Temporal Relation Identification with Endpoints

Temporal relation classification task has issues of fourteen target relations, skewed distribution of the target relations, and relatively small amount of data. To overcome the issues, methods such as merging target relations and increasing data size with closure algorithm have been used. However, the method using merged relations has a problem on how to recover original relations. In this paper, a new reduced-relation method is proposed. The method decomposes a target relation into four pairs of endpoints with three target relations. After classifying a relation of each endpoint pair, four classified relations are combined into a relation of original fourteen target relations. In the combining step, two heuristics are examined.