Based on Text Augmentation Personalized Dialog Generation With Persona-Sparse Data

Introducing persona information in dialogue generation helps to make replies more human-like. However, how to solve the problem of ignoring the persona of the speakers in dialogue generation on persona-spare data, and only focusing on dialogue information is still a problem. In this paper, a text augmentation method for persona-sparse data is proposed. By performing text augmentation operations such as text generation on sparse persona information, the problem of ignoring persona in reply generation is improved. Two encoders are used to encode the context dialogue and character information respectively, and the decoder use multi-head attention mechanism to fully integrate dialogue context and persona. Experiments show, compared with the baseline model, for persona-spares data, the text augmentation method and dialogue generation model adopted in this paper have more advantages, and can generate responses that are consistent with the dialogue context and have personal information.