Research On Key Technologies Of Natural Language Understanding For Task-based Human-machine Dialogue System

Task-based human-machine dialogue system has developed significantly in recent years due to continuous breakthroughs in computer hardware and software levels and advances in artificial intelligence science.The natural language understanding is a key component of the task-based human-machine dialogue system,the results of semantic understanding will have an important impact on the system.There are still two problems in task-based human-machine dialogue system.First,the recognition rate of traditional classifiers is different in different categories because of the existence of class data imbalance;second,the performance of existing solutions still needs to be improved due to the special corpus environment of Chinese task-based humancomputer dialogue system.This paper does the following to address the above issues.This paper proposes a Simple-Mix-Sampling algorithm to solve the problem of data imbalance in SMP2019-ECDT natural language understanding evaluation.Through the experiment of text classification,it is proved that the algorithm can alleviate the problem of data imbalance.This paper proposes multi-task natural language understanding model,including a multitask joint model of natural language understanding and a pre-training composite model of natural language understanding,which are used to process the SMP2019-ECDT natural language understanding evaluation.The experiment shows that the model has better performance in this evaluation.To sum up,this paper proposed Simple-Mix-Sampling algorithm to alleviate the imbalance in the data set.Multi-task natural language understanding model are proposed to solve the domain classification,intention classification and slot filling tasks in task-based humanmachine dialogue system.