Bridging BDI Multi-Agent Systems and the Semantic Web Through the

Triples-to-Beliefs-to-Triples Paradigm

Carmelo Fabio Longo

1 a

, Rocco Paolillo

1 b

, Misael Mongiov

2 c

, Andrea Giovanni Nuzzolese

1 d

Francesco Poggi

1 e

, Michele Geremia Ceriani

1 f

, Antonio Zinilli

1 g

, Giusy Giulia Tuccari

1 h

and Corrado Santoro

2 i

National Research Council, Italy

Department of Mathematic and Informatics, University of Catania, Italy

Keywords:

Artiﬁcial Intelligence, BDI Agents, Semantic Web, Multi-Agent System.

Abstract:

Well-established agent engineering frameworks from the state-of-the-art, due to their outdated designs, are not

thought to work in the perspective of a shared semantics, nor do they provide an agent modeling language and

environment that integrates seamlessly with them. This is especially challenging in dynamic, distributed envi-

ronments where new concepts, data sources, and agents can emerge at runtime, potentially leading to semantic

conﬂicts or inconsistencies. This paper proposes the novel paradigm Triples-to-Beliefs-to-Triples (T2B2T),

which is being ontologically described, enabling multi-agent systems with seamless and consistent integra-

tion with the Semantic Web. In order to validate the approach, this paper proposes also a framework called

SEMAS implementing the T2B2T paradigm, which provides a bridge between the mental attitudes Beliefs-

Desire-Intentions (BDI) and triples describing a domain with an abstraction over the SPARQL language that

feeds the inference process of agents. This enables more sophisticated forms of reasoning in the closed-world

assumption, by supporting predicates without any limitation on arity and compositional structures, allowing

also the employment of decentralized functions for the dynamic generation of new triples not included in the

origin ontologies. As a case-study, SEMAS was employed on decision-making applied to academic mobility

with real data coming from the SCOPUS database, demonstrating how the generated inferences can be tailored

to speciﬁc conditions of individual agents, and how new triples can be inferred to capture the impact of agents’

decisions on the evolution of the knowledge domain.

1 INTRODUCTION

A seamless integration with the Semantic Web plays

a fundamental role in Multi-Agent Systems (MAS),

whether they are virtual (as in social simulations),

physical, or both (digital twins). MAS and the Se-

mantic Web have traditionally evolved independently,

each addressing distinct challenges within the realm

https://orcid.org/0000-0002-2536-8659

https://orcid.org/0000-0001-9816-5839

https://orcid.org/0000-0003-0528-5490

https://orcid.org/0000-0003-2928-9496

https://orcid.org/0000-0001-6577-5606

https://orcid.org/0000-0002-5074-2112

https://orcid.org/0000-0001-8505-5040

https://orcid.org/0009-0008-3298-7168

https://orcid.org/0000-0003-1780-5406

of Artiﬁcial Intelligence and Linked Open Data.

The former focuses on the design and coordination

of autonomous agents capable of decision-making,

problem-solving, and interacting with other agents

or the environment. The Semantic Web envision

(Berners-Lee et al., 2001), on the other hand, empha-

sizes enriching web data with meaning, allowing ma-

chines to interpret and process information in a more

intelligent, context-aware manner. In such a scope,

the following research question arises: how to enable

dynamic and semantically consistent interoperability

between autonomous agents in heterogeneous Seman-

tic Web ecosystems? More speciﬁcally, the integra-

tion of MAS with the Semantic Web faces the chal-

lenge of ensuring that autonomous agents, each refer-

ring to potentially distinct ontologies, vocabularies,

and reasoning mechanisms, can dynamically interact

and exchange knowledge in a meaningful and seman-

Longo, C. F., Paolillo, R., Mongiovì, M., Nuzzolese, A. G., Poggi, F., Ceriani, M. G., Zinilli, A., Tuccari, G. G. and Santoro, C.

Bridging BDI Multi-Agent Systems and the Semantic Web Through the Triples-to-Beliefs-to-Triples Paradigm.

DOI: 10.5220/0013668800003985

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 21st International Conference on Web Information Systems and Technologies (WEBIST 2025), pages 327-334

ISBN: 978-989-758-772-6; ISSN: 2184-3252

327

tically consistent way. This is particularly difﬁcult

in open, distributed, and constantly evolving environ-

ments where new concepts, data sources, and agents

may appear at runtime, potentially causing semantic

conﬂicts or inconsistencies. Assuring scalable, dy-

namic, and reliable semantic interoperability in such

contexts remains an open research problem.

