Classifying Temporal Relations with Rich Linguistic Knowledge

Proceedings of NAACL-HLT 2013, pages 918–927, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Classifying Temporal Relations with Rich Linguistic Knowledge Jennifer D’Souza and Vincent Ng Human Language Technology Research Institute University of Texas at Dallas Richardson, TX 75083-0688 {jld082000,vince}@hlt.utdallas.edu Abstract We examine the task of temporal relation clas- sification. Unlike existing approaches to this task, we (1) classify an event-event or event- time pair as one of the 14 temporal relations defined in the TimeBank corpus, rather than as one of the six relations collapsed from the original 14; (2) employ sophisticated linguis- tic knowledge derived from a variety of se- mantic and discourse relations, rather than fo- cusing on morpho-syntactic knowledge; and (3) leverage a novel combination of rule-based and learning-based approaches, rather than re- lying solely on one or the other. Experiments with the TimeBank corpus demonstrate that our knowledge-rich, hybrid approach yields a 15–16% relative reduction in error over a state-of-the-art learning-based baseline sys- tem. 1 Introduction Recent years have seen a surge of interest in tem- poral information extraction (IE). Temporal relation classification, one of the most important temporal IE tasks, involves classifying a given event-event pair or event-time pair as one of a set of predefined temporal relations. The creation of the TimeBank corpus (Pustejovsky et al., 2003) and the organiza- tion of the TempEval-1 (Verhagen et al., 2007) and TempEval-2 (Verhagen et al., 2010) evaluation ex- ercises have facilitated the development and evalua- tion of temporal relation classification systems. Our goal in this paper is to advance the state of the art in temporal relation classification. Our work differs from existing work with respect to both the complexity of the task we are addressing and the ap- proach we adopt. Regarding task complexity, rather than focus on six temporal relations as is typically done in previous work (see Section 2 for more infor- mation), we address an arguably more challenging version of the task where we consider all the 14 re- lations originally defined in the TimeBank corpus. Our approach to temporal relation classification can be distinguished from existing approaches in two respects. The first involves a large-scale ex- pansion of the linguistic features made available to the classification system. Recall that exist- ing approaches have relied primarily on morpho- syntactic features as well as a few semantic fea- tures extracted from WordNet synsets and VerbO- cean’s (Chklovski and Pantel, 2004) semantic rela- tions. On the other hand, we propose not only novel lexical and grammatical features, but also sophis- ticated features involving semantics and discourse. Most notably, we propose (1) semantic features en- coding a variety of semantic relations, including PropBank-style predicate-argument relations as well as those extracted from the Merriam-Webster dictio- nary, and (2) discourse features encoding automat- ically computed Penn Discourse TreeBank (PDTB) style (Prasad et al., 2008) discourse relations. Second, while the vast majority of existing ap- proaches to temporal relation classification are learning-based, we propose a system architecture in which we combine a learning-based approach and …

Content truncated to 500 words