A Context-Sensitive Reasoning Framework for Knowledge Graphs

ww A Context-Sensitive Reasoning Framework for Knowledge Graphs Sara Todorovikj & Florian Hahn T Motivation • Real-world knowledge is contradictory and context-dependent • Traditional Semantic Web systems rely on monotonic logic → limits ability to adapt to contextual nuances • We propose a three-layered reasoning framework for KGs that supports context-sensitive reasoning Outlook • Our proposed framework reasons over claims, resolves contradictions and adapts conclusions based on context. • Allows for nonmonotonic reasoning and contextual filtering. • Modular reasoning layer - beyond deductive or defeasible reasoning. • Aligns with cognitive science findings showing how humans rank and revise beliefs. • Unification of semantic representation with “old-school-AI” reasoning and context-awareness. • Applicable beyond scholarly domains, a foundation towards trustworthy, adaptive and explainable knowledge systems. Professorship Data Management University of Technology Chemnitz Ontology Layer • Structured representation of claims and evidence • Captures contradictions without forcing consistency • Relations like contradicts, supports or refutes. Reasoning Layer • Supports different reasoning types (deductive, defeasible, abductive, etc.) • Encodes rule priorities and exceptions • Supports revision of tentative inferences Contextual Adaptation Layer • Metadata-aware priority modification • Context dynamically reshapes rule application • Multiple contexts supported over the same KG What is stored? Studies on Policy X: • A (Europe, 2018, observational) → Positive Effect • B (Asia, 2021, RCTs) → No Effect • C (America, 2019, RCTs) → Negative Effect Conflicting evidence stored as-is. What is inferred by default? Apply rules: • Prefer RCTs • Prefer recent studies Default outcome: Study B outranks A and C → Policy X: no effect Produces the context-free conclusion. How does the conclusion change with context? Context: Evaluating Policy X for Europe • Increase priority of local evidence • Downweight foreign RCTs Revised priorities: Local > RCT > Recency Contextual outcome: Study A now outranks B → Policy X: positive effect in Europe Same knowledge base, different conclusion because the context shifts the rule priorities. How do we explain conflicting results? Generate hypotheses consistent with metadata: • Regional implementation differences • Population differences • Variations in measurement or design Abductive reasoning: Provides explanation and hypotheses on contextual metadata [1] S. Al Manir, J. Niestroy, M. A. Levinson, and T. Clark (2021). Evidence Graphs: Supporting Transparent and FAIR Computation, with Defeasible Reasoning on Data, Methods, and Results. IPAW 2020-2021. [2] I. Asif, I. Tiddi, and A. J. G. Gray (2021). Using Nanopublications to Detect and Explain Contradictory Research Claims. eScience 2021. [3] T. Clark, P., and C. Goble (2014). Micropublications: a Semantic Model for claims, evidence, Arguments and Annotations in Biomedical Communications. Journal of Biomedical Semantics, vol. 5(28). [4] N. Fanizzi and C. D’amato (2025). The Blessing of Dimensionality Perspectives of Reasoning and Learning on Hyperdimensional computing/vector Symbolic Architectures. Neurosymbolic Artificial Intelligence, vol. 1. [5] A. Hogan et al. (2022). Knowledge Graphs. ACM Computing Surveys, vol. 54(4). [6] S. Todorovikj, G. Kern-Isberner, and M. Ragni (2021). On the Cognitive Adequacy of Non-monotonic Logics. NMR 2021 @ KR 2021. ✉ [email protected]