Classification and Acquisition of Contradictory Event Pairs using Crowdsourcing

We propose a taxonomy of contradictory event pairs and a method for building a database of such pairs. When a dialog system participates in an open-domain conversation with a human, it is important to avoid the generation of utterances that conflict with the context of the dialog. Here, we refer to a pair of events that are not able to co-occur or that are not inconsistent with each other as a contradictory event pair. In this study, we collected contradictory event pairs using crowdsourcing and constructed a taxonomy of such pairs. We also built a large-scale database of Japanese contradictory event pairs for each class using crowdsourcing. This database will be used for consistent utterance generation in dialog systems.