Xiaoyu Qi1,2, BoCheng1,2, Kang Yang3, Lili Zhong3 and Yan Tang3, 1Shenzhen University, China, 2Beijing University of Posts and Telecommunications, China, 3Ping An Bank Co. Ltd., China
Customer service text data is known as the dialogue text data between users and customer service provider, and it contains a large amount of user information. The effective use of customer service text content can bring great business plan optimization to the service provider. Based on the traditional machine reading comprehension model, this paper builds a customer service text user's attribute label recognition model, and proposes a model pre-training method based on sentence-level pre-training technology: aiming at the background of poor performance of the model in answering comprehensive full-text content analysis questions such as user intent and text sentiment analysis. This paper extracts text summaries based on the T5-pegasus model, constructing a text summaries dataset for model pre-training. Then build a text summarization model including an ERNIE pre-training model, train the model's ability to understand the full text, and improve the model's ability to answer questions that need to be combined with full-text content understanding, such as user intent and sentiment analysis. Use the pre-trained model to solve customer service text label recognition tasks based on machine reading comprehension tasks. The test results based on the data set show that the improved model has an improvement in performance of customer service text label recognition task.
Machine Reading Comprehension, Pre-trained model, Customer Service Text Analysis, Natural Language Generation, Attention Mechanism