Multiple granularity user intention fairness recognition of intelligent government Q & A system via three-way decision

Abstract

With respect to an intelligent government questions & answers (Q & A) system, user intention recognition for government affairs is a key issue. Accurate intention recognition can effectively reduce manual participation and improve the user satisfaction. However, the intention recognition model in the field of government affairs not only requires the recognition accuracy, but also needs to meet the fairness demands of users. In order to improve the fairness of the model, this paper makes the model focus on the unrecognizable intention samples in each intention type as much as possible. Hence, we firstly design a two-stage intention recognition method based on the idea of three-way decision (TWD). In the first stage, we use the Bert model as the intention recognition model and divide the samples with insufficient classification confidence into the boundary region. In the second stage, we combine the divide-and-rule idea of TWD with focal loss to suppress the easily recognized samples in the non-boundary region, so as to reduce the contribution of these samples to the loss of the classifier. Meanwhile, we can enhance the contribution of samples with insufficient classification confidence in the boundary region to the loss of the classifier, and then optimize the recognition ability of the classifier. Then, by utilizing sequential three-way decision (STWD), we recognize the user's intention types at multiple granularity. According to the recognition results of coarser granularity, we can optimize the recognition ability of the classifier for the intention that is difficult-to-recognize by adjusting the loss function. On the premise of ensuring that it has little impact on the intention recognition ability of other types of users, we improve the recognition ability of the intention that is difficult-to-recognize. Based on the above-mentioned methods, we further propose a multi-stage training method that can make the model focus on the unrecognizable intention text and the unrecognizable intention type. Finally, the effectiveness of the proposed method is verified through some series of experimental analysis.