ODKAR: “Ontology-Based Dynamic Knowledge Acquisition and

Automated Reasoning Using NLP, OWL, and SWRL”

Claire Ponciano

, Markus Schaffert

and Jean-Jacques Ponciano

i3mainz, University of Applied Sciences, Germany

{claire.ponciano, markus.schaffert, jean-jacques.ponciano}@hs-mainz.de

Keywords:

Ontology Generation, Natural Language Processing (NLP), OWL (Web Ontology Language),

SWRL (Semantic Web Rule Language), Text-to-Ontology, Knowledge Extraction,

ChatGPT Comparison, Knowledge Representation.

Abstract:

This paper introduces a novel approach to dynamic ontology creation, leveraging Natural Language Process-

ing (NLP) to automatically generate ontologies from textual descriptions and transform them into OWL (Web

Ontology Language) and SWRL (Semantic Web Rule Language) formats. Unlike traditional manual ontology

engineering, our system automates the extraction of structured knowledge from text, facilitating the develop-

ment of complex ontological models in domains such as ﬁtness and nutrition. The system supports automated

reasoning, ensuring logical consistency and the inference of new facts based on rules. We evaluate the per-

formance of our approach by comparing the ontologies generated from text with those created by a Semantic

Web technologies expert and by ChatGPT. In a case study focused on personalized ﬁtness planning, the system

effectively models intricate relationships between exercise routines, nutritional requirements, and progression

principles such as overload and time under tension. Results demonstrate that the proposed approach generates

competitive, logically sound ontologies that capture complex constraints.

1 INTRODUCTION

The integration of Natural Language Processing

(NLP)(Shamshiri et al., 2024; Chen et al., 2024; Yin

et al., 2024; Osman et al., 2024) with Semantic Web

technologies (Matthews, 2005) offers signiﬁcant po-

tential for creating intelligent systems that automat-

ically convert unstructured text into formal knowl-

edge representations. Ontologies (Fensel and Fensel,

2001), as a cornerstone of the Semantic Web, provide

a structured, machine-readable format for represent-

ing domain knowledge, while automated reasoning

(Wang et al., 2004) over these ontologies enables sys-

tems to infer new information, ensure logical consis-

tency, and support sophisticated decision-making pro-

cesses. This combination opens up new possibilities

for dynamically acquiring, managing, and reasoning

over knowledge extracted from natural language in-

puts.

Traditionally, the construction and management of

ontology-based systems have required expert knowl-

https://orcid.org/0000-0001-8883-8454

https://orcid.org/0000-0002-7970-9164

https://orcid.org/0000-0001-8950-5723

edge of formal description logic languages, such

as OWL (Web Ontology Language)(Antoniou and

Harmelen, 2009), RDF(Pan, 2009) and SWRL (Se-

mantic Web Rule Language)(Horrocks et al., 2004).

Creating ontological deﬁnitions, formulating logical

rules, and implementing automated reasoning mech-

anisms are technically demanding tasks that necessi-

tate a deep understanding of Semantic Web technolo-

gies. This complexity poses a barrier for domain ex-

perts who may excel in their ﬁelds but lack the spe-

cialized skills required to formalize knowledge in on-

tological formats. Consequently, there is a growing

need for tools that facilitate the automatic generation

of ontologies such as (Ponciano et al., 2022; Prud-

homme et al., 2020; Prudhomme et al., 2017), lower-

ing the technical barriers for non-experts in Semantic

Web technologies.

This paper addresses this challenge by proposing

a system for automatic ontology creation and reason-

ing. Instead of relying on manual input or conversa-

tional agents, the system uses NLP techniques to ex-

tract key concepts and relationships from unstructured

text, translating them into formal OWL and SWRL

representations. Automated reasoning is then applied

to maintain logical consistency and infer new knowl-

Ponciano, C., Schaffert, M. and Ponciano, J.

ODKAR: “Ontology-Based Dynamic Knowledge Acquisition and Automated Reasoning Using NLP, OWL, and SWRL”.

DOI: 10.5220/0013071500003825

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

In Proceedings of the 20th International Conference on Web Information Systems and Technologies (WEBIST 2024), pages 457-465

ISBN: 978-989-758-718-4; ISSN: 2184-3252

457

edge. The system’s ability to convert text into struc-

tured knowledge is evaluated through a comparative

analysis, contrasting the results of our approach with

those generated by ChatGPT and a Semantic Web

technologies expert.

To demonstrate the system’s capabilities, we ap-

ply it to the domain of personalized ﬁtness planning,

modeling intricate relationships between exercises,

nutrition, and training principles such as progressive

overload and time under tension. The resulting on-

tologies are used to generate tailored ﬁtness plans that

accommodate individual needs and goals. Our system

contributes to the ﬁeld of artiﬁcial intelligence and the

Semantic Web by providing a framework for the auto-

mated creation of ontologies, enabling real-time rea-

soning and supporting the development of expert sys-

tems from unstructured text, even by non-experts in

Semantic Web technologies.

