Fast and Robust Joint Models for Biomedical Event Extraction

Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 1–12, Edinburgh, Scotland, UK, July 27–31, 2011. c⃝2011 Association for Computational Linguistics Fast and Robust Joint Models for Biomedical Event Extraction Sebastian Riedel Andrew McCallum Department of Computer Science University of Massachusetts, Amherst {riedel,mccallum}@cs.umass.edu Abstract Extracting biomedical events from literature has attracted much recent attention. The best- performing systems so far have been pipelines of simple subtask-specific local classifiers. A natural drawback of such approaches are cas- cading errors introduced in early stages of the pipeline. We present three joint models of increasing complexity designed to overcome this problem. The first model performs joint trigger and argument extraction, and lends it- self to a simple, efficient and exact infer- ence algorithm. The second model captures correlations between events, while the third model ensures consistency between arguments of the same event. Inference in these models is kept tractable through dual decomposition. The first two models outperform the previous best joint approaches and are very competi- tive with respect to the current state-of-the- art. The third model yields the best results re- ported so far on the BioNLP 2009 shared task, the BioNLP 2011 Genia task and the BioNLP 2011 Infectious Diseases task. 1 Introduction Whenever we advance our scientific understanding of the world, we seek to publish our findings. The result is a vast and ever-expanding body of natural language text that is becoming increasingly difficult to leverage. This is particularly true in the context of life sciences, where large quantities of biomedi- cal articles are published on a daily basis. To sup- port tasks such data mining, search and visualiza- tion, there is a clear need for structured representa- tions of the knowledge these articles convey. This is indicated by a large number of public databases with content ranging from simple protein-protein interac- tions to complex pathways. To increase coverage of such databases, and to keep up with the rate of pub- lishing, we need to automatically extract structured representations from biomedical text—a process of- ten referred to as biomedical text mining. One major focus of biomedical text mining has been the extraction of named entities, such genes or gene products, and of flat binary relations be- tween such entities, such as protein-protein interac- tions. However, in recent years there has also been an increasing interest in the extraction of biomedi- cal events and their causal relations. This gave rise to the BioNLP 2009 and 2011 shared tasks which challenged participants to gather such events from biomedical text (Kim et al., 2009; Kim et al., 2011). Notably, these events can be complex and recursive: they may have several arguments, and some of the arguments may be events themselves. Current state-of-the-art event extractors fol- low the same architectural blueprint and divide the extraction process into a pipeline of three stages (Björne et al., 2009; Miwa et al., 2010c). First they predict a set of candidate event trigger words (say, tokens 2, 5 and 6 in figure 1), then …

Content truncated to 500 words