T2B2T: Bridging RDF and BDI Agent Cognition

• T2B2T is a paradigm defining a bidirectional mapping between external RDF triples and internal BDI cognitive states in agents.
• It operationalizes the conversion of RDF knowledge into beliefs, desires, and intentions using formal OWL and Description Logic patterns.
• Implemented in Semas, the framework ensures semantic interoperability and sets the stage for advances in automated planning and conflict resolution.

The Triples-to-Beliefs-to-Triples (T2B2T) paradigm defines a bidirectional mapping between external RDF knowledge graphs and the internal Belief-Desire-Intention (BDI) reasoning state of agents. Instantiated in the Semas reasoning platform, T2B2T operationalizes how knowledge encoded as RDF triples can drive agent cognition and, conversely, how agent commitments are externalized back to the semantic data layer. This paradigm is built atop a formal BDI ontology that specifies precise roles for beliefs, desires, and intentions, enabling semantic interoperability and explainability for multi-agent and neuro-symbolic systems engaging with the Web of Data (Zuppiroli et al., 21 Nov 2025).

1. Formalization of the T2B2T Cycle

The T2B2T paradigm centers around three core mappings that constitute a cycle between external data and internal cognitive states:

• Let $\mathcal{T}$ denote the set of RDF triples from an external knowledge base.
• $\mathcal{B}$ is the set of Belief instances, $\mathcal{D}$ the Desires, and $\mathcal{I}$ the Intentions in the agent’s mental state.

The T2B2T loop is defined via:

1. $f_1: \mathcal{T} \rightarrow \mathcal{B}$: Parsing RDF triples into beliefs. For each triple $t \in \mathcal{T}$ matching world-state patterns, a Belief $b$ is instantiated such that $bdi{:}refersTo(worldState(t))$, with temporal/provenance annotations.
2. $$f_2: \mathcal{B} \rightarrow \mathcal{D} \times \mathcal{I}$$: Deliberation transforms beliefs into desires and intentions. This step can be decomposed as $f_{21}: \mathcal{B} \rightarrow \mathcal{D}$ and $f_{22}: \mathcal{D} \rightarrow \mathcal{I}$, under the ontology’s axioms ($Belief \sqsubseteq \exists motivates.Desire$, $Intention \sqsubseteq \exists fulfils.Desire$).
3. $f_3: \mathcal{I} \rightarrow \mathcal{T}'$: Projection of intentions into new RDF triples. Each committed intention $i \in \mathcal{I}$ is rendered as a set of triples encoding the intention, associated plan, and addressed goals.

Semas interleaves these mappings in real time, ensuring continuous synchronization between the knowledge graph and agent cognition.

2. Core Ontology Structure and Description Logic Patterns

The T2B2T process is grounded in a modular BDI Ontology, whose main constructs and relations are formalized in OWL and Description Logic (DL):

Principal Classes:

• MentalEntity, MentalState, Belief, Desire, Intention
• WorldState, MentalProcess (+ BeliefProcess, DesireProcess, IntentionProcess)
• Plan, Goal, Task, PlanExecution, Action, Justification, TemporalEntity

Key Object Properties:

Property	Domain(s)	Range(s)
refersTo	MentalEntity	WorldState
motivates	Belief	Desire
fulfils	Intention	Desire
specifies	Intention	Plan
addresses	Plan	Goal
hasComponent	Plan	Task
bringsAbout	PlanExecution, Action	WorldState
atTime	MentalEntity, Action,<br>PlanExecution	TemporalEntity

Representative Axioms:

• Belief $\sqsubseteq$ MentalState
• Belief $\sqsubseteq \exists$ refersTo.WorldState
• Belief $\sqsubseteq \exists$ motivates.Desire
• Intention $\sqsubseteq \exists$ fulfils.Desire
• Intention $\sqsubseteq \exists$ specifies.Plan
• Plan $\sqsubseteq \exists$ addresses.Goal

This ontology provides the formal machinery for mapping, inference, and validation within T2B2T.

3. Operational Workflow in Semas

The Semas platform realizes T2B2T through a structured workflow:

Ingestion ($f_1$):

1. Semas loads RDF triples $\mathcal{T}$ from SPARQL endpoints or Turtle files.
2. SPARQL-based ingestion rules select triples describing world states, percepts, or mental entities (typically those with predicates such as $bdi{:}refersTo$, $bdi{:}atTime$).
3. Selected triples are transformed into Prolog-like facts: \texttt{belief(Agent, BeliefID)}, with auxiliary facts for properties and annotations. These correspond internally to OWL individuals of $bdi{:}Belief$.

