Dialogue model training method, training device, dialogue method and dialogue system

Abstract

The invention discloses a dialogue model training method, training device, dialogue method and dialogue system. In the total error construction step, a total error function including both the first error of the natural language understanding model and the second error of the policy generation model is constructed. In the joint training step, to reduce the total error function as the goal, use the corpus samples to jointly train the natural language understanding model and the policy generation model, wherein the input of the natural language understanding model is the dialogue sentence, and the output is the analysis of the dialogue sentence The internal representation obtained is that the input of the policy generation model includes at least the output of the natural language understanding model, and the output of the policy generation model is an action for the dialogue sentence. Compared with the existing technology, when there is an error in the natural language understanding model or the policy generation model, the system can carry on the dialogue normally, which solves the error transmission caused by the separate modeling of the natural language understanding model and the dialogue management model in the traditional method question.

Description

technical field [0001] The invention relates to the field of man-machine dialogue, in particular to a dialogue model training method, training device, dialogue method and dialogue system. Background technique [0002] Traditional dialogue systems can generally be summarized into the following three stages: [0003] 1) The user inputs through text or voice, and the natural language understanding (NLU) module analyzes the user's input to obtain a structured internal representation. [0004] 2) The dialogue management (DM) module receives the structured internal representation, updates it according to the historical system state, obtains the latest system state, and then selects the system action corresponding to the latest system state according to the dialogue strategy formulated in advance, to the next module. [0005] 3) Finally, the Natural Language Generation (NLG) module is responsible for converting system actions into natural language or speech output that humans can...

AI Technical Summary

Problems solved by technology

Since the natural language understanding module and the dialogue management module are modeled separately in traditional dialogue systems, errors in both will affect the final result. For example, when there is a problem with the natural language understanding module, the error generated by the natural language understanding module will be Passed to the dialogue management module, so that the dialogue management module cannot be carried out normally

In the actual process, errors in the natural language understanding module or dialogue management module are unavoidable, so the traditional training method obviously cannot solve this error transfer problem

Images