Identifying Temporal Relations by Sentence and Document Optimizations
Katsumasa Yoshikawa, Masayuki Asahara, Ryu Iida · International Conference on Computational Linguistics · 2012
This paper presents a temporal relation identification method optimizing relations at sentence and document levels. Temporal relation identification is to identify temporal orders between events and time expressions. Various approaches of this task have been studied through the shared tasks TempEval (Verhagen et al., 2007, 2010). Not only identifying each temporal relation independently, some works also try to find multiple temporal relations jointly by logical constraints in Integer Linear Programming (Chambers and Jurafsky, 2008; Do et al., 2012) or Markov Logic Networks (Yoshikawa et al., 2009; Ling and Weld, 2010; Ha et al., 2010). Though previous joint approaches optimize temporal relations in an entire document, we first optimize our model at sentence level and then extend it to document level. We consider that different types of temporal relations require different types of optimizations. By evaluating our sentence and document optimized model on the TempEval-2 data, we show that our approaches can achieve competitive performance in comparison to other state-of-the-art systems. We find that the sentence and document optimized model has strong tasks in TempEval-2, respectively.