A method and system for dialogue reply generation based on self-comment sequence learning

Abstract

The invention discloses a dialogue reply generation method and system based on self-comment sequence learning, wherein the dialogue reply generation method includes the following steps: (1) modeling the context of the current dialogue to obtain a contextual semantic vector; (2) ) According to the context semantic vector, establish a dialogue model based on self-review sequence learning; (3) train and test the dialogue model, and obtain the training reward value and test reward value respectively; (4) calculate the difference between the two reward values, Optimize the dialog model by calculating the policy gradient; (5) After the dialog model is trained, output the dialog reply. Utilizing the present invention, the dialog generation model can generate more substantive replies toward the direction of optimizing the evaluation index during the training process, and greatly reduce the instability of generated dialogs.

Description

technical field [0001] The invention relates to the field of natural language processing dialog systems, in particular to a method and system for generating dialog replies based on self-comment sequence learning. Background technique [0002] In recent years, dialogue systems, as an important technical support for human-computer interaction, have attracted the attention of researchers from industry and academia. Dialogue reply generation has always been one of the most popular and difficult tasks in dialogue systems. Dialogue reply generation technology, that is, given the specific context and the current chat dialogue given by the other party, the machine can generate a reply that conforms to the context and has substantive reply content through semantic understanding and semantic generation technology. This technology has been widely used in many human-computer interaction systems, such as intelligent customer service systems, chat robots, personal intelligent assistants,...

AI Technical Summary

Problems solved by technology

However, this technology still has some inherent defects: one is the generation distribution deviation, that is, the dialogue generation distribution on the training set and the test set is likely to be different

Second, the objective function does not match, and it is impossible to directly optimize the evaluation index during the training process.

However, there are still thorny problems in the dialogue reply generation algorithm based on reinforcement learning: one is based on the minimum batch gradient descent method, there are a lot of changes, if it is not well regularized, it is easy to become unstable; Under the characteristics of learning, the word space of text generation becomes extremely large, and it is difficult to get a good initialization strategy

Images