University of Massachusetts, Amherst

Biomedical Event Extraction, Joint Inference and Dual Decompostion

The cell is the core building block of life, and the subject of a large and ever-growing body of research publications. For life scientists it is hence becoming increasingly difficult to keep track of all information relevant to the cell processes of their interest. This in turn reduces the pace of progress in this field. In this work we show how information about cell processes, or so called biomedical events, can be automatically extracted from literature. While thistask has gathered much recent attention, most work has either used a pipeline of classifiers that is prone to cascading errors, or joint models for which inference is slow and which so far have failed to yield competitive results.

We present novel joint models of biomedical event extraction that address the cascading error problem and are very efficient. This is achieved through framing event extraction as a global optimization problem, and solving this problem through dual decomposition. This technique allows us to decompose the optimization problem into several tractable subproblems for which fast optimization sub-routines can be designed. Our proposed method achieves the best results on the current benchmark datasets for the task. It is also the basis of a joint UMass-Stanford entry to the 2011 Biomedical Event Extraction Shared Task. This entry ranked 1st in 3 of the 4 tasks it was submitted to.