Structured local exponential models for machine translation

Michael Vladimir Subotin

This thesis proposes a synthesis and generalization of local exponential translation models, the subclass of feature-rich translation models which associate probability distributions with individual rewrite rules used by the translation system, such as synchronous context-free rules, or with other individual aspects of translation hypotheses such as word pairs or reordering events. Unlike other authors we use these estimates to replace the traditional phrase models and lexical scores, rather than in addition to them, thereby demonstrating that the local exponential phrase models can be regarded as a generalization of standard methods not only in theoretical but also in practical terms. We further introduce a form of local translation models that combine features associated with surface forms of rules and features associated with less specific representation -- including those based on lemmas, inflections, and reordering patterns -- such that surface-form estimates are recovered as a special case of the model. Crucially, the proposed approach allows estimation of parameters for the latter type of features from training sets that include multiple source phrases, thereby overcoming an important training set fragmentation problem which hampers previously proposed local translation models. These proposals are experimentally validated. Conditioning all phrase-based probabilities in a hierarchical phrase-based system on source-side contextual information produces significant performance improvements. Extending the contextually-sensitive estimates with features modeling source-side morphology and reordering patterns yields consistent additional improvements, while further experiments show significant improvements obtained from modeling observed and unobserved inflections for a morphologically rich target language.