信息管理与商业智能系学术讲座

时间：2023年10月31日（周二）13:30-14:30

地点：线下史带楼501室；线上腾讯会议室229912278/密码722722

题目：A hierarchical multi-label classification method for question -answer in customer service

主讲人：复旦大学博士生汪雅婧

主持人：信息管理与商业智能系徐云杰教授

摘要：Automatically matching customer questions with the answers in the firm knowledge base can improve the quality of customer service. In this commercial scenario, the number of dialogue samples is limited and unstructured, while there are numerous answers in the knowledge base. These challenges hinder the deployment of existing text classification methods. Based on the theory of information foraging, this study proposes that the tree-like structure of the firm's knowledge base can be used in deep learning algorithms to improve the accuracy of customer question matching. To this end, this study organizes the firm knowledge base into a tree structure according to the problem similarity. According to the information foraging theory, this study builds the information diet of service agents through contextual encoding of customer questions and context. By representing the tree structure of the knowledge base as a matrix, the path between the searched information patches is determined. Information scent is established by encoding the answer keywords and the responses of service personnel corresponding to question labels in the training set. With these designs, this study constructs the Tree-seq2seq method, which transforms the customer question matching problem into the problem of finding a path to a leaf node from the tree root of the firm knowledge base. The experimental results on real business datasets show that the question classification method based on tree structure is better than regular text classification models. Ablation studies illustrate the effectiveness of the designed components as informative diet, informative patches, and informative scent in the Tree-seq2seq model, thus approving the informative foraging theory. These results prove that our Tree-seq2seq model gets higher accuracy in customer service question-answering.

Keywords: Customer service question answering; Natural language processing; Text matching; Deep Learning; Hierarchical multi-label classification