Self-Supervised Logic Induction for Explainable Fuzzy Temporal Commonsense Reasoning

Understanding temporal commonsense concepts, such as times of occurrence and durations is crucial for event-centric language understanding. Reasoning about concepts in a complex context requires reasoning over both the stated world knowledge that underlines it. A recent study shows massive pre-trained LM still struggle with under contexts (e.g., dialog) because they only implicitly encode relevant fail to explicitly uncover underlying logical compositions inference, thus may not be robust enough. In this work, we propose augment LMs logic induction ability, which frames by defining three modular components: dependency inducer concept defuzzifier validator. The former two components disentangle explicit/implicit between across (before, after, ...) specific meaning fuzzy respectively, while validator combines intermediate clues contextual concepts. Extensive experimental results on TIMEDIAL, challenging dataset dialog, show our method, Logic Induction Enhanced Contextualized TEmporal (LECTER), can yield great improvements traditional model reasoning.