This paper addresses the above challenge and pro-

poses a novel paradigm for integrating MAS and

the Semantic Web, called Triples-to-Beliefs-to-Triples

(T2B2T), which is ontologically described in this pa-

per. In this paradigm, RDF triples from a triple

store are processed and turned into beliefs, by pop-

ulating agents’ Knowledge Base (KB) and interacting

with goal oriented production rules systems to infer

newly introduced beliefs. Such inferred beliefs will

be possibly translated back into RDF triples to ei-

ther update the origin triple store or build new brand

derived Knowledge Graphs (KG). This cycle facili-

tates the seamless communication between agents and

the semantic data layer, ensuring that agents operate

based on a shared, evolving KB. The gain in such a

two-way translation of the T2B2T paradigm lies in

the chance of more sophisticated inferences involving

computations out of SPARQL and OWL-based rea-

soners, shifting from open- to closed-world assump-

tion. Moreover, this approach ensures results consis-

tency by coordinating inferences of multiple agents in

a deterministic way, relying on well-established and

shared criteria of interoperability.

The contribution of our work is twofold: the for-

malization of the T2B2T paradigm through an on-

tology, and its implementation with the SEMAS BDI

framework. The paper is organized as follows: Sec-

tion 2 provides an ontology modeling the T2B2T

paradigm; Section 3 delves into the framework mod-

ules; Section 4 provides an overview of the current

state-of-the-art in the topic; Section 5 offers a com-

prehensive overview of a case-study based on the aca-

demic mobility; ﬁnally, Section 6 concludes the paper

with some ﬁnal considerations. The code of SEMAS

is publicly available for research purposes through a

dedicated GitHub repository

2 THE ONTOLOGY

This Section shows how we ontologically modeled

the here-proposed T2B2T paradigm, whose simpli-

ﬁed functional schema is depicted in Figure 1. It is

important to clarify that such ontology must be dis-

tinguished from external KGs we want agents to in-

https://github.com/cfabiolongo/Semas

teract with seamlessly. T2B2T could support in case

agents need to make inference on KGs lacking re-

quired triples, through external functions outside of

SPARQL, e.g. outcomes from classiﬁers, large lan-

guage models, or even aggregated outcomes as in so-

cial simulation. The pipeline Triple-to-Beliefs (T2B)

of T2B2T, ﬁrst computes triples to produce beliefs in-

teracting with symbolic axioms. Beliefs can either

have symbolic notation or can be references to sub-

symbolic information. In this conﬁguration, we use

Prolog-like engines to combine predicates with ar-

ity greater than two, overcoming the limitations of

the SWRL (World Wide Web Consortium, 2004) lan-

guage, and by leveraging on inference criteria based

on the backward-chaining algorithm rather than the

forward-chaining of OWL-based reasoners like Pellet

(Sirin et al., 2007) or Hermit (Glimm et al., 2014).

After inference with axioms, inferred beliefs can

be translated into triples and fed either the starting

KG or newly introduced KGs, which is the Beliefs-

to-Triples (B2T) pipeline.

Inference must also be supported by an event

queue, in order to coordinate MAS and let them pro-

duce consistent/deterministic results in case of inter-

related inferences, i.e., when an agent’s inference (in

terms of new triples) can affect inferences of other

agents and subsequently their behavior.

We follow Rao et al. (Rao and Georgeff, 1995)

framework of Belief-Desires-Intentions (BDI), build-

ing on plans, which are abstract speciﬁcations repre-

senting both the means for achieving certain desires

and the options available to the agent. Desires are

implicitly described by plans aimed at achieving the

well-deﬁned goal. Moreover, each intention that the

system forms by adopting certain plans of action is

represented implicitly by using a conventional run-

time stack of hierarchically related plans, which is the

reason we omit to report them in the ontology. Here’s

some details of classes and properties modeling the

T2B2T paradigm (cf. Figure 2):

• Agent. Instances of this class represent a single

agent aimed to make inference on KGs.

• Belief. Instances of this class are referenced

by the object-property hasBelief, and refer to a

piece of information that an agent considers to

be true about the environment. Information can

be (but not limited to) a predicate with arbitrary

arity and compositional structures, and even a

sub-symbolic representation in a vectorial space.