2 RELATED WORK

The reviews and surveys about ontology learning

techniques (Wong et al., 2012; Asim et al., 2018;

Al-Aswadi et al., 2020) agreed on three main tech-

niques for ontology learning from text: linguistics-

based, statistics-based, and logic-based techniques.

Statistics-based and logic-based techniques are clas-

siﬁed as machine learning approaches in the review

of (Al-Aswadi et al., 2020).

Linguistics-based techniques are mainly based on

natural language processing (NLP) tools (Wong et al.,

2012). NLP-based ontology generation (Osman et al.,

2024; Yin et al., 2024; Chen et al., 2024; Shamshiri

et al., 2024; Sui et al., 2010; Zhang et al., 2023) fo-

cuses on the automatic extraction of entities, relation-

ships, and rules directly from unstructured text. +

Statistics-based techniques are mostly derived

from information retrieval, machine learning, and

data mining (Wong et al., 2012). They are mainly

used for term extraction, concept extraction and tax-

onomic relationship extraction and most make exten-

sive use of probabilities (Asim et al., 2018). They

include C/NC, contrastive analysis, clustering, co-

occurrence analysis, term subsumption and ARM

(Asim et al., 2018).

Logic-based techniques are presented as the least

common and based on knowledge representation and

rule-based reasoning by (Wong et al., 2012), whereas

in (Asim et al., 2018), the authors present Induc-

tive Logic Programming as a discipline of machine

learning that derives hypothesis based on background

knowledge and a set of examples using logic pro-

gramming, which is used to acquire general axioms

from schematic axioms. For example, the approach

(Shamsfard and Barforoush, 2004) is based on a Ker-

nel ontology and a rule-based system for handling

the linguistic, semantic, syntactic, morphological and

grammatical aspects. The texts processed by this ap-

proach enrich the Kernel’s knowledge.

In addition to these three techniques, (Wong et al.,

2012) presents also, hybrid approaches that combine

several of the previous presented techniques. These

hybrid approaches can be explained by the different

roles of the different techniques in the methodology

for ontology learning as presented by (Asim et al.,

2018). Linguistic-based techniques seem to be a must

for the pre-processing in ontology learning method.

Steps of concepts/terms and relations extraction are

generally done through statistics, linguistic- based

techniques or a combination of both. Finally, axiom

step is done through inductive logical programming

(Asim et al., 2018). One example of hybrid approach

is Text2Onto (Cimiano and V

olker, 2005), which in-

tegrates machine learning with basic linguistic pro-

cessing (such as tokenization and shallow parsing) to

model ontologies probabilistically. By adding proba-

bilistic reasoning, Text2Onto allows for scalable and

ﬂexible ontology creation, making it a hybrid method

between NLP, machine learning, and statistical anal-

ysis. In addition, methods like OpenIE (Etzioni et al.,

2011) employ NLP to extract relationships and enti-

ties, which are then processed using learning-based

systems to identify the arguments. These systems en-

hance the generation of knowledge graphs, but they

often lack built-in support for rule-based reasoning,

which limits their utility in domains requiring ad-

vanced logical consistency.

The review (Al-Aswadi et al., 2020) also high-

lights the beneﬁt of deep-learning in comparison to

machine learning (referred as a shallow learning).

Deep learning models, including BERT (Kenton

and Toutanova, 2019) and GPT (Brown et al., 2020),

have been applied to ontology generation and knowl-

edge extraction due to their ability to learn complex

patterns from large datasets. These models excel at

representing semantic context within text, perform-

ing well across diverse domains for tasks such as en-

tity extraction and relationship identiﬁcation. Recent

transformer-based models (Mihindukulasooriya et al.,

2023; Yenduri et al., 2024) have shown improvements

in domain-speciﬁc text understanding, but they of-

ten struggle to convert unstructured text into formal

ontological models that adhere to strict logical con-

straints. Large Language Models (LLMs), such as

GPT-4 (Baktash and Dawodi, 2023), can produce se-

mantically rich outputs but still face limitations in for-

mal reasoning and ontology-based rule enforcement

WEBIST 2024 - 20th International Conference on Web Information Systems and Technologies

458

(Mukanova et al., 2024).

Our approach presented in this paper is a hybrid

one, combining linguistic-based and logic-based tech-

niques. It seeks to bridge these gaps by offering a

fully automated system that uses semantic and rule-

based reasoning to ensure logical consistency in on-

tologies generated from unstructured text. Unlike ex-

isting approaches, which rely heavily on deep learn-

ing and machine learning models, our system inte-

grates real-time reasoning based on semantic rules,

producing formal OWL and SWRL ontologies with

minimal human intervention.

