Research On User Intent Recognition Towards Spoken Text From Mobile Customer Service Hotline

The customer service system is an important bridge between the enterprise and the customer.For a long time,mobile operators have accumulated a large amount of customer contact data in the manual customer service business,but these unstructured text data are often not used properly.In the context of urgent transformation and upgrading of traditional customer service,intelligent customer service is gradually replacing manual customer service.The key part of intelligent customer service is the recognition of user's call intent,how to apply data science method to fully exploit the semantic information of these text data,clarify the user's call request,accurately identify the customer's intention and call preferences,it is vital to save the operator's labor cost and relieve the customer service pressure.Therefore,research on user intent recognition has positive practical significance.At present,the intent recognition of multi-turn dialogue is a hot issue in spoken language understanding.Most of the existing researches are based on the English public corpus with low noise after manual finishing.Most of the existing researches on Chinese spoken language are directed to single-turn dialogue,and there are few researches on the intent recognition of multiturn dialogue text data.Therefore,this paper conducts empirical research on user intent recognition in the mobile customer service hotline text through the deep learning method of neural network structure.In order to avoid the subjective drawbacks of traditional artificial feature extraction,this paper uses Skip-Gram model to obtain the word vector representation of spoken text;the long short-term memory(LSTM)is adopted as the benchmark model of classification,and construct hierarchical network model and hierarchical attention network model respectively to perform intent recognition.After comparison,the hierarchical attention network model with better effect is selected to further identify the deep intention of the user to explore the robustness of this model.The research shows that:(1)In terms of text representation,using the spoken text of the mobile domain for word vector pre-training can enhance the effect of intent recognition;(2)In terms of model structure,the hierarchical attention network model can effectively capture the semantics of multi-turn dialogue,which has a better intent recognition effect than the long shortterm memory(LSTM).The innovation of this paper is to use multi-turn dialogue as the starting point,analyze the composition of the text and the characteristics of the dialogue,and establish a hierarchical attention network model for the spoken text.Similar research on such intent recognition of the spoken text has not been explored in the literature.Moreover,this paper still remains some points that need to be improved: In the case of large number of noisy data and unbalanced categories of spoken texts,the robustness of the model needs further study.