A Neuro-Symbolic Framework Learns Constraint Networks from Examples

Summary

Constraint acquisition traditionally depends on repeated interaction with a human oracle, making it costly in time and queries. This paper presents a neuro-symbolic framework for automating that process from previously available examples. Neural Oracle Transformer models learn to emulate user responses and generalize the underlying conceptual knowledge. Their responses are passed to FastCA, a dedicated constraint-acquisition engine that systematically refines them into a sound, consistent, and interpretable constraint network. The approach can recover structured symbolic models without prior domain knowledge. The reported results show that combining data-driven pattern recognition with symbolic reasoning can reduce user involvement while supporting automated model construction in combinatorial domains.