Survey of 50-plus neurosymbolic systems proposes provenance semirings as a unifying rule-language foundation and a scenario-to-feature decision matrix

It proposes a four-axis taxonomy (semantics, expressiveness, neural integration, evaluation mechanism) for rule-based neurosymbolic languages, with tables mapping six semantic categories, Boolean-probabilistic, differentiable-probabilistic, stable-model, soft-unification, fuzzy/t-norm, and algebraic-circuit, to representative systems. Earlier work typically presents individual systems or single formalism families; this survey separates semantic choice, syntactic features, neural interface style, and evaluation strategy into independently comparable dimensions, making trade-offs between dialects explicit. Based on organising and classifying over 50 systems and applications, presented as tables linking category, semantics, and representative systems.

It uses provenance semirings as a unifying algebra: changing the semiring changes the semantics without changing the program, so a single stratified Datalog engine with a differentiable semiring (Scallop, Lobster, DOLPHIN) already covers the visual-reasoning, planning, and program-analysis rows of the scenario matrix. It connects the provenance-semiring result of Green et al. with convergence conditions for recursive Datalog, and notes that possible-world probabilities do not arise from a direct semiring homomorphism but require weighted model counting over the Boolean provenance polynomial. The argument rests on the formal derivation-tree sum and the -continuity condition, citing convergence characterisation work; it also states explicitly that the temporal and soft-unification columns lack an obvious semiring formulation.

It provides a decision matrix from scenarios to features, showing that perception-driven tasks need probabilistic semantics, visual reasoning also needs aggregation and negation, spatio-temporal tasks need temporal operators, knowledge graphs rely on soft unification, and natural-language reasoning relies on foreign-function interfaces, with no single dialect covering all scenarios. It aligns language-feature requirements directly with concrete application scenarios, shifting selection from formalisation preference to task-constraint-driven choice. Based on a curated corpus across four research communities and six recurring scenarios, with a table marking required features per scenario.

It distils five open problems: static analysis and repair of LLM-generated rule sets, source-level unification of Boolean, probabilistic, fuzzy, and embedding semantics, bridging Datalog and ASP in one differentiable system, scalability and error bounds for probabilistic inference, and verification of neurosymbolic programs. It turns current capability boundaries into an actionable research agenda, for example noting that approximate inference currently returns point estimates without error bounds, and that trigger-graph materialisation could be lifted to propagate intervals or label distributions. Grounded in the survey of existing evaluation mechanisms; it is a problem statement and direction proposal rather than an experimental result.

The survey targets researchers and engineers who need an inspectable rule layer above neural components, covering Datalog and its extensions, answer set programs, Prolog-style probabilistic logic programs, and inductive logic programming methods that learn such rules; purely sub-symbolic models, including graph neural networks, are explicitly out of scope. The decision matrix gives feature requirements for six scenario classes and can guide trade-offs among semantics, expressiveness, neural integration, and evaluation mechanism; provenance semirings are positioned as the closest thing to a common foundation and a natural starting point for a unified engine.

The corpus centres on the past six years and over 50 systems, with AI dominating and vision and robotics less represented, so matrix coverage for those two may be thinner than for AI. Temporal operators and soft unification still lack an obvious semiring formulation, foreign-function interfaces are orthogonal to the semiring choice, and probabilistic inference remains the scalability bottleneck; approximate methods currently return point estimates without error bounds, and when computable bounds can be derived remains open. In addition, reasoning shortcuts and symbol grounding are noted as largely orthogonal to language design, and the generality of their mitigations still needs further observation.