The goal of entity linking (EL) is to automatically map mentions in natural language to corresponding entities in the knowledge graph (KG). Nonetheless, EL is replete with challenges due to the natural language’s diversity and ambiguity. Most advanced EL models do not fully utilize the valuable KG structure information for denoising context to disambiguate. This paper proposes an EL model, KRL-CoAtt, that encodes and integrates the KG structure information into the EL process for disambiguation based on knowledge representation learning (KRL), and introduces co-attention mechanism to filter the context information by obtaining weight coefficients for context words and triplets, improving local similarity calculation precision. KRL-CoAtt acquires EL results by adopting loopy belief propagation (LBP) to combine local similarity and global coherence. Extensive experimental results demonstrate the superiority and effectiveness of KRL-CoAtt compared with other state-of-the-art methods.