Deliberation ($f_2$):

1. The rule engine (production rules) matches belief facts, evaluates context conditions (e.g., absence of conflicts), and applies action rules.
2. Upon firing, rules generate new desires (\texttt{desire(Agent, D)}) linked by $isMotivatedBy$, and intentions (\texttt{intention(Agent, I)}) linked by $fulfils$, with optional references to plans.
3. Complex goal-selection and planning can be achieved by chaining rule invocations or invoking external planners, registering plans as $bdi{:}Plan$ individuals.

Projection ($f_3$):

1. At each tick, Semas identifies newly adopted intentions.
2. For each intention, a corresponding RDF graph $\mathcal{T}'$ is emitted, containing at least:
   • $:AgentX\ bdi{:}hasIntention\ :I$
   • $$:I\ a\ bdi{:}Intention;\ bdi{:}fulfils\ :D;\ bdi{:}specifies\ :PlanY$$
   • $$:PlanY\ a\ bdi{:}Plan;\ bdi{:}addresses\ :GoalZ;\ bdi{:}hasComponent\ :TaskA$$
3. The produced triples are either published into the shared RDF store or delivered as action requests.

This pipeline ensures coherence and operational closure between knowledge representation and agent deliberation (Zuppiroli et al., 21 Nov 2025).

4. Experimental Evaluation

Quantitative evaluation of the Semas T2B2T implementation is not provided in the principal reference. The only experiments reported involve coupling the BDI ontology with GPT-4 in Logic Augmented Generation (LAG), focusing on improvements to inferential coherence and consistency. No timing, throughput, or correctness metrics for T2B2T round-trips are reported. The operational integration with Semas has been documented elsewhere ([Longo 2025], not replicated in (Zuppiroli et al., 21 Nov 2025)), so the demonstration in this context should be understood as conceptual rather than empirically validated.

A plausible implication is that subsequent work must address systematic benchmarks to quantify round-trip latency, rule-firing rates, and the semantic coherence of agent projections via T2B2T in large, dynamic knowledge graphs (Zuppiroli et al., 21 Nov 2025).

5. Challenges, Limitations, and Extensions

Several operational and conceptual challenges are identified:

• Conflict Modelling: The current ontology lacks explicit axioms for intention–intention conflicts. Conflict detection is delegated to procedural rules or external modules. An enrichment with an "IncompatibleIntentions" pattern and supporting reasoning rules is a potential extension.
• Scalability: Continuous projection and re-ingestion of intentions as RDF triples in a multi-agent context can lead to bottlenecks. Proposed optimizations include delta-uploading and event-based RDF streams.
• Automated Planning: Although declarative linkage of plans and goals is supported, full integration of external planners (e.g., HTN or PDDL-based) is not yet realized.
• Quantitative Metrics: Standardized metrics such as round-trip T2B2T latency, rule-processing throughput, and semantic coherence (percent of BDI competency questions satisfied) on real-world graphs are not yet reported.
• Richer Mental Dynamics: Possible future directions include belief revision (AGM postulates), intention reconsideration, and utility-based desire ranking, each grounded in extended OWL patterns and Semas rules.

These challenges delimit the current operational envelope of T2B2T while indicating clear directions for research and engineering improvements (Zuppiroli et al., 21 Nov 2025).

6. Significance and Implications

T2B2T mediates between declarative knowledge (RDF triples) and procedural agent cognition (BDI), constituting an operational and conceptual bridge for neuro-symbolic and multi-agent AI. The paradigm enables explainable, semantically grounded reasoning, supports multi-agent interoperability via common ontologies, and can extend to real-time, web-scale data environments. The modularity of the underlying BDI ontology assures extensibility and alignment with foundational ontological frameworks, facilitating further integration across cognitive architectures and reasoning infrastructures (Zuppiroli et al., 21 Nov 2025).

In summary, the T2B2T paradigm as realized in Semas provides a systematic framework for closing the loop between open semantic data and agent intentionality, with the potential to undergird sophisticated, explainable, and interoperable AI systems operating over the evolving Web of Data.