From Large Language Model Predicates to Logic Tensor Networks: Neurosymbolic Offer Validation in Regulated Procurement

Hybrid AI approach improves transparency in high-stakes decisions: The paper demonstrates that combining language models with Logic Tensor Networks creates a system where public procurement decisions can be traced back to specific rules, text passages, and reasoning steps—addressing a critical need in regulated institutions where decisions must be legally defensible and auditable.

Interpretability doesn't require sacrificing performance: The neurosymbolic system achieved performance comparable to existing models while providing significantly better explainability, suggesting that transparency and accuracy are not necessarily trade-offs in AI systems used for important institutional decisions.

Domain knowledge can be preserved alongside AI automation: By explicitly incorporating existing legal and procurement rules into the AI system, the approach allows institutions to maintain their established expertise and compliance frameworks rather than replacing them entirely with black-box algorithms.

Explainability enables accountability in regulated contexts: The system can justify its decisions by showing which predicates (extracted facts), which rules were applied, and which text passages supported the conclusion—providing the kind of step-by-step reasoning that regulators, auditors, and affected parties need to verify fairness and correctness.

Modular design supports practical implementation: The separation of information extraction from decision-making allows institutions to update rules independently of the language model, making the system more maintainable and adaptable to changing regulations without requiring complete retraining.