CUAB: Supervised Learning of Disorders and their Attributes using Relations

Proceedings of the 9th International Workshop on Semantic Evaluation (SemEval 2015), pages 417–421, Denver, Colorado, June 4-5, 2015. c⃝2015 Association for Computational Linguistics CUAB: Supervised Learning of Disorders and their Attributes Using Relations James Gung University of Colorado 1111 Engineering Drive Boulder, Colorado 80309, USA james.gung@colorado.edu John David Osborne, Steven Bethard University of Alabama at Birmingham 1720 2nd Ave South Birmingham, AL, USA 35294 {ozborn,bethard}@cis.uab.edu Abstract We implemented an end-to-end system for dis- order identification and slot filling. For iden- tifying spans for both disorders and their at- tributes, we used a linear chain conditional random field (CRF) approach coupled with cTAKES for pre-processing. For combining disjoint disorder spans, finding relations be- tween attributes and disorders, and attribute normalization, we used l2-regularized l2-loss linear support vector machine (SVM) classifi- cation. Disorder CUIs were identified using a back-off approach to YTEX lookup (CUAB1) or NLM UTS API (CUAB2) if the target text was not found in the training data. Our best system utilized UMLS semantic type features for disorder/attribute span identification and the NLM UTS API for normalization. It was ranked 12th in Task 1 (disorder identification) and 6th in Task 2b (disorder identification and slot filling) with a weighted F Measure of 0.711. 1 Introduction One of the core problems in the field of clinical text processing is the identification and normalization of medical disorders (Pradhan et al., 2014). A sec- ondary problem is the identification of attributes for the identified disorders such as their severity or body location. Attribute identification and normalization helps to better describe the disorder context, allowing for a better determination of the appropriateness of the discovered disorder for the task at hand. SemEval-2015 Task 14 addresses these problems as separate tasks, assessing end to end systems capa- ble of identifying both disorders and attributes from unlabeled clinical text. The first task requires par- ticipants to identify discontinuous disorder spans in clinical text and normalize them to a UMLS Con- cept Unique Identifier (CUI) that is both within the disorder Semantic Group and present in SNOMED CT. The second task requires identification of disor- der CUIs as well as 8 additional attributes associated with each disorder as shown in Table 1 on the shared task page1. For each attribute, the span offset of the lexical cue must also be identified, which may be discontinuous. 2 Approach We combined and extended our previous work (Gung, 2014; Osborne et al., 2014) for the ShARe/CLEF 2013 eHealth Evaluation Lab (Suominen et al., 2013). Both previous systems and our base system for this task are based on the clinical Text Analy- sis and Knowledge Extraction System (cTAKES) (Savova et al., 2010), an open source pipeline for the natural language processing (NLP) of clinical text that utilizes the Unstructured Information Man- agement Architecture (UIMA) (Ferrucci and Lally, 2004) framework. Our combined system is avail- able for download at https://github.com/ jgung/ClearClinical. We developed two systems for this task that dif- fered in their method of CUI lookup and the presence of UMLS …

Content truncated to 500 words