UColorado_SOM: Extraction of Drug-Drug Interactions from Biomedical Text using Knowledge-rich and Knowledge-poor Features

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 684–688, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics UColorado SOM: Extraction of Drug-Drug Interactions from BioMedical Text using Knowledge-rich and Knowledge-poor Features Negacy D. Hailu Lawrence E. Hunter K. Bretonnel Cohen negacy.hailu@ucdenver.edu larry.hunter@ucdenver.edu kevin.cohen@gmail.com University of Colorado, Anschutz Medical Campus Abstract In this paper, we present our approach to SemEval-2013 Task 9.2. It is a feature rich classification using LIBSVM for Drug-Drug Interactions detection in the BioMedical do- main. The features are extracted considering morphosyntactic, lexical and semantic con- cepts. Tools like openDMAP and TEES are used to extract semantic concepts from the corpus. The best F-score that we got for Drug- Drug Interaction (DDI) detection is 50% and 61% and the best F-score for DDI detection and classification is 34% and 48% for test and development data respectively. Keywords: text mining, event extraction, ma- chine learning, feature extraction. 1 Introduction Our approach to the Semeval 2013 drug-drug in- teraction task explored the potential for integrat- ing knowledge-based approaches with supervised machine learning. In practice, most supervised machine learning systems are actually hybrids of machine learning and some knowledge-based ap- proach. However, the integration between the two is typically quite loose, with the knowledge-based approach being realized either as heuristic pre- processing or post-processing of the results. The work reported here is an attempt to make a tighter coupling between knowledge-based methods and machine learning. In particular, we took the ap- proach of using knowledge-based methods for fea- ture extraction. 2 Methodology In this challenge we approach the Drug-Drug inter- action task 9.2 as a binary classification problem. A pair of drugs is interacting if there is some kind of influence between the two. Our approach for Drug- Drug interaction extraction 2013 mainly makes use of domain specific morphosyntactic, lexical and se- mantic features between paired drugs. We applied Machine Learning classification tech- niques in order to determine whether a pair of drugs within a biomedical text is interacting or not. For a training set of labeled instances � Xi, yi � = 1, 2, ..., l where Xi ∈ Rn and y ∈ {1, −1}l, the support vector machines (SVMs) optimization problem is defined as(Boser et al., 1992) (Cortes and Vapnik , 1995): ˆα = arg max α,w,b �1 2W T W + C l � i=1 αi � (1) such that yi � W T φ(Xi) + b � > 1 − αi, αi ≥ 0. 2.1 Materials The corpus is provided from two data sources. There are 572 documents describing drug-drug interac- tions from the DrugBank database and 142 abstracts on the subject of drug-drug interactions from Med- Line (Isabel et.al., 2011). We prepared datasets for the entire corpus. Each instance in the dataset is a set of paired drugs. In our dataset, there are 27787 instances. 93.57% of them are from Drug- bank database and the remaining are from MedLine abstracts. DDI shared …

Content truncated to 500 words