PRIMAL: Prompting Multiple Language Models for Low-Resource Diverse Response Generation

Low-resource conversation models are becoming increasingly important. Existing conversation models tend to generate uninformative responses that lack diversity, especially when the training data are limited. Researchers address this issue by refining training objectives or incorporating additional data sources. Learned from masses of diverse texts, pre-trained language models show their potential in text generation. However, language models (LMs) learn by approximating the distribution of a single ground truth, which limits their ability to capture the ‘one-to-many’ characteristic essential for generating diverse utterances in conversation tasks. In this paper, we propose to leverage multiple pre-trained LMs in a unified framework to diversify low-resource dialogue generation. We apply a prompt-based method to exploit pre-trained LMs for more information. To generate responses referring to multiple LMs, we design a multi-LMs decoding algorithm where LMs interact with each other at each step. We also propose a multi-LMs-based entropy for response generation to further enhance diversity. Experiments on two datasets demonstrate that our method outperforms competitive baselines in low-resource settings.