An Uncertainty-Gated Neuro-Symbolic Framework for High-Coverage Topic Modeling and Trend Analysis in Scholarly Corpora with LLM Assistance

Abstract

The rapid growth of scientific literature demands scalable methods that can track research evolution, yet density-based topic models such as BERTopic systematically exclude low-density documents as outliers, obscuring emerging and niche research areas. We propose a Neuro-Symbolic, Uncertainty-Gated Framework that recovers these outliers through geometric centroid reassignment and an ontological entropy gate derived from the Computer Science Ontology (CSO), routing only genuinely ambiguous cases to a local Large Language Model (Qwen2.5-14B via Ollama). A controlled ablation study demonstrates that centroid reassignment provides the largest coverage gain (+ 22.9 percentage points (pp)), the CSO entropy gate preserves niche-topic integrity, and selective LLM routing adds an additional + 5.9 pp. On 12,535 Turkish computer engineering theses (TR-CS; 2001-2025), the full pipeline raises coverage from 75.5% +/- 1.2 % (Bare BERTopic) to 95.7% +/- 0.4% (five-seed means) while maintaining competitive coherence (NPMI = 0.112 +/- 0.006) and cross-seed stability (AMI = 0.832 +/- 0.015), at similar to 15x fewer LLM calls than a fully generative Pure-LLM baseline. Mann-Kendall trend tests on the high-coverage series identify 69 statistically significant trends (FDR q < 0.05), and cross-corpus validation on similar to 200K arXiv CS abstracts confirms that the architecture generalizes beyond the primary dataset. The framework offers a reproducible, cost-effective solution for monitoring scientific developments in rapidly evolving fields.