Event Coreference Resolution by Iteratively Unfolding Inter-dependencies among Events

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2124–2133 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Event Coreference Resolution by Iteratively Unfolding Inter-dependencies among Events Prafulla Kumar Choubey and Ruihong Huang Department of Computer Science and Engineering Texas A&M University (prafulla.choubey, huangrh)@tamu.edu Abstract We introduce a novel iterative approach for event coreference resolution that grad- ually builds event clusters by exploiting inter-dependencies among event mentions within the same chain as well as across event chains. Among event mentions in the same chain, we distinguish within- and cross-document event coreference links by using two distinct pairwise classifiers, trained separately to capture differences in feature distributions of within- and cross- document event clusters. Our event coref- erence approach alternates between WD and CD clustering and combines argu- ments from both event clusters after ev- ery merge, continuing till no more merge can be made. And then it performs fur- ther merging between event chains that are both closely related to a set of other chains of events. Experiments on the ECB+ corpus show that our model outper- forms state-of-the-art methods in joint task of WD and CD event coreference resolu- tion. 1 Introduction Event coreference resolution is the task of iden- tifying event mentions and clustering them such that each cluster represents a unique real world event. The capability of resolving links among co- referring event identities is vital for information aggregation and many NLP applications, includ- ing topic detection and tracking, information ex- traction, question answering and text summariza- tion (Humphreys et al., 1997; Allan et al., 1998; Daniel et al., 2003; Narayanan and Harabagiu, 2004; Mayfield et al., 2009; Zhang et al., 2015). Yet, studies on event coreference are few com- pared to the well-studied entity coreference reso- lution. Event mentions that refer to the same event can occur both within a document (WD) and across multiple documents (CD). One common practice (Lee et al., 2012) to approach CD coreference task is to resolve event coreference in a mega- document created by concatenating topic-relevant documents, which essentially does not distinguish WD and CD event links. However, intuitively, recognizing CD corefer- ent event pairs requires stricter evidence com- pared to WD event linking because it is riskier to link two event mentions from two distinct docu- ments rather than the same document. In a per- fect scenario where all WD event mentions are properly clustered and their participants and argu- ments are combined within a cluster, CD cluster- ing can be performed with ease as sufficient ev- idences are collected through initial WD cluster- ing. Therefore, another very common practice for event coreference is to first group event mentions within a document and then group WD clusters across documents (Yang et al., 2015). Nonetheless, WD coreference chains are equally hard to resolve. Event mentions in the same document can look very dissimilar (”killed/ VB” and ”murder/ NN”), have event arguments (i.e., participants and spatio-temporal information of an event (Bejan and Harabagiu, 2010)) partially …

Content truncated to 500 words