Classifying Temporal Relations with Simple Features

Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 308–317, Gothenburg, Sweden, April 26-30 2014. c⃝2014 Association for Computational Linguistics Classifying Temporal Relations with Simple Features Paramita Mirza Fondazione Bruno Kessler and University of Trento Trento, Italy paramita@fbk.eu Sara Tonelli Fondazione Bruno Kessler Trento, Italy satonelli@fbk.eu Abstract Approaching temporal link labelling as a classification task has already been ex- plored in several works. However, choos- ing the right feature vectors to build the classification model is still an open is- sue, especially for event-event classifica- tion, whose accuracy is still under 50%. We find that using a simple feature set re- sults in a better performance than using more sophisticated features based on se- mantic role labelling and deep semantic parsing. We also investigate the impact of extracting new training instances using in- verse relations and transitive closure, and gain insight into the impact of this boot- strapping methodology on classifying the full set of TempEval-3 relations. 1 Introduction In recent years, temporal processing has gained in- creasing attention within the NLP community, in particular since TempEval evaluation campaigns have been organized on this topic (Verhagen et al., 2007; Verhagen et al., 2010; UzZaman et al., 2013). In particular, the classification of tem- poral relations holding between entities such as events and temporal expressions (timex) is crucial to build event timelines and to reconstruct the plot of a story. This could be exploited in decision sup- port systems and document archiving applications, among others. In this work we focus on the problem of clas- sifying temporal relation types, assuming that the links between events and time expressions are al- ready established. This task is part of Tempeval-3 evaluation campaign, hence we follow the guide- lines and the dataset provided by the organizers, so that we can compare our system with other systems participating in the challenge. Recent works have tried to address this complex classifi- cation task by using sophisticated features, based on deep parsing, semantic role labelling and dis- course parsing (D’Souza and Ng, 2013; Laokulrat et al., 2013). We argue that a simpler approach, based on lexico-syntactic features, achieves com- parable results, while reducing the processing time needed to extract the features. Besides, the perfor- mance of complex NLP tools may strongly vary when moving to new domains, affecting in turn the classification performance, while our approach is likely to be more stable across different domains. Our features include some basic information on the position, the attributes and the PoS tags of events and timexes, as well as other information obtained from external lexical resources such as a list of typical event durations and a list of temporal signals. The few processing steps required include PoS-tagging, dependency parsing and the seman- tic tagging of connectives (based on the parser out- put). We also investigate the impact of extending the number of training instances through inverse rela- tions and transitive closure, which is a ‘simplified’ version of temporal closure covering only entities …

Content truncated to 500 words