Emotional Text Generation with Hard Constraints

Transformer-based pre-training language model can generate high-quality natural language. This kind of language model has also achieved good results in the task of generating emotional text However, most of the existing emotional text generation methods use soft constraints to guide the text to output the target emotion and focus on controlling the emotional attributes of the whole text, but there is still a lack of hard control over the aspect emotion at the word and phrase level In order to solve the above problems, this paper proposes an emotional text generation method under the restriction of hard constraints Firstly, aspect emotion analysis is used to extract the aspect words and emotion words of the sentence, and then the aspect words and emotion words of the target emotion are selected as the hard constraint input of the pre-training language model to reconstruct the complete sentence. In this paper, a new word weight calculation method is designed to make the model generate important words first Experimental results show that the sentences generated by this method not only have aspect-level emotion, but also significantly improve the evaluation indexes of text quality and diversity.