Joint Inference for Event Timeline Construction

Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 677–687, Jeju Island, Korea, 12–14 July 2012. c⃝2012 Association for Computational Linguistics Joint Inference for Event Timeline Construction Quang Xuan Do Wei Lu Dan Roth Department of Computer Science University of Illinois at Urbana-Champaign Urbana, IL 61801, USA {quangdo2,luwei,danr}@illinois.edu Abstract This paper addresses the task of construct- ing a timeline of events mentioned in a given text. To accomplish that, we present a novel representation of the temporal structure of a news article based on time intervals. We then present an algorithmic approach that jointly optimizes the temporal structure by coupling local classifiers that predict associations and temporal relations between pairs of tempo- ral entities with global constraints. Moreover, we present ways to leverage knowledge pro- vided by event coreference to further improve the system performance. Overall, our experi- ments show that the joint inference model sig- nificantly outperformed the local classifiers by 9.2% of relative improvement in F1. The ex- periments also suggest that good event coref- erence could make remarkable contribution to a robust event timeline construction system. 1 Introduction Inferring temporal relations amongst a collection of events in a text is a significant step towards vari- ous important tasks such as automatic information extraction and document comprehension. Over the past few years, with the development of the Time- Bank corpus (Pustejovsky et al., 2003) , there have been several works on building automatic systems for such a task (Mani et al., 2006; Chambers and Jurafsky, 2008; Yoshikawa et al., 2009; Denis and Muller, 2011). Most previous works devoted much efforts to the task of identifying relative temporal relations (such as before, or overlap) amongst events (Chambers • −∞ | • t1 | • t2 | • t3 • | t4 | • +∞ | Time −→ I1 I2 I3 e1 e2 e4 e3/e5 e7 − e6 Figure 1: A graphical illustration of our timeline representation. The e’s, t’s and I’s are events, time points and time intervals, respectively. and Jurafsky, 2008; Denis and Muller, 2011), with- out addressing the task of identifying correct asso- ciations between events and their absolute time of occurrence. Even if this issue is addressed, certain restrictions are often imposed for efficiency reasons (Yoshikawa et al., 2009; Verhagen et al., 2010). In practice, however, being able to automatically infer the correct time of occurrence associated with each event is crucial. Such information not only leads to better text comprehension, but also enables fusion of event structures extracted from multiple articles or domains. In this work, we are specifically interested in map- ping events into an universal timeline representa- tion. Besides inferring the relative temporal rela- tions amongst the events, we would also like to au- tomatically infer a specific absolute time of occur- rence for each event mentioned in the text. Unlike previous work, we associate each event with a spe- cific absolute time interval inferred from the text. An example timeline representation is …

Content truncated to 500 words