Under the open-world assumption (which under-

pins the Semantic Web) missing information is

treated as unknown. In contrast, the closed-

world assumption interprets unknown informa-

tion as false. A transition from the open- to

WEBIST 2025 - 21st International Conference on Web Information Systems and Technologies

328

Functions

Triples-to-Beliefs

DERIVED Knowledge Graph

Beliefs-to-Triples

SPARQL

Agent

BDI Agent + Knowledge Base

Inferences

Knowledge Graph

DERIVED

BDI Knowledge Base

Figure 1: The simpliﬁed functional schema of the process behind the T2B2T paradigm.

the closed-world assumption is valid if and only

if an equivalence is assumed between the pred-

icates Retrieved(X)/Non-Retrieved(X) with

Asserted(Y)/Non-Asserted(Y)

, where X de-

notes a triple extracted from the Semantic Web

and Y represents its corresponding belief. This

equivalence holds under the condition that the

mapping between X and Y is semantically coher-

ent.

• Action. Instances of this class represent a primi-

tive action or sub-goal that has to be achieved for

deﬁned plans execution to be successful, and they

are referenced by the object-property hasAction.

• Plan. Instances of this class represent a set of in-

tentions which link together triggering conditions

(beliefs) to instance of Action aimed at achiev-

ing goals, and they are referenced by the object-

property hasPlan.

• Goal. Instances of this class represents a desired

state or outcome that agent aims to achieve. Goals

drive agent’s decision-making process and inﬂu-

ence its intentions, and they are referenced by the

object-property hasGoal.

• Triple. Instances of this class represent triples

from KGs, whose properties, subjects and objects

are referenced by the data-properties hasProperty,

hasSubject, hasObject, respectively. The values

of these data-properties are URIs pointing to Web

resources, whose textual contents are to be ex-

tracted and utilized for the formulation of beliefs

within the system.

• Function. Instances of this class represent func-

tions computing instances of DerivedBelief (cf.

The Non-Asserted(Y) predicate is also subsumed by

Retracted(Y), which indicates a belief Y that has been

removed from the KB as a result of an inference, thus no

longer present.

below), having (or not) in input beliefs translated

from triples of the origin KG, and they are refer-

enced by the object-property hasFunction.

• Enricher. This is a subclass of Action, being part

of a speciﬁc plan aimed to enriching agents KB

with beliefs computed by functions referenced by

the object-property hasFunction. Among other

Action’s instances related to Goal, before any in-

ference by instances of Agent, an instance of the

class Enricher represents the action of populat-

ing the KB with additional beliefs derived from

triples not present in the origin KGs.

• DerivedBelief. This is a subclass of Belief, whose

instances represent a new belief not directly re-

lated to any triple present in the origin KG, but a

computation of triples by leveraging on instances

of Function, referenced by the object-property

hasFunction of Enricher. The latter carry out the

task of asserting derived beliefs in the agent’s KB,

which is referenced by the object-property assert-

Belief.

• Event. Each instance of the class Event aims to

model whatever kind of action aimed to change

agent’s KB content, and it is referenced by the

object-property hasEvent.

• Queue. Instances of this class represent a queue

to coordinate parallel agent’s interactions with

shared semantic resources, in multi-agent setting,

in order to achieve deterministic outcomes. It is

referenced by the object-property isQueuedOn of

instances of the class Event.

• Query. Instances of this class are SPARQL

queries that will feed agent’s KB, after triples-to-

beliefs translation, and they are referenced by the

object-property hasQuery.

• AgentType. Instances of this class represent enti-

ties whose behavior is simulated within the frame-

Bridging BDI Multi-Agent Systems and the Semantic Web Through the Triples-to-Beliefs-to-Triples Paradigm

329

work where inference takes place, and they are

referenced by the object-property hasAgentType.

• RESTful. Instances of this class represent remote

RESTful services, and they are referenced by the

object-property hasRESTful. Such class play an

important role, as it provides shared-decentralize

interfaces computing functions aimed to produce

instances of DerivedBelief.

• Code. Instances of this class represent code frag-

ment to compute locally new DerivedBelief, and

they are referenced by the object-property has-

Code.

