Research On Dialogue Management Base On Real User Interaction

In recent years,the rapid development of commercial spoken dialogue systems has attracted more attention from industry and academia.Dialogue management(DM),as the core part of the spoken dialogue system,plays the role of selecting execution actions given the context of the dialogue.In DM research based on real user interactions(such as Man-man dialogue corpus),most researchers treat each instance equally to learn different dialogue strategies based on finite corpora.However,multiple dialogue styles from system role players in Man-man dialogue corpus are likely to affect the training efficiency of dialogue management.In order to solve the above problems,we consider the corpus of non-target system role players as external corpus of the training target dialogue management model and study the system action prediction based on the training corpora discrepancy evaluation.The main research contents and innovations are as follows.(1)The system action prediction based on the discrepancies of training corpora is analyzed and designed.By analyzing the differences in dialogue styles between different system role players in Man-man dialogue corpus,and the feature expressions of training corpus,combined with the system architecture of the mainstream system action prediction,the training corpus discrepancy awareness scheme in system action prediction is designed.(2)This paper first proposes a framework called the Corpus-based Learning Dialogue Manager with Discrepancy Awareness in which adopted the selection/deprecation strategy of external corpora,and the consideration of less context will lead to smaller behavior pattern granularity which may lead to deviations.Two simple and effective methods of approximation difference measurement based on behavior are proposed.One is based on action,and the other is based on intent-action pairs.External corpus with high difference is eliminated which reduced the impact of different dialogue style.Experiments show that our approach greatly improved the model predictions.(3)This paper also proposes a framework called the Learning Rate based Dialogue Management with Discrepancy Awareness in which a more general method of external corpus-based discrepancy perception is proposed.This method evaluates the external corpus discrepancies based on similarity measuring of encoded context.Then adjusted the learning rate of the recurrent neural network model and achieved good predictions.This paper verifies the effectiveness of the proposed framework in the classic Man-man dialogue corpus DSTC4.Experiments show that our method greatly improves the accuracy of DM's execution of system action prediction.