LIMSI-COT at SemEval-2016 Task 12: Temporal relation identification using a pipeline of classifiers

Proceedings of SemEval-2016, pages 1136–1142, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics LIMSI-COT at SemEval-2016 Task 12: Temporal relation identification using a pipeline of classifiers Julien Tourille LIMSI, CNRS, Univ. Paris-Sud, Universit´e Paris-Saclay julien.tourille@limsi.fr Olivier Ferret CEA, LIST, Gif-sur-Yvette, F-91191 France. olivier.ferret@cea.fr Aur´elie N´ev´eol LIMSI, CNRS, Universit´e Paris-Saclay aurelie.neveol@limsi.fr Xavier Tannier LIMSI, CNRS, Univ. Paris-Sud, Universit´e Paris-Saclay xtannier@limsi.fr Abstract SemEval 2016 Task 12 addresses temporal reasoning in the clinical domain. In this paper, we present our participation for relation ex- traction based on gold standard entities (sub- tasks DR and CR). We used a supervised approach comparing plain lexical features to word embeddings for temporal relation identi- fication, and obtained above-median scores. 1 Introduction SemEval 2016 Task 12 offers 6 subtasks address- ing temporal reasoning in the clinical domain using the THYME corpus (Styler IV et al., 2014). This corpus provides annotated clinical and pathological notes from colon cancer patients. The first group of subtasks concerns the identification of time and event expressions within raw text. The second group of subtasks deals with the identification of tempo- ral relations. The latter consists of two subtasks. In the Document Creation Time Relation subtask (DR), participants are challenged to identify relations be- tween the events and the document creation time. For the Container Relation subtask (CR), partici- pants have to identity container relations between entities. Participants may submit either a complete system extracting entities and relations or focus on either the entity extraction or relation extraction (us- ing the gold standard entities provided by the orga- nizers). More details about the task and the defini- tion of each subtask can be found in Bethard et al. (2016). In this paper, we present our submission for the CR and DR subtasks based on gold-standard entities (phase 2). Our global approach, which is illustrated in Figure 1, tackles the identification of temporal relations as a set of supervised classification tasks. We submitted two runs, one using plain lexical fea- tures and one using word embeddings computed on a large clinical corpus. We obtained scores well above the median scores in both subtasks. The remainder of this paper is organized as fol- lows. Section 2 presents our system for the DR sub- task while Section 3 describes our system for the CR subtask. Section 4 gives an overview of the system implementation. Finally, Section 5 presents our re- sults. 2 Document Creation Time Relation (DR) Subtask We treated the subtask as a supervised classification problem where each EVENT entity was classified into four categories (Before, Before-Overlap, Over- lap, After). We extracted lexical, contextual and structural features from the texts. Regarding the lexical fea- tures of EVENT entities, we took their surface forms, their gold standard attributes (type, modality, de- gree and polarity), their lemma(s)1, as well as their Part-Of-Speech (POS) and Coarse Part-Of-Speech (CPOS) tags. We also extracted the semantic types 1The span of an EVENT entity can overlap with several to- kens. 1136 Figure 1: Overview of the processing chain …

Content truncated to 500 words