GLUON: A Reasoning-based and Natural Language Generation-based
System to Explicit Ontology Design Choices
Zakaria Mejdoul
a
and Ga
¨
elle Lortal
b
Thales Research & Technology, Palaiseau, France
Keywords:
Ontology, Reasoner, Controlled Natural Language, Verbalization, Sentences Generation.
Abstract:
The Semantic Web (SW) is an enhancement to the World Wide Web (WWW). It allows Humans to find, share
and integrate information more easily. One of the Knowledge Representation technologies related to the
SW standards is the ontology, which is an inventory of knowledge defining a universal or specific domain.
Ontology construction requires expertise related to logics and to the expert domain the SW application is
applied to. We aim to enable ontology wide adoption by bringing the end-users closer to their own ontology
building choices, providing them with the possibility to build their ontology, to validate its consistency and
to formally represent their knowledge without formal methods knowledge. In this paper, we detail our tool
architecture combining SW technologies and Natural Language Generation (NLG) to support users in creating
consistent ontologies.
1 INTRODUCTION
Knowledge sharing is key to support intelligent pro-
cess automation and collaborative problem solving
(Gruber, 1995) in large organizations. THALES, a
Worldwide company, needs efficient critical data and
knowledge management to support daily business
applications in Air Traffic Management (ATM) and
Avionics. An efficient tool to explicit information and
allow its sharing (Wankhade and Raut, 2021) is ontol-
ogy. Thus, the solution we consider to improve busi-
ness application is ontology-based application appro-
priately managing domain knowledge (e.g. (Wu et al.,
2020)). This process involves knowledge modeling as
a ontology (Shimizu et al., 2020).
Despite, ontologies are expected to be widely used
in SW applications to bridge the gap among machines
and humans, their acceptance and use is hampered by
their complexity and the need of skilled knowledge
engineers to develop them. The human interaction as-
pects of these formalisms (ease of learning, reading,
writing) have received little attention. Ontology ed-
itors such as Prot
´
eg
´
e (Musen, 2015), Neon (Su
´
arez-
Figueroa et al., 2015), etc. still require knowledge
and training in ontology and formal logics domains,
as well as a good grasp of the editor’s plugins and
a
https://orcid.org/0000-0003-4341-2139
b
https://orcid.org/0000-0001-5374-6584
widgets, which makes the handling of these software
difficult. Likewise, keeping track of the ontology
design decisions is another missing aspect, blurring
design rationale consequences and disorienting non-
ontology-experts.
In section 2, we present the Air Mission domain
we want to improve. In section 3, we present the
use cases and the main scenarios we considered to
evaluate the solution. In section 4, we detail existing
systems of ontology, reasoners and verbalizers, then
in section 5, the technical architecture including our
conceptual design. Section 6 covers results of vali-
dation scenarios simulation as described in section 3.
Then, we conclude and state the future work.
2 CONTEXT
With increasing airspace traffic congestion, ATM and
Avionics need more advances in massive data pro-
cessing, sophisticated edge avionics (coordination
with weather updates, estimation of multiple uncer-
tainty types,...). The complexity of the situation over-
whelms the skills of pilots or ATM controllers. In
order to provide automatic and Artificial Intelligence
(AI) decision assistance for Air Missions, ontologies
are an attractive knowledge technology. Air domains
information management systems must be efficient
with larger amounts of data, effective in combining
228
Mejdoul, Z. and Lortal, G.
GLUON: A Reasoning-based and Natural Language Generation-based System to Explicit Ontology Design Choices.
DOI: 10.5220/0011587900003335
In Proceedings of the 14th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2022) - Volume 2: KEOD, pages 228-236
ISBN: 978-989-758-614-9; ISSN: 2184-3228
Copyright
c
2022 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
information from different sources, and relevant in
reducing uncertainty in decision support systems (In-
saurralde and Blasch, 2018).
An attractive approach to supporting decision
making in advanced Air Mission systems is the imple-
mentation of ontologies for avionics systems (Pala-
cios et al., 2016), (Palacios et al., 2018). Ontolo-
gies are intended to model cognitive processes by
representing and reasoning about knowledge. In air
Mission field, two ontologies are widely used: AT-
MONTO
1
(The NASA Air Traffic Management On-
tology) which describes classes, properties and rela-
tionships related to ATM domain and represents in-
formation about a large and diverse set of interacting
components in the U.S. and global airspace; AIRM
2
(ATM Information Reference Model) which is an
OWL ontology derived from Air Traffic Management
(ATM) and was developed in the framework of the
BEST project of the Single European Sky ATM Re-
search (SESAR) program (Standar, 2012).
