WBI-NER: The impact of domain-specific features on the performance of identifying and classifying mentions of drugs

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 356–363, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics WBI-NER: The impact of domain-specific features on the performance of identifying and classifying mentions of drugs Tim Rockt¨aschel Torsten Huber Michael Weidlich Ulf Leser Humboldt-Universit¨at zu Berlin Knowledge Management in Bioinformatics Unter den Linden 6 Berlin, 10099, Germany {trocktae,thuber,weidlich,leser}@informatik.hu-berlin.de Abstract Named entity recognition (NER) systems are often based on machine learning tech- niques to reduce the labor-intensive devel- opment of hand-crafted extraction rules and domain-dependent dictionaries. Nevertheless, time-consuming feature engineering is often needed to achieve state-of-the-art performance. In this study, we investigate the impact of such domain-specific features on the performance of recognizing and classifying mentions of pharmacological substances. We compare the performance of a system based on general fea- tures, which have been successfully applied to a wide range of NER tasks, with a system that additionally uses features generated from the output of an existing chemical NER tool and a collection of domain-specific resources. We demonstrate that acceptable results can be achieved with the former system. Still, our ex- periments show that using domain-specific fea- tures outperforms this general approach. Our system ranked first in the SemEval-2013 Task 9.1: Recognition and classification of pharma- cological substances. 1 Introduction The accurate identification of drug mentions in text is an important prerequisite for many applications, in- cluding the retrieval of information about substances in drug development (e.g. Roberts and Hayes (2008)), the identification of adverse drug effects (e.g. Lea- man et al. (2010)) and the recognition of drug-drug interactions (e.g. Thomas et al. (2011)). Given that most of the information related to drug research is covered by medical reports and pharmacological pub- lications, computational methods for information ex- traction should be used to support this task. The SemEval-2013 Task 9.1 competition1 (Segura- Bedmar et al., 2013) aims at a fair assessment on the state-of-the-art of tools that recognize and classify mentions of pharmacological substances in natural language texts – a task referred to as drug named entity recognition (NER). The goal of participating teams is to recreate the gold annotation on a held- out part of an annotated corpus. Four classes of en- tities have to be identified: Drug, DrugN, Group and Brand. Entities of class Drug denote any kind of drug that is approved for use in humans, whereas DrugN denotes substances that are not approved. Group are terms describing a group of drugs and Brand stands for drug names introduced by a phar- maceutical company. The aim of this study is to examine whether it is worthwhile to implement domain-specific features for supporting drug NER. The question we attempt to answer is whether such features really help in iden- tifying and classifying mentions of drugs or whether a mostly domain-independent feature set, which can be applied to many other tasks, achieves a compara- ble performance. 2 Related work Various NER systems for identifying different classes of chemical …

Content truncated to 500 words