DOMAIN ONTOLOGY GENERATION USING LMF

STANDARDIZED DICTIONARY STRUCTURE

Feten Baccar Ben Amar, Bilel Gargouri

MIRACL Laboratory, University of Sfax, FSEGS, B.P. 1088, 3018, Sfax, Tunisia

Abdelmajid Ben Hamadou

MIRACL Laboratory, University of Sfax, ISIMS, B.P. 242, 3021 Sakiet-Ezzit, Tunisia

Keywords: Core Domain Ontology Generation, LMF Standardized Dictionary, Ontology Quality, Arabic Language.

Abstract: The present paper proposes a methodology for generating core domain ontology from LMF standardized

dictionary (ISO-24613). It consists in deriving the ontological entities systematically from the explicit

information, taking advantage of the LMF dictionary structure. Indeed, such finely-structured source

incorporates multi-domain lexical knowledge of morphological, syntactic and semantic levels, lending itself

to ontological interpretations. The basic feature of the proposed methodology lies in the proper building of

ontologies. To this end, we have integrated a validation stage into the suggested process in order to maintain

the coherence of the resulting formalized ontology core during this process. Furthermore, this methodology

has been implemented in a rule-based system, whose high-performance is shown through an experiment

carried out on the Arabic language. This choice is explained not only by the great deficiency of work on

Arabic ontology building, but also by the availability within our research team of an LMF standardized

Arabic dictionary.

1 INTRODUCTION

Domain ontologies are “engineering artifacts”

describing a set of relevant domain-specific concepts

and their relationships in a formal way. Although the

area of ontology learning aiming to automate the

ontology creation process has been dealt with by

plenty of work, it is still a long way from being fully

automatic and deployable on a large scale (Cimiano

et al., 2009). This is essentially because it is a time-

consuming and difficult task that requires significant

human involvement for the validation of each step

throughout this process.

In order to reduce the costs, research on (semi-)

automatic ontology building from scratch has been

conducted using a variety of resources, such as raw

texts (Aussenac-Gilles et al., 2008), Machine-

Readable Dictionaries (MRDs) (Kurematsu et al.,

2004; Rigau et al., 1998), and thesauri (Chrisment et

al., 2008). Obviously, these resources have different

features, and therefore, each proposed process is

based on a different approach with respect to

principles, design criteria, NLP techniques, etc.

As linguistic information is increasingly required

in ontologies mainly in NLP applications (Buitelaar

et al., 2009), among the considered terminological

resources, MRDs represent one of the most likely

and suitable sources promoting the knowledge

extraction both at ontological and lexical levels.

However, since much information has not yet been

encoded, the access to the potential wealth of

information in dictionaries remains limited to

software applications.

Recently, Lexical Markup Framework (LMF)

(ISO 24613, 2008), which is a standard for the

representation and construction of lexical resources,

has been defined. Basically, its meta-model provides

a common and shared representation of lexical

objects that allows the encoding of rich linguistic

information, including morphological, syntactic, and

semantic aspects (Francopoulo and George, 2008).

It is in this context that we have proposed a new

approach that makes use of LMF standardized

dictionaries to generate domain ontologies (Baccar

396

Baccar Ben Amar F., Gargouri B. and Ben Hamadou A..

DOMAIN ONTOLOGY GENERATION USING LMF STANDARDIZED DICTIONARY STRUCTURE.

DOI: 10.5220/0003512803960401

In Proceedings of the 6th International Conference on Software and Database Technologies (ICSOFT-2011), pages 396-401

ISBN: 978-989-8425-77-5

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

et al., 2010). Such resource incorporates widely-

accepted and commonly-referenced diversified

linguistic knowledge lending itself to ontological

interpretations. Indeed, finely-structured knowledge

in an LMF-standardized dictionary paves the way

for the constitution of core domain ontology.

Furthermore, the abundance of texts available in the

definitions, explanations and examples are very

interesting to realize the core enrichment and above

all to provide the ontology elements with linguistic

grounding.

