Research On Natural Language Understanding Of Task-specific Dialogue System

Recent years,with the continuous development of artificial intelligence technology,human-computer interaction(HCI)systems are widely used in all walks of life,and have greatly changed people’s lifestyles.However,the personal privacy of users may be involved and compromised,when the key entity information required by the dialogue system in the process of human-computer,interaction.Therefore,the dialogue system needs to process the user’s commands and give feedback in real time on the local.As an important part of HCI system,the natural language understanding module plays a key role in the performance of the subsequent dialogue state tracking,dialogue policy and dialogue generation modules.The current mainstream method is to complete the natural language understanding task through a pre-trained model,but the pre-trained model has a huge amount of parameters and calculated quantities,which limits its application in edge devices.And when the task scene is relatively single,using the pre-trained model to build the natural language understanding module has large parameter redundancy,which causes additional memory waste and computational overhead for the hardware.Therefore,how to design an efficient and low-power natural language understanding model is the key to improving the application of task-based dialogue systems in real life.The thesis proposes one models to solve the above problems,and implements a Task-oriented dialogue system based on the proposed model design.The specific work content can be summarized as follows:(1)A parameter sharing for joint intent classification and slot filling is proposed to reduce the number of model parameters and obtain a better semantic representation of the intent by using parameter sharing for multiple encoders in the model,and to encode the output intent of the model and interact the encoded intent information with the semantic slot hidden vector.Following that,the constraint of the semantic slot output by intent is achieved.Compared with the Slot-gate model,this model has 53%smaller parameters than the Slot-gate model.And the model’s intent recognition accuracy,slot filling F1 score and Semantic accuracy on the ATIS dataset are improved by 0.3%,2.4%,and 2.2%.(2)A pipelined task-oriented dialogue system is designed and implemented to apply the natural language comprehension model proposed in(1)and combined with dialogue state tracking,dialogue strategy,natural language generation,speech recognition and language synthesis modules to realize a complete human-computer interaction process.According to the function of each module in the system,the corresponding algorithm is implemented.At the same time,in solve to reduce the excessive redundant parameters and memory consumption of the complex system,the model parameter quantization and parameter clustering are carried out without affecting the model effect.