Robust biomedical event extraction with dual decomposition and minimal domain adaptation

We present a joint model for biomedical event extraction and apply it to four tracks of the BioNLP 2011 Shared Task. Our model de-composes into three sub-models that concern (a) event triggers and outgoing arguments, (b) event triggers and incoming arguments and (c) protein-protein bindings. For efficient de-coding we employ dual decomposition. Our results are very competitive: With minimal adaptation of our model we come in second for two of the tasks—right behind a version of the system presented here that includes pre-dictions of the Stanford event extractor as fea-tures. We also show that for the Infectious Diseases task using data from the Genia track is a very effective way to improve accuracy.