On the other hand, the nature of ontologies as

reference models for a domain requires a high

degree of quality of the respective model. Indeed,

several approaches have been considered in

literature in order to assess ontology construction

methodologies. However, a comprehensive and

consensual standard methodology seems to be out of

reach (Almeida, 2009). Yet, evaluating the ontology

as a whole is a costly and challenging task especially

when the reduction of human intervention is sought.

This can be deemed as a major impediment that may

elucidate the ontologies‟ failure not only to be

reused in others but also to be exploited in final

applications.

The ultimate objective of this paper is to show

the way in which an LMF-compliant MRD is

exploited for the core domain ontology generation.

In fact, its systematic organization allowed us to

implement a fully automatic and iterative process for

a direct dictionary transformation of some particular

information into ontological elements. Additionally,

the suggested process includes a validation phase in

order to preserve the quality of the produced

ontology throughout its development life cycle.

Furthermore, the proposed methodology has been

implemented in a rule-based system, whose high-

performance has proven to be trustworthy through

an experiment carried out on an Arabic dictionary

(Baccar et al., 2008).

The remainder of this paper is structured as

follows: Section 2 gives a related work overview of

ontology construction based on (semi-) structured

resources. Section 3 presents the proposed

methodology for generating the core domain

ontologies from LMF standardized dictionaries.

Section 4 presents the details of implementation. As

for Section 5, it is devoted to describe our

experimentation as well as to discuss the results

quality. Finally, Section 6 concludes the paper with

opening perspectives for future work.

2 RELATED WORK

Since ontology engineering has long been a tedious

task requiring considerable human involvement and

effort, many proposals have been suggested to

facilitate knowledge domain acquisition. It is also

widely recognized that taking into account relevant

resources from the very beginning of the ontology

development process yields more effective results.

Accordingly, recent approaches based on a variety

of structured or semi-structured resources, such as

XML documents (Aussenac-Gilles and Kamel,

2009), UML models (Na et al., 2006) and so on,

have been proposed with the aim of producing an

early stage of domain ontology through rule-based

learning techniques. The resulting primary

ontologies, also named core or kernel ontologies,

can help in the quick modeling of the domain

knowledge and could be further extended to obtain a

complete ontology.

Considered as a large repository of quasi-

structured knowledge about language and the world,

MRDs illustrate the building stone for the generation

of conceptual structures ranging from concept

hierarchies, thesauri to ontologies (Jannink, 1999).

Indeed, as noted in (Hirst, 2009), word senses can be

seen as the equivalent of ontological categories, and

lexical relations (e.g., synonymy, antonymy,

hyponymy, meronymy and so on) would correspond

to ontological relations (for example, hypernymy

would stand for subsumption). Nevertheless, since

the MRDs are oriented towards human reader, much

information is not well-structured and therefore its

machine interpretation might not be evident.

Consequently, systems relying on MRDs incorporate

two major problems, one of which is their need for

massive human intervention and the other is their

confinement to limited relations, in almost all cases,

the taxonomic ones.

From another standpoint, being a newly

emerging standard for the creation and use of

computational lexicons, LMF has recently been

defined (Francopoulo and Georges, 2008). Its meta-

model allows the representation of NLP and MRD

lexicons in a systematic organization. Indeed, such

model contains much explicit linguistic information

as well as a lot of implicit information included in

the definitions and examples.

After the introduction of LMF standard, a good

deal of active work, among which we can mention

LexInfo (Buitelaar et al., 2009), LIR (Montiel-

Pensoda et al., 2008) and (Pazienza and Stellato,

2006) models, has been undertaken in response to

the need of increasing the linguistic expressivity of

DOMAIN ONTOLOGY GENERATION USING LMF STANDARDIZED DICTIONARY STRUCTURE

397

given ontologies (Buitelaar et al., 2009). The

proposed models try to associate lexical information

with ontological entities, which is a heavy and time

consuming activity considering the plurality and the

heterogeneity of the resort sources. In addition, some

complexity rises when linguistic information is

involved in ontology reasoning (Ma et al., 2010).

3 METHODOLOGY FOR CORE

DOMAIN ONTOLOGY

GENERATION

3.1 Basic Idea

Thanks to its encompassing of both ontological and

lexical information, an LMF standardized dictionary

offers a very suitable primary knowledge resource to

learn domain ontologies (Baccar et al., 2010).

