MWU-aware Part-of-Speech Tagging with a CRF model and lexical resources

Abstract : This paper describes a new part-of-speech tagger including multiword unit (MWU) identification. It is based on a Conditional Random Field model integrating language-independent features, as well as features computed from external lexical resources. It was implemented in a finite-state framework composed of a preliminary finite-state lexical analysis and a CRF decoding using weighted finite-state transducer composition. We showed that our tagger reaches state-of-the-art results for French in the standard evaluation conditions (i.e. each multiword unit is already merged in a single token). The evaluation of the tagger integrating MWU recognition clearly shows the interest of incorporating features based on MWU resources.