An Emotionally Controllable Reply Generation Method Using Fine-tuning and Re-ranking Strategies

Abstract

An emotion-controllable reply generation method using fine-tuning and re-ranking strategies, using a fine-tuning algorithm to train a large-scale open-domain language pre-training model GPT‑2, that is, using open-domain multi-round dialogue corpus training with emotional category labels GPT-2, through the mixed training of dialogue sentences and their corresponding emotional category labels, the language model learns the dependency relationship between semantics and emotion, and obtains the EmoGPT model. Using the EmoGPT model, according to the specified emotional category labels, the generated K replies Perform emotion reordering to obtain the final generated reply. The present invention solves the problem that the emotion of the generated reply is inconsistent with the specified emotion due to the lack of emotion control method in the dialog generation model in the case of multiple rounds of dialogue, and realizes the emotion of multiple rounds of dialogue. Controlling reply generation improves the emotional consistency of dialogue-generated replies.

Description

technical field [0001] The invention belongs to the technical field of artificial intelligence and relates to the generation of emotional dialogues, in particular to an emotion controllable reply generation method using fine-tuning and reordering strategies. Background technique [0002] Dialogue generation refers to the generation of reply sentences word by word or word by word according to the given dialogue context. The most basic requirement for the generated reply statement is that the grammar is correct and the language expression is natural and fluent. Emotional dialogue generation puts higher requirements on the reply sentences generated by the model. In addition to meeting the basic grammar and language expression requirements, it also needs to be full of emotion. The emotional controllability of generated responses is an important research direction in emotional dialogue generation. [0003] Previous research work was mainly based on the Seq2Seq model, which impr...

AI Technical Summary

Problems solved by technology

[0005] In order to overcome the shortcomings of the above-mentioned prior art, the object of the present invention is to provide a method for generating emotionally controllable responses using fine-tuning and reordering strategies. Reply, the lack of emotion control method in the dialogue generation model leads to the problem that the emotion of the generated reply is inconsistent with the specified emotion, and the "pre-training + fine-tuning" strategy is applied to the generation of emotion controllable reply

Images