Negative Training for Neural Dialogue Response Generation

TL;DR

This paper introduces a 'Negative Training' framework that fine-tunes dialogue models by penalizing undesirable responses, significantly reducing malicious outputs and increasing response diversity.

Contribution

The paper proposes a novel negative training method that identifies and penalizes undesirable responses to improve dialogue response quality.

Findings

Reduces malicious response generation

Increases response diversity

Decreases generic responses

Abstract

Although deep learning models have brought tremendous advancements to the field of open-domain dialogue response generation, recent research results have revealed that the trained models have undesirable generation behaviors, such as malicious responses and generic (boring) responses. In this work, we propose a framework named "Negative Training" to minimize such behaviors. Given a trained model, the framework will first find generated samples that exhibit the undesirable behavior, and then use them to feed negative training signals for fine-tuning the model. Our experiments show that negative training can significantly reduce the hit rate of malicious responses, or discourage frequent responses and improve response diversity.