Sieve-based Coreference Resolution in the Biomedical Domain

An Investigation of Coreference Phenomena in the Biomedical Domain Dane Bell, Gus Hahn-Powell, Marco A. Valenzuela-Esc´arcega, Mihai Surdeanu University of Arizona Tucson, AZ 85721, USA {dane, hahnpowell, marcov, msurdeanu}@email.arizona.edu Abstract We describe challenges and advantages unique to coreference resolution in the biomedical domain, and a sieve-based architecture that leverages domain knowledge for both entity and event coreference resolution. Domain-general coreference resolution algorithms perform poorly on biomedical documents, because the cues they rely on such as gender are largely absent in this domain, and because they do not encode domain-specific knowledge such as the number and type of participants required in chemical reactions. Moreover, it is difficult to directly encode this knowledge into most coreference resolution algorithms because they are not rule-based. Our rule-based architecture uses sequentially applied hand-designed “sieves”, with the output of each sieve informing and constraining subsequent sieves. This architecture provides a 3.2% increase in throughput to our Reach event extraction system with precision parallel to that of the stricter system that relies solely on syntactic patterns for extraction. Keywords: coreference resolution, information extraction, biomedical text mining 1. Introduction Cancer is triggered by extremely complex protein signaling pathways that interact in an intricate network, leading to accordingly complex reactions to cancer treatment drugs. This complexity, combined with the rapid pace of publish- ing in the biomedical domain, makes these systems effec- tively intractable for human experts, and so studies are typ- ically reductionist, focusing on a single pathway at a time. The Defense Advanced Research Projects Agency (DARPA) created the Big Mechanism program (Cohen, 2015) to develop a holistic view of these cancer pathways through automated, large-scale reading and pathway as- sembly in amendment of and extension to already-existing human-made models of protein reaction networks. Two major requirements to accomplish this ambitious goal are effective automatic reading and assembly, i.e., extracting individual biochemical interactions, linking these isolated interactions together to form larger pathways, and merging them into an existing cancer model. Coreference resolution is a fundamental requirement for effective reading and assembly, as it is crucial for decid- ing which spans of text refer to the same entity in the real world, which can have such disparate labels as “the pro- tein”, “both”, or “ASPP1”. However, with the notable ex- ception of the BioNLP shared task (Kim et al., 2013), coref- erence has rarely been a focus of reading-based work. Domain-general coreference systems perform poorly in this domain, because they fail to capitalize on domain-specific constraints on possible coreference relations. This is illus- trated in example 1a, in which a domain-general corefer- ence resolution system (Lee et al., 2013) would link its to GSK3β because of a generally trustworthy heuristic that the earliest named entity in the sentence is likely to be the an- tecedent of a pronoun if they match grammatically (Hobbs, 1978). The domain-specific knowledge that a protein bind- ing to itself would not be referred to this way allows us to rule this link out and instead correctly choose Axin GBD, as in 1b. (1) …

Content truncated to 500 words