3 METHODOLOGY

This section details the methodology used to de-

velop the ontology-based approach proposed, includ-

ing the system architecture, natural language process-

ing techniques, ontology management, and reasoning

mechanisms. To illustrate the methodology, we take

the following text that describes concept in human

language as input:

The physical training component is detailed

through various exercises and sessions. Exercises like

the Bench Press and Squats are included in this struc-

ture, each targeting speciﬁc muscle groups. For ex-

ample, the Bench Press targets the chest, and Squats

target the legs. These exercises are categorized based

on whether they are compound exercises, such as

Bench Press and Squats, which engage multiple mus-

cle groups, or isolation exercises like Bicep Curls,

which target speciﬁc muscles.

3.1 System Architecture

The system architecture is designed with modularity

and scalability in mind, ensuring that the various com-

ponents responsible for natural language processing,

ontology management, and reasoning can be indepen-

dently developed, maintained, and extended. The ar-

chitecture consists of three primary components: the

Natural Language Processing (NLP) module, the On-

tology Management module, and the SWRL-based

Reasoning module. The two last modules use a rea-

soning engine for inference and consistency checking

tasks. The ontology is stored into a knowledge base.

3.2 Natural Language Processing

Module

The Natural Language Processing (NLP) module is

essential for transforming natural language input into

structured data with an ontological form. This mod-

ule allows users to interact with the system using ev-

eryday language, without requiring expertise in on-

tology creation. It encompasses several stages that

translate user inputs into formal knowledge represen-

tations, mapped to Web Ontology Language (OWL)

classes, properties, and individuals, ensuring consis-

tent and accurate updates to the knowledge base.

The NLP pipeline processes natural language in-

puts by breaking down text, identifying entities and

relationships, and converting these into structured

data. The steps involved are described in this section.

3.2.1 Step 1: Tokenization and POS Tagging

The input sentence is ﬁrst broken into smaller units

(tokens) during tokenization. Each token is assigned

a Part-of-Speech (POS) tag, identifying its grammat-

ical role, such as noun, verb, or adjective. For exam-

ple, the sentence ”Exercises like the Bench Press and

Squats are included in this structure, each targeting

speciﬁc muscle groups” would be tokenized as:

TOKENS = ["Exercises", "like", "the",

"Bench", "Press", "and", "Squats", "are",

"included", "in", "this", "structure", ",",

"each", "targeting", "specific", "muscle",

"groups", "."]

Each token is POS-tagged as follows: "Bench"

(noun), "Press" (noun), "targeting" (verb), and

so on. This establishes the syntactic structure of the

sentence, which is further analyzed in the following

steps.

3.2.2 Step 2: Named Entity Recognition (NER)

Once tokenization and POS tagging are complete,

Named Entity Recognition (NER) identiﬁes key enti-

ties in the text. Entities generally represent objects,

people, or domain-speciﬁc terms. In this example,

"Bench Press", "Squats", and "muscle groups"

are identiﬁed as key entities. This step narrows down

the critical elements of the sentence, helping the sys-

tem focus on the most meaningful components.

The identiﬁed entities are:

ENTITIES = ["Bench Press", "Squats",

"muscle groups"]

3.2.3 Step 3: Dependency Parsing

Dependency parsing is then applied to analyze

the grammatical structure and relationships between

words. It builds a graph where tokens are connected

by grammatical dependencies. For example, in the

sentence ”Bench Press targets the chest”, the sys-

tem identiﬁes that "Bench Press" is the subject,

"targets" is the action, and "chest" is the object.

ODKAR: “Ontology-Based Dynamic Knowledge Acquisition and Automated Reasoning Using NLP, OWL, and SWRL”

459

These relationships provide deeper insights into the

meaning of the sentence.

DEP_GRAPH = [

{" token ": " B e n c h P r ess " , " dep ": " nsubj

", " head ": " target s "} ,

{" to k en ": " t a rgets " , " dep ": " RO OT "} ,

{" to k en ": " ches t " , " dep ": " dob j " , "

he ad ": " targets "} ,

]

3.2.4 Step 4: Entity and Relation Extraction

Next, the system extracts entities and their relation-

ships based on the dependency parsing results. For

instance, in the sentence ”Bench Press targets the

chest”, the system extracts that "BenchPress" is the

subject, "targets" is the relation, and "chest" is the

object. This extraction forms the basis for structuring

the sentence into meaningful data.

