Resolution of Coordination Ellipses in Biological Named Entities Using Conditional Random Fields

Coordination ellipsis is a linguistic phenomenon where relevant information is eliminated from the linguistic surface expression of a coordination. Thereby, the assumption is that readers can easily reconstruct the eliminated material given appropriate background knowledge. This phenomenon is particularly abundant in science jargon. As far as named entity recognition (NER) is concerned, failing resolution of coordination ellipsis leads to serious degradation in the performance of NER systems. Usually, they recognize elliptical coordinations, wrongly, as complex single entity mentions or classify correctly only non-elliptical parts of coordinations. As an alternative, we propose a methodology for decomposing complex coordinated entity expressions into constituent conjuncts, for determining the missing elements and thus reconstructing explicitly all of the single entity mentions. We present here a novel supervised machine learning-based approach to the resolution of elliptical coordinations in noun phrases. For the task of conjunct identification the model achieves performance of 93% on the biomedical GENIA corpus.