A Structured Learning Approach to Temporal Relation Extraction

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 1027–1037 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics A Structured Learning Approach to Temporal Relation Extraction Qiang Ning1 and Zhili Feng 2 and Dan Roth 1,2 1Department of Electrical and Computer Engineering 2Department of Computer Science University of Illinois, Urbana, IL 61801 {qning2,zfeng6,danr}@illinois.edu Abstract Identifying temporal relations between events is an essential step towards nat- ural language understanding. However, the temporal relation between two events in a story depends on, and is often dic- tated by, relations among other events. Consequently, effectively identifying tem- poral relations between events is a chal- lenging problem even for human annota- tors. This paper suggests that it is im- portant to take these dependencies into ac- count while learning to identify these re- lations and proposes a structured learning approach to address this challenge. As a byproduct, this provides a new perspective on handling missing relations, a known is- sue that hurts existing methods. As we show, the proposed approach results in sig- nificant improvements on the two com- monly used data sets for this problem. 1 Introduction Understanding temporal information described in natural language text is a key component of nat- ural language understanding (Mani et al., 2006; Verhagen et al., 2007; Chambers et al., 2007; Bethard and Martin, 2007) and, following a se- ries of TempEval (TE) workshops (Verhagen et al., 2007, 2010; UzZaman et al., 2013), it has drawn increased attention. Time-slot filling (Surdeanu, 2013; Ji et al., 2014), storyline construction (Do et al., 2012; Minard et al., 2015), clinical narra- tives processing (Jindal and Roth, 2013; Bethard et al., 2016), and temporal question answering (Llorens et al., 2015) are all explicit examples of temporal processing. The fundamental tasks in temporal process- ing, as identified in the TE workshops, are 1) time expression (the so-called “timex”) extraction and normalization and 2) temporal relation (also known as TLINKs (Pustejovsky et al., 2003a)) ex- traction. While the first task has now been well handled by the state-of-the-art systems (Heidel- Time (Str¨otgen and Gertz, 2010), SUTime (Chang and Manning, 2012), IllinoisTime (Zhao et al., 2012), NavyTime (Chambers, 2013), UWTime (Lee et al., 2014), etc.) with end-to-end F1 scores being around 80%, the second task has long been a challenging one; even the top systems only achieved F1 scores of around 35% in the TE work- shops. The goal of the temporal relation task is to gen- erate a directed temporal graph whose nodes rep- resent temporal entities (i.e., events or timexes) and edges represent the TLINKs between them. The task is challenging because it often re- quires global considerations – considering the en- tire graph, the TLINK annotation is quadratic in the number of nodes and thus very expensive, and an overwhelming fraction of the temporal relations are missing in human annotation. In this paper, we propose a structured learning approach to tem- poral relation extraction, where local models are updated based on feedback from global inferences. The …

Content truncated to 500 words