UWM: Disorder Mention Extraction from Clinical Text Using CRFs and Normalization Using Learned Edit Distance Patterns

Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), pages 828–832, Dublin, Ireland, August 23-24, 2014. UWM: Disorder Mention Extraction from Clinical Text Using CRFs and Normalization Using Learned Edit Distance Patterns Omid Ghiasvand University of Wisconsin-Milwaukee Milwaukee, WI ghiasva2@uwm.edu Rohit J. Kate University of Wisconsin-Milwaukee Milwaukee, WI katerj@uwm.edu Abstract This paper describes Team UWM’s sys- tem for the Task 7 of SemEval 2014 that does disorder mention extraction and nor- malization from clinical text. For the dis- order mention extraction (Task A), the sys- tem was trained using Conditional Ran- dom Fields with features based on words, their POS tags and semantic types, as well as features based on MetaMap matches. For the disorder mention normalization (Task B), variations of disorder mentions were considered whenever exact matches were not found in the training data or in the UMLS. Suitable types of variations for disorder mentions were automatically learned using a new method based on edit distance patterns. Among nineteen partic- ipating teams, UWM ranked third in Task A with 0.755 strict F-measure and second in Task B with 0.66 strict accuracy. 1 Introduction Entity mention extraction is an important task in processing natural language clinical text. Disor- ders, medications, anatomical sites, clinical pro- cedures etc. are among the entity types that pre- dominantly occur in clinical text. Out of these, the Task 7 of SemEval 2014 concentrated on ex- tracting (Task A) and normalizing (Task B) dis- order mentions. Disorder mention extraction is particularly challenging because disorders are fre- quently found as discontinuous phrases in clinical sentences. The extracted mentions were then to be normalized by mapping them to their UMLS CUIs if they were in the SNOMED-CT part of UMLS 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/ and belonged to the “disorder” UMLS seman- tic group, otherwise they were to be declared as “CUI-less”. This normalization task is challeng- ing because disorder names are frequently men- tioned in modified forms which prevents their ex- act matching with concept descriptions in UMLS. Our team, UWM, participated in both Task A and Task B. We modelled disorder mention extraction as a standard sequence labeling task. The model was trained using Conditional Ran- dom Fields (Lafferty et al., 2001) with various types of lexical and semantic features that in- cluded MetaMap (Aronson, 2001) matches. The model was also inherently capable of extracting discontinuous disorder mentions. To normalize disorder mentions, our system first looked for ex- act matches with disorder mentions in the train- ing data and in the UMLS. If no exact match was found, then suitable variations of the disorder mentions were generated based on the commonly used variations of disorder mentions learned from the training data as well as from the UMLS syn- onyms. We developed a novel method to automat- ically learn such variations based on edit distances (Levenshtein, 1966) which is described in the next …

Content truncated to 500 words