To set up SW applications in air Mission field, we
built scenarios answering realistic Air Mission use-
cases (UCs) thanks to these two open-source ontolo-
gies (ATMONTO and AIRM). Our scenarios cover
events that can occur during an air mission as: a
weather event, a destination diversion, a systematic
failure.
Table 1: Ontologies metrics extracted from Prot
´
eg
´
e.
Ontologies ATMONTO AIRM
Metrics
axiom 1644946 34576
logical axiom count 1386774 15735
declaration axioms count 2768 9605
class count 2461 1177
object property count 127 2727
data property count 189 1972
individual count 226516 3727
loading time (ms) 18966 837
The table 1 shows the ontologies metrics AT-
MONTO and AIRM which are composed of domains
and concepts similar to our working context (air mis-
sion management). These ontologies metrics come
from the Prot
´
eg
´
e ontology editor, giving a schema on
the richness and density of the two: ATMONTO (with
a total of 1644946 axioms) is richer and more com-
plex compared to AIRM.
1
https://data.nasa.gov/ontologies/atmonto/index.html
2
https://airm-o.github.io/airm-o/
3 USE-CASES
3.1 Purposes
To bring end-users closer to their ontology building
choices and to provide them with formalization and
validation means depends on end-users experience.
We consider three main interaction levels in a SW ap-
plication to fit with our end-users interaction needs:
Average users with no computer science back-
ground, who want to contribute to the SW
with knowledge about their domains of interest
(Berners-Lee et al., 2001).
Domain experts, who create formalized knowl-
edge to be used in a SW application. Tradi-
tionally, these ontologies are created by ontol-
ogy engineers, who interview domain experts in
an iterative process. Allowing domain experts to
create initial versions of these ontologies them-
selves lower the costs compared to the traditional
method.
Ontology engineers using our approach to rapid
ontology prototyping.
We want to achieve our goal of increasing the
ontology acceptance by enabling end-users to han-
dle with full transparency the formalisms required to
build them. As our aim is to support users, we will
first propose the most widely used format of ontol-
ogy to our users, namely OWL2. The bridge we build
is between the OWL logical and formal format and
an easy and natural interaction means: Natural Lan-
guage. These end-users act in a domain and for our
works we gathered Air Mission needs. In the next
section, we present our scenarios to test the use of on-
tologies.
3.2 Description of UCs
To be able to evaluate the usability and performances
of our solution, we set up validation scenarios, based
on various UCs given by Air Mission experts. The
identified UCs are represented in our air mission on-
tology and are:
System failure scenario;
Diversion due to arrival runway closed scenario;
Meteorological event scenario.
They are formally described below. Each scenario re-
flects an incident, which can occur during a flight or
as we call it, air mission.
GLUON: A Reasoning-based and Natural Language Generation-based System to Explicit Ontology Design Choices
229
3.2.1 System Failure Scenario
This system failure UC, due to an air mission oper-
ated with an aircraft having a low fuel level (value of
the level is equal to 0.0) can trigger an emergency. An
emergency mission implies that the mission is unsat-
isfiable, therefore, the mission is classified as unsatis-
fiable mission, by semantic rules (‘SWRL rule). This
System Failure scenario is represented as follows:
Property Assertions:
Individual: MissionSMAScenarioA
MissionSMAScenarioA Type Mission
MissionSMAScenarioA isFlownWith F-DEV1
F-DEV1 hasFuel FDEV1CurrentFuel
FDEV1CurrentFuel Type Fuel
FDEV1CurrentFuel hasFuelValue ’0.0’
Mission DisjointWith UnsatisfiableMission
LowFuel SWRL Rule:
Mission(?m) ˆ Fuel(?fcv) ˆ IsFlownWith(?m, ?a)
ˆ
hasFuel(?a, ?fcv) ˆ hasFuelValue(?fcv, 0.0)
-> hasUrgency(?m, "3")
Urgency SWRL Rule:
Mission(?m) ˆ hasUrgency(?m, ?urg) ->
UnsatisfiableMission(?m)
Explanation: F-DEV1 has fuel with fuel value 0.0
Implies: MissionSMAScenarioA Type UnsatisfiableMis-
sion
3.2.2 Diversion Due to Arrival Runway Closed
Scenario