SU B J E C T = [" BenchPress "]

REL A T I O N = [" targ e t s "]

OB J E C T = [" c hes t "]

3.2.5 Step 5: Mapping to OWL

The ﬁnal step involves mapping the extracted enti-

ties and relationships to OWL. The system translates

subjects, objects, and relationships into correspond-

ing OWL classes, properties, and individuals. For ex-

ample, "BenchPress" is mapped as an individual of

the ex:Exercise class, "chest" as an individual of

the ex:MuscleGroup class, and the targets relation-

ship is mapped as an OWL object property. This step

results in the following mappings that is referenced as

(∆O) later in this paper.

• Classes:

– ex:Exercise

– ex:MuscleGroup

– ex:ExerciseType

• Individuals (of Exercises):

– ex:BenchPress

– ex:Squats

– ex:BicepCurls

• Individuals (of Muscle Groups):

– ex:chest

– ex:legs

• Individuals (without class):

– ex:SpecificMuscles

• Object Properties:

– ex:targets (connects exercises to muscle

groups)

• Relationships:

– ex:BenchPress ex:targets ex:chest

– ex:Squats ex:targets ex:legs

– ex:BicepCurls ex:targets

ex:SpecificMuscles

3.3 Ontology Management Module

The Ontology Management Module (OMM) is a core

component of the system, responsible for oversee-

ing the continuous management and evolution of the

OWL ontology, which acts as the formal knowledge

representation. This module dynamically integrates

new information into the ontology, ensures logical

consistency, and supports versioning and rollback

mechanisms. By managing the dynamic construction

of the ontology, enforcing logical constraints, and per-

forming consistency checks, the OMM ensures the

system remains robust and adaptable to new data,

maintaining the overall integrity and coherence of the

knowledge base.

The OMM integrates dynamically new knowledge

(new classes, properties, individuals and relationship

resulting from the NLP module) into the ontology. As

each new text is processed by the NLP module, the

ontology evolves without requiring manual interven-

tion or redesign. This section presents the detailed

functionalities for the process of dynamic construc-

tion and the integration of new knowledge into the

existing ontology. These functionalities are illustrated

through the process of the NLP module output (∆O)

that includes new classes, individuals, and properties,

such as exercises and muscle group.

3.3.1 Functionality 1: Dynamic Ontology

Construction

New knowledge from ∆O, such as new classes (e.g.,

ex:Exercise), new properties (e.g., ex:targets),

new individuals (e.g., ex:Bench Press) and new

relationships (e.g., ex:Bench Press ex:targets

ex:chest) is added to the current ontology (O) dy-

namically.

3.3.2 Functionality 2: Consistency Checking

Each time that a new element (such as a new ele-

ment from ∆O, a relationship or a constraint) is inte-

grated, the OMM performs consistency checks. This

is achieved by employing a description logic reasoner

(e.g., Pellet or HermiT):

• If the consistency check passes, the newly added

elements are fully integrated into the ontology.

WEBIST 2024 - 20th International Conference on Web Information Systems and Technologies

460

• If an inconsistency is detected, the new element is

rejected, ensuring the ontology remains logically

sound.

3.3.3 Functionality 3: Versioning and Rollback

To maintain the integrity of the knowledge base, the

OMM supports version control. After consistency

checks, the updated ontology is saved as a new ver-

sion. This feature ensures that the ontology’s history

is preserved and can be rolled back to a previous, sta-

ble version if inconsistencies arise in the future.

• Version Control: A new version of the ontology

is created, incorporating the latest addition.

• Rollback: In case of error or inconsistency, the

system can revert to the previous version, preserv-

ing the integrity of the knowledge base.

3.3.4 Functionality 4: Enforcing Logical

Constraints

Once the new consistent elements from ∆O are inte-

grated to the knowledge base, the OMM enforces log-

ical constraints, through class hierarchies and prop-

erty restrictions. This reinforces the logical structure

of the ontology.

1. Class Hierarchy:

• The class hierarchy construction uses the NER

result from the NLP module (c.f. 3.2.2). Each

entity (E) composed of several words and clas-

siﬁed as a class is analyzed. If a class (C) exists

whose name is the last word of entity E, then

the class resulting from entity E is deﬁned as a

subclass of C.

• For example, the two classes

ex:CompoundExercise and

ex:IsolationExercise (resulting from

the two entities ”Compound Exercise” and

”Isolation Exercise” respectively) are deﬁned

as subclass of ex:Exercise.

2. Property Restrictions:

• The property restriction construction uses the

dependency parsing result from the NLP mod-

ule (c.f. 3.2.3) to identify classes that ﬁt the

domain and range of a property.

• For example, this process deﬁnes

ex:Exercise as the domain of the prop-

erty ex:targets and ex:Muscle Group as its

range.

