Exploiting timegraphs in temporal relation classification

Abstract

Most of the recent work on machine learning-based temporal relation classification has been done by considering only a given pair of temporal entities (events or temporal expressions) at a time. Entities that have temporal connections to the pair of temporal entities under inspection are not considered even though they provide valuable clues to the prediction. In this paper, we present a new approach for exploiting knowledge obtained from nearby entities by making use of timegraphs and applying the stacked learning method to the temporal relation classification task. By performing 10-fold cross validation on the Timebank corpus, we achieved an F1 score of 59.61% based on the graph-based evaluation, which is 0.16 percentage points higher than that of the local approach. Our system outperformed the state-of-the-art system that utilizes global information and achieved about 1.4 percentage points higher accuracy.