LASIGE: using Conditional Random Fields and ChEBI ontology

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 660–666, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics LASIGE: using Conditional Random Fields and ChEBI ontology Tiago Grego Dep. de Inform´atica Faculdade de Ciˆencias Universidade de Lisboa Portugal tgrego@fc.ul.pt Francisco Pinto Dep. de Qu´ımica e Bioqu´ımica Faculdade de Ciˆencias Universidade de Lisboa Portugal frpinto@fc.ul.pt Francisco M. Couto Dep. de Inform´atica Faculdade de Ciˆencias Universidade de Lisboa Portugal fcouto@di.fc.ul.pt Abstract For participating in the SemEval 2013 chal- lenge of recognition and classification of drug names, we adapted our chemical en- tity recognition approach consisting in Condi- tional Random Fields for recognizing chemi- cal terms and lexical similarity for entity res- olution to the ChEBI ontology. We obtained promising results, with a best F-measure of 0.81 for the partial matching task when us- ing post-processing. Using only Conditional Random Fields the results are slightly lower, achieving still a good result in terms of F- measure. Using the ChEBI ontology allowed a significant improvement in precision (best precision of 0.93 in partial matching task), which indicates that taking advantage of an ontology can be extremely useful for enhanc- ing chemical entity recognition. 1 Introduction Most chemical named entity recognition systems can be classified in two approaches: dictionary based and machine learning based approaches. Dic- tionary based approaches are usually easier to im- plement and maintain, but require a reference chem- ical term dictionary and are dependent on its com- pleteness and quality. The availability of public chemical databases has been an issue until recently, when several publicly available databases such as PubChem (Wang et al., 2009), DrugBank (Wishart et al., 2006) and ChEBI (Degtyarenko et al., 2007) were released. An example of a popular system that uses this approach is Whatizit (Rebholz-Schuhmann et al., 2008). Machine learning based approaches are not limited to a terminology and are thus better suited for finding novel chemical terms that are yet to be inserted in reference databases. However this approach requires training data for a classifier to be able to successfully learn and perform the chemi- cal entity recognition task. Some methods combine both approaches and thus are hybrid systems that aim to take the best out of both approaches (Jessop et al., 2011; Rockt¨aschel et al., 2012). An annotated corpus of patent documents was re- leased by ChEBI, and using such corpus as train- ing data we developed an chemical entity recogni- tion system (Grego et al., 2009) that uses a ma- chine learning approach based on Conditional Ran- dom Fields (CRF) (Lafferty et al., 2001). We fur- thermore expanded our method to allow resolution of recognized entities to the ChEBI ontology (Grego et al., 2012). This paper describes how our system (Grego et al., 2012) was adapted to perform the task of recog- nition and classification of drug names, and presents the results obtained in the task 9.1 of the 7th Interna- tional Workshop on Semantic Evaluation (SemEval 2013). 2 Task …

Content truncated to 500 words