Abstract

In the biomedical information extraction community, the focus has largely been on binary protein-protein (PPIs) interactions, for a long time. Quite recently, text analytics have been developed dealing with bio-event extraction. In essence, this means that the general PPI problem is broken down into much more specific subtasks. A bio-event is then defined as a change of the biological state, the properties, or the location of a bio-molecule. The most recent BioNLP 2009 Shared Task on Event Extraction required to determine, from a sample of Medline abstracts, all mentioned events of nine bio-molecular interaction types, including, e.g., Binding, Transcription and Regulation events. Interestingly, the three top-performing systems of that competition all rely on dependency graphs for solving this event extraction task. However, while the UTurku and ConcordU systems exploit Stanford grammatical relations, the JulieLab system (U Jena) uses CoNLL dependencies. Obviously, the question turns up as to what extent the performance of these systems depends on proper choices of the available parsers and dependency output representations.

I will discuss the Shared Task solution of the JulieLab system which relies on ‘trimming’ dependency graphs by syntactic simplification and semantic enrichment operations. Furthermore, the role of different representation formats of dependency graphs in the event extraction task (basically, Stanford vs. CoNLL encodings) will be considered. Given JulieLab’s high-performance event extraction system and taking considerations of its methodological underpinnings into account, we currently expand our approach into more real-life information extraction scenarios, e.g., the support of human database curators.