3 THE SEMAS FRAMEWORK

The SEMAS framework presented in this paper imple-

menting the T2B2T paradigm, whose acronym stands

for SEmantic Multi-Agent System, is built on top of

the BDI architecture Phidias (D’Urso et al., 2019),

which enables programs with the ability to perform

logic-based reasoning (in Prolog-style). By follow-

ing the T2B2T paradigm in Figure 1, triples retrieved

from a KG via a SPARQL query are being turned into

beliefs within the SEMAS KB, by interacting with its

inference system made of production rules as follows:

[BELIEF] / [CONDS] >> [RULE PLAN],

where the [BELIEF] placeholder refers to runtime

events asserting a belief in the agent KB, which trig-

gers the [RULE PLAN] execution in case the content

of [CONDS] is satisﬁed, i.e., the presence of other be-

liefs in the KB; [RULE PLAN] contains a list of ac-

tions which implicitly implement the designed goal,

where each action can either invoke functions in high

level language or assert/retract beliefs. At any time,

the whole content of the KB can be newly translated

(with speciﬁc built-in functions) in triples, either to

update the origin KG or to build locally novel derived

KGs. The SEMAS framework enables efﬁcient pop-

ulation of its KB according to the speciﬁc use case,

also by feeding agents through sequential contextual-

interrelated

SPARQL queries (cf. Section 5), by

preventing overpopulation with triples that are irrel-

evant to the inference process. As a result, it becomes

possible to populate KBs directly derived from vir-

tual goal-focused KGs of arbitrary size, by overcom-

ing the current limitations of SPARQL in handling

views

, which is a feature typically associated with re-

Where results from one or more query are used as pa-

rameters to subsequent query.

A view in a relational database is a virtual table deﬁned

by a saved SQL query: it doesn’t store data but dynamically

shows results from one or more tables each time it’s queried.

lational databases. The mapping between RDF prop-

erties and beliefs is internally declared into the SE-

MAS’s conﬁguration, whereas each desire is mapped

into the the so-deﬁned procedure, which can be used

to trigger manually part of the production rules stack,

taking in account (or not) of one or more arguments.

Furthermore, purely reactive events linkable to inten-

tions can be mapped into the so-called reactors, i.e.

beliefs that after assertion do not remain resident in

the KB, but interacting with the inference system as

well as beliefs.

4 RELATED WORKS

A few attempts have been made to integrate MAS

with ontologies. The authors of OASIS (Cantone

et al., 2019; Cantone et al., 2022) proposed an OWL-

based agent model language, endowed with ﬁne-

grained descriptions of agents’ behaviors. However,

their approach requires the executive grounding to be

delegated to other frameworks. In contrast, in this pa-

per we address both the semantic agents modeling and

the reactive reasoning on the Semantic Web. The au-

thors of SW-CASPAR (Longo et al., 2022) propose

a BDI framework based on Natural Language Pro-

cessing (NLP), capable of meta-reasoning in the Se-

mantic Web. Although such a framework can be con-

sidered a move towards interoperability among NLP-

based BDI agents, it does not provide templates to

assist engineers in implementing multi-agent coordi-

nation protocols and is not compliant with the well-

known FIPA

agents interoperability guidelines. The

authors of AJAN (Antakli et al., 2023) propose a mod-

ular framework for building Semantic Web-enabled

intelligent agents, using Semantic Web standards and

Behavior Tree technology. It supports SPARQL-

extended behavior trees for agent scripting, multi-

agent coordination, and is extensible with additional

modules and communication layers. The advantage

of using behavior trees over production rule systems

(which is the core of the SEMAS inference) has not yet

been documented. On one hand, production rule sys-

tems have historically been a foundational approach

to artiﬁcial intelligence, such as in the General Prob-

lem Solver (GPS) (Newell and Simon, 1961); on the

other hand, behavior trees are primarily designed for

robotic applications (

Ogren and Sprague, 2022), but

in other domains, they may introduce signiﬁcant over-

head as their complexity increases. A special case of

MAS is social simulation, where agents represent au-

tonomous virtual entities, rather than physical ones as

http://www.ﬁpa.org/

WEBIST 2025 - 21st International Conference on Web Information Systems and Technologies

330

Figure 2: The ontology modeling the T2B2T paradigm.

in FIPA, capable of processing information and inter-

acting with their surrounding environment, by repli-

cating collective behavior and changes in the system

due to the decision of individual agents. The appli-

cation of T2B2T in Section 5 aligns with this view-

point. Farrenkopf and colleagues used ontologies to

model agents’ cognitive architecture through BDI and

their communication in a model of business decisions

(Farrenkopf et al., 2016). They classify between an

Individual Domain Layer (IDL) of knowledge of lo-

cal agents, a Speciﬁc Domain Layer (SDL) on the

market sector, and an Abstract Domain Layer (ADL)

of knowledge shared by all agents. The authors focus

on updating of the last two layers through direct com-

munication between agents. Our contribution with the

implementation of T2B2T goes beyond these studies,

using BDI to formalize the cognitive architecture of

agents but also modeling the processes of elaboration

of inferences from semantic ontologies and associated

effects through update of shared knowledge domain.

