Abstract

Consistency is a long standing issue faced by dialogue models. In this paper,we frame the consistency of dialogue agents as natural language inference (NLI)and create a new natural language inference dataset called Dialogue NLI. Wepropose a method which demonstrates that a model trained on Dialogue NLI can beused to improve the consistency of a dialogue model, and evaluate the methodwith human evaluation and with automatic metrics on a suite of evaluation setsdesigned to measure a dialogue model's consistency.

 

Quick Read (beta)