It reflects a diversion of the aircraft’s destination dur-
ing a mission, due to a closure of the runway intended
for the destination, which also implies that the current
mission is unsatisfiable, by an implication semantic
rule. This Diversion scenario is represented as fol-
lows:
Property Assertions:
Individual: MissionSMAScenarioB
MissionSMAScenarioB Type Mission
MissionSMAScenarioB hasActiveAirportDeparture
LFPB
MissionSMAScenarioB hasActiveAirportArrival
KTEB
MissionSMAScenarioB hasActiveArrivalRunway
KTEB19
KTEB19 isOpen 0
Mission DisjointWith UnsatisfiableMission
ArrivalRunwayAvailability SWRL Rule:
Mission(?m) ˆ hasActiveArrivalRunway(?m, ?arw)
ˆ isOpen(?arw, 0) -> UnsatisfiableMission(?m)
Explanation: KTEB19 is not open
Implies: MissionSMAScenarioB Type UnsatisfiableMis-
sion
3.2.3 Meteorological Event Scenario
It describes an air mission failure due to a weather
event that can endanger the mission aircraft during its
landing on its destination runway: the wind speed of
the destination runway exceeds the speed limit that
the aircraft can tolerate, which also trigger a pre-
landing condition alert. This event classifies an air
mission as unsatisfiable mission by a rule containing
a conditional form (swrlb:greaterThan) and then a di-
rect implication. The meteorological event scenario is
represented as follows:
Property Assertions:
Individual: MissionSMAScenarioC
MissionSMAScenarioC Type Mission
MissionSMAScenarioC hasActiveArrivalRunway KTEB06
MissionSMAScenarioC isFlownWith F-DEV1
F-DEV1 hasLandingCapacity F-DEV1LandingCapacity
F-DEV1LandingCapacity hasCrosswindLimitation ’5.0’
KTEB06 hasWeather Weather_Crosswind_KTEB
Weather_Crosswind_KTEB hasWindMagnitude ’20.0’
Mission DisjointWith UnsatisfiableMission
CrosswindSatisfiability SWRL Rule:
Mission(?m) ˆ IsFlownWith(?m, ?a) ˆ
hasLandingCapacity(?a, ?lc) ˆ
hasCrosswindLimitation(?lc, ?xwindLim) ˆ
hasActiveArrivalRunway(?m, ?arw) ˆ
hasWeather(?arw, ?w) ˆ Crosswind(?w) ˆ
hasWindMagnitude(?w, ?xwm) ˆ
swrlb:greaterThan(?xwm, ?xwindLim) ->
UnsatisfiableMission(?m)
Explanation: KTEB06 has wind magnitude 20.0 greater
than F-DEV1 limitation 5.0
Implies: MissionSMAScenarioC Type UnsatisfiableMis-
sion
The proposed scenarios lead to classify a mission
into ”missions that cannot be accomplished”, through
inconsistencies detection in the air missions ontology
and declaring also an expert intervention. The incon-
sistency is triggered by a disjoint relation between
the concepts ”Mission” and ”UnsatisfiableMission”.
For this function, we developed a solution, on which
we can test scenarios based on the main UCs of air
mission applications. Nevertheless, several solutions
close to our needs exist. The next section presents
them.
4 RELATED WORKS
Our basic needs are to 1. Create ontology, 2. En-
sure users their modeling actions are consistent and 3.
Ensure users their modeling actions answer their de-
sign choices. Existing systems answer partially these
KEOD 2022 - 14th International Conference on Knowledge Engineering and Ontology Development
230
needs using Ontologies engines, Semantic Reasoners
and Verbaliser. We present a review below.
4.1 Ontology and Semantic Reasoning
Ontology Overview. In less than thirty years, the
term ontology has gained considerable popularity
in the field of computer science and information sys-
tems. This popularity is due to the ability of ontolo-
gies to achieve interoperability among multiple rep-
resentations of reality (e.g., data or business process
models) residing in computer systems, and among
these representations and reality, i.e., human users
and their perception of the domain. Surprisingly for a
tool, which aims at reconcialiting people from various
communities, the term ontology has different, even
incompatible, definitions. It is something of a para-
dox that the starting term of a research field, which
aims to reduce ambiguity about the meaning of sym-
bols, is understood and used so inconsistently. From
the early years of ontology research, Guarino and Gi-
aretta (Guarino et al., 1995) were concerned about the
inconsistent use of the term ontology”. They found
at least seven different notions attributed to the term.
