Research On The Semantic-based User Intention Recognition Algorithms

The Semantic-based user intention recognition problem belongs to the category of natural language processing problem,which is essentially a multi-classification problem.With the development of the Internet,the development of human-machine conversation system is also more and more rapid and the demand is more and more clear.In recent years,the research of human-machine conversation system always receives the academic to take important.The user intention recognition is a very important module in the human-machine conversation system.At present,most methods of intention recognition are based on semantics,and multi-classification results are obtained by analyzing text content and applying deep learning algorithm,such as long short-term memory neural network?recurrent neural network?gate recurrent unit neural network and other deep learning models or their variants.Meanwhile,traditional machine learning models such as support vector machine and XGBOOST are also used to study the problem of text multi-classification.Aiming at the problem of user intention classification,this paper analyzed the performance of traditional machine learning model and common deep learning model on this task by taking the dialog text with intention label in SMP2017 and SMP2018 competitions as data.At the same time,a combined CNN-BGRU-ATTEN model based on character level vector is proposed to solve this task.This model combines the convolutional neural network(CNN)and the bidirectional gated recurrent unit network(BGRU)with ATTENTION mechanism,and the accuracy is improved to 93.3% compared with other models.The model converts the input text into character-level vector representation and then carries out the convolution operation in the dimension of sentence order through the convolutional neural network.The obtained output is spliced through feature fusion,and the position information of the vector is still retained.Next,the attention weight of the nodes is calculated according to the attention mechanism method through the bidirectional gated recurrent unit network with attention mechanism.Such combined network structure design can not only extract the deep semantics of text through convolutional neural network,but also take into account the sequence information of context through the bidirectional gated recurrent unit network,and finally add attention mechanism to optimize the distribution of attention.In the end,a series of comparative experiments are designed,and the results show that the combined model has a good performance in user intention recognition based on semantics.The comparison experiment of text feature vector(word level and character level)shows that the effect is best when the character vector dimension is 200,and the accuracy is 4.8% higher than that of random initialization of vector.The comparison experiments also include the model parameters,the recurrent unit type,the effectiveness of bidirectional structure and the necessity of attention mechanism.Finally,the combined model is compared with traditional machine learning models and common deep learning models.It shows that the accuracy of the combined model is 10.7% higher than that of the traditional machine learning model(Xgboost)and 5.8% higher than that of the commonly used deep learning model(BGRU),which verifies the rationality and effectiveness of the combined model.