Large Language Model-Enhanced Symbolic Reasoning for Knowledge Base Completion

Structured knowledge bases (KBs) enable a range of language understanding tasks including question answering, dialogue modelling, and semantic retrieval. Knowledge base completion (KBC) supports these tasks by inferring missing facts from partially observed relational graphs, improving coverage and consistency in structured language representations. While rule-based KBC methods offer verifiable reasoning, they often lack flexibility and coverage; in contrast, large language models (LLMs) demonstrate strong semantic generalization but are prone to hallucination and lack transparency. To integrate the linguistic capacity of LLMs with the structural clarity of symbolic reasoning, we propose LeSR, a hybrid framework composed of a Subgraph Extractor, an LLM Proposer, and a Rule Reasoner. The Subgraph Extractor identifies context-rich substructures within the KB, which the LLM uses to synthesize diverse candidate logic rules. These proposed rules are then refined and scored by the Rule Reasoner, which filters unreliable patterns and supports interpretable inference over the KB. LeSR combines LLM-enhanced rule diversity with symbolic validation, improving both robustness and transparency. Empirical results across five diverse KB benchmarks demonstrate that LeSR delivers strong predictive performance and high rule interpretability, establishing it as a generalizable framework for structured reasoning in language-based knowledge tasks.