This thesis describes a machine learning approach to temporal relation identification. The task is also called temporal ordering, which identifies the relation between two event or time expressions in documents. For document understanding, it is essential to understand the temporal relation between expressions in documents. Because some critical information such as causal information are based on temporal information in understanding. Although recent works have tried various methods for the tasks, any works have not obtained enough accuracy. Our goal is to reconsider the three relations defined in the TempEval 2007 shared task: temporal relations between an event and a time expression appearing in the same sentence; temporal relations between an event and the document creation time (DCT); and temporal relations between the main events of adjacent sentence. %a contiguous pair of matrix verbs.
One of the reasons why recent works does not work well is that recent works focus on only two target expressions and choose local optimization for these pairs. In this thesis, we focus on the dependency among the three tasks and try to optimize our models globally. We propose two models for the global optimization. The first model is to use a new feature called temporal relation paths. The paths work as semi-transitive constraints and contribute an improvement of the performance. The second is to apply Markov Logic Networks(MLNs) to the tasks and to solve them by a joint learning method. This approach includes logically full-transitive constraints. Due to the advantages of MLNs, we can use these constraints as soft-constraints. These constraints relatively work well notwithstanding the smallness of the training data. Our approaches are innovative in that we jointly solve the three tasks. We compare our results with the final TempEval results to show the effectiveness of our methods.