Accordingly, we have proposed an approach

consisting in firstly building the core of the target

ontology, taking advantage of the LMF standardized

dictionary structure. Secondly, it consists in

enriching such core starting from textual sources

with guided semantic fields available in the

definitions and the examples of lexical entries.

Within our context, the core provides all possible

sets of basic objects in a specific domain that could

be directly derived from systematic organization of

linguistic objects in an LMF standardized dictionary.

But before proceeding to the core building, we have

to create a dictionary fragment by extracting the

relevant part of the whole dictionary. It gathers

lexical entries of related senses to the domain of

interest as well as their semantically related words.

This dictionary fragment represents then the

privileged initial source for generating the target

ontology. Besides, when handling the obtained

ontology, conceptual nodes always keep reference to

lexical information included in the dictionary

fragment (henceforth dictionary).

In order to identify the concepts of a particular

domain, we consider the domain information given

in the dictionary by the SubjectField instances. Since

a concept corresponds to a meaning of a word, we

can directly deduce the concepts of the domain

ontology from particular instances (e.g., Context,

SenseExample) attached to the Sense class. With

regard to concepts properties, the generic LMF

meta-model allows for defining any type of

semantic relationship (e.g. synonymy, hypernymy,

meronymy) between the senses of two or several

lexical entries by means of the SenseRelation class.

Consequently, a relation that connects two or several

senses in the dictionary leads to an ontological

relation linking the corresponding concepts.

3.2 The Proposed Methodology

For the construction of core domain ontology, we

propose an automatic and incremental process. It

consists of three main stages (Figure 1). Firstly, we

identify candidates of concepts and relations relying

on some identification rules that we defined in

advance. Secondly, we check for duplicated

candidates by means of two lists of previously

identified and validated concepts and their

relationships. Finally, through a validation stage

supported by some validation rules, we check

whether the current change is still coherent after the

current change.

Figure 1: The core domain ontology generation process.

3.2.1 Concepts and Relations Identification

In this stage, we identify the concepts and their

relationships from the lexical entries in the LMF

standardized dictionary. According to our

investigation on LMF structure we managed to

define a set of identification rules allowing for the

elicitation of ontological entities. As a result of this

stage, we acquire all candidates of the core elements;

each one is assigned to a given signature. In the

present work, we formally define a signature of a

candidate concept and a candidate relation as follow:

Definition 1. A concept, denoted by C, is defined as

a couple, C = (N, S), where N is the name of C and S

denotes its binary tag whose value is equal either

“0” if C has no relation with other concepts, or “1”

if C is linked to another concept.

Definition 2. A relation, denoted by R, is defined by

a triplet, R = (N, CD, CR), where N is the name of

the relation, CD is the domain of R and CR is the

range of R.

ICSOFT 2011 - 6th International Conference on Software and Data Technologies

398

3.2.2 Duplication Check

As its name indicates, the goal of this stage is to

check for duplicated candidates. Such verification is

very important since the identification task may

imply several inter-related lexical entries per

iteration. In order to detect duplicated candidates,

the proposed process is based on two lists of

concepts and relations, which are initially empty.

They are intended to contain the concept as well as

the relation signatures, emanated from valid

introduction of new concepts and/or relations into

the resulting core. In addition, we distinguish three

types of duplication: exact duplication, quasi-exact

duplication and implicit duplication. Let £

and £

be the lists of concepts and relations, respectively.

Exact duplication. It refers to the identification of

the same copy of a previously identified candidate.

