Using shallow linguistic features for relation extraction in bio-medical texts

National audienceIn this paper2 , we model the corpus-based relation extraction task as a classification problem. We show that, in this framework, standard machine learning systems exploiting representations simply based on shallow linguistic information can rival state-of-the-art systems that rely on deep linguistic analysis. Even more effective systems can be obtained, still using these easy and reliable pieces of information, if the specifics of the extraction task and the data are taken into account. Our original method combining lazy learning and language modeling out-performs the existing systems when evaluated on the LLL2005 protein-protein interaction extraction task data