SCAI: Extracting drug-drug interactions using a rich feature vector

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 675–683, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics SCAI: Extracting drug-drug interactions using a rich feature vector Tamara Bobi´c1,2, Juliane Fluck1, Martin Hofmann-Apitius1,2 1Fraunhofer SCAI Schloss Birlinghoven 53754 Sankt Augustin Germany 2B-IT, Bonn Universit¨at Dahlmannstraße 2 53113 Bonn Germany {tbobic, jfluck, hofmann-apitius}@scai.fraunhofer.de Abstract Automatic relation extraction provides great support for scientists and database curators in dealing with the extensive amount of biomed- ical textual data. The DDIExtraction 2013 challenge poses the task of detecting drug- drug interactions and further categorizing them into one of the four relation classes. We present our machine learning system which utilizes lexical, syntactical and semantic based feature sets. Resampling, balancing and en- semble learning experiments are performed to infer the best configuration. For general drug- drug relation extraction, the system achieves 70.4% in F1 score. 1 Introduction Drug-drug interactions (DDI) describe possible in- terference between pharmacological substances and are of critical importance in drug development and administration (August et al., 1997). A drug may alter the metabolism of another, thus causing an en- hanced, reduced or even toxic effect in certain med- ical treatments. For example: “Fluvoxamine in- hibits the CYP2C9 catalyzed biotransformation of tolbutamide.” Automated extraction of DDI from biomedical literature allows for a more efficient maintenance of the drug knowledge databases and is beneficial for patients, health care professionals and the pharmaceutical industry. Having in mind their biomedical importance, the objective of the first DDIExtraction challenge1 in 1http://labda.inf.uc3m.es/ DDIExtraction2011/ 2011 was to motivate the development and to eval- uate the automatic relation extraction (RE) systems for DDI. Given annotated drug entities, the partic- ipants addressed the task of identifying undirected binary relations among them. The knowledge ex- traction was performed on the sentence level and the best system achieved 65.74% F1 score (Thomas et al., 2011a). The 2013 DDIExtraction challenge2 (organized as Task 9 of SemEval 2013 (Segura-Bedmar et al., 2013)) is based on a similar task definition, but ad- ditionally includes the disambiguation between four types of interaction: mechanism, effect, advise and int. The evaluation of participating systems is two- fold, i. e. partial and strict. Partial evaluation con- siders that a prediction is correct when the pair la- bel matches the gold annotation, while strict eval- uation requires also a correct relation type to be assigned. The train and test corpora were gener- ated from textual resources of DrugBank (Knox et al., 2011) database and MedLine3 abstracts, dealing with the topic of DDI. In the following sections we describe our super- vised machine learning based approach for the ex- traction of DDI, using a rich feature vector (see Sec- tion 2.1). The base system employed LibLINEAR classifier, generating the first run submitted to the DDIExtraction challenge. Configurations coming from the two ensemble strategies (Section 2.2) pro- duced the remaining prediction runs. Furthermore, we experimentally investigated the impact of train 2http://www.cs.york.ac.uk/semeval-2013/ task9/ 3http://www.ncbi.nlm.nih.gov/pubmed/ 675 corpora imbalance on DDI detection …

Content truncated to 500 words