Classifying Relations for Biomedical Named Entity Disambiguation

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 1513–1522, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Classifying Relations for Biomedical Named Entity Disambiguation Xinglong Wang†‡ Jun’ichi Tsujii∗†‡ Sophia Ananiadou†‡ †School of Computer Science, University of Manchester, UK ‡National Centre for Text Mining, UK ∗Department of Computer Science, University of Tokyo, Japan {xinglong.wang,j.tsujii,sophia.ananiadou}@manchester.ac.uk Abstract Named entity disambiguation concerns linking a potentially ambiguous mention of named entity in text to an unambigu- ous identifier in a standard database. One approach to this task is supervised classifi- cation. However, the availability of train- ing data is often limited, and the avail- able data sets tend to be imbalanced and, in some cases, heterogeneous. We pro- pose a new method that distinguishes a named entity by finding the informative keywords in its surrounding context, and then trains a model to predict whether each keyword indicates the semantic class of the entity. While maintaining a compara- ble performance to supervised classifica- tion, this method avoids using expensive manually annotated data for each new do- main, and thus achieves better portability. 1 Introduction While technology on named entity recognition (NER) matures, many researchers in the field of information extraction (IE) gradually shifted their focus to more complex tasks such as named en- tity disambiguation and relation extraction. Both tasks are particularly important for biomedical text mining, which concerns automatically extracting facts from the exponentially growing biomedical literature (Hunter and Cohen, 2006). One type of facts is relations between biomedical named en- tities, such as disease-drug relation, gene-disease relation, protein-protein interaction (PPI), etc. To automatically extract these facts, advanced natu- ral language processing techniques such as parsing have been adopted to analyse the syntactic and se- mantic structure of text. The idea is that linguistic structures between the interacting biological enti- ties may have common characteristics that can be exploited by similarity measures or machine learn- ing algorithms. For example, Erkan et al. (2007) used the shortest path between two genes accord- ing to edit distance in a dependency tree to de- fine a kernel function for extracting gene interac- tions. Miwa et al. (2008) comparably evaluated a number of kernels for incorporating syntactic fea- tures, including the bag-of-word kernel, the subset tree kernel (Moschitti, 2006) and the graph ker- nel (Airola et al., 2008), and they concluded that combining all kernels achieved better results than using any individual one. Miyao et al. (2008) used syntactic paths as one of the features to train a support vector machines (SVM) model for PPIs and also discussed how different parsers and out- put representations affected the end results. Another crucial IE task is named entity disam- biguation, which concerns grounding mentions of named entities in text to unambiguous concepts as defined in some standard dictionary or database. For instance, given a search term Python, users may like to see the results grouped into the fol- lowing categories: a type of snake, a programming language, or a film (Bunescu and Pas¸ca, 2006). One approach …

Content truncated to 500 words