3.3.5 Functionality 5: Inference

Once the consistent constraints have been added

to the knowledge base, an inference process is

triggered by using Pellet reasoner engine. It al-

lows for adding new knowledge. For exam-

ple, the individual ex:specificMuscle becomes

an instance of the class ex:MuscleGroup, due

to ex:MuscleGroup is deﬁned as the range of

the property ex:targets and the knowledge base

contains the axiom ex:BicepCurls ex:targets

ex:SpecificMuscles.

3.4 SWRL-Based Reasoning Module

The SWRL-Based Reasoning Module plays a criti-

cal role in enhancing the system’s ability to infer new

knowledge, going beyond the standard reasoning ca-

pabilities provided by OWL alone. This module de-

ﬁnes and applies logical rules, speciﬁed using the Se-

mantic Web Rule Language (SWRL). It allows the

system to autonomously generate new facts by evalu-

ating the SWRL rules in conjunction with the ontol-

ogy’s structural aspects. By applying these rules, the

system can infer additional knowledge that may not

be explicitly stated in the ontology.

The SWRL-Based Reasoning Module follows a

structured process to apply rules and infer new knowl-

edge from the ontology. The key steps in the execu-

tion are outlined in this section.

3.4.1 Step 1: Rule Deﬁnition

This module deﬁnes SWRL rule, through a pattern

search based on structures and keywords such as:

• if {A} then {B} that produces

A → B

• whether {A}, which/that {B} or {C},

which/that {D} that produces

B → A

D → C

In the text provided in section 3, the analysis of the

sentence:”These exercises are categorized based on

whether they are compound exercises, such as Bench

Press and Squats, which engage multiple muscle

groups, or isolation exercises like Bicep Curls, which

target speciﬁc muscles.” produces the deﬁnition of

the following SWRL rules:

• Rule R1:

ex : targets(?x, ?y) ∧ ex : MuscleGroup(?y)

→ ex : CompoundExercise(?x)

• Rule R2:

ex : targets(?x, ex : SpecificMuscles)

→ ex : IsolationExercise(?x)

ODKAR: “Ontology-Based Dynamic Knowledge Acquisition and Automated Reasoning Using NLP, OWL, and SWRL”

461

The rule R1 means that if X targets multiple mus-

cle groups, X should be classiﬁed as an instance of

compound exercise. Similarly, the rule R2 might in-

fer that if X targets a speciﬁc muscle, it should be

classiﬁed as an instance of isolation exercise.

3.4.2 Step 2: Rule Application

Once the rules are deﬁned, the module applies them to

the ontology using the Pellet reasoner engine. Based

on the rules R1 and R2, the following inferences are

made:

• ex:Bench Press a ex:CompoundExercise

(from R1)

• ex:Squats a ex:CompoundExercise (from

R1)

• ex:BicepCurls a ex:IsolationExercise

(from R2)

3.4.3 Step 3: Consistency Checking

After new facts are inferred, the module performs

consistency checks to ensure that the inferred knowl-

edge does not introduce any logical contradictions. If

inconsistencies are detected, the conﬂicting inferred

facts and the rule it comes from are discarded, pre-

serving the integrity of the ontology.

3.4.4 Real-Time Inference and Scalability

Handling

The SWRL-Based Reasoning Module supports real-

time inference, applying checked and consistent rules

immediately after new data is integrated into the on-

tology. This real-time inference allows the system to

update the ontology with inferred knowledge as new

facts are added, enabling dynamic decision-making.

For example, when a new exercise is added to the on-

tology, the system automatically infers its classiﬁca-

tion as either a compound or isolation exercise, de-

pending on the muscle group it targets.

As the ontology grows, the number of SWRL

rules may increase, requiring the system to han-

dle reasoning efﬁciently. To address scalability, the

SWRL-Based Reasoning Module employs incremen-

tal reasoning techniques, which focus on evaluating

only the parts of the ontology that are affected by re-

cent changes. This approach minimizes computation

and ensures that the system remains responsive, even

as the ontology becomes more complex.

4 EVALUATION

To assess the performance of the three ontologies gen-

erated from the same source text, a comprehensive

evaluation was conducted comparing the ontologies

produced by ChatGPT

, an expert in Semantic Web

technologies, and the proposed approach. The eval-

uation was based on several critical criteria, includ-

ing completeness, consistency, correctness, richness,

reasoning capabilities, and the level of human effort

required to create each ontology. These criteria pro-

vide a detailed view of each approach’s strengths and

weaknesses. The input text and the corresponding on-

tologies used in this evaluation are publicly accessi-

ble via the following repository: https://github.com/

JJponciano/ODKAR.

4.1 Completeness

Completeness assesses how well each ontology cap-

tures the concepts and relationships described in the

input text (cf. Physical Training Structure).

