Research And Implement On Temporal Relation Identification Between Events

The research on temporal relation between events is becoming more and more important in Natural Language Processing, such as Question and Answering System, Text Summarization. This paper is originated from the Nature Science Foundation Project―Research on Modeling Typical Event Process‖. The aim of this research is to establish the time sequence relationship between events.First of all, we used statistical machine learning methods to construct an English event temporal relation classification model based on OTC corpus. Then about the data sparse problem of the OTC corpus, the paper mitigate it with temporal reasoning. Make the event temporal relation classification model run on a larger raw corpus, build an event temporal relation repository is the ultimate goal of this paper. The major contents of this paper are described as followings:(1)Construct an English event temporal relation classification model based on OTC corpus. Including temporal relation mapping, extract the pure text from XML file and the construction of feature space. Finally, the average accuracy of the model is 60.15%. Then we analyze each feature in the feature space, make sure what they can contribute to the classification model.(2) Use temporal reasoning methods to mitigate data sparse problem. About the data sparse problem of OTC corpus, the paper makes use of temporal reasoning to mitigate it. The discussion of temporal reasoning focuses on two points: the construction of temporal reasoning rules and Constraint Propagation Algorithms. The extended training samples are nine times as many as OTC's. To test the effect of increasing training samples to the experiment, the research selects 12124 samples from the extended dataset as OTC's temporal relation distribution. The average accuracy of the new model is 67.57%.(3) Try to use Wiki corpus as the fundamental dataset to construct event temporal relation repository. The paper will process Wiki corpus to consistent with the standard input specification of Evita. Three steps are taken to process the corpus: extract single document from the corpus, pure the single document, output the single document to XML file.(4) Construct event temporal relation repository. After processing the Wiki Corpus, we will take the following steps: Event identification, construct event temporal relation chain in a single document, the patient of event identification, events-based text similarity computing, consolidate lists of single document's temporal relation chain. Finally, we get several event temporal relation directed graphs, each graph contains some associated events.