LIMSI at SemEval-2016 Task 12: machine-learning and temporal information to identify clinical events and time expressions

Proceedings of SemEval-2016, pages 1225–1230, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics LIMSI at SemEval-2016 Task 12: machine-learning and temporal information to identify clinical events and time expressions Cyril Grouin LIMSI, CNRS, Universit´e Paris-Saclay Bˆat 508, rue John von Neumann, Campus Universitaire, F-91405 Orsay cyril.grouin@limsi.fr V´eronique Moriceau LIMSI, CNRS, Univ. Paris-Sud, Universit´e Paris-Saclay Bˆat 508, rue John von Neumann, Campus Universitaire, F-91405 Orsay veronique.moriceau@limsi.fr Abstract Our experiments rely on a combination of machine-learning (CRF) and rule-based (Hei- delTime) systems. First, a CRF system iden- tifies both EVENTS and TIMEX3, along with polarity values for EVENT and types of TIMEX. Second, the HeidelTime tool identi- fies DOCTIME and TIMEX3 elements, and computes DocTimeRel for each EVENT iden- tified by the CRF. Third, another CRF sys- tem computes DocTimeRel for each pre- viously identified EVENT, based on Doc- TimeRel computed by HeidelTime. In the first submission, all EVENTS and TIMEX3 are identified through one general CRF model while in the second submission, we combined two CRF models (one for both EVENT and TIMEX3, and one only for TIMEX3) and we applied post-processing rules on the outputs. 1 Introduction In this paper, we present the methods we used while participating in the 2016 Clinical TempE- val task as part of the SemEval-2016 challenge. A few recent NLP challenges focused on tempo- ral expressions within clinical records, such as the 2014 i2b2/UTHealth1 challenge (Stubbs et al., 2015) or the second task from the 2014 ShARe/CLEF eHealth2 evaluation lab (Mowery et al., 2014). 1https://www.i2b2.org/NLP/ TemporalRelations/ 2http://clefehealth2014.dcu.ie/task-2 2 Task description 2.1 Presentation The 2016 Clinical TempEval track3 proposed six tasks (Bethard et al., 2016). We participated in the first five tasks which concern the identification of: (i) spans of time expressions (TS task), (ii) spans of event expressions (ES task), (iii) the attribute of time expressions (TA task), (iv) attributes of event expressions (EA task), and (v) the relation between each event and the document creation time (DR task). We did not participate in the narrative con- tainer relation task (CR). In both TS and ES tasks, spans of time and event expressions are represented using start and end off- sets of characters. In the TA task, time expres- sions are related to the attribute “class” which speci- fies the type of time expressions among six possible values: date, duration, prepostexp, quantifier, set, time. In the EA task, event expressions are related to four attributes: “polarity” (either positive or neg- ative), “modality” (actual, hedged, hypothetical or generic), “degree” (most, little or N/A) and “type” (aspectual, evidential or N/A). Two phases were proposed. In the first phase, all tasks were proposed, based on raw texts. We thus participated to TS ES TA EA and DR tasks. In the second phase, reference annotations were given for tasks TS ES TA and EA, and participants were expected to identify the relations from DR and CR tasks. We also participated in this DR task. 3http://alt.qcri.org/semeval2016/task12/ 1225 2.2 Corpora The given training corpus was divided into two …

Content truncated to 500 words