Reducing Semantic Drift in Biomedical Lexicon Bootstrapping S ID ERE · ME NS ·EAD

Lexical-semantic resources are crucial in many Natural Language Processing (NLP) tasks. These resources are difficult to create and maintain, and their bias, inconsistency and lack of coverage limits their usefulness. This issue is apparent in specialist domains with limited re-sources, such as biomedicine, where NLP is now being applied. Developing automated methods for extracting such resources is critical to overcoming this knowledge bottleneck. Biomedical information extraction will greatly improve when systems can exploit full-text articles and process a wider range of linguistic phenomena. I demonstrate the importance of these challenges by analysing the Molecular Interaction Map (MIM) corpus I have created. This corpus maps known interaction facts to passages within full-text articles, which are anno-tated with the phenomena that must be resolved for fact extraction. Unfortunately, the lexical-semantic resources needed for this are limited to specific sub-domains. Minimally supervised bootstrapping algorithms are commonly used to extract semantic lex-icons from text. These methods iteratively expand lexicons from a small number of seed terms. Existing bootstrappers are prone to semantic drift, where the original meaning associated wit