Semas Reasoning Platform

• Semas Reasoning Platform is a cognitively grounded system that integrates structured RDF knowledge graphs with BDI mental state representations.
• It employs a modular T2B2T cycle to transform triples into beliefs, desires, and intentions through distinct Prolog-based components.
• The platform offers explainable, provenance-rich outputs, ensuring semantic interoperability and effective integration in neuro-symbolic and multi-agent systems.

The Semas Reasoning Platform is a research-grade, cognitively grounded reasoning system for agent-based AI, explicitly designed to provide full bidirectional semantic interoperability between structured RDF knowledge graphs and agent mental states as formalized in the Belief-Desire-Intention (BDI) paradigm. Semas operationalizes the T2B2T (“Triples-to-Beliefs-to-Triples”) cycle to drive symbolic deliberation over Web and IoT data, producing explainable and provenance-rich outputs suitable for integration in neuro-symbolic, conversational, and multi-agent systems. The platform is implemented in Prolog and fully reifies its internal reasoning steps as ontologically principled RDF graphs, leveraging the BDI OWL ontology to guarantee alignment with declarative, interoperable knowledge resources (Zuppiroli et al., 21 Nov 2025).

1. System Architecture and Principal Modules

The Semas platform is structured into five architecturally distinct but interoperating modules:

• RDF Ingestor: Implements a SPARQL-based loader (via Apache Jena or custom Prolog RDF libraries) to ingest one or more RDF graphs $G_0,\dots,G_n$. Triples $(s,p,o)$ are parsed and stored as both Prolog facts rdf(s,p,o) and as raw RDF annotations in an internal triple store.
• T2B2T Mapper: Encodes the mapping functions $\varphi: \text{Triples} \rightarrow \text{Beliefs}$, transforming selected RDF properties into agent beliefs, desires, and intentions. Mappings are user-configurable via declarative mapping files determining which RDF schema elements become mental state predicates (e.g., belief/3, desire/2, intention/2).
• BDI Reasoning Engine: A Prolog production-rule interpreter acting on mental-state predicates. Rules are specified in the form $$[ \text{HEAD} ] / [ \text{CONDITIONALS} ] \gg [ \text{TAIL} ]$$, with schedule-driven alternation between belief, desire, intention, and plan phases. The engine supports depth-first rule firing, backtracking, and meta-level management of deliberative cycles.
• Triple Emitter: Defines the inverse mapping $\psi: (B, D, I) \rightarrow \text{Triples}$ to emit updated or newly generated mental states as BDI-ontology-typed RDF triples.
• Explanation Generator: Whenever a mental-process rule fires, generates explicit :MentalProcess RDF individuals with causal and process links (e.g., bdi:generates, bdi:modifies, bdi:suppresses), providing a trace-annotated provenance chain for explainability (Zuppiroli et al., 21 Nov 2025).

2. The T2B2T Reasoning Paradigm

The T2B2T cycle provides a semantically principled and operational flow for integrating Web of Data knowledge into agentic reasoning:

1. Mapping into Mental States: $\varphi : G_0 \rightarrow (B_0, D_0, I_0)$. For each RDF triple where $p$ is declared a “belief-property,” Semas asserts Prolog facts such as belief(s,p,o).
2. BDI Reasoning: $$\langle B_0, D_0, I_0 \rangle \vdash_R \langle B_1, D_1, I_1 \rangle$$ via a set of inference rules $R$. Rules take the form

$$\frac{\text{BeliefProcess}(P)\;\land\;\text{refersTo}(P, W)} {\text{generates}(P, b)\;\land\;\text{Belief}(b)\;\land\;\text{refersTo}(b, W)}$$

and manage belief, desire, and intention formation and update, with fixpoint progression until exhaustivity or new inputs.

1. Projecting Back into RDF: $$\psi : \langle B_1, D_1, I_1 \rangle \rightarrow G_1$$. All generated mental state individuals and process/justification resources are translated to RDF according to BDI ontology URI templates.

