Comparison of different algebras for inducing the temporal structure of texts

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 250–258, Beijing, August 2010 Comparison of different algebras for inducing the temporal structure of texts Pascal Denis† † Alpage Project-Team INRIA & Université Paris 7 pascal.denis@inria.fr Philippe Muller†,⋄ ⋄ IRIT Université de Toulouse muller@irit.fr Abstract This paper investigates the impact of us- ing different temporal algebras for learn- ing temporal relations between events. Specifically, we compare three interval- based algebras: Allen (1983) algebra, Bruce (1972) algebra, and the algebra de- rived from the TempEval-07 campaign. These algebras encode different granular- ities of relations and have different infer- ential properties. They in turn behave dif- ferently when used to enforce global con- sistency constraints on the building of a temporal representation. Through various experiments on the TimeBank/AQUAINT corpus, we show that although the TempE- val relation set leads to the best classifica- tion accuracy performance, it is too vague to be used for enforcing consistency. By contrast, the other two relation sets are similarly harder to learn, but more use- ful when global consistency is important. Overall, the Bruce algebra is shown to give the best compromise between learn- ability and expressive power. 1 Introduction Being able to recover the temporal relations (e.g., precedence, inclusion) that hold between events and other time-denoting expressions in a docu- ment is an essential part of natural language un- derstanding. Success in this task has important implications for other NLP applications, such as text summarization, information extraction, and question answering. Interest for this problem within the NLP com- munity is not new (Passonneau, 1988; Webber, 1988; Lascarides and Asher, 1993), but has been recently revived by the creation of the TimeBank corpus (Pustejovsky et al., 2003), and the orga- nization of the TempEval-07 campaign (Verhagen et al., 2007). These have seen the development of machine learning inspired systems (Bramsen et al., 2006; Mani et al., 2006; Tatu and Srikanth, 2008; Chambers and Jurafsky, 2008). Learning the temporal stucture from texts is a difficult problem because there are numerous in- formation sources at play (in particular, seman- tic and pragmatic ones) (Lascarides and Asher, 1993). An additional difficulty comes from the fact that temporal relations have logical proper- ties that restrict the consistent graphs that can be built for a set of temporal entities (for instance the transitivity of inclusion and temporal prece- dence). Previous work do not attempt to directly predict globally coherent temporal graphs, but in- stead focus on the the simpler problem of label- ing pre-selected pairs of events (i.e., a task that directly lends itself to the use of standard classifi- cation techniques). That is, they do not consider the problem of linking pairs of events (i.e., of de- termining which pairs of events are related). Given the importance of temporal reasoning for determining the temporal structure of texts, a natural question is how to best use it within a machine-based learning approach. Following (Mani et al., 2006), prior approaches exploit tem- poral inferences to enrich the set of training …

Content truncated to 500 words