VUACLTL at SemEval 2016 Task 12: A CRF Pipeline to Clinical TempEval

Proceedings of SemEval-2016, pages 1241–1247, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics VUACLTL at SemEval 2016 Task 12: A CRF Pipeline to Clinical TempEval Tommaso Caselli and Roser Morante CLTL Lab - Vrije Univeristeit Amsterdam De Boelelaan 1105 1081 HV Amsterdam, The Netherlands {t.caselli,r.morantevallejo}@vu.nl Abstract This paper describes VUACLTL, the system the CLTL Lab submitted to the SemEval 2016 Task Clinical TempEval. The system is based on a purely data-driven approach based on a cascade of seven CRF classifiers which use generic features and little domain knowledge. The challenge consisted in six subtasks related to temporal processing clinical notes from raw text (event and temporal expression detection and attribute classification, temporal relation classification between events and the Docu- ment Creation Time, and narrative container detection). The system was initially developed to process newswire texts and then re-trained to process clinical notes. This had an impact on the results, which are not equally competi- tive for all the subtasks. 1 Introduction Temporal Processing is becoming more and more important for improving access to content. The availability of timelines (either event-centric or entity-centric) can help improving more complex semantically-focused tasks such as Question An- swering, Text Summarization, and Textual Entail- ment, among others. Furthermore, timelines can be further exploited for monitoring the development in time of different phenomena, e.g. the opinions in debates. Temporal Processing research has mainly focused on the newswire domain (Verhagen et al., 2007; Verhagen et al., 2010; UzZaman et al., 2013) in the framework of several shared tasks where sys- tems were challenged to extract the relevant compo- nents of a document timeline: temporal expressions, event mentions, and temporal relations. Several evaluations have shown the capabilities and limits of both the annotated resources and the systems. For instance, the best system in TempEval-3 (Bethard, 2013) reports 0.398 F1 on Temporal Relation Detec- tion and Classification from raw text. The develop- ment of temporally annotated corpora has boosted research in languages other than English such as Italian (Caselli et al., 2014), French (Arnulphy et al., 2015), and Spanish (Llorens et al., 2010), among others. Recently, interest in temporal processing has moved forward in two directions: cross-document timeline extraction (Minard et al., 2015) and domain adaptation (Sun et al., 2013; Bethard et al., 2015). The setting of the 2015 and 2016 SemEval Clini- cal TempEval Tasks is similar to previous TempEval campaigns, with the two main differences: i.) the domain , i.e. (colon) cancer clinical notes; and ii.) the annotation scheme, i.e. the THYME annotation scheme (Styler IV et al., 2014), an extended version of TimeML (Pustejovsky et al., 2003a). Similarly to the previous edition, the SemEval 2016 Clinical TempEval task (Bethard et al., 2016) consists of the following six subtasks: temporal expression detec- tion (TS) and attribute classification (TA), event de- tection (ES) and attribute classification (EA), tem- poral relation detection and classification of an event with respect to the Document Creation Time (DR), and, finally, narrative container relation identifica- tion (CR). Systems are evaluated …

Content truncated to 500 words