Knowledge graphs and rules for compliance
Knowledge graphs appear on 141 recent papers and neuro-symbolic AI on 14, almost all in AI journals; only 34 of 481 such papers concern accounting, tax, legal, regulatory, financial or audit uses. Compliance is a domain of explicit rules and entities, which language models alone handle badly. Do graphs and symbolic rules make them reliable enough for audits and tax?
What prompted this
Since October 2025, 481 papers carry a knowledge-graph, neuro-symbolic, symbolic-reasoning, rule-based, ontology, formal-verification or logical-reasoning keyword (451 in AI journals): "knowledge graphs" on 141, "knowledge graph completion" on 41, "temporal knowledge graphs" on 38, "neuro-symbolic AI" on 14, "symbolic reasoning" on 8. Only 34 of the 481 also carry an accounting, tax, legal, regulatory, compliance, finance, contract, audit, law, bank, fraud or anti-money-laundering keyword, although "regulatory compliance" appears on 60 papers in the archive.
Questions
- On tasks such as checking a transaction against sanctions lists and ownership chains, or applying a tax rule to a set of facts, do systems that combine a language model with a knowledge graph and explicit rules make fewer errors than the model alone, and are their errors explainable?
- How much effort does it take to keep the rules and graphs current when regulations change, and who maintains them?
- Would auditors and regulators accept outputs with symbolic justifications where they reject model-only outputs?
Data that could help
Public regulatory texts and sanctions lists, company ownership graphs (Open Ownership, Orbis), and test cases with known answers from tax and compliance training materials.
Next step
Build a benchmark of 200 compliance cases (sanctions screening with ownership chains, tax rule application) and compare a language model alone with the same model backed by a graph and rules.