Classifying the Semantic Relations in Noun Compounds via a Domain-Specific Lexical Hierarchy

Classifying the Semantic Relations in Noun Compounds via a Domain-Specific Lexical Hierarchy Barbara Rosario and Marti Hearst School of Information Management & Systems University of California, Berkeley Berkeley, CA 94720-4600 {rosario,hearst}@sims.berkeley.edu Abstract We are developing corpus-based techniques for iden- tifying semantic relations at an intermediate level of description (more specific than those used in case frames, but more general than those used in tra- ditional knowledge representation systems). In this paper we describe a classification algorithm for iden- tifying relationships between two-word noun com- pounds. We find that a very simple approach using a machine learning algorithm and a domain-specific lexical hierarchy successfully generalizes from train- ing instances, performing better on previously un- seen words than a baseline consisting of training on the words themselves. 1 Introduction We are exploring empirical methods of determin- ing semantic relationships between constituents in natural language. Our current project focuses on biomedical text, both because it poses interesting challenges, and because it should be possible to make inferences about propositions that hold between sci- entific concepts within biomedical texts (Swanson and Smalheiser, 1994). One of the important challenges of biomedical text, along with most other technical text, is the proliferation of noun compounds. A typical article title is shown below; it consists a cascade of four noun phrases linked by prepositions: Open-labeled long-term study of the effi- cacy, safety, and tolerability of subcuta- neous sumatriptan in acute migraine treat- ment. The real concern in analyzing such a title is in de- termining the relationships that hold between differ- ent concepts, rather than on finding the appropriate attachments (which is especially difficult given the lack of a verb). And before we tackle the prepo- sitional phrase attachment problem, we must find a way to analyze the meanings of the noun com- pounds. Our goal is to extract propositional information from text, and as a step towards this goal, we clas- sify constituents according to which semantic rela- tionships hold between them. For example, we want to characterize the treatment-for-disease relation- ship between the words of migraine treatment ver- sus the method-of-treatment relationship between the words of aerosol treatment. These relations are intended to be combined to produce larger proposi- tions that can then be used in a variety of interpreta- tion paradigms, such as abductive reasoning (Hobbs et al., 1993) or inductive logic programming (Ng and Zelle, 1997). Note that because we are concerned with the se- mantic relations that hold between the concepts, as opposed to the more standard, syntax-driven com- putational goal of determining left versus right as- sociation, this has the fortuitous effect of changing the problem into one of classification, amenable to standard machine learning classification techniques. We have found that we can use such algorithms to classify relationships between two-word noun com- pounds with a surprising degree of accuracy. one-out-of-eighteen classification using a neural netA achieves accuracies as high as 62%. By taking ad- vantage of lexical ontologies, we achieve strong re- sults on noun compounds for which neither word …

Content truncated to 500 words