MFSF-CEA: A Multi-Feature Similarity Fusion Model for Chinese Entity Alignment

INTRODUCTION: Entity alignment across multi-source encyclopedic knowledge bases is crucial for constructing high-quality knowledge graphs. This task is particularly challenging in specialized vertical domains like Chinese cultural relics, where heterogeneous data sources and diverse descriptive patterns render single-feature alignment methods inadequate. OBJECTIVES: To address this challenge, we propose MFSF-CEA, a Multi-feature Similarity Fusion model featuring dual-layer optimization: multi-granularity semantic modeling and domain-adaptive dynamic weight fusion. METHODS: This approach employs a three-tiered semantic capture structure: character-level similarity using Longest Common Subsequence for variant character matching; word-level similarity via TF-IDF for core concept association; and sentence-level semantic similarity through Latent Dirichlet Allocation for deep topic alignment. Beyond feature extraction, we introduce an Entropy-AHP combined weighting mechanism that dynamically balances objective information contribution and domain expert knowledge, overcoming limitations of fixed-weight fusion strategies. Experimental evaluation on a Chinese cultural relics dataset demonstrates that MFSF-CEA significantly outperforms baseline methods in precision, recall, and F1-score. The sentence-level contextual features contribute most substantially to alignment accuracy, while the multi-feature fusion effectively compensates for the limitations of any single feature type, particularly the sparsity of word-level abstract features. RESULTS: The proposed framework successfully addresses the unique challenges of entity alignment in cultural relic texts by leveraging complementary features across multiple linguistic levels. CONCLUSION: This work provides an effective and extensible solution for knowledge fusion in vertical domains, advancing entity alignment from traditional string matching toward deeper semantic integration.