Classifying Semantic Relations in Bioscience Texts

Classifying Semantic Relations in Bioscience Texts Barbara Rosario SIMS UC Berkeley Berkeley, CA 94720 rosario@sims.berkeley.edu Marti A. Hearst SIMS UC Berkeley Berkeley, CA 94720 hearst@sims.berkeley.edu Abstract A crucial step toward the goal of au- tomatic extraction of propositional in- formation from natural language text is the identification of semantic relations between constituents in sentences. We examine the problem of distinguishing among seven relation types that can oc- cur between the entities “treatment” and “disease” in bioscience text, and the problem of identifying such entities. We compare five generative graphical mod- els and a neural network, using lexical, syntactic, and semantic features, finding that the latter help achieve high classifi- cation accuracy. 1 Introduction The biosciences literature is rich, complex and continually growing. The National Library of Medicine’s MEDLINE database1 contains bibli- ographic citations and abstracts from more than 4,600 biomedical journals, and an estimated half a million new articles are added every year. Much of the important, late-breaking bioscience infor- mation is found only in textual form, and so meth- ods are needed to automatically extract semantic entities and the relations between them from this text. For example, in the following sentences, hep- atitis and its variants, which are DISEASES, are found in different semantic relationships with var- ious TREATMENTs: 1http://www.nlm.nih.gov/pubs/factsheets/medline.html (1) Effect of interferon on hepatitis B (2) A two-dose combined hepatitis A and B vac- cine would facilitate immunization programs (3) These results suggest that con A-induced hep- atitis was ameliorated by pretreatment with TJ-135. In (1) there is an unspecified effect of the treat- ment interferon on hepatitis B. In (2) the vaccine prevents hepatitis A and B while in (3) hepatitis is cured by the treatment TJ-135. We refer to this problem as Relation Classifi- cation. A related task is Role Extraction (also called, in the literature, “information extraction” or “named entity recognition”), defined as: given a sentence such as “The fluoroquinolones for uri- nary tract infections: a review”, extract all and only the strings of text that correspond to the roles TREATMENT (fluoroquinolones) and DISEASE (urinary tract infections). To make inferences about the facts in the text we need a system that accomplishes both these tasks: the extraction of the semantic roles and the recognition of the rela- tionship that holds between them. In this paper we compare five generative graph- ical models and a discriminative model (a multi- layer neural network) on these tasks. Recogniz- ing subtle differences among relations is a diffi- cult task; nevertheless the results achieved by our models are quite promising: when the roles are not given, the neural network achieves 79.6% accu- racy and the best graphical model achieves 74.9%. When the roles are given, the neural net reaches 96.9% accuracy while the best graphical model gets 91.6% accuracy. Part of the reason for the Relationship Definition and Example Cure TREAT cures DIS 810 (648, 162) Intravenous immune globulin for recurrent spontaneous abortion Only DIS TREAT not mentioned 616 (492, 124) Social ties and susceptibility to the common cold …

Content truncated to 500 words