题目：Learning Modality-specific Representations with Self-Supervised Multi-Task Learning for Multimodal Sentiment Analysis

现状、痛点：

Yu et al. need additional labor costs,and for Hazarika et al. research, spatial differences are difficult to represent the modality-specific difference.

They require manually balanced weights between constraint components in the global loss function.

方法重要细节、创新点：

1.the paper design a label generation module:    

    (1) relative distance value based on the distance between modality representations an class centers

2.an adaptive loss function

结果、结论：

    (1) SOTA on three benchmarks

    (2) save human costs

存在问题：

    modality-specific really make sense?