A Document-level Emotion Change Analysis Method Based on DualGATs for Prose

In Prose, the author's emotions evolve within passages or sentences, transitioning between sadness, happiness, joy and anger. This paper constructs a dataset of prose emotion changes and proposes an emotion analysis method based on Dual-Graph Attention Network. Firstly, this paper incorporates image knowledge to get the deep representation of sentence. Secondly, the context-weighted dependency graph and the discourse-weighted dependency graph are constructed, and the sentence representation is established by integrating context knowledge and discourse structure. Additionally, a pleasure valence recognition layer is designed to capture the integrate pleasure valence information. Experimental results indicate the method is effective for identifying emotion changes.