Dealing with bilingual divergences in MT using target language N-gram models

Abstract

In this paper we present a prototype translation system that uses only a sourcelanguage (SL) tagger, a bilingual dictionary and a lemmatised target-language (TL) corpus. In our approach, the TL corpus is innovatively exploited both for lexical selection (selecting among the different translations proposed by the dictionary) and for structure building of the output. To that end a series of n-gram model over lemmas and POS tags are built from the TL corpus, which are then searched at run-time. The system presented here uses Spanish as SL and English as TL but the architecture is language independent and translatable to languages with very little NLP development.