UTU: Disease Mention Recognition and Normalization with CRFs and Vector Space Representations

Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), pages 807–811, Dublin, Ireland, August 23-24, 2014. UTU: Disease Mention Recognition and Normalization with CRFs and Vector Space Representations Suwisa Kaewphan1,2,3∗, Kai Hakaka1∗, Filip Ginter1 1 Dept. of Information Technology, University of Turku, Finland 2 Turku Centre for Computer Science (TUCS), Turku, Finland 3 The University of Turku Graduate School (UTUGS), University of Turku, Finland sukaew@utu.fi, kahaka@utu.fi, ginter@cs.utu.fi Abstract In this paper we present our system par- ticipating in the SemEval-2014 Task 7 in both subtasks A and B, aiming at recognizing and normalizing disease and symptom mentions from electronic medi- cal records respectively. In subtask A, we used an existing NER system, NERsuite, with our own feature set tailored for this task. For subtask B, we combined word vector representations and supervised ma- chine learning to map the recognized men- tions to the corresponding UMLS con- cepts. Our system was placed 2nd and 5th out of 21 participants on subtasks A and B respectively showing competitive perfor- mance. 1 Introduction The SemEval 2014 task 7 aims to advance the de- velopment of tools for analyzing clinical text. The task is organized by providing the researchers an- notated clinical records to develop systems that can detect the mentions of diseases and symptoms in medical records. In particular, the SemEval task 7 comprises two subtasks, recognizing the men- tions of diseases and symptoms (task A) and map- ping the mentions to unique concept identifiers that belong to the semantic group of disorders in the Unified Medical Language System (UMLS). Our team participated in both of these sub- tasks. In subtask A, we used an existing named entity recognition (NER) system, NERsuite, sup- plemented with UMLS dictionary and normaliza- tion similarity features. In subtask B, we com- bined compositional word vector representations ∗These authors contributed equally. This work is licensed under a Creative Commons At- tribution 4.0 International Licence. Page numbers and pro- ceedings footer are added by the organisers. Licence details: http://creativecommons.org/licenses/by/4.0/ with supervised machine learning to map the rec- ognized mentions from task A to the UMLS con- cepts. Our best systems, evaluated on strict match- ing criteria, achieved F-score of 76.6% for the sub- task A and accuracy of 60.1% for the subtask B, showing competitive performance in both tasks. 2 Task A: Named Entity Recognition with NERSuite The ML approach based on conditional random fields (CRFs) has shown to have state-of-the-art performance in recognizing the biological entities. We thus performed task A by using NERsuite, an existing NER toolkit with competitive perfor- mance on biological entity recognition (Campos et al., 2013). NERsuite is a NER system that is built on top of the CRFsuite (Okazaki, 2007). It con- sists of three language processing modules: a to- kenizer, a modified version of the GENIA tagger and a named entity recognizer. NERsuite allows user-implemented features in addition to dictio- nary matching and features shown to benefit the systems such as raw token, lemma, part-of-speech (POS) and text chunk. Prior …

Content truncated to 500 words