Using SVMs with the Command Relation features to identify negated events in biomedical literature

In this paper we explore the identification of negated molecular events (e.g. protein binding, gene expressions, regulation, etc.) in biomedical research abstracts. We construe the problem as a classification task and apply a machine learning (ML) approach that uses lexical, syntactic, and semantic features associated with sentences that represent events. Lexical features include negation cues, whereas syntactic features are engineered from constituency parse trees and the command relation between constituents. Semantic features include event type and participants. We also consider a rule-based approach that uses only the command relation. On a test dataset, the ML approach showed significantly better results (51% F-measure) compared to the commandbased rules (35-42% F-measure). Training a separate classifier for each event class proved to be useful, as the micro-averaged F-score improved to 63% (with 88% precision), demonstrating the potential of task-specific ML approaches to negation detection.