Determining the Temporal Order of Events in Natural Language Using Learned Commonsense Knowledge
David A. Ziegler · 2014
Natural language understanding systems, for example question answering or document summarization, can be expected to benefit from knowledge about the temporal order of events in a text. Traditionally, only linguistic features (for example verb tense, syntactic structure etc.) are used to determine the temporal order of events. However, when humans read a piece of text they use commonsense knowledge additional to the information in the text to reason about temporal order (for example in a restaurant setting the event eat usually precedes pay even if this information cannot be deduced from the text). This thesis tests whether using commonsense knowledge about the prototypical order of events can improve the performance of a temporal ordering system. To this end, the output of Modi & Titov’s (2014a) classifier for learning prototypical temporal order of events is used as an additional feature for Bethard’s (2013) temporal order classifier ‘ClearTK TimeML’ and is evaluated on the TempEval-3 task (Uzzaman et al., 2013). The performance using both the ‘commonsense’ features as well as ClearTK’s linguistic features is similar to using linguistic features only. However, the classifier learning ‘commonsense knowledge’ is only trained on the same data that is used as training data for ClearTK while it should be trained on a much larger separate data set in order to be able to call it ‘commonsense knowledge’. The ‘commonsense’ feature taken by itself reaches a similar performance as the linguistic features, although using a very different approach, which shows that this feature is predictive. The results suggest that using ‘commonsense’ knowledge is a promising approach to further pursue, however, this knowledge should be learned from much larger data sets.