Emotion regression approach with both forward and reverse values based on multi-task learning

Abstract

An emotion values regression approach based on multi-task learning was proposed. Firstly, forward score and reverse score were designed for each dimension of emotion. Secondly, the regression task was divided into the forward score regression subtask and the reverse score regression subtask. Finally, a multi-task learning approach was proposed to jointly learn both the main task (forward score regression subtask) and the auxiliary task (reverse score regression subtask). In order to improve the performance of the main task, three sharing models were designed to share different kinds of information between the main and auxiliary tasks. The results showed that the proposed multi-task learning method achieved better regression performance than the baseline. 提出一种基于多任务学习的情绪分值回归方法。首先，针对每一种情绪分值设计了正向打分和逆向打分; 其次，将每一种分值的回归任务分为正向打分回归子任务和逆向打分回归子任务; 最后，提出一种多任务学习方法用于主任务( 正向打分回归子任务) 和辅助任务( 逆向打分回归子任务) 的共同学习。该方法通过 3 种不同的共享机制实现中间特征信息共享，从而提升主任务的性能。结果表明，所提出的多任务学习方法能比基准方法获得更好的回归性能。