An Evaluation and Possible Improvement Path for Current SMT Behavior on Ambiguous Nouns

Mistranslation of an ambiguous word can have a large impact on the understandability of a given sentence. In this article, we describe a thorough evaluation of the translation quality of ambiguous nouns in three different setups. We compared two statistical Machine Translation systems and one dedicated Word Sense Disambiguation (WSD) system. Our WSD system incorporates multilingual information and is independent from external lexical resources. Word senses are derived automatically from word alignments on a parallel corpus. We show that the two WSD classifiers that were built for these experiments (English– French and English–Dutch) outperform the SMT system that was trained on the same corpus. This opens perspectives for the integration of our multilingual WSD module in a statistical Machine Translation framework, in order to improve the automated translation of ambiguous words, and by consequence make the translation output more understandable.