Unsupervised training of maximum-entropy models for lexical selection in rule-based machine translation

This article presents a method of training maximum-entropy models to perform lexical selection in a rule-based machine translation system. The training method described is unsupervised; that is, it does not require any annotated corpus. The method uses source-language monolingual corpora, the machine translation (MT) system in which the models are integrated, and a statistical target-language model. Using the MT system, the sentences in the sourcelanguage corpus are translated in all possible ways according to the different translation equivalents in the bilingual dictionary of the system. These translations are then scored on the target-language model and the scores are normalised to provide fractional counts for training source-language maximum-entropy lexical-selection models. We show that these models can perform equally well, or better, than using the target-language model directly for lexical selection, at a substantially reduced computational cost.