Fostering the interoperability and integration between

ontologies and MAS, we show how to interrogate a

remote graph empirically grounded through a remote

SPARQL. In the case-study in Section 5, we apply

T2B2T and SEMAS to a scenario of academic career

decision making, empirically grounded in the Italian

landscape.

5 CASE-STUDY

As case-study, we ﬁrst initialized a KG about Italian

scholars in 2024, extracting raw data from the SCO-

PUS database through its Search API, including au-

thors’ name, afﬁliations, and topic of articles. Data

were elaborated to identify co-authorships through

shared publications and top-authors. Co-authorship is

counted as one, regardless of the number of shared

publications. A scholar is considered a top-author

if published at least 10 papers and have a h-index

equal or higher than 10. Data collected were trans-

lated into RDF triples and stored in a GRAPHDB

triple store to communicate with SEMAS via URIs,

resulting in a KG containing up to 100.000 triples.

The beliefs obtained from co-authorships and top-

authorships can be ascribed to instances of the class

DerivedBelief in Figure 2. In Figure 3, we show the

SEMAS production rules applied to the case, where

rules are grouped in three distinct stages: Acquiring

Triples Stage (ATS), Inference Stage (IS), and Updat-

The repository for this extract is available at

https://github.com/RoccoPaolillo/Semas/tree/webist

https://graphdb.ontotext.com/

Bridging BDI Multi-Agent Systems and the Semantic Web Through the Triples-to-Beliefs-to-Triples Paradigm

331



1 # Acq u i r in g Tri p le s Stage ( A T S )

2 Be T o p A u t h o r s h i p ( X ) >> [ sh ow _ li ne (' Pl a nn i ng to be top - a u th o r i n ' , X ,' ...' ) ,

3 loa d _o bj (' acad : is To pA ut ho rI n' , X) , F i n d R e la te d () , P ub li ca ti on sh ip ( X ) ]

5 Fin dR el at e d () / C o n s i d e r To pA ut ho r ( X , Y ) > > [ - C on si de r T o p A u t h o r ( X , Y ) , + T op A u t h o r s h i p (X , Y ) ,

6 sho w _ l in e (' F i nd ing t r ip l es re la t ed with ' , X ,' ...' ), l oa d _ s ub j (' aca d : h a s A f f il ia ti o n W i t h ' , X) ,

7 loa d _ s ub j (' ac a d : c o A u t ho rW it h' , X) , l o ad _o b j (' ac a d : co A u t h o r W it h' , X) , F in dR e l a t e d () ]

8 Fin dR el at e d () > > [ s h o w _l in e (' Re lat ed tr i pl es re tr i ev ed .') ]

10 # Inf e r e nc e Stag e (IS )

11 Pu b l i c a t i o n s h i p ( X ) / ( C oA ut ho r s h i p (Z , Y ) & To pA ut ho rs hi p ( Y ,X ) & A f f i li at io n ( Z , U ) & S e l e c t i o n s h i p ( S , U )

12 & Aff il ia t i o n ( S , T ) ) >> [- Co Au th or sh ip (Z , Y) , + P ro p o s e C o a u t ho rs hi p ( Z ,U , Y , X ,S , T ) , Pub l i c a t i on sh ip ( X ) ]

14 # Upd a ti ng Tr i pl es Sta g e ( U T S )

15 + P r o po se Co a u t h o r s hi p ( Z ,U ,Y , X , S , T ) > > [ s h ow _l i n e (Z , ' at O r g a ni za ti on ' , U , ' is co - au t ho r wi t h ' , Y ,

16 ' top - a ut h or in the topi c ' , X ) , - Af fi li a t i o n (S , T ) ,- Se le ct io ns hi p ( S ,U ) , + Af fi l i a t i on (S , U ) ,

17 De l e t e A l te rn at iv e ( S ) ]

18 De l e t e A l te rn at iv e ( S ) / ( S e l e c t i o n s h i p ( S , P ) ) >> [- S e l e c t io ns hi p ( S , P ) , D el e t e A l t e r n a t iv e ( S )]



 

Figure 3: The SEMAS production rules applied to SCOPUS database.



