Inferring Implicit Causal Relationships in Biomedical Literature

Biomedical relations are often expressed between entities occurring within the same sentence through syntactic means. However, a significant portion of such relations (in particular, causal relations) are expressed implicitly across sentence boundaries. Inferring these discourse-level relations can be challenging in the absence of syntactic clues. In this paper, we present a study of textual characteristics that contribute to expression of implicit causal relations across sentence boundaries. Focusing on a chemical-disease relationship corpus, we identify and investigate the contribution of various features that can assist in identifying such inter-sentential relations. Using these features for supervised learning, we were able to improve previously reported best results by more than 13%. Our results demonstrate the usefulness of the proposed features and the importance of using a balanced dataset for this task.