UWM: A Simple Baseline Method for Identifying Attributes of Disease and Disorder Mentions in Clinical Text

Proceedings of the 9th International Workshop on Semantic Evaluation (SemEval 2015), pages 385–388, Denver, Colorado, June 4-5, 2015. c⃝2015 Association for Computational Linguistics UWM: A Simple Baseline Method for Identifying Attributes of Disease and Disorder Mentions in Clinical Text Omid Ghiasvand Rohit J. Kate University of Wisconsin-Milwaukee University of Wisconsin-Milwaukee Milwaukee, Wisconsin Milwaukee, Wisconsin ghiasva2@uwm.edu katerj@uwm.edu Abstract In this paper the system that was developed by Team UWM for the Task 14 of SemEval 2015 competition is described. Task 14 included two tasks: Task 1 was identification of disorder mentions and their normalization, and Task 2 was identification of the following attributes for disorder mentions: the CUI of the disorder, negation indicator, subject, uncertainty indicator, course, severity, conditional, generic indicator, and body location. For Task 1, an earlier system was applied that uses Conditional Random Fields (CRFs) for disorder recognition and learned edit distance patterns for normalization. Task 2 was implemented by a simple method that finds the attribute terms around the disease mentions by matching them in the training data. Among all participants Team UWM was ranked fourth in Task 1, fourth in Task 2A (over gold-standard mentions) and third in Task 2B (over extracted mentions). 1 Introduction Automated extraction tools are crucial for managing huge amount of clinical texts. These tools have the potential to enable many automated applications in healthcare. The Task 14 of SemEval 2015 was designed to serve as a platform for evaluating one such extraction tool. Its Task 1 involved extracting and normalizing disorder mentions from clinical text and its Task 2 involved assertion identification for the mentions. Task 1 is challenging because there is a lot of variability in which diseases and disorders are mentioned in clinical text and hence a pre-defined list of mentions is not sufficient to extract them. The task also required normalizing the extracted mentions by mapping them to UMLS CUIs if they exist in the SNOMED-CT part of UMLS and are marked as disease/disorder, otherwise they were to be declared as “CUI-less.” This normalization process is also challenging because disorder names are frequently mentioned in modified forms which prevents them from exactly matching the concepts in UMLS. Task 2 required finding certain attributes for the mentions and finding the spans of these attributes in text. This task is also challenging due to the variability in which attributes are attributed to disease and disorder mentions in clinical text. Our team, UWM, participated in both Task 1 and Task 2. For Task 1, we used the same system that we had previously developed for the Task 7 of SemEval 2014 (Ghiasvand and Kate 2014). For Task 2, we used a simple method that finds attributes of mentions by first collecting lists of attribute terms from the training data and then matching in this list. The nearest attribute terms to a mention are assigned to that mention. The attribute terms are normalized by finding their normalized values in the training data. Despite being simple, this method gave competitive results. The methods …

Content truncated to 500 words