1 # ass e r t io n bel i ef s for the hy po th e t i c a l scena r io , sele ct i o n of i d e nt if ie d s ch o la r S by two

or ga ni za ti o n s S and P , he r e r e po rt e d with the cor r e s p o n d i n g URIs

2 eShe l l : main >

3 + S e l e c t i o ns hi p (' BASE - UR I / a u th o rs /5 72 01 11 7 4 0 1 ' ,' BASE - URI / or ga ni za ti on s / 6 0 0 00 48 1 ' )

4 + S e l e c t i o ns hi p (' BASE - UR I / a u th o rs /5 72 01 11 7 4 0 1 ' ,' BASE - URI / or ga ni za ti on s / 1 0 5 93 72 50 ' )

6 # kno w l e dg e grap h s h ow i ng the aff il ia ti o n of s c ho l ar S and sel e c t i o n s h i p by r e se ar c h or ga ni za t i o n s

7 eShe l l : main > kb

8 Aff il ia ti o n (' BASE - U R I / a uth or s /5 72 01 11 74 01 ' , ' BASE - U R I / o rg an iz at io ns /6 0 0 2 46 90 ' )

9 Se l ec ti on sh ip (' BASE - U R I / a u th ors /5 72 01 11 7 4 0 1 ' , ' BASE - U R I / o rg an iz at io ns /6 0 00 04 8 1 ' )

10 Se l ec ti on sh ip (' BASE - U R I / a u th ors /5 72 01 11 7 4 0 1 ' , ' BASE - U R I / o rg an iz at io ns /1 0 5 9 37 25 0' )

12 # T he desi r e Be T o p A ut ho rs hi p () in a s pe c if ic top i c i s invoked , which en a ct s ACS , IS an d UTS stag e s .

13 eShe l l : main > B e T o p A u t h o r s h i p (' BASE - URI / to p ic s /20 0 3' )

15 # Resu lt s of IS s ta g e rep o r t in g co - au th o rs to top - au t ho r s fou n d in the two or ga ni za ti on s .

16 # In this case only one o rg an iz at io n mat c he s the c on di t i o ns re q u ired , so to be c h ose n by sch ol a r S

17 BASE - URI / au t ho r s / 16 02 11 98 60 0 at Org a n i z a t i o n BASE - URI / o rg a n i z a t i o n s / 60 00 04 8 1 is co - au t ho r

18 with B ASE - URI / au t ho rs / 5 5 90 43 39 60 0 top - aut h or in the topi c BASE - URI / t op i cs / 200 3

20 BASE - URI / or ga ni za ti on s / 6 0 0 00 48 1 is co - aut hor with BASE - U R I / a u th or s /4 80 61 34 05 0 0 t op - au tho r in t h e to p ic

21 BASE - URI / top ics / 200 3

23 # kno w l e dg e grap h s h ow i ng the new a ff il i a t i o n of sc ho l ar S afte r UTS stag e

24 eShe l l : main > kb

25 Aff il ia ti o n (' BASE - U R I / a uth or s /5 72 01 11 74 01 ' , ' BASE - U R I / o rg an iz at io ns /6 0 0 0 04 81 ' ) , ...



 

Figure 4: Initialized case study with SCOPUS data and running of SEMAS.

ing Triples Stage (UTS). Figure 4 shows the initializa-

tion and running of SEMAS with SCOPUS data. To en-

sure consistency and determinism each group can be

run accordingly to the stage; afterward, the SEMAS’

built-in event queue (inherited form Phidias) will do

implicitly the rest of the work, taking into account of

stage’s priorities of other agents through messages ex-

change.

SELECT ? su b j ? p r op ? o b j WHERE {

<[ S UBJ ] > [ PR O P ] ? obj .

BIND([ P R O P ] AS ? p rop )

BIND( <[ SUB J ]> AS ? subj )

}

Listing 1: The parameterized SPARQL query executed

by the action load obj([PROP],[SUBJ]), with property

[PROP] and subject [SUBJ] as placeholders.

WEBIST 2025 - 21st International Conference on Web Information Systems and Technologies

332

The production rules leverage on the objects pro-

cedures and reactors introduced in Section 3, to im-

plement the cognitive qualities of agents’ mental at-

titudes. We build on the recognized role of co-

authoring with top-authors and networking as a driver

to academic career (Letina, 2016).

