LGCFPEL: local-global correlation fusion for probabilistic collective entity linking

Entity linking, as a key technology for heterogeneous information networks modeling and semantic knowledge retrieval, has attracted significant attention. However, current entity linking methods suffer from three main limitations: (1) They generally rely on a single type of information from the network ( e.g ., short-path relationships or knowledge representations), leading to insufficient utilization of information and incomplete extraction of both similarity and difference features among entities. (2) They frequently employ deep learning approaches, which lack interpretability. (3) They often depend on additional information from third-party bases, which limits their applicability in some domains’ scenarios. To solve these problems, this article innovatively proposes a probabilistic entity linking method that does not rely on external knowledge bases—Local-Global Correlation Fusion for Probabilistic Collective Entity Linking (LGCFPEL). Specifically, LGCFPEL first constructs a probabilistic model to solve for the optimal linked entities. Second, it treats all entity mentions in a text as a whole for collective linking. Third, it skillfully avoids the direct computation of entity mentions’ correlation by transforming the problem into calculating correlation among their candidate entities. Finally, LGCFPEL builds a local-global fusion correlation metric model based on short-path relationships and knowledge representation learning, simultaneously capturing both local and global features of entities to highlight the tight contextual relationships among different entities. Experimental results indicate that LGCFPEL demonstrates competitive performance with both classical and state-of-the-art models in entity linking performance.