Improving Neural Response Diversity with Frequency-Aware Cross-Entropy Loss

Sequence-to-Sequence (Seq2Seq) models have achieved encouraging performance on the dialogue response generation task. However, existing Seq2Seq-based response generation methods suffer from a low-diversity problem: they frequently generate generic responses, which make the conversation less interesting. In this paper, we address the low-diversity problem by investigating its connection with model overconfidence reflected in predicted distributions. Specifically, we first analyze the influence of the commonly used Cross- Entropy (CE) loss function, and find that CE prefers high-frequency tokens, which results in low-diversity responses. We propose a Frequency-Aware Cross-Entropy (FACE) loss function that improves over the CE loss by incorporating a weighting mechanism conditioned on token frequency. Extensive experiments on benchmark datasets show that FACE is able to substantially improve the diversity of existing state-of-the-art Seq2Seq response generation methods, in terms of both automatic and human evaluations.