Learning to improve persona consistency in conversation generation with information augmentation

Abstract

In an open-domain conversation system, maintaining consistent persona is a key factor to earn trust from users and engage users in the conversation. Existing methods suffer from the issue that only sparse persona-relevant signals are available in the target responses, leading to the generation of responses with inconsistent persona. To address the issue, in this paper, we propose two methods to augment persona learning signals for persona preservation. At the sentence level, we develop a dual variational learning model based on the bidirectional encoder representations from transformers (i.e., BERT), which enriches persona signals with relevant persona sentences, in addition to target responses. Therefore, both the encoder part and the latent variable can be guided to learn consistent persona features through back-propagation of losses, which will drive response decoding towards consistent persona expression. At the word level, we propose a persona-based calibration network, which is used to amplify the influence of persona-relevant words in target responses. The experimental results show that our developed model outperforms the strong baseline algorithms by large margins and effectively promotes persona consistency in conversation generation.