Statistical Anaphora Resolution in Biomedical Texts

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 257–264 Manchester, August 2008 Statistical anaphora resolution in biomedical texts Caroline Gasperin Ted Briscoe Computer Laboratory University of Cambridge Cambridge, UK {cvg20,ejb}@cl.cam.ac.uk Abstract This paper presents a probabilistic model for resolution of non-pronominal anaphora in biomedical texts. The model seeks to find the antecedents of anaphoric expres- sions, both coreferent and associative ones, and also to identify discourse-new expres- sions. We consider only the noun phrases referring to biomedical entities. The model reaches state-of-the art performance: 56- 69% precision and 54-67% recall on coref- erent cases, and reasonable performance on different classes of associative cases. 1 Introduction Inspired by Ge et al. (1998) probabilistic model for pronoun resolution, we have developed a model for resolution of non-pronominal anaphora in biomed- ical texts. The probabilistic model results from a simple decomposition process applied to a conditional probability equation that involves several param- eters (features). The decomposition makes use of Bayes’ theorem and independence assumptions, and aims to decrease the impact of data sparse- ness on the model, so that even small training cor- pora can be viable. The decomposed model can be understood as a more sophisticated version of the naive-Bayes algorithm, since we consider the dependence among some of the features instead of full independence as in naive Bayes. Proba- bilistic models can return a confidence measure (probability) for each decision they make, while c⃝ 2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported li- cense (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. decision trees, for example, cannot. Another ad- vantage of this type of model is the fact that they consider the prior probability of each class, while other machine-learning techniques such as SVMs and neural networks do not. Our model seeks to classify the relation between an anaphoric expression and an antecedent candi- date as coreferent, associative or neither. It com- putes the probability of each pair of anaphor and candidate for each class. The candidate with the highest overall probability for each class is se- lected as the antecedent for that class, or no an- tecedent is selected if the probability of no relation overcomes the positive probabilities; in this case, the expression is considered to be new to the dis- course. Coreferent cases are the ones in which the anaphoric expression and its antecedent refer to the same entity in the real world, as below: (1) “The expression of reaper has been shown ... the gene encodes ...” Associative cases are the ones in which the anaphoric expression and its antecedent refer to different entities, but where the antecedent plays a role in defining the meaning of the anaphoric ex- pression, as in Example 2: (2) “Drosophila gene Bok interacts with ... expression of Bok protein promotes apoptosis ...” Discourse new cases usually consist of the first mention of an entity in the text, so no antecedent can be found for it. We have focused on the biomedical domain for two reasons. Firstly, there is a …

Content truncated to 500 words