Abstract: 【Purpose/significance】This study aims to conduct fine-grained emotion recognition on multi-source conversational data, track the dynamic evolution of emotions in conversations, and support research in areas such as public opinion analysis, user emotion recognition, and intelligent library interaction.【Method/process】We propose a dialogue emotion recognition model named HIS-ERC, which incorporates speaker information and leverages deep graph convolution to capture both global and local contextual information in conversations. Additionally, a three-level cross-modal attention mechanism is designed to capture information interactions between different modalities, enabling accurate recognition in multi-party dialogues.【Result/conclusion】Experimental results demonstrate that deep and highly interactive methods can effectively enhance the emotion perception performance of the model. Compared to the classical multimodal dialogue emotion recognition model (MMGCN), HIS-ERC improves accuracy and F1-score by 3.51% and 3.57% on the IEMOCAP dataset and by 2.15% and 1.97% on the MELD dataset, respectively.【Innovation/limitation】This study proposes a multimodal information fusion and emotion perception model applicable to multi-party dialogue scenarios, providing methodological support for research in the field of information science. Future research will focus on applying the technology to real-world scenarios, such as intelligent library interactions and online public opinion analysis.