Penalty Prediction in Discipline Reinforcement Based on Multi-Task Learning

Abstract

Penalty Prediction in Discipline Reinforcement aims to automatically determine how to punish a party member who violates the party's discipline. The task is challenged by two main issues. The first is the party affiliation of the accused cannot be effectively learned by conventional models, which has a significant impact on the result. The second is the case number of various penalties is highly imbalanced, which reduces the accuracy of the model due to insufficient training data. To handle these issues, we formalize the predicting task as a problem of multi-label classification and define the prediction of each label to be a single task. Thus, we introduce a multi-task learning framework. Besides, we take party affiliation into consideration and define an auxiliary task which predicts party affiliation. More specifically, we propose a model which integrates the contextual representations of fact descriptions and the party affiliation to predict penalties to be given. To verify the proposed approach, we conduct experiments on a dataset relevant to party discipline reinforcement. Empirical results demonstrate that the proposed model outperforms some prior models significantly.