Verifiable Knowledge Expansion with Retrieval-Grounded FCA

Constructing ontologies and expanding knowledge bases requires careful validation of objects, attributes, and their structural relationships. While language models can propose such structures from text, their outputs often lack verifiable support or consistency. To address this, researchers introduce a novel framework that combines a retrieval-augmented small language model (SLM) with Formal Concept Analysis (FCA) as a symbolic verification loop for knowledge expansion. The framework begins with seed attributes, from which FCA proposes logical implications over a growing formal context. A retrieval-grounded SLM oracle then validates each proposed implication, providing either confirmation or a counterexample. This oracle also handles incidence judgments, consistency checks, and attribute proposals, ensuring that all accepted implications, counterexamples, contradictions, and corrections are inspectable and traceable. In a case study focused on rare ataxia, using resources from Orphadata, the system achieved relation F1 scores of 0.29-0.52 and closure-based implication F1 scores of 0.22-0.30. Larger seed sets generally improved implication F1, though lower scores for implications reflect a stricter evaluation. Ablation studies confirmed that incidence judgments can improve scores in fixed object-attribute settings, but identifying positive object-attribute pairs remains challenging.