Learning the Scope of Negation in Biomedical Texts

Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 715–724, Honolulu, October 2008. c⃝2008 Association for Computational Linguistics Learning the Scope of Negation in Biomedical Texts Roser Morante†, Anthony Liekens‡, Walter Daelemans† CNTS - Language Technology Group†, Applied Molecular Genomics Group‡ University of Antwerp Prinsstraat 13, B-2000 Antwerpen, Belgium {Roser.Morante,Anthony.Liekens,Walter.Daelemans}@ua.ac.be Abstract In this paper we present a machine learning system that finds the scope of negation in biomedical texts. The system consists of two memory-based engines, one that decides if the tokens in a sentence are negation signals, and another that finds the full scope of these nega- tion signals. Our approach to negation detec- tion differs in two main aspects from existing research on negation. First, we focus on find- ing the scope of negation signals, instead of determining whether a term is negated or not. Second, we apply supervised machine learn- ing techniques, whereas most existing systems apply rule-based algorithms. As far as we know, this way of approaching the negation scope finding task is novel. 1 Introduction In this paper we present a machine learning system that finds the scope of negation in biomedical texts. The system consists of two classifiers, one that de- cides if the tokens in a sentence are negation sig- nals (i.e., words indicating negation), and another that finds the full scope of these negation signals. Finding the scope of a negation signal means deter- mining at sentence level which words in the sentence are affected by the negation. Our approach differs in two main aspects from existing research. First, we focus on finding the scope of negation signals, in- stead of determining whether a term is negated or not. Second, we apply supervised machine learn- ing techniques, whereas most existing systems apply rule-based algorithms. Predicting the scope of negation is important in information extraction from text for obvious rea- sons; instead of simply flagging the sentences con- taining negation as not suited for extraction (which is currently the best that can be done), correct se- mantic relations can be extracted when the scope of negation is known, providing a better recall. Not being able to recognize negation can also hinder automated indexing systems (Mutalik et al., 2001; Rokach et al., 2008). As Mutalik et al. (2001) put it, “to increase the utility of concept indexing of medical documents, it is necessary to record whether the concept has been negated or not”. They highlight the need to detect negations in examples like “no ev- idence of fracture”, so that an information retrieval system does not return irrelevant reports. Szarvas et al. (2008) report that 13.45% of the sentences in the abstracts section of the BioScope corpus and 13.76% of the sentences in the full papers section contain negations. A system that does not deal with negation would treat these cases as false positives. The goals of this research are to model the scope finding task as a classification task similar to the se- mantic role labeling task, and to …

Content truncated to 500 words