Based On The Semantic Field Of User Intent Classification Algorithm Analysis

In recent years,with the rise of the AI craze,the technology of man-machine dia-logue has attracted wide attention from the academia and industry field.Man-machine dialogue is one of the most natural way for human-computer interaction,whose develop-ment influences and promotes the development process of studies on natural language understanding,speech recognition and synthesis,dialogue management and natural language generation,etc.Human-computer interaction has penetrated into every as-pect of our lives,including our daily use of WeChat voice input function,apple' s Siri query function,as well as functional robots that could accurately identify and follow the human instructions,etc.,in order to realize these functions,one of the important directions is user intention recognition classification,namely sentiment analysis task in NLP field.In the application process of man-machine dialogue system,there may be multiple user intents that will trigger the accordingly many fields in man-machine dia-logue system,which includes task-vertical field(such as query airfare,hotel,bus,etc.),knowledge-quiz field,chatting field and other fields.One of the core tasks of the man-machine dialogue system is to accurately identify the user's input,then accurately categorize it into the corresponding field,and finally return the correct reply result.At present,more and more attentions are paid to the semantic analysis of user intention.Through automatic analysis of text content,assigning the user' s search or query intention or precise to a specific domain has become a popular NLP multi-ple classification problem.Traditional classification algorithm like Softmax and SVM,have been applied to text categorization in succession,which all achieved good results.However,as people' s requirements of classification accuracy is higher and higher in recent years,deep learning is increasingly applied to this field,which establishes a com-plex neural network model for natural language processing,such as Convolution Neural Network(CNN),Recurrent Neural Network(RNN)model,etc.,within which one vari-ant of RNN,long-term and short-term memory network(LSTM)model,overcomes the drawback of RNN that gradient disappears,and achieved extremely good effect on the field of text classification.A deep learning model based on LSTM is proposed in this paper to solve the prob-lem of user intention classification,and ultimately LSTM model with hidden layers and bidirectional LSTM(BI-LSTM)model were adopted.After the Jieba participle and word2vec preprocessing operations of Chinese participle and term vectors,the term vectors were conducted dimensionality reduction by self-coding,then parameters like appropriate learning rate were selected after multiple parameter adjustments,and ad-justment results were input into complex LSTM neural network.Cleverly setting input,output and forgetting threshold,controlling the information transformation and out-put by weight coefficient,LSTM algorithm output multiple classifications by Softmax classifier in the end.Various optimization methods,including preventing overfitting by dropout,setting custom rule aided identification on the basis of word frequency,selecting the Attention mechanism optimization model,integrating votes by interior extrapolation method for getting the accuracy of algorithm,are used to reach the final classification accuracy of the test set reaching up to 94%,which is much more accurate than the result of the champion of Chinese man-machine dialogue technology evalua-tion(ECDT),92.7%classification accuracy.The model hereof attempted various deep learning optimization technologies in the process of system implementation,which have been all briefly introduced and analyzed by experiments.