Explore the scholarly record.

Abstract

Abstract Technical knowledge exploration demands both the precision of symbolic reasoning over structured data and the flexibility of natural language interaction. We present a loosely-coupled neuro-symbolic architecture that integrates explicit symbolic reasoning through ontology-driven knowledge graphs with neural language understanding and generation. Our architecture maintains distinct symbolic and neural subsystems that communicate through structured evidence transfer, preserving the interpretability and auditability of graph-based reasoning while leveraging the adaptability of large language models for conversational interaction. The symbolic subsystem performs data integration across heterogeneous repositories through focused crawling, ontological schema unification and graph materialization with rule-based inference. The neural subsystem handles natural language query interpretation, semantic retrieval through embeddings and contextual response generation. The subsystems remain modular by design: symbolic graph operations produce provenance-preserving evidence artifacts that the neural components consume for language generation, without joint training or differentiable symbolic computation within the language model. We use our framework for cybersecurity vulnerability analysis, integrating four authoritative repositories into a unified knowledge graph containing 50,000+ entities with 200,000+ materialized relationships. Our evaluation shows that the loosely-coupled hybrid approach improves response correctness by 79%, completeness by 83%, analytical depth by 82% and cross-source synthesis by 100% over neural-only baselines with web search. Head-to-head comparisons against a dense-retrieval RAG pipeline and an LLM-driven GraphRAG system built over the same corpus confirm these gains. A provenance audit of the two graph substrates shows that 100% of the edges in our materialised graph are traceable to authoritative source records, against 83% for the LLM-extracted GraphRAG graph. These results indicate that modular neuro-symbolic integration benefits conversational knowledge systems and that separating symbolic reasoning from neural language processing improves explainability without compromising analytical quality. The architecture extends to technical domains that require conversational access to distributed authoritative knowledge, among them healthcare informatics, legal research and scientific literature analysis. We contribute architectural patterns for building loosely-coupled neuro-symbolic systems where symbolic knowledge structures guide and constrain neural generation without sacrificing the interpretability that pure neural approaches typically lack.