Traditional literature feature extraction methods typically rely on single-dimensional domain features, making it difficult to accurately predict the relatedness of fine-grained literature research domains. The multi-dimensional extraction process often faces the over-smoothing problem, leading to inaccurate predictions of relatedness and lower quantization accuracy. To address these issues, a method is proposed to quantify the relatedness of literature research based on multi-dimensional feature fusion. First, based on the traditional Doc2Vec model for extracting semantic content features from literature, multiple related dimension graphs are constructed and assigned corresponding weights to enhance the comprehensiveness of structural related features. Second, a multi-channel propagation strategy and adaptive aggregation mechanism are incorporated into the graph learning module, which mitigates the over-smoothing problem in traditional GCN by optimizing the aggregation of related node features, thus enabling precise domain-relatedness prediction among different literature. Finally, a minimum n-dimensional sphere model is constructed to cover the multi-dimensional related feature vector space of scholars, enabling the quantitative evaluation of cross-domain scientific research abilities. Experimental results on a large-scale real literature dataset show that the tolerance-aware accuracy (TAA) of the proposed method reaches 0.68, outperforming Doc2Vec, GCN, and Sentence-BERT models by 0.67, 0.08, and 0.02, respectively. Moreover, the performance fluctuation across different graph neural network models is minimal, demonstrating that the proposed method outperforms mainstream baseline models in terms of both accuracy and stability.