Research Of Spoken Language Understanding Method Based On Deep Neural Network

The spoken dialogue system is that a computer can understand human natural expression through computation processing and make a corresponding answer.Spoken Language Understanding(SLU)is the key to the dialogue system.It enables the computer to "understand" human language.Therefore,the performance of the SLU directly determines the performance of spoken language on the system.The traditional methods to solve spoken language understanding are mainly conditional random fields,Hidden Markov Models,etc.,which are based on traditional statistical methods.With the improvement of deep learning in recent years,deep learning has achieved great success in the field of natural language processing(NLP),and Recurrent Neural Networks(RNN)is extracting sequence information of natural language texts and used establishing mathematical models to solve the corresponding problems has great advantages.The purpose of spoken language understanding is to automatically recognize the domain,category,and intention of the user's spoken language,and then extract related concepts to achieve the system's purpose of understanding the user's language.Computers convert speech into corresponding textual information,while spoken language understanding handles textual information.The RNN predicts the next textual information by affecting the input and output of text sequence information.The traditional feedforward neural network is not suitable for processing text sequences,mainly because the text information is variable and the text information is related to each other,but the feedforward neural network can only deal with fixed length and all input nodes are mutually independent,so this article uses RNN to solve spoken language understanding problems.In the RNN,the nodes between the hidden layers of the network progressively advance the sequence of time steps,and parameter sharing is also very important in the RNN.However,RNN has the problem of vanishing gradient.Therefore,the introduction of long-term memory networks and gated recurrent unit through the addition of gate control elements and cell memory states in the unit nodes can make long-term information for the sequences preserved,thus solving the RNN vanishing gradient.Based on RNN and its variants,this paper proposes a deep neural network based on feature fusion to directly stack multiplenodes at a time node to deepen the complexity of nonlinear transformation.Through the feature fusion network architecture,the spoken language understanding experiment was performed on the Airline Travel Information System(ATIS)data set.The experimental results prove that the feature-fusion RNN network can further improve the effectiveness of the RNN network.