WBI-DDI: Drug-Drug Interaction Extraction using Majority Voting

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 628–635, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics WBI-DDI: Drug-Drug Interaction Extraction using Majority Voting Philippe Thomas Mariana Neves Tim Rockt¨aschel Ulf Leser Humboldt-Universit¨at zu Berlin Knowledge Management in Bioinformatics Unter den Linden 6 Berlin, 10099, Germany {thomas,neves,trocktae,leser}@informatik.hu-berlin.de Abstract This work describes the participation of the WBI-DDI team on the SemEval 2013 – Task 9.2 DDI extraction challenge. The task con- sisted of extracting interactions between pairs of drugs from two collections of documents (DrugBank and MEDLINE) and their clas- sification into four subtypes: advise, effect, mechanism, and int. We developed a two-step approach in which pairs are initially extracted using ensembles of up to five different clas- sifiers and then relabeled to one of the four categories. Our approach achieved the sec- ond rank in the DDI competition. For interac- tion detection we achieved F1 measures rang- ing from 73 % to almost 76 % depending on the run. These results are on par or even higher than the performance estimation on the train- ing dataset. When considering the four inter- action subtypes we achieved an F1 measure of 60.9 %. 1 Introduction A drug-drug interaction (DDI) can be described as interplay between drugs taken during joint adminis- tration. DDIs usually lead to an increase or decrease in drug effects when compared to isolated treatment. For instance, sildenafil (Viagra) in combination with nitrates can cause a potentially live-threatening de- crease in blood pressure (Cheitlin et al., 1999). It is therefore crucial to consider potential DDI effects when co-administering drugs to patients. As the level of medication generally is raising all over the world, the potential risk of unwanted side effects, such as DDIs, is constantly increasing (Haider et al., 2007). Only a fraction of knowledge about DDIs is contained in specialized databases such as Drug- Bank (Knox et al., 2011). These structured knowl- edge bases are often the primary resource of infor- mation for researchers. However, the majority of new DDI findings are still initially reported in scien- tific publications, which results in the situation that structured knowledge bases lag behind recently pub- lished research results. Thus, there is an urgent need for researchers and database curators to cope with the fast growth of biomedical literature (Hunter and Cohen, 2006). The SemEval 2013 – Task 9.2 (Extraction of Drug-Drug Interactions from BioMedical Texts) is a competitive evaluation of methods for ex- tracting mentions of drug-drug interactions from texts (Segura-Bedmar et al., 2013). For training, the organizers provide a corpus annotated with drug- names and interactions between them. This corpus is composed of 572 articles collected from Drug- Bank and 142 PubMed abstracts. Interactions are binary (always between two drugs) and undirected, as target and agent roles are not annotated. Fur- thermore, the two interacting drugs are always men- tioned within the same sentence. In contrast to the previous DDI-challenge 2011 (Segura-Bedmar et al., 2011), four different …

Content truncated to 500 words