UC3M: A kernel-based approach to identify and classify DDIs in bio-medical texts.

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 617–621, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics UC3M: A kernel-based approach to identify and classify DDIs in biomedical texts. Daniel Sanchez-Cisneros Universidad Carlos III de Madrid Avda. de la Universidad, 30 28911 Leganés - Madrid - Spain dscisner@inf.uc3m.es Abstract The domain of DDI identification is constantly showing a rise of interest from scientific community since it represents a decrease of time and healthcare cost. In this paper we purpose a new approach based on shallow linguistic kernel methods to identify DDIs in biomedical manuscripts. The approach outlines a first step in the usage of semantic information for DDI identification. The system obtained an F1 measure of 0.534. 1 Introduction In recent years a new discipline appeared in the biomedical domain for processing pharmacological manuscripts related to drug substances. This discipline is the so called Pharmacovigilance, and takes care of the management and control of Drug- Drug interactions (DDI) among other faculties. A DDI occurs when one drug influences the effect level or activity of another drug. Some events such as BioCreative1 and BioNLP2 establish a benchmark of comparison in the field of natural language processing applied to biomedical domain. This is the case of Semeval 2013: Extraction of Drug-Drug Interactions from BioMedical Texts3, where our system has been evaluated. The field of DDI extraction from biomedical text has been faced from different perspectives such as rule-based approaches, SVM approaches and kernel-methods approaches, among others. 1 http://www.biocreative.org/ 2 http://2013.bionlp-st.org/ 3 http://www.cs.york.ac.uk/semeval-2013/task9/ Segura-Bedmar et al. (2010) proposed an approach to extract DDI from biomedical texts based on Shallow Linguistic (SL) Kernel (Giuliano et al., 2006) methods obtaining an F1 measure of 60,01%. The system was evaluated over a DrugDDI dataset created in 2010 that contains 579 biomedical documents collected from the pharmacological database DrugBank4. The dataset contains a total of 3,160 DDIs. Recently, the DDIExtraction2011 task5 compared the latest advances in Information Extraction techniques applied to the DDI identification. The event provided a benchmark forum of 10 different approaches. The evaluation of the systems was made over the DrugDDI dataset. We now describe the most relevant works. Thomas et al. (2011) developed a system by combining a preprocessing phase based on Charniak-Lease (Lease, Charniak, 2005) and Stanford (Marneffe et al., 2006) parsers, with a classification phase based on SL kernel (Giuliano et al., 2006), k-Band Shortest Path Spectrum (kBSPS) kernel (Airola et al., 2008), All Path Graphic (APG) kernel (Tikk et al., 2010) and case- based reasoning (CBR) (Aamodt, Plaza, 1994) techniques. The system obtained a F1 measure of 65.7%. Chowdhury et al. (2011) presented a system combining a preprocessing phase based on Stanford parser and SPECIALIST (Browne, 2000) lexicon tool, with a classification phase based on Featured-Based kernel such as SL kernel and Tree- Based kernel such as Dependency tree (DT) kernel (Culotta and Sorensen, 2004) and Phrase Structure Tree (PST) kernel (Moschitti, 2004). The system achieved …

Content truncated to 500 words