We consider this general definition for our works: An
ontology is a structured set of insights in a particular
domain of knowledge (Guarino et al., 1995). We also
consider the use of a W3C recommendation as OWL2
(Hitzler et al., 2012) to ensure interoperability base on
a valid and shared syntax and hence semantics.
To encapsulate OWL possibilities in a new appli-
cation, the use of OWL API seems inescapable despite
some lacks. Most of reasoners are accessible via the
OWL API, which is advantageous for applications that
want to access multiple reasoners via the same inter-
face. The use of the OWL API greatly facilitated the
execution of our experiments. The design of the OWL
API is directly based on the OWL2 structural specifi-
cation described in (Motik and al., 2008).
Other Ontology Frameworks. A number of other
similar developments have been initiated to provide
application interfaces for OWL: Jena toolkit (Car-
roll and al., 2004) provides practical ontological in-
terfaces around RDF interfaces. Comparisons of the
OWL API and Jena’s triplet-based approach for some
tasks are discussed in (Bechhofer and Carroll, 2004);
The KAON toolkit (Bozsak and al., 2002) is an open
source ontology management system for commercial
applications. The KAON toolkit includes an API
for RDF graph processing and differs from the OWL
API in that KAON doesn’t offer support for process-
ing OWL ontologies, and the OWL API doesn’t pro-
vide direct support for processing RDF graphs. Thus,
OWL API is a good candidate for our works, yet our
system needs a reasoner to ensure consistency and ex-
plain users their modeling actions.
Reasoner Overview. A reasoner” is a program that
performs reasoning on an ontology or a knowledge
base (KB). It relies on the KB, as well as on the set of
rules of the ontology to infer (logically deduce) new
knowledge updating the KB content.
However, rules, common to all reasoners, may not
be sufficient for all application cases. Moreover, it is
possible to define rules specific to a given application,
in a language like SWRL (Semantic Web Rule Lan-
guage). In this case, the reasoner assimilates these
new rules and applies them to the KB in the same
way they pre-established rules. Advanced verification
techniques are imperative to have a consistent ontol-
ogy. The reasoner then composes the core of the func-
tional validation module.
The logics we rely on is Description Logics in
general to ensure OWL adequacy with reasoning as
well as termination of valid inferences. Description
logics (DLs) are a family of knowledge representa-
tion formalisms based on classes (concepts). They
are characterized by the use of various constructors to
build complex concepts from simpler concepts, by the
decidability of key reasoning tasks, and by providing
correct, complete and feasible reasoning. DLs have
been used in a range of applications, for example con-
figuration and reasoning about schemas and database
queries (Toman and Weddell, 2022).
Reasoners Benchmark. To specify the reasoners,
we relied on a benchmark based on reasoning fea-
tures: Reasoning Algorithm, Expressivity, Rule Sup-
port, Justification (provides explanations for inconsis-
tencies that exist in ontologies) and ABox reasoning
task support (reasoning with individuals for instance
checking, answering queries and ABox consistency
checking), the benchmark is also based on usability
features: support for OWL API and reasoner license
(open source/commercial).
Table 2: Reasoner Benchmark (HermiT, Pellet and ELK).
HermiT Pellet ELK
Format OWL2 OWL2 OWL2
DL DL/EL EL
Reasoning Hyper- Tableau CB
Algorithm Tableau
Expressivity SROIQ (D) SROIQ (D) EL
Rule Support SWRL SWRL Own
rule format
Justifications - + +
Abox + + -
Reasoning (SPARQL)
OWL API + + +
License Open- Open- Open-
source source source
+ stands for yes and - for no.
GLUON: A Reasoning-based and Natural Language Generation-based System to Explicit Ontology Design Choices
231
After a detailed study of ontological reasoners and
our benchmark results, the table 2 shows us a good
compromise among speed, expressivity and precision:
Pellet, based on the Tableau algorithm and supporting
OWL2 DL. Nevertheless, for unit tests we use also
ELK based on the Consequence-based algorithm, al-
though it supports a less expressive profile of OWL,
OWL2 EL. Now we set a reasoner to ”understand”,
users should be able to understand too. We checked
the different verbalizers and Controlled Natural Lan-
guages (CNLs) systems.
4.2 Verbalizers
Overview. Natural language is very expressive and
doesn’t require additional learning effort to present
Human knowledge. To represent knowledge in com-
puters, people use formal languages (Fuchs et al.,
2008) as CNLs. To make ontologies easier to un-
derstand, several ontology verbalizers have been de-
veloped (Schwitter, 2010). The verbalizers typically
translate the ontology axioms (here the OWL state-
ments) one by one into CNLs statements. Often not
quite fluent, they generally don’t pay attention to the
resulting texts consistency. There are different syn-
tax possibilities to express a specific OWL axiom in
CNL.
