Research On Deep Semantic Analysis Technologies For Home Scenes

With the rapid development of smart home,natural language semantic analysis technologies are used to understand user needs and provide corresponding services accordingly.Due to the complex natural language structure and diverse semantic expressions for home scenes,related technologies are far from achieving semantic understanding.There are still many problems that need to be solved in the semantic analysis of natural language in the home scenes.The main works of thesis are given as follows:I.The basic theories of semantic analysis on natural language are researched.Firstly,the vectorized representation models of natural language are researched.Then,the attention mechanism models are researched.Finally,the neural network models related to natural language semantic analysis are researched.II.A multi-intention recognition algorithm based on Bi-Ind RNN and Attention mechanism is proposed for home scenes.Firstly,based on the thought of migration learning,a model of pre-trained BERT is migrated to multi-intent recognition tasks,and then the pre-trained model is used to vectorize the input texts.Secondly,the temporal and semantic features of input texts are extracted by constructing the multilayer Bi-Ind RNN neural network.Finally,in order to better solve the problem of multi-intent recognition in home scenes,combined with the characteristics of user language in home scenes,the traditional Attention mechanism is improved.And a scene-adaptive Attention semantic enhancement method is proposed.The method is able to improve the generalization ability of semantic analysis algorithms for home scenes.Experimental results show that the algorithm can improve the accuracy of multi-intent recognition effectively.III.A semantic analysis algorithm for joint recognition of multiple intents and semantic slots is proposed for home scenes.Firstly,the two subtasks of multi-intention recognition and semantic slot filling share the semantic features of BERT.Secondly,a Slot-Gated correlation mechanism gate is used to pass the recognition result of the previous tasks to the next tasks,thereby improving the performance of semantic slot filling.In addition,in order to improve the overall performance of the joint recognition model,in view of the uncertainty of the number of user intentions in home scenes,the multi-intent recognition part of the joint model is optimized.Then,a multi-intent recognition model based on clustering pre-analysis is designed.The number of intents is judged before intents is recognized and then processed separately.In particular,the traditional method of measuring semantic similarity is improved in the pre-analysis method.And a new measurement method is proposed to reliably achieve intent clustering.Experimental results show that the joint recognition algorithm proposed in this thesis can effectively improve the accuracy of semantic analysis.