Proposed Approach: The ontology generated by

our system includes 15 classes, 3 object properties,

22 data properties, 11 individuals, and 158 axioms. It

captures key entities such as Person, Exercise, Muscle

Group. It includes object properties, such as targets

and partOfSession, and data properties (e.g., curren-

tWeight, currentReps).

ChatGPT: ChatGPT’s ontology includes 9

classes, 3 object properties, 13 data properties, 11

individuals, and 84 axioms. It covers basic concepts

such as Person, Exercise, and MuscleGroup, but

lacks concepts such as Workout routine or time under

tension for example. It includes properties such as

hasWeight and targetsMuscleGroup, but misses a lot

of data properties such as volume and start date for

example.

Expert: The expert ontology is the most com-

prehensive, capturing all major concepts and relation-

ships. It includes 10 classes, 10 object properties, 38

data properties, 21 individuals, and 178 axioms.

4.2 Consistency

Consistency measures whether the ontology is logi-

cally sound and free of contradictions.

Proposed Approach: The ontology produced by

the presented approach is consistent with OWL re-

strictions that are included. It covers some interest-

ing restrictions such as domain and range of proper-

ties. Processes of consistency checking intervening in

ChatGPT, the ofﬁcial app by OpenAI, version

1.2024.275

WEBIST 2024 - 20th International Conference on Web Information Systems and Technologies

462

Table 1: Comparison of Ontologies Generated by Different Approaches.

Evaluation Criteria Proposed Approach ChatGPT Expert Approach

Completeness High Medium Very High

Consistency High High Very High

Correctness High Medium Very High

Richness Medium-High Poor Very High

Reasoning Capabilities Medium No High

Human Effort Required Low Medium High

OMM and in SWRL-based reasoning module allow

to add elements, restrictions and constraints to an on-

tology with a better logically sound, while ensuring

consistency.

ChatGPT: The ChatGPT-generated ontology has

a good consistency due to the lack of logical restric-

tions. There is no inconsistency without restrictions

and constraints.

Expert: The expert ontology enforces strict con-

sistency through OWL restrictions. It ensures that

logical constraints like currentWeight and proteinIn-

take follow domain-speciﬁc constraints and that no

contradictory information is present.

4.3 Correctness

Correctness evaluates how well the ontology adheres

to domain-speciﬁc knowledge.

Proposed Approach: The generated ontology

captures domain knowledge well, including correct

classiﬁcations of exercises as compound or isolation

exercises. It also correctly models exercise instances

(e.g., Bench Press, Squats, and BicepCurls). How-

ever, it has some gaps, such as incomplete mod-

eling of ﬁner distinctions in muscle groups, like

SpecificMuscle.

ChatGPT: While ChatGPT correctly identiﬁes

basic entities like Bench Press and Squats as in-

stances of ex:CompoundExercise, it also creates

ex:CompoundExercise and ex:IsolationExercise as

two instances of ex:Exercise rather than creating a

class hierarchy between the three classes.

Expert: The expert ontology reﬂects domain

knowledge with high accuracy. It includes all clas-

siﬁcations, logical relationships, and constraints (e.g.,

protein intake per body weight), making it the most

correct and comprehensive.

4.4 Richness

Richness measures the complexity and detail within

the ontology, including class hierarchy, constraints

and restrictions to capture a ﬁne domain knowledge.

Proposed Approach: Although it retains a lower

level of reﬁnement than that of an expert, the ontology

captures some ﬁne knowledge such as the hierarchy

of Exercise classes, some OWL restrictions such as

domain and range of properties and some SWRL rules

(R1 and R2, c.f. subsection 3.4.1).

ChatGPT: ChatGPT’s ontology does not include

class hierarchy, logical constraints and restrictions,

limiting the detail and depth of its representation.

Expert: The expert ontology is the richest, with a

well class hierarchy, an extensive use of SWRL rules

and OWL restrictions. It models complex relation-

ships and behaviors, such as exercise type classiﬁca-

tions, progressive overload, and protein intake con-

straints.

4.5 Reasoning Capabilities

Reasoning Capabilities assesses the ontology’s abil-

ity to support automated reasoning and infer new

knowledge.

Proposed Approach: The ontology includes

some OWL restrictions such as domain and range of

properties that allow for some inferences (c.f. speciﬁc

muscle in subsection 3.3.5) and some SWRL rules

(R1 and R2 to infer the classiﬁcation of exercises as

compound or isolation exercise, c.f. subsection 3.4.1).

However, it lacks more advanced SWRL rules com-

pared to the expert ontology.

ChatGPT: ChatGPT’s ontology does not support

reasoning. It lacks OWL restrictions or SWRL rules,

meaning no logical inferencing or automated checks

are possible.

Expert: The expert ontology includes advanced

reasoning capabilities, using both OWL restrictions

and SWRL rules to infer new knowledge and enforce

domain-speciﬁc rules like progressive overload and

protein intake ranges.