Several works were led to develop CNLs, mainly
from English, which can be used as alternative OWL
syntaxes (Khabarov et al., 2019). Sydney OWL Syn-
tax (sos) (Cregan et al., 2007) is an English-like CNL
with a bidirectional mapping to and from OWL syn-
tax; sos is based on peng (Schwitter, 2010). A similar
bidirectional mapping has been defined for ACE ”At-
tempto Controlled English” (Fuchs et al., 2008).
The verbalization approach proposed by (Vidan-
age et al., 2021) gave good results and verbalized be-
yond the level of the CNL. Its free configuration, be-
ing domain and schema independent and its overall
accuracy results, close to 82%, is interesting. How-
ever, Semantic Level Refinement (Venugopal and Ku-
mar, 2021), aiming at transforming knowledge repre-
sented as a combination of low expressive (and non-
specific) logical expressions into high expressive and
specific expressions is more precise. This technique
uses a predefined set of rules that are repeatedly ap-
plied to the restrictions associated with individuals
(and concepts) to obtain a refined set of restrictions.
These refined restrictions sets are then verbalized to
obtain concise descriptions of the ontology’s enti-
ties. Nevertheless, SLR is not particularly adpated to
OWL2 and needs heavy preprocessing.
ACE is particularly, well adapted to the verbal-
ization of OWL ontologies, because the axioms are
expressed in compact ready-made sentences, which,
besides, do not offer the user the possibility to ex-
press variable patterns that OWL would not be able
to express. In other words, ACE makes explicit the
sentences that are compatible with the expressivity of
OWL and those that are not. ACE is formally de-
scribed by a Definite Clause Grammar (DCG) which
also provides a mapping of ACE to (a part of) OWL
which can be a set of axioms (Fuchs et al., 2008). The
main focus is on the transition from CNL to OWL.
Evaluation of existing systems. This section
presents a comparison between existing systems and
OWL verbalisation libraries.
Table 3: Comparisons of OWL-based verbalizers.
System Tbox Abox Cove- Aggre- Lexi- Dom-
rage gation con ain
ACE + + OWL2 - Auto Gen
SWOOP + - OWL - Auto Gen
DL
MIAKT - + RDF + Hand- Spec
crafted Spec
Natural- + + OWL + User- Spec
OWL DL defined
SWAT + + OWL2 + Auto Gen
+ stands for yes and - for no. Auto stands for automatic. Gen
and Spec stands respectively for Generic and Specific.
The table 3 shows a comparison of the different
OWL verbalizers, according to several axes: ’TBox’
and ’ABox’. The coverage is approximate: for ex-
ample, ACE and SWAT cover almost all of OWL2.
The lexicon indicates the source of lexical entries for
atomic entities. ’Domain’ is generic if the system can
produce text for any ontology (of the indicated OWL
coverage) with English names, and specific if hand-
crafted lexical entries or grammar rules are needed.
ACE (Fuchs et al., 2008), SWAT (Power, 2012)
and SWOOP (Narasimha et al., 2022) are suitable
for generic domains as lexicons are created automati-
cally, which is good for a larger ontology with a gen-
eral domain. Compared to MIAKT (Bontcheva and
Wilks, 2004), ’TBox is not supported, which is a cru-
cial factor when generating text from an ontology.
Then, both NaturalOWL and ACE could be used
for our UCs as they provide more functionality and
are relevant for our topic. The exception is the in-
tegrability and the fact that both systems offer APIs
exploitable by our system and adequate with the ar-
chitecture and flows of our solution. For technical
reasons, then, we chose the ACE-based OWL verbal-
izer.
KEOD 2022 - 14th International Conference on Knowledge Engineering and Ontology Development
232
5 SOLUTION ARCHITECTURE
This section discusses the proposed solution of
GLUON (Generation of Links Unifying an ONtol-
ogy) and in figure 1 shows the high level architecture
of GLUON.
To solve this problematic, we have developed a
module named GLUON which is the acronym of
Generation of Links Unifying an ONtology”, a user-
friendly ontology design and creation methodology
assisted by software components, each one dedicated
to a task of the scenario of iterative ontology creation
and evaluation, with the aim of making ontologies us-
able by non-experts. We want to make everyone in-
terested in ontology creation understand that ontol-
ogy design does not have to be complex and that by
using our guide-assisted methodology, the user will
be more productive than with existing methodologies
and tools. The novelty of the solution resides in the
Figure 1: High level architecture of GLUON.
integration of an OWL verbalizer and in the fact that
informative sentences are generated with the ontology
itself by the system equipped with an ontology con-
structor, a reasoning system and a natural language
generator or verbalizer.
