@inproceedings{wiedmer_neurosymbolic_2026a,
   abstract = {Natural language interfaces promise to simplify access to enterprise data, but their use in regulated environments is complicated by large heterogeneous schemas, domain-specific semantics, and the limited reasoning capacity of lightweight on-premises language models. This paper presents a neurosymbolic approach to enterprise Text-to-SQL that combines such models with symbolic reasoning over a schema-level meta knowledge graph, separating natural language interpretation from deterministic query construction. A prototype is evaluated on semi-automatic knowledge graph construction, symbolic join reasoning, and end-to-end Text-to-SQL generation. The generated knowledge graph achieves high retrieval coverage, and symbolic join reasoning substantially improves execution accuracy while reducing performance variability compared with a purely neural baseline. Combining lightweight language models with symbolic reasoning therefore offers a robust route to natural language access to heterogeneous enterprise data under on-premises deployment constraints.},
   address = {Thessaloniki, Greece},
   author = {Wiedmer, Silvan, Weichselbraun, Albert, Brasoveanu, Adrian M.P. and Süsstrunk, Norman},
   booktitle = {11th {International} {Conference} on {Web} {Intelligence}, {Machine} {Intelligence} and {Semantics} ({WIMS} 2026)},
   month = {October},
   title = {A {Neurosymbolic} {Approach} to {Enterprise} {Text}-to-{SQL} {Using} a {Schema}-{Level} {Meta} {Knowledge} {Graph}},
   year = {2026}
}