UTurku: Drug Named Entity Recognition and Drug-Drug Interaction Extraction Using SVM Classification and Domain Knowledge

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 651–659, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics UTurku: Drug Named Entity Recognition and Drug-Drug Interaction Extraction Using SVM Classification and Domain Knowledge Jari Bj¨orne, Suwisa Kaewphan and Tapio Salakoski Turku Centre for Computer Science (TUCS) Department of Information Technology University of Turku Joukahaisenkatu 3-5B, 20520 Turku, Finland firstname.lastname@utu.fi Abstract The DDIExtraction 2013 task in the SemEval conference concerns the detection of drug names and statements of drug-drug interac- tions (DDI) from text. Extraction of DDIs is important for providing up-to-date knowl- edge on adverse interactions between co- administered drugs. We apply the machine learning based Turku Event Extraction Sys- tem to both tasks. We evaluate three fea- ture sets, syntactic features derived from deep parsing, enhanced optionally with features de- rived from DrugBank or from both DrugBank and MetaMap. TEES achieves F-scores of 60% for the drug name recognition task and 59% for the DDI extraction task. 1 Introduction Drug-drug interactions (DDI) refer to one drug af- fecting the function of another when they are co- administered. These interactions are often adverse, frequently not well known and a source of poten- tially life-threatening unintended consequences for the patients. Databases such as DrugBank and Mi- cromedex have been developed to store informa- tion about known DDIs, but at present their cover- age remains limited and there can be inconsistencies in supplementary information (Knox et al., 2011; Wong et al., 2008). Text mining has been proposed as a solution for providing not only lists of DDIs but also a connection to the scientific evidence and supplementary information in the literature (Tari et al., 2010). Several groups of researchers are devel- oping text-mining techniques to extract DDIs from literature and pharmaceutical documents (Tari et al., 2010; Segura-Bedmar et al., 2011a). The DDIExtraction 2013 shared task concerns the detection of drug mentions and statements of DDIs from unannotated text (Segura-Bedmar et al., 2013). The first version of the DDIExtraction shared task was organized in 2011, with 10 teams participat- ing from various universities (Segura-Bedmar et al., 2011b). The best result of 65.74% was achieved by team WBI of Humboldt University of Berlin (Thomas et al., 2011). University of Turku partic- ipated also in this task, placing 4th with an F-score of 62.99%, using the Turku Event Extraction System (Bj¨orne et al., 2011). The Turku Event Extraction System (TEES)1 is an open source program for extracting events and re- lations from biomedical texts. It was originally de- veloped for extracting events in the BioNLP Shared Task scheme, and it models event extraction as a graph generation task, where keywords are nodes and the event arguments connecting them are edges. The system can be directly applied to pairwise re- lation extraction, representing relations as edges and the words they connect as nodes. The node detection system is somewhat similar to named entity recog- nition (NER) tools, and while quite flexible, can in …

Content truncated to 500 words