Generating adversarial negative responses by using large-scale language model

Abstract

Dialogue response selection models typically predict an appropriate response relying on the context-response content similarity. However, the selection model with over-reliance only on superficial features is vulnerable to adversarial responses that are semantically similar but irrelevant to dialogue context. Recent studies have shown that leveraging these adversarial responses as negative training samples is useful for improving the robustness of the selection model. Nevertheless, existing methods often require further fine-tuning for data creation or have limited scalability. To overcome these limitations, this paper proposes a simple but effective method for generating adversarial negative responses leveraging a large-scale language model. Our method can generate realistic negative responses only with a few human-written examples and a prompt designed to optimize generation quality. Experimental results on the dialogue selection task show that our method outperforms existing synthesizing methods for creating negative responses. Synthetic quality analyses and ablation studies prove that our method is scalable and can generate high-quality negative responses. These results suggest that our method can be an effective alternative to human annotators in generating adversarial responses.