Human–machine dialogue modelling with the fusion of word- and sentence-level emotions

Abstract

Emotion intelligence plays an important role in building a successful human–machine dialogue system. However, the extreme difficulty of capturing emotional information of social text, as well as the weakness of generative models for learning emotional expression, limits the performance of existing dialogue system. Combining dictionary matching and machine learning, this paper proposes a generative model, which fuses the word- and sentence-level emotions, to model the dialogue text and learn emotional expression. The model first obtains the emotional embedding of each word through dictionary matching, then concatenates the emotional word embedding with its traditional word embedding, and the finally formed vector is taken as the input of the encoder. In order to control the emotional feature of the generated response, our model employs a BernoulliNB-based classifier to extract the emotional feature of the post, which is used as the attribute of the original text, and subsequently adds it to the decoder. For further significantly improving the emotional expression of the response, the model leverages a discriminator to constrain the latent variable, which enables the latent variable to encode better the information of emotional feature of the post. With this model, we can generate the dialogue text that is consistent with the original emotion. Experimental results on our Twitter dataset demonstrate that our model outperforms several state-of-the-art methods in the emotion accuracy and quality of generated texts.