TeamHCMUS: Analysis of Clinical Text

Proceedings of the 9th International Workshop on Semantic Evaluation (SemEval 2015), pages 370–374, Denver, Colorado, June 4-5, 2015. c⃝2015 Association for Computational Linguistics TeamHCMUS: Analysis of Clinical Text Nghia Huynh Quoc Ho Faculty of Information Technology Faculty of Information Technology University of Science, Ho Chi Minh City, Vietnam University of Science, Ho Chi Minh City, Vienam huynhnghiavn@gmail.com hbquoc@fit.hcmus.edu.vn Abstract We developed a system to participate in shared tasks on the analyzing clinical text. Our system approaches are both machine learning-based and rule-based. We applied the machine learning-based approach for Task 1: disorder identification, and the rule-based approach for Task 2: template slot filling for the disorder. In Task 1, we developed a supervised conditional random fields model that was based on a rich set of features, and used for predicting disorder mentions. In Task 2, we based on the de- pendency tree to build a rule set. This rule set was extracted from the training data and applied to fill values of disorder attribute types on the test data. The evaluation on the test data showed that our system achieved the F-score of 0.656 (0.685 in case of re- laxed score) for Task 1 and the F*WA of 0.576 for Task 2A and the F*WA of 0.671 for Task 2B. 1 Introduction SemEval-2015 Task 14 is a continuation of pre- vious tasks such as: CLEF eHealth Evaluation Labs 20131 (Hanna Suominen et al., 2013), CLEF eHealth Evaluation Labs 20142 (Liadh Kelly et al., 2014), and SemEval-2014 task 73 (Sameer Pradhan et al., 2014). The aim of the tasks is to improve the methods of natural lan- guage processing (NLP) of the clinical domain 1 https://sites.google.com/site/shareclefehealth/ 2 http://clefehealth2014.dcu.ie/ 3 http://alt.qcri.org/semeval2014/task7/ and to widely introduce the clinical text pro- cessing to the community of NLP research. The clinical narrative is abundant in mentions of clinical conditions, anatomical sites, medica- tions and procedures. It is completely different from the newswire domain where text is domi- nated by mentions of countries, locations and people. Many surface forms represent the same concept. Unlike the general domain, in biomedi- cine which are rich lexical and ontology re- sources that can be leveraged when applications are built. The SemEval-2015 Task 14 is split into two tasks: 1) Task 1 is disorder identification, and its goal is to recognize the span of disorder men- tions, the named entity recognition, and the normalization to a unique CUI in a SNOMED- CT terminology in a set of clinical notes. The SNOMED-CT is a resource provided by the or- ganizers for the normalization of Task 1; and 2) Task 2 is disorder slot filling; it focuses on iden- tifying the normalized value for nine modifiers in a disorder mentioned in a clinical note: the CUI of the disorder (much similar to Task 1), as well as the potential attributes (e.g. negation in- dicator, subject, uncertainty indicator, course, severity, conditional, generic indicator, and body location). Participants can submit to either or both of the tasks. We participated in both tasks. In this …

Content truncated to 500 words