A hybrid approach to recognising discourse causality in the biomedical domain

Whilst current domain-specific information extraction systems represent an important resource for biomedical researchers, the increasing amount of knowledge published daily is still overwhelming them. As such, automatic discourse causality recognition can further improve the search for relevant information by suggesting possible causal connections. We describe here an approach to the automatic recognition of discourse causality in the biomedical domain using a combination of machine learning and rules. We test and evaluate our system on BioCause, a corpus containing gold standard annotations of causal relations. The best performance in identifying triggers is achieved by CRFs with 79.35% F-score. We then locate the arguments using naïve syntactic rules, achieving F-scores of around 90% in most cases. Determining which argument plays which role is performed by a group of machine learners with an F-score of 84.35%.