Information retrieval framework using knowledge graph embeddings and uncertainty modelling using probabilistic soft logic

The growing complexity and uncertainty inherent in modern information retrieval tasks necessitate systems that can reason probabilistically while leveraging rich semantic structures. This research explores the core question: How can uncertainty-aware retrieval be enhanced through the integration of probabilistic logic and knowledge graph semantics? To address this, we introduce a novel hybrid framework that fuses probabilistic soft logic (PSL) with transformer-enhanced knowledge graph embeddings (TEKGE), further empowered by a dynamic uncertainty quantification layer (DUQL). DUQL enables granular modeling of both epistemic and aleatoric uncertainties across graph entities and relationships. Additionally, a multi-hop probabilistic graph traversal (MPGT) mechanism, informed by Bayesian-regularized contextual embeddings, guides the retrieval process. Empirical evaluations on two benchmark datasets—CN15k and O*NET20k—demonstrate the system’s effectiveness, with notable gains including an 11.6% increase in nDCG@20, a 14.2% improvement in mean reciprocal rank (MRR), and an uncertainty-aware precision (UAP) score of 0.87. The use of conformal prediction ensures statistically valid confidence calibration across retrieval outputs. This work presents a scalable and interpretable approach that combines symbolic reasoning with neural representation learning, achieving both robustness and trustworthiness in uncertain environments, the study reaffirms the importance of quantifiable confidence in semantic retrieval, contributing a resilient solution to the broader challenge of explainable and reliable information systems.