A Weakly Supervised Approach to Train Temporal Relation Classifiers and Acquire Regular Event Pairs Simultaneously

摘要

Capabilities of detecting temporal relations between two events can benefitmany applications. Most of existing temporal relation classifiers were trainedin a supervised manner. Instead, we explore the observation that regular eventpairs show a consistent temporal relation despite of their various contexts,and these rich contexts can be used to train a contextual temporal relationclassifier, which can further recognize new temporal relation contexts andidentify new regular event pairs. We focus on detecting after and beforetemporal relations and design a weakly supervised learning approach thatextracts thousands of regular event pairs and learns a contextual temporalrelation classifier simultaneously. Evaluation shows that the acquired regularevent pairs are of high quality and contain rich commonsense knowledge anddomain specific knowledge. In addition, the weakly supervised trained temporalrelation classifier achieves comparable performance with the state-of-the-artsupervised systems.