FBK-irst : A Multi-Phase Kernel Based Approach for Drug-Drug Interaction Detection and Classification that Exploits Linguistic Information

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 351–355, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics FBK-irst : A Multi-Phase Kernel Based Approach for Drug-Drug Interaction Detection and Classification that Exploits Linguistic Information Md. Faisal Mahbub Chowdhury † ‡ and Alberto Lavelli ‡ ‡ Fondazione Bruno Kessler (FBK-irst), Italy † University of Trento, Italy fmchowdhury@gmail.com, lavelli@fbk.eu Abstract This paper presents the multi-phase relation extraction (RE) approach which was used for the DDI Extraction task of SemEval 2013. As a preliminary step, the proposed approach in- directly (and automatically) exploits the scope of negation cues and the semantic roles of in- volved entities for reducing the skewness in the training data as well as discarding possible negative instances from the test data. Then, a state-of-the-art hybrid kernel is used to train a classifier which is later applied on the in- stances of the test data not filtered out by the previous step. The official results of the task show that our approach yields an F-score of 0.80 for DDI detection and an F-score of 0.65 for DDI detection and classification. Our sys- tem obtained significantly higher results than all the other participating teams in this shared task and has been ranked 1st. 1 Introduction Drug-drug interaction (DDI) is a condition when one drug influences the level or activity of another. The extraction of DDIs has significant importance for public health safety. It was reported that about 2.2 million people in USA, age 57 to 85, were taking potentially dangerous combinations of drugs (Lan- dau, 2009). Another report mentioned that deaths from accidental drug interactions rose by 68 percent between 1999 and 2004 (Payne, 2007). The DDIEx- traction 2011 and DDIExtraction 2013 shared tasks underline the importance of DDI extraction. The DDIExtraction 2013 task concerns the recog- nition of drugs and the extraction of drug-drug in- teractions from biomedical literature. The dataset of the shared task is composed by texts from the Drug- Bank database as well as MedLine abstracts in or- der to deal with different type of texts and language styles. Participants were asked to not only extract DDIs but also classify them into one of four pre- defined classes: advise, effect, mechanism and int. A detailed description of the task settings and data can be found in Segura-Bedmar et al. (2013). The system that we used in this shared task combines various techniques proposed in our re- cent research activities for relation extraction (RE) (Chowdhury and Lavelli, 2012a; Chowdhury and Lavelli, 2012b; Chowdhury and Lavelli, 2013).1 2 DDI Detection Our system performs DDI detection and classifica- tion in two separate steps. In this section, we explain how DDI detection (i.e. whether two drug mentions participate in a DDI) is accomplished. DDI classifi- cation will be described in Section 3. There are three phases for DDI detection: (i) dis- card less informative sentences, (ii) discard less in- formative instances, and (iii) train the system (a …

Content truncated to 500 words