Better Learning and Decoding for Syntax Based SMT Using PSDIG

Better Learning and Decoding for Syntax Based SMT Using PSDIG Yuan Ding Martha Palmer Department of Computer and Information Science University of Pennsylvania Philadelphia, PA 19104, USA {yding, mpalmer}@cis.upenn.edu Abstract As an approach to syntax based statistical machine translation (SMT), Probabilistic Synchronous Dependency Insertion Grammars (PSDIG), introduced in (Ding and Palmer, 2005), are a version of syn- chronous grammars defined on depend- ency trees. In this paper we discuss better learning and decoding algorithms for a PSDIG MT system. We introduce two new grammar learners: (1) an exhaustive learner combining different heuristics, (2) an n-gram based grammar learner. Com- bining the grammar rules learned from the two learners improved the performance. We introduce a better decoding algorithm which incorporates a tri-gram language model. According to the Bleu metric, the PSDIG MT system performance is sig- nificantly better than IBM Model 4, while on par with the state-of-the-art phrase based system Pharaoh (Koehn, 2004). The improved integration of syntax on both source and target languages opens door to more sophisticated SMT processes. 1 Introduction Syntax-based statistical machine translation (SMT) aims at applying statistical models to structural data, has begun to emerge. With the advances in the broad-coverage parsers trained from treebanks, e.g. (Collins, 1999) and (McDonald et al., 2005), structural analysis of different languages has been made possible. Ideally, by combining the natural language syntax and machine learning methods, a broad-coverage and linguistically well-motivated statistical MT system can be constructed. However, because of structural divergences be- tween languages (Dorr, 1994)，due to either sys- tematic differences between languages or loose translations in real corpora, the syntax based MT systems have to transduce between non-isomorphic tree structures, which is a major challenge. (Wu, 1997) and (Alshawi et al., 2000) learn the tree representations directly from parallel sen- tences, and do not make allowances for non- isomorphic structures. (Yamada and Knight, 2001) used a sequence of tree operations transforming a syntactic tree into a string of the target language. (Quirck et al., 2005) used syntax structures to guide the phrase reorder process. When researchers try to use syntax trees (here we think of syntax as linguistic syntax, in contrast to formal syntax) in both languages, the problem of non-isomorphism must be addressed. In theory, stochastic tree transducers and some versions of synchronous grammars provide solutions for the non-isomorphic tree based transduction problem and hence possible solutions for MT. Synchronous Tree Adjoining Grammars was proposed by (Shie- ber and Schabes, 1990). Eisner (2003) proposed viewing the MT as synchronous tree substitution grammar parsing. Melamed (2003) formalized the MT as synchronous parsing based on multi-text grammars. Graehl and Knight (2004) defined train- ing and decoding algorithms for both generalized tree-to-tree and tree-to-string transducers. Lin (2004) proposed to base the MT process on parallel dependency paths. Ding and Palmer (2005) intro- duced the use of Probabilistic Synchronous De- pendency Insertion Grammars (PSDIG) to model machine translation. These approaches model the two languages using tree transduction rules or syn- chronous grammars, possibly with multi-lemma elementary structures as atomic …

Content truncated to 500 words