Integrating Phrase-based Reordering Features into a Chart-based Decoder for Machine Translation

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 1587–1596, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics Integrating Phrase-based Reordering Features into a Chart-based Decoder for Machine Translation ThuyLinh Nguyen Language Technologies Institute Carnegie Mellon University Pittsburgh, PA 15213, USA thuylinh@cs.cmu.edu Stephan Vogel Qatar Computing Research Institute Tornado Tower Doha, Qatar svogel@qf.org.qa Abstract Hiero translation models have two lim- itations compared to phrase-based mod- els: 1) Limited hypothesis space; 2) No lexicalized reordering model. We pro- pose an extension of Hiero called Phrasal- Hiero to address Hiero’s second problem. Phrasal-Hiero still has the same hypoth- esis space as the original Hiero but in- corporates a phrase-based distance cost feature and lexicalized reodering features into the chart decoder. The work consists of two parts: 1) for each Hiero transla- tion derivation, find its corresponding dis- continuous phrase-based path. 2) Extend the chart decoder to incorporate features from the phrase-based path. We achieve significant improvement over both Hiero and phrase-based baselines for Arabic- English, Chinese-English and German- English translation. 1 Introduction Phrase-based and tree-based translation model are the two main streams in state-of-the-art machine translation. The tree-based translation model, by using a synchronous context-free grammar for- malism, can capture longer reordering between source and target language. Yet, tree-based trans- lation often underperforms phrase-based transla- tion in language pairs with short range reordering such as Arabic-English translation (Zollmann et al., 2008; Birch et al., 2009). We follow Koehn et al. (2003) for our phrase- based system and Chiang (2005) for our Hiero sys- tem. In both systems, the translation of a source sentence f is the target sentence e∗ that maximizes a linear combination of features and weights: ⟨e∗, a∗⟩ = argmax ⟨e,a⟩∈H(f) � m∈M λmhm (e, f, a) . (1) where • a is a translation path of f. In the phrase- based system, aph represents a segmentation of e and f and a correspondance of phrases. In the Hiero system, atr is a derivation of a parallel parse tree of f and e, each nontermi- nal representing a rule in the derivation. • H (f) is the hypothesis space of the sentence f. We denote Hph (f) as the phrase-based hypothesis space of f and Htr (f) as its tree- based hypothesis space. Galley and Manning (2010) point out that due to the hard con- straints of rule combination, the tree-based system does not have the same excessive hy- pothesis space as the phrase-based system. • M is the set of feature indexes used in the decoder. Many features are shared between phrase-based and tree-based systems includ- ing language model, word count, and trans- lation model features. Phrase-based systems often use a lexical reordering model in addi- tion to the distance cost feature. The biggest difference in a Hiero system and a phrase-based system is in how the reordering is modeled. In the Hiero system, the reordering de- cision is encoded in weighted translation rules, de- termined by nonterminal mappings. For example, the rule …

Content truncated to 500 words