Research And Application Of Unsupervised Dialogue Task Mining Technology

Task-oriented dialogue systems have been widely used in various fields.However,there are still two problems in the construction and operation of traditional task-oriented dialogue systems.One is that the construction of task-oriented dialogue systems usually requires developers to first define the dialogue task,which requires experts to analysis the existing human-to-human dialogues,which is time-consuming and labor-intensive.The second is that the dialogue tasks involved in the dialogue system often change rapidly with the development of the business,and the dialogue system cannot be adapted and updated in time.Therefore,if the elements necessary for the construction of the dialogue system can be automatically mined from the human-human dialogues to assist the construction and update of the dialogue system,it can greatly reduce the cost of system construction and improve the efficiency of system update and operation.Aiming at this problem,this paper adopts a top-down strategy to mine dialogue tasks,that is,firstly clustering out different dialogue tasks contained in dialogues,and secondly mining dialogue subtasks contained in dialogue tasks.For this reason,this paper proposes a Dialogue Task Mining Network(DTMN),which is used to simultaneously perform dialogue task clustering and dialogue subtask mining.The model fully considers the internal structure existing in the dialogue,and models the context-aware utterance representation based on the adjacency relationship between utterances;at the same time,based on the dialogue subtask structure existing in the dialogue,it is divide into three levels,which are utterance-level,subtask-level and dialogue-level to model dialogue representation,and an iterative training strategy is used to continuously optimize dialogue representation and context-aware utterance representation,and then continuously optimize dialogue clustering and dialogue subtask mining results.Experimental results on three public dialogue datasets show that the performance of our proposed model on dialogue clustering is significantly better than the existing strong clustering algorithms.At the same time,we qualitatively analyzed the nature of the dialogue subtasks of the mining,and the results show that the dialogue subtasks of the mining are easy to understand,and there is a clear logical transfer relationship between the dialogue subtasks.Finally,this paper builds a dialogue task mining system based on the proposed dialogue task mining model to assist users in automatically mining dialogue tasks from human-to-human dialogues.