In line 2 of Figure 3, the procedure

BeTopAuthorship(X) ﬁrst starts the pipeline

triples-to-beliefs (T2B) of T2B2T with the ac-

tion load obj(’acad:isTopAuthorIn’, X),

which executes a parameterized SPARQL query

(cf. Listing 1)

to ﬁlter all triples’ subjects hav-

ing acad:isTopAuthorIn as property and topic

X as object, that are used to assert the symbolic

belief ConsiderTopAuthor, which is a special

belief used to avoid an inﬁnite loop in line 5.

In this line, the procedure FindRelated() exe-

cutes more SPARQL queries through the action

load subj(’acad:hasAffiliationWith’, X),

which ﬁlters all triples having hasAffiliationWith

as property and author X as subject, in order to assert

beliefs with Affiliation as label; similarly, the

actions load subj(’acad:coAuthorWith’, X)

and load obj(’acad:coAuthorWith’, X) aim to

extract triples with property coAuthorWith, where

author X is either subject or object of the triple,

both to assert beliefs with CoAuthorship as label.

Likewise to load obj, load subj(X, Y) executes

parameterized SPARQL queries (cf. Listing 2) to

extract all triples having X as property and Y as

subject.

SELECT ? su b j ? prop ? obj WHERE {

? su b j [ P R OP ] <[ OBJ ] > .

BIND([ P R O P ] AS ? p rop )

BIND( <[ OBJ ] > AS ? obj )

}

Listing 2: The parameterized SPARQL query executed

by the action load subj([PROP], [OBJ]), with property

[PROP] and object [OBJ] as placeholders.

The pipeline thus composes the KG used for the

inferences upon beliefs assertions. Line 11 shows the

actual reactive plan in the inference stage (IS) ac-

tivated by the procedure Publicationship() with

top-authors, to which a set of condition beliefs is

applied. In the scenario, a scholar S who is al-

ready afﬁliated with the university T, is offered a

position by universities U and P with neither of

them hosting a top-author in the elective ﬁeld. Ap-

plying a strategy based on small world networks

(Koseoglu, 2016), the inference identiﬁes the orga-

The PREFIX declarations in the query have been inten-

tionally omitted.

nization U with co-authors Z to a top-author Y they

could connect S with in case the offer is accepted.

The proposed intention is implemented by the reac-

tor ProposeCoauthorship() (line 15), which im-

plies accepting the offer by the organization match-

ing the conditions. This translates into the retrac-

tion of the belief of scholar S selected by the chosen

organization U (-Selectionship(S,U)), and retrac-

tion of the previous afﬁliation of S with the organiza-

tion T (-Affiliation(S,T)), updating the assertion

belief to the new afﬁliation (+Affiliation(S,U)).

Also the information of S being selected by the al-

ternative organization P is eliminated with the re-

traction of -Selectionship(S,P) activated by the

procedure DeleteAlternative(). In Figure 4,

authors/57201117401

is selected from SCOPUS

database as scholar S, with lines 8 to 10 show-

ing their real afﬁliation to an Italian university and

the two alternative institutions (equivalent to U,P

in Figure 3) we selected. Neither of them hosts

a top-author in the elective ﬁeld topics/2003 (ﬁ-

nance), but organizations/60000481 hosts coau-

thors. Line 17 shows the inference produced by SE-

MAS based on scanning the SCOPUS database, while

line 25 shows the new hypothetical afﬁliation of the

scholar based out of the inference.

6 CONCLUSIONS

In this paper, we addressed the challenge of enabling

dynamic and semantically consistent interoperabil-

ity between Multi-Agent Systems (MAS) and het-

erogeneous Semantic Web environments. We pro-

posed the T2B2T paradigm as a novel integration

model, aimed at overcoming the limitations of tradi-

tional MAS frameworks in processing and interacting

with semantic data. By ontologically formalizing the

T2B2T cycle and implementing it within the SEMAS

framework, we demonstrated how agents can reason

upon RDF KGs, update their internal beliefs accord-

ingly, and propagate new inferences back into the Se-

mantic Web ecosystem. In our work we showed the

soundness of SEMAS by feeding its KB with empir-

ical data from SCOPUS database, by adding new be-

liefs computed from existing triples, which take part

in the inference process to produce newly introduced

triples. The proposed paradigm goes beyond conven-

tional SPARQL-based querying by allowing agents

to perform richer symbolic inference, while ensur-

ing consistency and determinism across distributed

We report URIs and not actual name to guarantee

anonymity

Bridging BDI Multi-Agent Systems and the Semantic Web Through the Triples-to-Beliefs-to-Triples Paradigm

333

agents. This approach also provides a practical mech-

anism for moving between open- and closed-world

assumptions depending on the context of reasoning,

thereby enhancing agents’ cognitive capabilities in

dynamic and evolving environments.

