Flexible and Efficient Hypergraph Interactions for Joint Hierarchical and Forest-to-String Decoding

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 545–555, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics Flexible and Efficient Hypergraph Interactions for Joint Hierarchical and Forest-to-String Decoding∗ Martin ˇCmejrek†‡ †IBM Prague Research Lab V Parku 2294/4 Prague, Czech Republic, 148 00 martin.cmejrek@us.ibm.com Haitao Mi‡ and Bowen Zhou‡ ‡IBM T. J. Watson Research Center 1101 Kitchawan Rd Yorktown Heights, NY 10598 {hmi,zhou}@us.ibm.com Abstract Machine translation benefits from system combination. We propose flexible interaction of hypergraphs as a novel technique combin- ing different translation models within one de- coder. We introduce features controlling the interactions between the two systems and ex- plore three interaction schemes of hiero and forest-to-string models—specification, gener- alization, and interchange. The experiments are carried out on large training data with strong baselines utilizing rich sets of dense and sparse features. All three schemes signif- icantly improve results of any single system on four testsets. We find that specification—a more constrained scheme that almost entirely uses forest-to-string rules, but optionally uses hiero rules for shorter spans—comes out as the strongest, yielding improvement up to 0.9 (Ter-Bleu)/2 points. We also provide a de- tailed experimental and qualitative analysis of the results. 1 Introduction Recent years have witnessed the success of var- ious statistical machine translation (SMT) mod- els using different levels of linguistic knowledge– phrase (Koehn et al., 2003), hiero (Chiang, 2005), and syntax-based (Liu et al., 2006; Galley et al., 2006). System combination became a promising way of building up synergy from different SMT sys- tems and their specific merits. Numerous efforts that have been proposed in this field recently can be broadly divided into two cat- ∗M. ˇC and H. M. contributed equally to this work. egories: Offline system combination (Rosti et al., 2007; He et al., 2008; Watanabe and Sumita, 2011; Denero et al., 2010) aims at producing consensus translations from the outputs of multiple individ- ual systems. Those outputs usually contain k-best lists of translations, which only explore a small por- tion of the entire search space of each system. This issue is well addressed in joint decoding (Liu et al., 2009), or online system combination, showing comparable improvements to the offline combina- tion methods. Rather than finding consensus trans- lations from the outputs of individual systems, joint decoding works with different grammars at the de- coding time. Although limited to individual systems sharing the same search paradigm (e.g. left-to-right or bottom-up), joint decoding offers many poten- tial advatages: search through a larger space, bet- ter efficiency, features designed once for all subsys- tems, potential cross-system features, online sharing of partial hypotheses, and many others. Different approaches have different strengths in general–hiero rules are believed to provide reliable lexical coverage, while tree-to-string rules are good at non-local reorderings. Different contexts present different challenges–noun phrases usually follow the adjacency principle, while verb phrases require more challenging reorderings. In this work, we study different schemes of interaction between translation models, reflecting their specific strengths at differ- ent (syntactic) contexts. …

Content truncated to 500 words