Open domain dialogue reply method and system based on deep reinforcement learning

Abstract

The invention belongs to the field of artificial intelligence, relates to natural language generation and emotion analysis, and provides an open domain dialogue reply method and system based on deep reinforcement learning in order to introduce content coherence control and emotion rationality control to an open domain dialogue system. The method comprises the steps that dialogue input content is acquired and preprocessed; the preprocessed information is input into a dialogue reply control model to be processed; the dialogue reply control model comprises a dialogue generation module, a contentcoherence control module and an emotion rationality control module, wherein the dialogue generation module is used for generating a dialogue reply, the content coherence control module is used for enabling dialogue context content to be coherent, and the emotion rationality control module is used for sentence emotion classification and judging whether reply emotion is reasonable or not; and dialogue replies with coherent content and reasonable emotion are output.

Description

technical field [0001] The invention belongs to the field of artificial intelligence and relates to natural language generation and sentiment analysis. In order to introduce content coherence control and emotional rationality control to an open domain dialogue system, an open domain dialogue reply method and system based on deep reinforcement learning is proposed. Background technique [0002] Open-domain dialog system control aims to add more control and determinism to the process of generating dialog responses by neural networks. Before the neural network-based generative dialogue system, the mainstream dialogue system was retrieval-based, and all candidate dialogue responses came from the database. For a new dialogue context, when there is no suitable context in the database, the system’s The performance will be reduced, and the lack of diversity of responses is not suitable for open-domain dialogue scenarios. The generative dialogue system can generate responses outside...

AI Technical Summary

Problems solved by technology

[0004]Using real replies as supervision information enables the neural network to learn how to generate complete and fluent reply sentences. paid more attention to the reply

The supervised method that indirectly introduces emotion annotation can achieve better results on specific dialogue datasets, but it is difficult to extend to other datasets. At present, there is no method that can be applied to any open domain dialogue dataset.

And directly use the emotion of the dialogue reply as supervisory information to control the emotion of the reply. Another problem is that it leads to the generation of an emotionally safe reply, that is, the generated reply meets the expected emotion, but the content is relatively vague, which is not conducive to the progress of the dialogue.

Images