ACKNOWLEDGEMENTS

This work was supported by FOSSR (Fostering Open

Science in Social Science Research), funded by the

European Union - NextGenerationEU under NRRP

Grant agreement n. MUR IR0000008.

REFERENCES

Antakli, A., Kazimov, A., Spieldenner, D., Rojas, G. E. J.,

Zinnikus, I., and Klusch, M. (2023). Ajan: An engi-

neering framework for semantic web-enabled agents

and multi-agent systems. In Mathieu, P., Dignum,

F., Novais, P., and De la Prieta, F., editors, Advances

in Practical Applications of Agents, Multi-Agent Sys-

tems, and Cognitive Mimetics. The PAAMS Collec-

tion, pages 15–27, Cham. Springer Nature Switzer-

land.

Berners-Lee, T., Hendler, J., and Lassila, O. (2001). Web

semantic. Scientiﬁc American, 284(5):34–43.

Cantone, D., Longo, C. F., Nicolosi-Asmundo, M., San-

tamaria, D., and Santoro, C. (2019). Towards an

Ontology-Based Framework for a Behavior-Oriented

Integration of the IoT. In Proceedings of the 20th

Workshop From Objects to Agents, 26-28 June, 2019,

Parma, Italy, CEUR Workshop Proceeding Vol. 2404,

pages 119–126.

Cantone, D., Longo, C. F., Nicolosi Asmundo, M., San-

tamaria, D. F., and Santoro, C. (2022). Ontological

smart contracts in oasis: Ontology for agents, systems,

and integration of services. In Camacho, D., Rosaci,

D., Sarn

e, G. M. L., and Versaci, M., editors, In-

telligent Distributed Computing XIV, pages 237–247,

Cham. Springer International Publishing.

D’Urso, F., Longo, C. F., and Santoro, C. (2019). Pro-

gramming intelligent iot systems with a python-based

declarative tool. In The Workshops of the 18th Inter-

national Conference of the Italian Association for Ar-

tiﬁcial Intelligence.

Farrenkopf, T., Guckert, M., Urquhart, N., and Wells, S.

(2016). Ontology based business simulations. Journal

of Artiﬁcial Societies and Social Simulation, 19(4).

Glimm, B., Horrocks, I., Motik, B., Stoilos, G., and Wang,

Z. (2014). HermiT: An OWL 2 Reasoner. Journal of

Automated Reasoning, 53(3):245–269.

Koseoglu, M. A. (2016). Growth and structure of au-

thorship and co-authorship network in the strategic

management realm: Evidence from the strategic man-

agement journal. BRQ Business Research Quarterly,

19(3):153–170.

Letina, S. (2016). Network and actor attribute effects on

the performance of researchers in two ﬁelds of social

science in a small peripheral community. Journal of

Informetrics, 10(2):571–595.

Longo, C. F., Santoro, C., Nicolosi-Asmundo, M., Cantone,

D., and Santamaria, D. F. (2022). Towards ontolog-

ical interoperability of cognitive iot agents based on

natural language processing. Intelligenza Artiﬁciale,

16(1):93–112.

Newell, A. and Simon, H. A. (1961). Gps, a program

that simulates human thought. In Billing, H., editor,

Lernen und automatische Informationsverarbeitung,

pages 109–124. Springer.

Rao, A. and Georgeff, M. (1995). BDI agents: From theory

to practice. In Proceedings of the ﬁrst international

conference on multi-agent systems (ICMAS-95), pages

312–319. San Francisco, CA.

Sirin, E., Parsia, B., Grau, B. C., Kalyanpur, A., and Katz,

Y. (2007). Pellet: A practical OWL-DL reasoner. Web

Semantics, 5(2):51–53.

World Wide Web Consortium (2004). SWRL: A Semantic

Web Rule Language Combining OWL and RuleML.

Ogren, P. and Sprague, C. I. (2022). Behavior trees in robot

control systems. Annual Review of Control, Robotics,

and Autonomous Systems, 5(Volume 5, 2022):81–107.

WEBIST 2025 - 21st International Conference on Web Information Systems and Technologies

334