Planning emotional education strategies for college students using a graph attention network and BERT-BiLSTM

To address the perceptual lag and online psychological attribution challenges in traditional university emotion monitoring, this study constructs an emotional intelligence framework for college students. This framework employs a dual-mode model combining sequence and graph structures: a Transformer disambiguation network for slang, a bidirectional long short-term memory network for tracking emotional evolution, and a dependency syntax combined with a graph attention network to locate key emotional points, mapping emotional states to predefined hierarchical intervention strategies. The framework achieves a validation set accuracy of 0.960, outperforming mainstream models, and processes a single sample in just 0.08 seconds, enabling efficient processing of large-scale text inputs. Simultaneously, the framework effectively optimizes the allocation efficiency of psychological intervention resources, significantly reducing the resource mismatch rate to 9.6% in simulated evaluation environments. The research effectively addresses the pain points of risk tracing in complex semantic environments, provides interpretable intelligent decision support tools for ideological and political education in universities, and establishes a standardized, data-driven evaluation framework for university emotional risk screening. Not applicable.