Adaptive Transfer Learning for Multi-Label Emotion Classification

Abstract

In this study, we explore how data annotated with different taxonomies can be used to improve multi-label emotion classification. We propose a novel transfer learning framework to model the interaction between emotion categories, and introduce an adaptive aggregation mechanism to fuse the information from different taxonomies. The cross-taxonomy emotion interaction allows the source and target tasks to collaborate effectively, resulting in more accurate predictions. The experimental results on the SemEval-2018 dataset show that our approach can effectively boost the performance gain brought by transfer learning, and significantly outperforms existing methods.

Tasks

Reproductions