LIMSI-COT at SemEval-2017 Task 12: Neural Architecture for Temporal Information Extraction from Clinical Narratives

Proceedings of the 11th International Workshop on Semantic Evaluations (SemEval-2017), pages 597–602, Vancouver, Canada, August 3 - 4, 2017. c⃝2017 Association for Computational Linguistics LIMSI-COT at SemEval-2017 Task 12: Neural Architecture for Temporal Information Extraction from Clinical Narratives Julien Tourille1,2,3, Olivier Ferret4, Xavier Tannier1,2,3, Aurélie Névéol1,3 1 LIMSI, CNRS 2 Univ. Paris-Sud 3 Université Paris-Saclay 4 CEA, LIST, Gif-sur-Yvette, F-91191 France. firstname.lastname@limsi.fr, olivier.ferret@cea.fr Abstract In this paper we present our participation to SemEval 2017 Task 12. We used a neural network based approach for entity and temporal relation extraction, and ex- perimented with two domain adaptation strategies. We achieved competitive per- formance for both tasks. 1 Introduction SemEval 2017 Task 12 offers 6 subtasks address- ing medical event recognition and temporal rea- soning in the clinical domain using the THYME corpus (Styler IV et al., 2014). Similarly to the two previous editions of the challenge (Bethard et al., 2015, 2016), the first group of subtasks con- cerns medical event (EVENT) and temporal ex- pression (TIMEX3) extraction from raw text. In a second group of subtasks, participants are chal- lenged to extract containment (CONTAINS) re- lations between EVENT and/or TIMEX3 as well as Document Creation Time (DCT) relations be- tween EVENT entities and documents in which they are embedded. The novelty of the 2017 edi- tion lies in the difference of domains between train and test corpora. More details about the task and the definition of each subtask can be found in Bethard et al. (2017). The task has been offered by SemEval over the past two years. Concerning the first group of subtasks, different approaches have been im- plemented by the participants including Condi- tional Random Fields (CRF) (AAl Abdulsalam et al., 2016; Caselli and Morante, 2016; Chikka, 2016; Cohan et al., 2016; Grouin and Moriceau, 2016; Hansart et al., 2016) and deep learning mod- els (Fries, 2016; Chikka, 2016; Li and Huang, 2016). Similarly, CRF and neural networks mod- els have been used for the second group of sub- tasks (AAl Abdulsalam et al., 2016; Cohan et al., 2016; Lee et al., 2016). Other approaches in- clude Support Vector Machines (SVM) (AAl Ab- dulsalam et al., 2016; Tourille et al., 2016). 2 Methodology The EVENT and TIMEX3 entity extraction sub- tasks can be seen as two sequence labeling prob- lems where each token of a given sentence is as- signed a label. Entities can spread over several tokens and therefore, we used the IOB format (Inside, Outside, Beginning) for label represen- tation. Each token can be at the beginning of an entity (B), inside an entity (I) or outside (O). EVENT entities are characterized by a type at- tribute that we used in our IOB scheme resulting in 7 possible labels. Similarly, TIMEX3 entities are characterized by a class attribute that we used in our IOB scheme resulting in 13 possible labels. The container relation extraction task can be cast as a 3-class classification problem. For each combination E1 – E2 of EVENT and/or TIMEX3 from left to right, …

Content truncated to 500 words