KULeuven-LIIR at SemEval 2016 Task 12: Detecting Narrative Containment in Clinical Records

Proceedings of SemEval-2016, pages 1280–1285, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics KULeuven-LIIR at SemEval 2016 Task 12: Detecting Narrative Containment in Clinical Records Artuur Leeuwenberg and Marie-Francine Moens Department of Computer Science KU Leuven, Belgium {tuur.leeuwenberg, sien.moens}@cs.kuleuven.be Abstract In this paper, we describe the KULeuven- LIIR system at the Clinical TempEval 2016 Shared Task for the narrative container re- lation sub-task (CR). Our approach is based on the cTAKES Temporal system (Lin et al., 2015). We explored extending this system with different features. Moreover, we provide an error analysis of the submitted system, and report on some additional experiments done after submission. 1 Introduction We describe the KULeuven-LIIR submissions for the Clinical TempEval 2016 shared task (Bethard et al., 2016). Our motivation for this first participa- tion is to gain insight into the task, and the data, as a basis for future work. We participated in the nar- rative container relation (CR) task. In the CR task narrative containment relations between events, and events and temporal expressions are to be extracted. Two examples of such relations are given in Sen- tence 1. (1) A colonoscopy on 27 September 2008 revealed a circumferential lesion. The relations that are to be extracted are • CONTAINS(27 September 2008, colonoscopy) • CONTAINS(colonoscopy, lesion) In the shared task, the clinical records on colon cancer from the THYME corpus are used (Styler IV et al., 2014). We participated in phase 2 evaluation, which means that manually annotated EVENT and TIMEX3 spans with their attributes are provided as part of the input. We submitted two runs that are both based on the cTAKES-Temporal system (Lin et al., 2015). For the first run, we used cTAKES- Temporal with a set of already provided features. We further explain the cTAKES-Temporal system and the features that we used in Section 2. For the second run, we added new features, that are based on an error analysis of a small sample from the re- sults of running the first system on the development set. These additional features are described in Sec- tion 3. Moreover, we describe a more elaborate error analysis of the second system in Section 4, and some experiments that we performed with the aim to han- dle these errors in Section 5. 2 The cTAKES Temporal System (run 1) The first version of the temporal module in cTAKES1, an extensive open source information ex- traction system for clinical free-texts (Savova et al., 2010), is described by (Lin et al., 2015). We used the currently available version in our experiments. The current version of cTAKES Temporal consists of two SVM classifiers to detect containment. One for containment between events (EE), and one for con- tainment between temporal expressions and events (TE), each using different features, as the nature of the two types of relations is quite different. The features we used for each classifier are described in Table 2, and will be explained further in the next paragraph. Note that with entity we refer …

Content truncated to 500 words