4.6 Human Effort Required

Human Effort Required compares the amount of

manual work needed to produce the ontology.

Proposed Approach: Our approach signiﬁcantly

reduces manual effort by automatically generating the

ontology from input text. It requires minimal human

intervention, making it a scalable solution.

ODKAR: “Ontology-Based Dynamic Knowledge Acquisition and Automated Reasoning Using NLP, OWL, and SWRL”

463

ChatGPT: ChatGPT’s ontology requires more

manual intervention. While it can generate a basic

structure, the lack of logical constraints and reason-

ing capabilities means that signiﬁcant human effort is

required to reﬁne and complete the ontology.

Expert: The expert ontology requires the most

human effort, as it is manually created by a domain

expert. This results in high accuracy and complete-

ness but is time-consuming and requires specialized

knowledge.

In this evaluation, the term ”expert” refers specif-

ically to an expert in Semantic Web technologies and

ontology engineering, rather than a domain expert in

the ﬁeld of ﬁtness or physical training. This distinc-

tion is important when considering the complexity

and richness of the expert-generated ontology. It is

also critical to note that the techniques applied in this

evaluation are not limited to ﬁtness but are broadly

applicable to any domain requiring ontology-driven

knowledge representation and reasoning.

5 DISCUSSION

The evaluation of the proposed approach to ontol-

ogy generation reveals several strengths, particularly

in terms of completeness, consistency, and correct-

ness, achieved with minimal human intervention. By

automating the extraction of structured knowledge

from unstructured text, our system signiﬁcantly re-

duces the manual effort typically required in ontol-

ogy creation. While the expert-generated ontology

demonstrates superior richness and advanced reason-

ing capabilities, it requires extensive human effort and

domain expertise. This labor-intensive process is of-

ten not feasible for dynamic applications, where rapid

ontology updates are necessary. In contrast, our ap-

proach effectively balances automation and accuracy,

allowing domain specialists who may lack deep ex-

pertise in Semantic Web technologies to contribute to

ontology development. The comparison with Chat-

GPT underscores the added value of our system. Al-

though ChatGPT can identify key concepts, it lacks

the intricate reasoning mechanisms and formal logical

structures required for a fully coherent and logically

sound ontology. The necessity for signiﬁcant manual

reﬁnement with ChatGPT limits its usability in com-

plex domains. Our system introduces domain-speciﬁc

logical constraints and consistency checks, which en-

hance the quality of the generated ontologies and re-

duce reliance on manual post-processing.

However, these observations are only an initial as-

sessment of the results of our approach. As the evalu-

ation is based on a single text source and not various

corpora, it would be necessary to broaden and deepen

the evaluation of our approach on several text sources

representing various domains in order to further val-

idate and reﬁne its performance. Therefore, our fu-

ture work will aim to improve the experimental results

by including quantitative measures, such as precision,

recall and F1 scores, as well as qualitative analyses

to support our conclusions. This dual approach will

provide a better understanding of the performance of

our system compared with existing methods, includ-

ing other ontology generation systems. In addition,

these initial results have enabled us to identify ar-

eas for improvement in future research, such as as-

sertion enrichment, SWRL rule creation, and the sys-

tem’s ability to handle complex, multi-layered con-

straints. With these advances, we aim to increase the

practical applicability and efﬁciency of our approach,

ultimately contributing to the broader landscape of

knowledge management and representation.

REFERENCES

Al-Aswadi, F. N., Chan, H. Y., and Gan, K. H. (2020).

Automatic ontology construction from text: a review

from shallow to deep learning trend. Artiﬁcial Intelli-

gence Review, 53(6):3901–3928.

Antoniou, G. and Harmelen, F. v. (2009). Web ontology

language: Owl. Handbook on ontologies, pages 91–

110.

Asim, M. N., Wasim, M., Khan, M. U. G., Mahmood,

W., and Abbasi, H. M. (2018). A survey of ontol-

ogy learning techniques and applications. Database,

2018:bay101.

Baktash, J. A. and Dawodi, M. (2023). Gpt-4: A review on

advancements and opportunities in natural language

processing. arXiv preprint arXiv:2305.03195.

Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D.,

Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G.,

Askell, A., et al. (2020). Language models are few-

shot learners. Advances in neural information pro-

cessing systems, 33:1877–1901.

Chen, S., Zhang, Y., and Yang, Q. (2024). Multi-task

learning in natural language processing: An overview.

ACM Computing Surveys, 56(12):1–32.

Cimiano, P. and V

olker, J. (2005). Text2onto: A framework

for ontology learning and data-driven change discov-

ery. In International conference on application of nat-

ural language to information systems, pages 227–238.

Springer.

