Neuro-symbolic AI for proactive risk forecasting: bridging large language models and dynamic Bayesian networks

Purpose Periodic risk reports constitute a substantial repository of operational data for project management. Despite the inherent informational value, archived records remain largely unstructured and highly fragmented. This study proposes a fully local neuro-symbolic approach for dynamic risk forecasting by translating qualitative project narratives into a quantifiable predictive engine to support decision-making. Design/methodology/approach The proposed methodology adopts a mixed-methods architecture including four layers: (1) spatial morphology reconstruction using a Y-axis density clustering algorithm; (2) a hybrid pseudonymization filter; (3) a local large language model (LLM) with chain-of-thought reasoning; and (4) a symbolic causal engine powered by a dynamic Bayesian network (DBN) to quantify causal propagation into a dynamic risk index. Findings An empirical evaluation across 28 reporting cycles spanning a 33-month large-scale infrastructure project indicates that the symbolic layer provides a temporally coherent aggregation of the output of a zero-shot perception layer operating at a macro F1-score of 53.62%. By quantifying causal risk propagation, the DBN captured a 93.59% risk-saturation state, for which the retrieved evidence identifies chronic site-clearance bottlenecks. Notably, using an alert threshold selected by sensitivity analysis together with an analytic constraint on the posterior mass attributable to the critical state (talert = 70.0%), the model outputs Bayesian probabilities alongside LLM-extracted textual evidence, yielding an anticipation interval of 1.33 reporting cycles relative to the documented administrative response. Originality/value The neuro-symbolic framework bridges the gap between deep learning perception and causal inference in construction management. By coupling LLM-driven zero-shot ordinal extraction with DBN causal reasoning, the research establishes an auditable framework for risk documentation. The integrated framework transitions AI from an opaque prediction model into an accountable decision-support tool, providing stakeholders with transparent explainable AI diagnostics and supporting a shift from reactive mitigation to proactive, evidence-based risk prevention.