By repeating this loop on new data arrivals or temporal triggers, Semas maintains a live synchronization between agentic deliberation and its semantic web environment (Zuppiroli et al., 21 Nov 2025).

3. BDI Ontology Integration and Semantic Interoperability

Semas incorporates the formal BDI Ontology—developed as an OWL 2 DL schema capturing bdi:Belief, bdi:Desire, bdi:Intention, bdi:Justification, and process/agent relations—both as type system and as operational mapping specification:

• Every Prolog predicate directly matches a BDI ontology class or property.
• Output RDF is strictly typed using canonical URIs (e.g., https://w3id.org/fossr/ontology/bdi/).
• All mental state individuals are linked by BDI properties (e.g., bdi:motivates, bdi:generates, bdi:justifies), enabling seamless SPARQL/OWL consumption.
• SPARQL endpoint exposes live agent mental-states for external query.
• Alignment with W3C DOLCE and TimeIndexedSituation patterns is maintained for interoperability with third-party ontologies and tools (Zuppiroli et al., 21 Nov 2025).

Semantic interoperability is assured by validating all inputs and outputs against BDI OWL axioms, and by applying W3C best practices for modular, DUL-compliant ontological integration.

4. Explainability and Provenance

Explainability is natively supported at both the process and outcome levels:

• Traceability: Each reasoning step generates distinct :MentalProcess RDF individuals linked via bdi:reasonsUpon, bdi:generates, bdi:modifies, or bdi:suppresses properties to explicit mental states.
• Justification: Every belief, desire, or intention is annotated with a :Justification resource, containing structured or textual rationales.
• Provenance: Uniform naming conventions for internal (Prolog) and output (RDF) resource URIs allow automatic construction of provenance and explanation chains with no ad hoc logging (Zuppiroli et al., 21 Nov 2025).

This architecture enables full reconstruction of deliberative traces and rationales at the level of Web queries and facilitates integration with explainability-driven applications.

5. Illustrative Use Cases and Examples

Semas demonstrates its design via two representative examples:

• Hotel Check-In Contradiction: Upon RDF input asserting “isCurrentlyAt ex:Location_home” during a weekend-morning, Semas generates a belief, forms a location-appropriate desire, then evaluates intentions for “Check into the hotel,” generating a justification for infeasibility and finally forming a new home-appropriate intention. The entire inferential episode is mirrored as typesafe BDI RDF individuals, processes, and justifications in the output graph.
• Zelle Payment Notification: Ingesting a payment request triple, Semas fires a process yielding a new payment belief, triggers a desire to pay, then forms an intention, each with associated process and justification nodes, all exported as BDI-typed RDF.

All reasoning products are addressable and queryable via SPARQL as first-class Web data (Zuppiroli et al., 21 Nov 2025).

6. Performance, Scalability, and Applicability

Prototypical performance evaluation indicates:

• Rule firing and triple projection cycles on the order of tens of milliseconds for RDF graphs of several hundred triples.
• Linear scalability with the number of concurrently managed mental state individuals (dozens practical).
• Negligible computational overhead leveraging Prolog’s efficient indexing and backtracking, with incremental SPARQL ingestion.
• Sustained real-time BDI reasoning over live RDF streams typical of IoT or conversational applications, with full OWL-DL compliance and rich provenance (Zuppiroli et al., 21 Nov 2025).

A plausible implication is that larger-scale deployments would benefit from horizontal partitioning, but such extension is not detailed in current reports.

7. Relation to Broader Research and Applications

Semas directly instantiates the conceptual and operational bridge between declarative (triple-based) and procedural (agentic/mental-state-based) intelligence required in neuro-symbolic and multi-agent systems. Its strict adherence to OWL, RDF, and BDI schema patterns positions it as a reference implementation for explainable, cognitively grounded, and interoperable Web agents in domains such as IoT, conversational systems, and cognitive robotics (Zuppiroli et al., 21 Nov 2025).