From Graphs to Events: A Subgraph Matching Approach for Information Extraction from Biomedical Text

Abstract

We participated in the BioNLP Shared Task 2011, addressing the GENIA event extraction (GE) and the Epigenetics and Post-translational Modifications (EPI) tasks. A graph-based approach is employed to automatically learn rules for detecting biological events in the life-science literature. The event rules are learned by identifying the key contextual dependencies from full syntactic parsing of annotated text. Event recognition is performed by searching for an isomorphism between event rules and the dependency graphs of sentences in the input texts. While we explored methods such as performance-based rule ranking to improve precision, we merged rules across multiple event types in order to increase recall. We achieved a 41.13% F-score in detecting events of nine types in the Task 1 of the GE task, and a 52.67% F-score in identifying events across fifteen types in the core task of the EPI task. Our performance on both tasks is comparable to the state-of-the-art systems. Our approach does not require any external domain-specific resources. The consistent performance on the two tasks supports the claim that the method generalizes well to extract events from different domains where training data is available.