FSS-TimEx for TempEval-3: Extracting Temporal Information from Text

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 58–63, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics FSS-TimEx for TempEval-3: Extracting Temporal Information from Text Vanni Zavarella Joint Research Centre European Commission 21027 Ispra, Italy vanni.zavarella @jrc.ec.europa.eu Hristo Tanev Joint Research Centre European Commission 21027 Ispra, Italy hristo.tanev @jrc.ec.europa.eu Abstract We describe FSS-TimEx, a module for the recognition and normalization of temporal ex- pressions we submitted to Task A and B of the TempEval-3 challenge. FSS-TimEx was developed as part of a multilingual event ex- traction system, Nexus, which runs on top of the EMM news processing engine. It consists of finite-state rule cascades, using minimalis- tic text processing stages and simple heuris- tics to model the relations between events and temporal expressions. Although FSS-TimEx is already deployed within an IE application in the medical domain, we found it useful to customize its output to the TimeML standard in order to have an independent performance measure and guide further developments. 1 Introduction The FSS-TimEx (Finite State-based Shallow Time Extractor) system participating in TempEval-3 is in- tegrated in the event extraction engine Nexus (Tanev et al., 2008), developed at the EC’s Joint Research Center for extracting event information from on-line news articles gathered by the Europe Media Mon- itor (EMM) news aggregation and analysis family of applications(Steinberger et al., 2009). Nexus is highly multilingual1 and easily portable across do- mains through semi-automatic learning of lexical re- sources. In the domain of epidemiological surveil- lance, the event extraction task required a particu- larly deep temporal information analysis, in order to 1Currently, it covers English, French, Italian, Spanish, Por- tuguese, Turkish, Russian, Arabic. detect temporal relations among event reports and mitigate the classical event duplication problem. As an example, from a report like: The overall death toll has risen to 160 since the beginning of the year, after 2 patients in Gulu and 2 in Masindi died on Tue 5 Dec 2000. a system might be prevented to wrongly sum up the two victim counts (160+4) only if it is made aware of the inclusion relation between the first time interval and the date, which in turn implies normalizing the two temporal expressions. Currently, FSS-TimEx is deployed for French, English and Italian and extensions are foreseen for further languages. Given such requirements for mul- tilinguality, we developed FSS-TimEx using a lin- guistically light-weight approach, applying shallow processing modules only. On the other hand, as we need to extract highly structured information out of the detected temporal expressions, to be used in the subsequent normalization phase, we mostly opted for a rule-based approach, using finite-state gram- mar cascades, rather than machine learning meth- ods. Nonetheless, some of the required lexicons were semi-automatically learned. In our participation in Tasks A and B of the TempEval-3, we experimented with adapting an ex- isting timex recognition module for the English lan- guage, to Spanish. We first describe our system in 2,3 and 4, …

Content truncated to 500 words