Search-based Structured Prediction applied to Biomedical Event Extraction

Proceedings of the Fifteenth Conference on Computational Natural Language Learning, pages 49–57, Portland, Oregon, USA, 23–24 June 2011. c⃝2011 Association for Computational Linguistics Search-based Structured Prediction applied to Biomedical Event Extraction Andreas Vlachos and Mark Craven Department of Biostatistics and Medical Informatics University of Wisconsin-Madison {vlachos,craven}@biostat.wisc.edu Abstract We develop an approach to biomedical event extraction using a search-based structured pre- diction framework, SEARN, which converts the task into cost-sensitive classification tasks whose models are learned jointly. We show that SEARN improves on a simple yet strong pipeline by 8.6 points in F-score on the BioNLP 2009 shared task, while achieving the best reported performance by a joint inference method. Additionally, we consider the issue of cost estimation during learning and present an approach called focused costing that improves improves efficiency and predictive accuracy. 1 Introduction The term biomedical event extraction is used to re- fer to the task of extracting descriptions of actions and relations involving one or more entities from the biomedical literature. The recent BioNLP 2009 shared task (BioNLP09ST) on event extraction (Kim et al., 2009) focused on event types of varying com- plexity. Each event consists of a trigger and one or more arguments, the latter being proteins or other events. Any token in a sentence can be a trigger for one of the nine event types and, depending on their associated event types, triggers are assigned appro- priate arguments. Thus, the task can be viewed as a structured prediction problem in which the output for a given instance is a (possibly disconnected) di- rected acyclic graph (not necessarily a tree) in which vertices correspond to triggers or protein arguments, and edges represent relations between them. Despite being a structured prediction task, most of the systems that have been applied to BioNLP09ST to date are pipelines that decompose event extrac- tion into a set of simpler classification tasks. Clas- sifiers for these tasks are typically learned indepen- dently, thereby ignoring event structure during train- ing. Typically in such systems, the relationships among these tasks are taken into account by incor- porating post-processing rules that enforce certain constraints when combining their predictions, and by tuning classification thresholds to improve the ac- curacy of joint predictions. Pipelines are appealing as they are relatively easy to implement and they of- ten achieve state-of-the-art performance (Bjorne et al., 2009; Miwa et al., 2010). Because of the nature of the output space, the task is not amenable to sequential or grammar-based ap- proaches (e.g. linear CRFs, HMMs, PCFGs) which employ dynamic programming in order to do ef- ficient inference. The only joint inference frame- work that has been applied to BioNLP09ST to date is Markov Logic Networks (MLNs) (Riedel et al., 2009; Poon and Vanderwende, 2010). However, MLNs require task-dependent approximate infer- ence and substantial computational resources in or- der to achieve state-of-the-art performance. In this work we explore an alternative joint in- ference approach to biomedical event extraction us- ing a search-based structured prediction framework, SEARN (Daum´e III et al., 2009). …

Content truncated to 500 words