Extracting Biomedical Events and Modifications Using Subgraph Matching with Noisy Training Data

Abstract

The Genia Event (GE) extraction task of the BioNLP Shared Task addresses the extraction of biomedical events from the natural language text of the published literature. In our submission, we modified an existing system for learning of event patterns via dependency parse subgraphs to utilise a more accurate parser and significantly more, but noisier, training data. We explore the impact of these two aspects of the system and conclude that the change in parser limits recall to an extent that cannot be offset by the large quantities of training data. However, our extensions of the system to extract modification events shows promise.