Knowledge fusion combines multi-source data to address the inherent incompleteness of knowledge graphs (KGs), with conflict resolution serving as a critical technique for resolving knowledge conflicts during the fusion process. However, current advanced methods often suffer from inadequate representation learning for KGs, which results in suboptimal performance. To tackle these challenges, we propose RANA, a framework that combines self-adversarial negative sampling and learnable weight matrices for learning better representation of KGs, thus benefiting conflict resolution. RANA uses a graph neural network (GNN) based encoder with a self-adversarial negative sampling mechanism to enhance representation learning quality and improve the model’s ability to identify noise. Furthermore, learnable parameters were introduced to project entity representations, enabling better capture of interactions between different attributes and entities, thus improving the accuracy of representation learning for attributed KGs. By integrating source credibility with the high-quality representations, RANA utilizes a truth inference module to update the KG, thus achieving effective conflict resolution. Experimental results demonstrate the effectiveness of RANA in conflict resolution tasks, achieving an F1-score of 0.715 and provide deeper insights into the proposed method.