Structured Learning for Temporal Relation Extraction from Clinical Records

Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers, pages 1150–1158, Valencia, Spain, April 3-7, 2017. c⃝2017 Association for Computational Linguistics Structured Learning for Temporal Relation 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 We propose a scalable structured learning model that jointly predicts temporal rela- tions between events and temporal expres- sions (TLINKS), and the relation between these events and the document creation time (DCTR). We employ a structured per- ceptron, together with integer linear pro- gramming constraints for document-level inference during training and prediction to exploit relational properties of temporal- ity, together with global learning of the re- lations at the document level. Moreover, this study gives insights in the results of in- tegrating constraints for temporal relation extraction when using structured learning and prediction. Our best system outper- forms the state-of-the art on both the CON- TAINS TLINK task, and the DCTR task. 1 Introduction Temporal information is critical in many clinical areas (Combi and Shahar, 1997). A big part of this temporal information is captured in the free text of patient records. The current work aims to im- prove temporal information extraction from such clinical texts. Extraction of temporal information from clinical text records can be used to construct a time-line of the patient’s condition (such as in Figure 1). The extracted time-line can help clini- cal researchers to better select and recruit patients with a certain history for clinical trials. Moreover, the time-line is crucial for making a good patient prognosis and clinical decision support (Onisko et al., 2015; Stacey and McGregor, 2007). Temporal information extraction can be divided into three sub-problems: (1) the detection of events (Ee); (2) the detection of temporal expres- sions (Et); and (3) the detection of temporal rela- tions between them. In the clinical domain, events include medical procedures, treatments, or symp- toms (e.g. colonoscopy, smoking, CT-scan). Tem- poral expressions include dates, days of the week, months, or relative expressions like yesterday, last week, or post-operative. In this work, we focus on the last sub-problem, extraction of temporal rela- tions (assuming events and temporal expressions are given). As a small example of the task we aim to solve, given the following sentence: In 1990 the patient was diagnosed and received surgery directly afterwards. in which we assume that the events diagnosed and adenocarcinoma, and the temporal expression 1990 are given, we wish to extract the following relations: • CONTAINS(1990, diagnosed) • CONTAINS(1990, surgery) • BEFORE(diagnosed, surgery) • BEFORE(diagnosed, d) • BEFORE(surgery, d) where d stands for the document creation time. Our work leads to the following contributions: First, we propose a scalable structured learning model that jointly predicts temporal relations be- tween events and temporal expressions (TLINKS), and the relation between these events and the doc- ument creation time (DCTR). In contrast to ex- isting approaches which detect relation instances separately, our approach employs a structured per- ceptron (Collins, …

Content truncated to 500 words