Inducing Temporal Relations from Time Anchor Annotation

Proceedings of NAACL-HLT 2018, pages 1833–1843 New Orleans, Louisiana, June 1 - 6, 2018. c⃝2018 Association for Computational Linguistics Inducing Temporal Relations from Time Anchor Annotation Fei Cheng and Yusuke Miyao Research Center for Financial Smart Data National Institute of Informatics 2-1-2 Hitotsubashi, Chiyoda-ku, Tokyo, Japan {fei-cheng, yusuke}@nii.ac.jp Abstract Recognizing temporal relations among events and time expressions has been an essential but challenging task in natural language process- ing. Conventional annotation of judging tem- poral relations puts a heavy load on annota- tors. In reality, the existing annotated corpora include annotations on only ”salient” event pairs, or on pairs in a fixed window of sen- tences. In this paper, we propose a new ap- proach to obtain temporal relations from abso- lute time value (a.k.a. time anchors), which is suitable for texts containing rich temporal in- formation such as news articles. We start from time anchors for events and time expressions, and temporal relation annotations are induced automatically by computing relative order of two time anchors. This proposal shows sev- eral advantages over the current methods for temporal relation annotation: it requires less annotation effort, can induce inter-sentence re- lations easily, and increases informativeness of temporal relations. We compare the em- pirical statistics and automatic recognition re- sults with our data against a previous tempo- ral relation corpus. We also reveal that our data contributes to a significant improvement of the downstream time anchor prediction task, demonstrating 14.1 point increase in overall accuracy. 1 Introduction Temporal information extraction is becoming an active research field in natural language process- ing (NLP) due to the rapidly growing need for NLP applications such as timeline generation and question answering (Llorens et al., 2015; Meng et al., 2017). It has great potential to create many practical applications. For example, SemEval- 2015 Task 4 (Minard et al., 2015) collects news articles about a target entity and the task required participants automatically ordering the events in- volving that entity in a timeline. The timeline rep- resentation of news can help people more easily comprehend a mass of information. This work aims to contribute to such timeline applications by extracting temporal information in specific do- mains like news articles. TimeBank1 (Pustejovsky et al., 2003) is the first widely used corpus with temporal informa- tion annotated in the NLP community. It contains 183 news articles that have been annotated with events, time expressions and temporal relations between events and time expressions. The anno- tation follows the TimeML2 specification (Saurı et al., 2006). Along with the TimeBank and other temporal information corpora, a series of competitions on temporal information extraction (TempEval-1,2,3) (Verhagen et al., 2009, 2010; UzZaman et al., 2012) are attracting growing re- search efforts. A majority of temporal information corpora adopt temporal links (TLINKs) to encode tem- poral information in documents. A TLINK de- notes a temporal relation between mentions, i.e., events, time expressions and document creation time (DCT) (Setzer, 2002). However, annotat- ing TLINKs is a painful work, because annotation candidates are quadratic to the number …

Content truncated to 500 words