Research On Key Technologies Of Intelligent Voice Interaction For Railway Voice Ticketing Scenario

The railway 12306 mobile client application only supports manual ticketing,which has a long operation process and contains more information.This means that self-screening is required,and it is not friendly enough for the elderly group.Intelligent voice interaction is easy to operate and becomes a link to realize the communication between human and machine using language.Therefore,one of the ways to age-appropriate transformation of the railway 12306 mobile client application is to use intelligent voice interaction technology to add the function of voice ticketing on the mobile application,so as to better solve the problem of the elderly group’s difficulty in using smartphones alone to purchase tickets.At present,speech recognition technology has reached a high level,but is usually based on experimental environments,and suffers from the inability to accurately recognize domain-specific proper names when implemented for applications.The limitation of natural language processing technology is the limited ability to predict future events,thus making it difficult for machines to respond as humans do in interactive dialogue systems.In this paper,the key technologies of intelligent voice interaction for railway voice ticketing scenario are studied by combining speech recognition technology and natural language processing technology.The main research contents and contributions of this paper include the following two major aspects.（1）Design and implementation of a speech recognition model based on Encoder-Decoder neural network.The requirements of railway voice ticketing are mainly reflected in the need for accurate and fast recognition of user’s input speech.Combining the above requirements,this paper proposes a speech recognition model based on Encoder-Decoder neural network.For the accurate recognition problem,we propose the solution of multiple language model fusion and hot word assignment to improve the recognition accuracy of railway proper nouns and optimize the recognition performance of speech recognition model.For the fast recognition problem,we propose the solution of parallel streaming and non-streaming recognition.We make streaming improvements to the Conformer encoding structure,and design the streaming decoding based on Connectionist Temporal Classification and non-streaming decoding based on Attention mechanism to improve the recognition speed of the model.Experiments show that the speech recognition model based on Encoder-Decoder neural network proposed in this paper improves 2.126% in recognition accuracy over the baseline model,and deploys it to practical applications to achieve the same good performance.（2）Design and implementation of a task-orient multi-round dialogue system based on rule matching and RASA.Voice ticketing is a process of continuously collecting information provided by users and making corresponding feedback.In order to extract information accurately and make feedback flexibly,this paper designs a task-orient multi-round dialogue system based on rule matching and RASA.To address the problem of ticketing information extraction,the solution of rule-based matching dialogue system is proposed to improve the accuracy of extracting key ticketing information such as time,location and train number.To address the problem of flexibility of multi-round dialogue,we propose a solution of RASA-based multi-round dialogue system,which extends the pipeline structure of RASA dialogue system,and introduces a Jieba-based word segmentation model,a Bert-BiLSTM-CRF-based named entity recognition model and a BiGRU-CRF-based intention recognition model in the structure to reduce the probability of incorrectly slicing railway proper nouns,and achieve accurate positioning of user’s ticketing intention and accurate extraction of key ticketing information.The experiments show that the F1 values of the above three models reach 91.48%,92.64%,and 92.49%,respectively,and show good performance in practical applications.In summary,according to the process of railway ticketing and the characteristics of voice ticketing,this paper researches and designs a speech recognition model based on Encoder-Decoder neural network and a task-orient multi-round dialogue system based on rule matching and RASA for railway ticketing scenarios.It solves the problem that the machine can recognize and understand the user’s demand and completes the research on key technologies of intelligent voice interaction for railway voice ticketing scenario.