. (2026). “Neurosymbolic Join Resolution for Enterprise Text-to-SQL”. Posters, Demos, Blue Sky, and Tutorials at SEMANTiCS 2026, Ghent, Belgium
Enterprise Text-to-SQL systems enable natural-language access to relational databases but remain challenging in real-world environments, where large schemas, undocumented business semantics, and implicit relationships frequently lead to incorrect SQL joins. These limitations are particularly problematic in high-stakes domains requiring reliable, explainable, and on-premises AI solutions. This paper presents a neurosymbolic approach that combines lightweight language models with a symbolic Semantic Join Path Resolver operating on a schema-level knowledge graph. Rather than requiring the language model to infer complex join sequences, the symbolic component deterministically computes valid join paths, allowing the model to focus on semantic interpretation and generation of the remaining SQL clauses. We evaluate the approach on an enterprise benchmark derived from a production database containing nearly 300 tables, comparing it with a purely neural Text-to-SQL baseline across five open-weight models ranging from 1.7B to 31B parameters. Symbolic join resolution raises execution accuracy substantially for models of at least 8B parameters, holding Soft F1 close to 1 where neural baselines degrade, and allows an 8B model to match a 31B one. Below that capability threshold the effect reverses: the smallest models fail to exploit the resolver’s output and score lower with it than without. The proposed architecture shows both how neurosymbolic AI can improve the reliability of enterprise natural-language interfaces under practical deployment constraints, and where that benefit runs out.