5.1 Components
5.1.1 Ontology Constructor
This component allows to create an ontology from
scratch or to load an already designed ontology via
the ontology handler, which is a group of instructions
from an ontology engine.
This block allows the creation of concepts, rela-
tions and individuals to be added to the ontology in
question, or even to apply CRUD (Create, Read, Up-
date, Delete) operations on ontology’s content to en-
sure what we can call persistence of ontology data
(mechanism responsible for the backup and restora-
tion of the ontology’s data or content). The user also
has the possibility to add rules (e.g. SWRL rules)
to his ontology by adapting them to the domain and
the business logic that organize his field of expertise.
About ontology engine, we opted for OWL API as an
ontology manipulation library. This module is present
in many uses and has a large community that uses it,
so documentation and references can also complete
our need when using it.
5.1.2 Reasoning System
This sub-module is a set of subsystems allowing the
use of a standalone reasoner or several parallel rea-
soning processes. In addition to reasoning on an on-
tology (inferences, detection of inconsistencies and
unsatisfiabilities), this module also includes the func-
tion of generating explanations in the form of justi-
fications expressed in axioms (OWL axioms). This
module also provides a function that manages the rea-
soner in use by using the ontology defined by the
ontology constructor module and checks its consis-
tency after each operation or change that the user ap-
plies to the ontology in use. Explanation is a func-
tion that also extends this module and generates ex-
planations in the form of an owl axiom or a list of
OWL axioms after each operation applied on the cur-
rent ontology, also using the default reasoner (possi-
bility to use more than one reasoner, e.g. Pellet, ELK,
...) for the generation of explanations. For the se-
lection of reasoners, two reasoners were chosen that
are aligned with two different types of ontology for-
mats and reasoning algorithms: Pellet based on the
Tableau algorithm and supports OWL2 DL and ELK
based on the Consequence-based algorithm and sup-
ports OWL2 EL. This choice allows us to cover both
DLs and ELs formats and also allows justification,
GLUON: A Reasoning-based and Natural Language Generation-based System to Explicit Ontology Design Choices
233
SWRL rules support, Abox reasoning and the use of
OWL API as part of reasoning processes.
5.1.3 Verbalisation System
The verbalization system is an integral component of
the system where sentences are generated from this
module with the content of the ontology in question.
In view of the flexibility of creating NL resources in
NaturalOWL, it is more appropriate to create an NLG
system for small domains. Since the user has to de-
fine the NL resources, this type of library is not suit-
able for large-scale ontologies, where the author will
have extra work to generate the texts. Furthermore,
this library does not provide any support for integra-
tion with other systems. Considering a potential com-
parison, NaturalOWL and ACE could be used for our
case because they provide more functionalities and
are relevant for our subject. The exception is the in-
tegrability and the fact that both systems offer APIs
that can be used by GLUON and that are adequate for
the architecture and flow of our solution, technically
viewed. In this project, we chose OWL Verbalizer as
the ACE-based OWL verbalizer. For this project, we
used the verbalizer as an HTTP server and at the same
time exploited the module’s Java methods for hybrid
verbalization. The lexicons can either be generated
from the names associated to concepts, relations and
individuals (e.g. #Aircraft, #flownWith, #MissionA
.etc) or can be defined by the user through the lexicon
handler module by adapting it to the different gram-
matical forms it can take (singular form, plural form,
past participle form for verbs .etc). Nouns, adjectives
and verbs must be defined for better form of genera-
tion or have the possibility to import a predefined NL
resource thanks to lexicon handler methods. The ben-
efits of OWL Verbalizer include: expressing the con-
tent of OWL ontologies in a natural way; naturally
adapts to expressing the content of rules and business
queries; textualizing an OWL ontology in concise, un-
derstandable English, on condition to adopt a specific
naming style for OWL individuals, classes and prop-
erties; allows people with no formal logic background
to read and understand ontologies written by others,
provided the readers are experts in the domain de-
scribed by the ontology; uses the syntax of the ACE
controlled sub-language; Provides interfaces (APIs)
that can be used in other software. On disadvantages:
difficulty in aggregating sentences in the case of ver-
balizing a list of axioms; generic and non-customized
lexicon ( fixed with the lexicon handler).
