A Simple and Interactive Transformer for Fine-Grained Emotion Detection

Fine-grained emotion detection is a long-standing task in sentiment analysis. One of the crucial challenges in this task is to capture emotion correlations. Previous studies either ignore this challenge or resort to a chain-link structure (e.g. seq2seq), both of which encounter some thorny issues in reality. Unlike previous studies, we approach this challenge by proposing a masked language modeling like mechanism, named masked emotion modeling (MEM), which can capture emotion correlations in a simple but effective manner. Specifically, we augment the transformer model with emotion states and require the transformer model to recover masked emotion states. Furthermore, to fully enhance the effect of the proposed MEM, we also introduce a taxonomy-guided sampling strategy at the training phase and an iterative decoding strategy at the test phase. On two widely used datasets, we demonstrate that our method excels at capturing emotion correlations and achieves state-of-the-art results.