Etzioni, O., Banko, M., Soderland, S., and Weld, D. S.

(2011). Open information extraction: The second gen-

eration. In Proceedings of the Twenty-Second Inter-

national Joint Conference on Artiﬁcial Intelligence,

pages 3–10.

Fensel, D. and Fensel, D. (2001). Ontologies. Springer.

WEBIST 2024 - 20th International Conference on Web Information Systems and Technologies

464

Horrocks, I., Patel-Schneider, P. F., Boley, H., Tabet, S.,

Grosof, B., Dean, M., et al. (2004). Swrl: A semantic

web rule language combining owl and ruleml. W3C

Member submission, 21(79):1–31.

Kenton, J. D. M.-W. C. and Toutanova, L. K. (2019). Bert:

Pre-training of deep bidirectional transformers for lan-

guage understanding. In Proceedings of naacL-HLT,

volume 1, page 2.

Matthews, B. (2005). Semantic web technologies. E-

learning, 6(6):8.

Mihindukulasooriya, N., Tiwari, S., Enguix, C. F., and Lata,

K. (2023). Text2kgbench: A benchmark for ontology-

driven knowledge graph generation from text. In

International Semantic Web Conference, pages 247–

265. Springer.

Mukanova, A., Milosz, M., Dauletkaliyeva, A., Nazyrova,

A., Yelibayeva, G., Kuzin, D., and Kussepova, L.

(2024). Llm-powered natural language text process-

ing for ontology enrichment. Applied Sciences (2076-

3417), 14(13).

Osman, M., Noah, S., and Saad, S. (2024). Identifying

terms to represent concepts of a work process on-

tology. Indian Journal of Science and Technology,

17(24):2519–2528.

Pan, J. Z. (2009). Resource description framework. In

Handbook on ontologies, pages 71–90. Springer.

Ponciano, C., Schaffert, M., W

urriehausen, F., and Pon-

ciano, J.-J. (2022). Publish and enrich geospatial data

as linked open data. In WEBIST, pages 314–319.

Prudhomme, C., Homburg, T., Ponciano, J.-J., Boochs, F.,

Cruz, C., and Roxin, A.-M. (2020). Interpretation and

automatic integration of geospatial data into the se-

mantic web: Towards a process of automatic geospa-

tial data interpretation, classiﬁcation and integration

using semantic technologies. Computing, 102:365–

391.

Prudhomme, C., Homburg, T., Ponciano, J.-J., Boochs, F.,

Roxin, A., and Cruz, C. (2017). Automatic integration

of spatial data into the semantic web. In WEBIST,

pages 107–115.

Shamsfard, M. and Barforoush, A. A. (2004). Learning

ontologies from natural language texts. International

journal of human-computer studies, 60(1):17–63.

Shamshiri, A., Ryu, K. R., and Park, J. Y. (2024). Text min-

ing and natural language processing in construction.

Automation in Construction, 158:105200.

Sui, Z., Liu, Y., Zhao, J., and Zhang, H. (2010). The de-

velopment of an nlp-based chinese ontology construc-

tion platform. In 2010 IEEE/WIC/ACM International

Conference on Web Intelligence and Intelligent Agent

Technology, volume 3, pages 190–194. IEEE.

Wang, X. H., Zhang, D. Q., Gu, T., and Pung, H. K. (2004).

Ontology based context modeling and reasoning using

owl. In IEEE annual conference on pervasive comput-

ing and communications workshops, 2004. Proceed-

ings of the second, pages 18–22. Ieee.

Wong, W., Liu, W., and Bennamoun, M. (2012). Ontology

learning from text: A look back and into the future.

ACM computing surveys (CSUR), 44(4):1–36.

Yenduri, G., Ramalingam, M., Selvi, G. C., Supriya, Y.,

Srivastava, G., Maddikunta, P. K. R., Raj, G. D.,

Jhaveri, R. H., Prabadevi, B., Wang, W., et al. (2024).

Gpt (generative pre-trained transformer)–a compre-

hensive review on enabling technologies, potential ap-

plications, emerging challenges, and future directions.

IEEE Access.

Yin, M., Tang, L., Webster, C., Yi, X., Ying, H., and Wen,

Y. (2024). A deep natural language processing-based

method for ontology learning of project-speciﬁc prop-

erties from building information models. Computer-

Aided Civil and Infrastructure Engineering, 39(1):20–

45.

Zhang, Z., Elkhatib, Y., and Elhabbash, A. (2023). Nlp-

based generation of ontological system descriptions

for composition of smart home devices. In 2023 IEEE

International Conference on Web Services (ICWS),

pages 360–370. IEEE.

ODKAR: “Ontology-Based Dynamic Knowledge Acquisition and Automated Reasoning Using NLP, OWL, and SWRL”

465