Study On Voice Interaction System Of Emergency Individual Equipment For Highway Traffic Emergency

Based on the current development results of artificial intelligence technology and mobile Internet of Things technology,the paper starts to study the key technologies of the voice interaction system,aiming to provide emergency personnel assistant services deployed on intelligent terminals for emergency personnel under road traffic emergencies to complete the operation method query and information upload functions.One of the key points in designing a dialogue system in a specific field is the definition of specific scenarios and the collection of scene dialogue corpora.However,the lack of real dialogue corpora causes cold boot for dialogue systems.Based on the concrete task situation and speaker habits in the survey results,this paper uses Chatito’s online IDE to apply grammatical symbols to manually construct corpus data in order to obtain corpus data and structured data in the dictionary form.Natural language understanding and dialogue management are the main components of a dialogue system.In the natural language understanding part,the pre-trained model BERT is used to solve the problem of small training data.At the same time,the task of intent classification and entity recognition are performed by transforming the output layer of the BERT model.Both tasks took a single sentence as the model input.The intent classification task used the [CLS] sign as the output vector of the intent category.The entity recognition task obtained the entity category of each word vector through the Soft Max output layer.The BERT models were packaged in the form of components respectively.Based on the Rasa chat robot framework,the components were configured to files using the pipeline method to accomplish the training of the natural language understanding model.In the dialogue management part,the dialogue data is time series data.The intents,entities and slots appearing in the dialogue data are part of the dialogue state.The One-hot Encoding method was adopted to complete the state characterization of the dialogue corpus.The conversation history and the current input codes were concatenated to sequence data in the form of binary matrix.The dialogue strategy learning task is transformed into a classification task for system actions based on historical dialogue and current input.The LSTM model took the action probabilities as the output,and the classification accuracy reached 94.6%.It also used the Rasa framework to configure the strategy model to the pipeline configuration file,and trained the dialog management model.Combined the model with the natural language understanding model,the effect of the dialog system was tested by using the command line to achieve good test results.Finally,the dialogue system based on the Rasa framework was deployed to the We Chat subscription number.Also,it was connected to the speech recognition and speech synthesis interfaces to achieve voice interaction.