Long-Distance Time-Event Relation Extraction

International Joint Conference on Natural Language Processing, pages 1330–1338, Nagoya, Japan, 14-18 October 2013. Long-Distance Time-Event Relation Extraction Alessandro Moschitti QCRI, Qatar Foundation, Doha, Qatar DISI, University of Trento, Povo (TN), Italy amoschitti@qf.org.qa moschitti@disi.unitn.it Siddharth Patwardhan and Chris Welty IBM T. J. Watson Research Center Yorktown Heights NY 10598 siddharth@us.ibm.com welty@us.ibm.com Abstract This paper proposes state-of-the-art mod- els for time-event relation extraction (TERE). The models are specifically de- signed to work effectively with relations that span multiple sentences and para- graphs, i.e., inter-sentence TERE. Our main idea is: (i) to build a computational representation of the context of the two target relation arguments, and (ii) to en- code it as structural features in Support Vector Machines using tree kernels. Re- sults on two data sets – Machine Read- ing and TimeBank – with 3-fold cross- validation show that the combination of traditional feature vectors and the new structural features improves on the state of the art for inter-sentence TERE by about 20%, achieving a 30.2 F1 score on inter- sentence TERE alone, and 47.2 F1 for all TERE (inter and intra sentence combined). 1 Introduction Time-Event Relation Extraction (TERE) is the task of linking event mentions and relation men- tions to occurrences of “time stamps” in text. We define it as follows: given a set of textual expres- sions denoting events and relations, and a set of time expressions in the same text document, find all instances of temporal relations between ele- ments of the two input sets. A relation between an event and a time expression indicates that the event occurs within the temporal context specified by the time expression, for example, the following sentence: He succeeded James A. Taylor, who stepped down as chairman, president and chief executive in March for health reasons; the appointment took effect Nov. 13. conveys two different events, succession and step- ping down, linked to the time stamps, Nov. 13 and March, respectively. In this paper, we focus on the task of linking time expressions to events, i.e., we carry out a classification task, where, for each possible pair of (event/relation, time) in a document, the clas- sifier decides whether there exists a link between the two. In particular, we assume that the event mentions, relation mentions and time expressions are given to us by an external process. There is a large body of work on the above topics and they remain difficult problems, but we use human an- notated mentions and expressions as input to our models since TERE itself is a relatively new prob- lem in this context. Previous work in TempEval-2 (Verhagen et al., 2010) and our work (Hovy et al., 2012) have shown that accurate relation classifiers can be modeled with supervised approaches, pro- vided that the expressions are limited to be in the same sentence. In contrast, there is almost no pre- vious work on inter-sentence TERE (ISTERE), for three main reasons: • Across a document, the number of time-event pairs to consider is large, as they are …

Content truncated to 500 words