6 EVALUATION
GLUON enables the outputs in CNL from OWL ax-
ioms presented below. Our evaluation plan is in two-
folds. The first part is unit test for consistence check-
ing of the scenarios based on avionics experts needs.
Here we present the Alpha scenario i.e. the consistent
ontology and then the 3 scenarios revealing failures.
The second part will be to set up a full scenario for
avionics experts, who will play an ontology building
according to a new mission (as Sick Passenger On-
board).
6.1 Scenario Alpha (Consistent
Ontology)
Consistent Ontology Verbalization:
F-DEV1 is an Aircraft.
F-DEV1 has fuel FDEV1 current fuel.
FDEV1 current fuel is a fuel.
FDEV1 current fuel have fuel value 0.0.
KTEB is an airport.
KTEB06 is an operational runway.
LFPB is an airport.
Mission SMA A is a mission.
Mission SMA A has active airport arrival KTEB.
...
6.2 Scenario A (System Failure)
System Failure Inconsistency Explanation: You
know that Mission SMA A is flown with F-DEV1 and
FDEV1 current fuel has fuel value 0.0 and Mission
SMA A is a mission and No mission is an unsatis-
fiable mission and F-DEV1 has fuel FDEV1 current
fuel and FDEV1 current fuel is a fuel.
6.3 Scenario B (Diversion for Closure)
Diversion Inconsistency Explanation: You know
that KTEB19 is open 0 and Mission SMA B has active
arrival runway KTEB19 and No mission is an unsat-
isfiable mission and Mission SMA B is a mission.
6.4 Scenario C (Meteorological Event)
Meteorological Event Inconsistency Explanation:
You know that Mission SMA C has active arrival run-
way KTEB06 and Mission SMA C is flown with F-
DEV1 and KTEB06 has weather weather-crosswind
of KTEB and F-DEV1 has landing capacity F-DEV1
landing-capacity and No mission is an unsatisfiable
mission and weather-crosswind of KTEB has wind
KEOD 2022 - 14th International Conference on Knowledge Engineering and Ontology Development
234
magnitude 20.0 and Mission SMA C is a mission and
F-DEV1 landing-capacity has crosswind limitation
5.0.
6.5 Discussion
The results show that despite the simplicity of most
axioms in real-world ontologies, these ontologies can
contain very complex axioms. Only a few of these
axioms are so complex that the verbalization method
proposed by OWL Verbalizer cannot handle them, the
rest, however, can be presented in a way that is more
readable than the standard syntax of the description
logic and its variants.
By feeding the verbalizer a lexicon of morpho-
logical correspondences, we can further improve the
verbalization. Note also that we have presented each
verbalization as a simple sequence of words. Finally,
we note that the integration of the verbalizer is easy
to adapt to GLUON and can translate large currently
available OWL ontologies in seconds.
Figure 2: GLUON UI Prototype (Ontology construction).
Figure 3: GLUON UI Prototype (Inconsistency check).
The semantics of the sentences that composes the
generated text is identically fine-grained and related
to the main causes of inconsistencies that the three
scenarios from the aerial missions domain declare.
A first GUI of GLUON (Figure 2 and 3) has been
developed to support the user in the ontology creation,
what remains to be done is to adapt this interface to
the ontology validation events through the verbaliza-
tion methods of GLUON, developed and tested
with the defined aeromission scenarios.
7 CONCLUSIONS
In this paper, we promoted ontology broad acceptance
by bringing the various formats closer to end-users,
who may have no knowledge of formal methods. On
the one hand, our tool allows user to create SW ap-
plication backbone, i.e. ontology, in a user-friendly
way. On the other hand, it provides an unambiguous
language to explain (verbalize) the content of exist-
ing ontologies in CNL. In order to support ontology
creation and to formally represent the consequences
of each step of ontology design in a CNL, GLUON
(Generation of Links Unifying ONtology) has been
developed. The GLUON system, which supports the
creation of ontologies and more specifically the ver-
balization of inferences in the ontology, is a system
presenting a methodology for the creation of ontolo-
gies, a set of components necessary to ontology build-
ing and a principle for the validation of the ontology
created. It has been evaluated through Air Mission
UCs given by Air Mission experts. As future work,
we intend to improve the developed GLUON HMI of
section 6.5 to adapt it to scenarios of ontology cre-
ation and clarification of user’s design choices. Such
a graphical interface should answer interaction lev-
els and be integrated with the methods developed in
GLUON, in order to be able to identify the reasons
for inconsistencies or invalid actions, and to justify
them in verbalized texts.
