Research On Semantic Parsing Methods Of Natural Language Interaction For Service Robot

Because service robots need to communicate directly with humans and provide services for humans,human-computer interaction modules have an important position in service robots.In order to improve user satisfaction,natural language interaction is the mainstream research direction of human-computer interaction.Semantic pars-ing,as a task in the natural language interaction module,aims to directly map natural language sentences into semantically complete,computer-executable meaning repre-sentation.The research on semantic parsing has a long history.From early methods based on manual rules and template matching,to statistical learning methods based on training data,to deep learning model methods,and weak supervision methods combined with reinforcement learning without labeled data,these methods have reflected their respective advantages and disadvantages.In this thesis,two models are proposed from different angles to solve the problems that annotation for semantic parsing data is diffi-cult,size of data sets are small,and deep learning models can't realize its potential.The first model is a Tree2Tree model that combines syntactic knowledge.By encoding the constituency-based parse tree,syntactic imformation is introduced into the Tree2Tree model as a priori knowledge,thereby enhancing the model's ability to encode natural language.In addition,the Tree2Tree model uses a tree decoder to perform hierarchi-cal decoding to obtain a logical form,which can ensure the grammatical correctness of the logical form.The second model is a semantic parsing model based on pre-trained language models and knowledge distillation.It uses the general semantic knowledge contained in the large-scale pre-trained model to complete the semantic parsing task.Although the pre-trained model can get higher accuracy after fine-tuning on the se-mantic parsing task,its large scale makes it difficult to embed it in the service robot.Therefore,in this thesis,knowledge distillation is adopted to transfer the general knowl-edge and domain knowledge learned by the pre-training model to a small-scale model,so that a lightweight semantic parsing model with significantly improved accuracy can be obtained.This thesis introduces and implements the above two semantic parsing models,and conducts experiments on the robot command data set GPSR and two other public data sets ATIS and GEO.The experimental results show that the two models proposed in this thesis alleviate the problem of insufficient semantic parsing data from different angles.Without increasing the training data,both models can effectively improve the accuracy of semantic parsing.