A Prompt-based Diverse Response Generation Approach via Multiple Language Model Cooperation

Anonymous

Abstract: Neural conversation models are of growing significance and achieve remarkable performance. However, existing conversation models tend to generate uninformative responses that lack diversity. Researchers alleviate this issue by refining the training objective or importing randomness into generation. 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 the single ground truth, thereby failing to model the “one-to-many” characteristic of conversation tasks, which is crucial to generate diverse utterances. In this paper, we propose to leverage multiple pre-trained LMs in a unified framework to improve dialogue diversity. We apply a prompt-based method to exploit pre-trained LMs. To generate responses referring to multiple LMs, we design a multi-LMs decoding algorithm where LMs interact with each other at each step. To further enhance the diversity, we propose a multi-LMs-based entropy with the help of multiple LMs and apply it to the training and inference. Experiments on two datasets verify that our method outperforms competitive baselines.

Paper Type: long