There Is No Standard Answer: Knowledge-Grounded Dialogue Generation with Adversarial Activated Multi-Reference Learning

TL;DR

This paper addresses the one-to-many challenge in knowledge-grounded dialogue generation by creating a multi-reference dataset and proposing a span-based variational model to improve diversity and relevance in responses.

Contribution

It introduces a multi-reference dataset, new evaluation metrics, and a span-based variational model optimized with a wake-sleep approach for better knowledge diversity.

Findings

The proposed model outperforms baselines in automatic metrics.

Human evaluations favor the diversity and relevance of generated responses.

New metrics effectively assess one-to-many knowledge grounding.

Abstract

Knowledge-grounded conversation (KGC) shows excellent potential to deliver an engaging and informative response. However, existing approaches emphasize selecting one golden knowledge given a particular dialogue context, overlooking the one-to-many phenomenon in dialogue. As a result, the existing paradigm limits the diversity of knowledge selection and generation. To this end, we establish a multi-reference KGC dataset and propose a series of metrics to systematically assess the one-to-many efficacy of existing KGC models. Furthermore, to extend the hypothesis space of knowledge selection to enhance the mapping relationship between multiple knowledge and multiple responses, we devise a span-based variational model and optimize the model in a wake-sleep style with an ameliorated evidence lower bound objective to learn the one-to-many generalization. Both automatic and human evaluations demonstrate the efficacy of our approach.