseline SAMT results on two Chinese–English test sets; they also outper- form a Hiero baseline by up to 0.60 BLEU on one of the sets. Aside from improved translation quality, in analysis we find significant reductions in deriva- tional ambiguity and rule …

Content truncated to 500 words