GFDC: Graph Function Dependence for Logically Consistent Dialogue Response Beyond Persona Data

Abstract

The rigorous logical consistency between current and historical responses during dialogue is the basis of an excellent human speaker, and it is also essential for dialogue agents. Although maintaining consistent personas has made tremendous advancements, they are still limited to the scale of persona data. Then it is necessary to study further how to lead to generate consistent responses when there are beyond personas or even without personas. In this work, we show how the problems can be tackled by consistent dialogue generation turn into Graph Functional Dependencies of a consistent graph schema representation and utilized, called GFDC. Specifically, the model consists of an XLNet-based consistency module and a Transformer-based generation module. The consistency module is for learning the representation of the consistent graph schema. The generation module is for generating responses using consistent graph schema representations. In particular, to learn a consistent graph schema in the dialogue data, we construct the dialogue data as a graph. Then introduce the graph of general knowledge for constructing the consistent graph schema. Both automatic and human evaluations show that our model outperforms strong baselines in response quality and consistency under fewer persona data settings.

Tasks

Reproductions