KULeuven-LIIR at SemEval-2017 Task 12: Cross-Domain Temporal Information Extraction from Clinical Records

Proceedings of the 11th International Workshop on Semantic Evaluations (SemEval-2017), pages 1030–1034, Vancouver, Canada, August 3 - 4, 2017. c⃝2017 Association for Computational Linguistics KULeuven-LIIR at SemEval-2017 Task 12: Cross-Domain Temporal Information Extraction from 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 system of the KULeuven-LIIR submission for Clinical TempEval 2017. We participated in all six subtasks, using a combination of Support Vector Machines (SVM) for event and temporal expression detection, and a structured perceptron for extracting temporal relations. Moreover, we present and analyze the results from our submis- sions, and verify the effectiveness of sev- eral system components. Our system per- formed above average for all subtasks in both phases. 1 Introduction In this paper, we describe the system used for the KULeuven-LIIR submissions at SemEval task 12, named Clinical TempEval 2017 (Bethard et al., 2017), which is concerned with temporal infor- mation extraction from clinical records. In Clin- ical TempEval extraction of temporal information is split into six subtasks. Our system participated in all tasks: 1. Detection of event spans (ES) 2. Identification of event attributes (EA) 3. Detection of temporal expressions (TS) 4. Attribute identification of temporal expres- sions (TA) 5. Extraction of document-creation-time rela- tions for events (DR) 6. Extraction of narrative container relations (CR) This year, a new aspect of Clinical TempEval is that systems will be evaluated across domains, which involves two phases: Firstly, unsupervised domain adaptation (Phase I), where the training data is in the colon cancer domain, and the test data in the brain cancer domain. And secondly, su- pervised domain adaptation (Phase II), where the vast majority of the training data are colon cancer reports, and a small number of brain cancer reports is made available for training as well. The test data is again in the brain cancer domain. Our system consist of a combination of linear Support Vector Machines (SVM) for entity span and attribute recognition (tasks ES, EA, TS and TA), and a document-level structured perceptron (Leeuwenberg and Moens, 2017) for relation ex- traction tasks (tasks DR and CR). We used three system components for the domain adaptation: (1) assigning more weight to target-domain training data, (2) introduction of a UNK (unknown) to- ken to model out-of-vocabulary words, and (3) ex- ploitation of relational properties of temporality during prediction. In Section 2, we provide a detailed description of our full system, and in Section 3 we discuss the results from our submissions. 2 Our System Our system consist of three main components (1) preprocessing, (2) entity detection, and (3) rela- tion extraction. In Figure 1, we show a schematic overview of our system. 2.1 Preprocessing The corpus used in Clinical TempEval 2017 is the THYME corpus (Styler IV et al., 2014). For the unsupervised domain adaptation phase (Phase I), we use all colon cancer sections for training. For the supervised domain adaptation (Phase II) par- ticipants also received a small training section in …

Content truncated to 500 words