Automatic Extraction of Time Expressions Accross Domains in French Narratives

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 492–498, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. Automatic Extraction of Time Expressions Across Domains in French Narratives Mike Donald Tapi Nzali Université de Montpellier LIRMM, I3M Montpellier, France Xavier Tannier LIMSI-CNRS UPR 3251 Univ. Paris-Sud 91403 Orsay, France xtannier@limsi.fr Aurélie Névéol LIMSI-CNRS UPR 3251 Rue John von Neuman 91403 Orsay, France neveol@limsi.fr Abstract The prevalence of temporal references across all types of natural language utter- ances makes temporal analysis a key is- sue in Natural Language Processing. This work adresses three research questions: 1/is temporal expression recognition spe- cific to a particular domain? 2/if so, can we characterize domain specificity? and 3/how can subdomain specificity be inte- grated in a single tool for unified temporal expression extraction? Herein, we assess temporal expression recognition from doc- uments written in French covering three domains. We present a new corpus of clin- ical narratives annotated for temporal ex- pressions, and also use existing corpora in the newswire and historical domains. We show that temporal expressions can be ex- tracted with high performance across do- mains (best F-measure 0.96 obtained with a CRF model on clinical narratives). We argue that domain adaptation for the ex- traction of temporal expressions can be done with limited efforts and should cover pre-processing as well as temporal specific tasks. 1 Introduction References to phenomena occurring in the world and their temporal characterization can be found in natural language utterances across domains, genres and languages. Temporal analysis is a key issue in natural language processing that has been receiving increasing attention in recent years. Many efforts in this direction focused on newswire text in English. The focus on this language and domain was in part guided by the availability of the TimeBank corpus (Pustejovsky et al., 2003) used in evaluation campaigns such as TempE- val (Verhagen et al., 2007). More recent efforts have extended the initial work on English and ad- dressed other languages such as Chinese (Li et al., 2014), French (Moriceau and Tannier, 2014), Ara- bic, Italian, Spanish, and Vietnamese (Strötgen et al., 2014a). A study of three domain corpora in English in addition to the newswire domain (SMS, historical narratives and clinical trial abstracts) yielded interesting insight to extend the normal- ized representation of temporal expressions (Ströt- gen and Gertz, 2012). This work was then ap- plied to cover historical narratives in an additional seven languages. One key finding was that domain specificity could differ between languages (Ströt- gen et al., 2014b). This prompts the need to study temporal analysis across domains in a variety of languages in order to adequately characterize each domain and language pairs. The clinical domain has been addressed dur- ing the 2012 i2b2 challenge (Sun et al., 2013b), with a task on temporal relation extraction from clinical narratives. This task used a corpus of clinical notes in English annotated with tempo- ral information (Sun et al., 2013a) based on ISO- TimeML (Pustejovsky et al., …

Content truncated to 500 words