Preference Grammars and Decoding Algorithms for Probabilistic Synchronous Context Free Grammar Based Translation.

摘要

Probabilistic Synchronous Context-free Grammars (PSCFGs) [Aho and Ullmann, 1969, Wu, 1996] define weighted transduction rules to represent translation and reordering operations. When translation models use features that are defined locally, on each rule, there are efficient dynamic programming algorithms to perform translation with these grammars [Kasami, 1965]. In general, the integration of non-local features into the translation model can make translation NP-hard, requiring decoding approximations that limit the impact of these features.;In this thesis, we consider the impact and interaction between two non-local features, the n-gram language model (LM) and labels on rule nonterminal symbols in the Syntax-Augmented MT (SAMT) grammar [Zollmann and Venugopal, 2006]. While these features do not result in NP-hard search, they would lead to serious increases in wall-clock runtime if naive dynamic programming methods are applied.;We develop novel two-pass algorithms that make strong decoding approximations during a first pass search, generating a hypergraph of sentence spanning translation derivations. In a second pass, we use knowledge about non-local features to explore the hypergraph for alternative, potentially better translations. We use this approach to integrate the n-gram LM decoding feature as well as a non-local syntactic feature described below.;We then perform a systematic comparison of approaches to evaluate the relative impact of PSCFG methods over a strong phrase-based MT baseline with a focus on the impact of n-gram LM and syntactic labels. This comparison addresses important questions about the effectiveness of PSCFG methods for a variety of language and resource conditions. We learn that for language pairs that exhibit long distance reordering, PSCFG methods deliver improvements over comparable phrase-based systems and that SAMT labels result in additional small, but consistent improvements even in conjunction with strong n-gram LMs.;Finally, we propose a novel approach to use nonterminal labels in PSCFG decoding by extending the PSCFG formalism to represent hard label constraints as soft preferences. These preferences are used to compute a new decoding feature that reflects the probability that a derivation is syntactically well formed. This feature mitigates the effect of the commonly applied maximum a posteriori (MAP) approximation and can be discriminatively trained in concert with other model features. We report modest improvements in translation quality on a Chinese-to-English translation task.