REFERENCES
Bechhofer, S. and Carroll, J. J. (2004). Parsing owl dl: trees
or triples? In WWW ’04.
Berners-Lee, T., Hendler, J. A., and Lassila, O. (2001). The
semantic web” in scientific american.
Bontcheva, K. and Wilks, Y. (2004). Automatic report
generation from ontologies: The miakt approach. In
NLDB.
Bozsak, E. and al. (2002). Kaon - towards a large scale
semantic web. In EC-Web.
Carroll, J. J. and al. (2004). Jena: implementing the seman-
tic web recommendations. In WWW Alt. ’04.
Cregan, A., Schwitter, R., and al. (2007). Sydney owl syn-
tax - towards a controlled natural language syntax for
owl 1.1. In OWLED.
Fuchs, N. E., Kaljurand, K., and Kuhn, T. (2008). Attempto
controlled english for knowledge representation. In
Reasoning Web.
GLUON: A Reasoning-based and Natural Language Generation-based System to Explicit Ontology Design Choices
235
Gruber, T. R. (1995). Toward principles for the design of
ontologies used for knowledge sharing? Int. J. Hum.
Comput. Stud., 43:907–928.
Guarino, N., Uniti, S., and Giaretta, P. (1995). Ontologies
and knowledge bases. towards a terminological clari-
fication.
Hitzler, P., Kr
¨
otzsch, M., Parsia, B., Patel-Schneider, P. F.,
and Rudolph, S. (2012). Owl 2 web ontology language
primer (second edition).
Insaurralde, C. C. and Blasch, E. P. (2018). Uncertainty in
avionics analytics ontology for decision-making sup-
port. Journal of Advances in Information Fusion, 13.
Khabarov, V., Stepanov, I., and Serenko, A. (2019). Con-
trolled natural language for ontology editing. Science
Bulletin of the Novosibirsk State Technical University.
Motik, B. and al. (2008). Owl 2 web ontology language:
structural specification and functional-style syntax.
Musen, M. A. (2015). The prot
´
eg
´
e project: a look back and
a look forward. AI matters, 1 4:4–12.
Narasimha, V. B., Sujatha, B. M., and al. (2022). An open-
source web-based owl ontology editing and browsing
tool: Swoop. The Role of IoT and Blockchain.
Palacios, L., Lortal, G., and al. (2016). Avionics mainte-
nance ontology building for failure diagnosis support.
In International Conference on Knowledge Engineer-
ing and Ontology Development, volume 3, pages 204–
209. SCITEPRESS.
Palacios, L., Lortal, G., and al. (2018). Knowledge dis-
covery for avionics maintenance support. In 2018
IEEE/AIAA 37th Digital Avionics Systems Conference
(DASC), pages 1–8. IEEE.
Power, R. J. D. (2012). Owl simplified english: A finite-
state language for ontology editing. In CNL.
Schwitter, R. (2010). Controlled natural languages for
knowledge representation. In COLING.
Shimizu, C., Hitzler, P., and al. (2020). Modular ontology
modeling: A tutorial. In Applications and Practices
in Ontology Design, Extraction, and Reasoning.
Standar, M. (2012). Sesar and the european atm master plan
- cooperation with the us/nextgen. In ICNS 2012.
Su
´
arez-Figueroa, M. C., G
´
omez-P
´
erez, A., and al. (2015).
The neon methodology framework: A scenario-based
methodology for ontology development. Appl. Ontol-
ogy, 10:107–145.
Toman, D. and Weddell, G. E. (2022). First order rewritabil-
ity in ontology-mediated querying in horn description
logics. In AAAI.
Venugopal, V. E. and Kumar, P. S. (2021). Verbalizing but
not just verbatim translations of ontology axioms. In
BNAIC/BENELEARN.
Vidanage, K., Mohemad, R., and al. (2021). Fully auto-
mated ontology increment‘s user guide generation.
Wankhade, S. R. and Raut, A. B. (2021). A review on
ontology-based semantic web information retrieval:
Techniques, weight functions. Algorithms for Intel-
ligent Systems.
Wu, Z., Shen, Y., and al. (2020). An ontology-based frame-
work for heterogeneous data management and its ap-
plication for urban flood disasters. Earth Science In-
formatics, 13:377–390.
KEOD 2022 - 14th International Conference on Knowledge Engineering and Ontology Development
236