This type of duplication is denoted by the „ ≡ „

symbol which is formally stated as follow:

Let < C2 = (Name_2, 0) > be a concept candidate,

<C2 = (Name_2, 0) > ≡ <C1=(Name_1, ?) > Iff

< C1 = (Name_1, ?) > ϵ £

such that

(Name_1=Name_2)

Let < R

= (relation

, C1

, C2

) > be a relation

candidate and the concepts <C1

=(Name_1

,1) >

and <C2

=(Name_2

, 1) > be its two arguments,

< R

= (relation

, C1

, C2

) > ≡ < R

= (relation

, C2

) > Iff

< R

= (relation

, C1

, C2

) > ϵ £

;

 <C1

=(Name_1

, 1) > ϵ £

and

<C2

= (Name_2

, 1 )> ϵ £

such that

< C1

=(Name

, 1) > ≡ < C1

= (Name

, 1) > and

< C2

=(Name

, 1) > ≡ < C2

= (Name

, 1) > and

relation

= relation

Quasi-exact duplication. This duplication denoted

by the „„ symbol concerns only relation candidates.

A quasi-exact duplicated relation candidate might

not be identical to an already identified candidate

but it represents an equivalent. Formally,

Let < R

=(relation

, C1

, C2

)> be a relation

candidate and < C1

=(Name_1

, 1)> and < C2

(Name_2

, 1) > be its two arguments,

= (relation

, C1

, C2

) >  <R

= (relation

, C2

) > Iff

< R

= (relation

, C1

, C2

) > ϵ £

;

 < C1

= (Name_1

, 1) > ϵ £

and

< C2

= (Name_2

, 1) > ϵ £

such that

< C1

=(Name

, 1) > ≡ < C2

= (Name

, 1) > and

< C1

=(Name

, 1) > ≡ < C2

= (Name

, 1) > and

(relation

= relation

= relation) and

symmetric (relation)

To illustrate the quasi-exact duplication with a

concrete example, we consider the case of “married-

to” symmetric relationship, for instance R1 =

(married-to, Man, Woman). Hence, a candidate

relation with the R2 = (married-to, Woman, Man)

signature is considered as a quasi-exact duplicated

relationship and should be ignored.

Implicit duplication. It also concerns only relation

candidates. An implicit duplicated candidate is a

completely different candidate but whose knowledge

can be inferred from existing core elements.

Formally, let < R

= (relation

, C1, C2) > be a

relation candidate and < C1= (Name_1

, 1) > and

< C2 = (Name_2

, 1) > be its arguments,

< R

= (relation, C1, C2) > is an implicit duplicated

relation Iff

< C3 = (Name_3

, ?) > ϵ £

;

<R

=(relation, C2, C3)> ϵ £

and

=(relation, C3, C1

)> ϵ £

such that

(transitive (relation) ) and ( R

and R

 R

)

For example, if we have two identified relations,

R1 = (is-a, Dog, Pet) and R2 = (is-a, Pet, Animal),

then we can derive the R3 = (is-a, Dog, Animal)

relation candidate. Hence, a candidate with R3

signature is an implicit duplicated relationship that

must be removed from the relations list.

In all duplication types, the duplicated candidates

should be ignored. Therefore, the constructed core

domain ontology does not store unnecessary or

useless entities. This quality criterion is also called

conciseness (Gómez-Pérez, 2004).

It is worth mentioning that the final lists of

concepts and relations are very helpful not only for

the core construction, but also for its enrichment.

Particularly, in the enrichment task, we will consider

only the orphan concepts (i.e., whose binary tag is

equal to 0) in order to link them to either old or new

concepts. Indeed, the first stage of this process may

introduce a good number of concepts that are not

involved in any relations. Likewise, the list of

relations is needed for the enrichment stage so as to

check the coherence of the whole ontology.

Once duplication check is performed, a further

validation stage is required to verify whether the

resulting core remains coherent when the candidate

is added to it.

3.2.3 Validation Stage

The automatic addition of non-duplicated candidates

to the ontology core could bring about errors. In

order to maintain the coherence of the built core, the

DOMAIN ONTOLOGY GENERATION USING LMF STANDARDIZED DICTIONARY STRUCTURE

399

integration of a validation stage into the proposed

process is necessary. In other words, a concept or a

relation is automatically added to the output core

structure only when the latter is still coherent.

Gómez-Pérez has identified different kinds of errors

in taxonomies: inconsistency, incompleteness, and

redundancy errors (Gómez-Pérez, 2004).

Incompleteness Error. It occurs if the domain of

interest is not appropriately covered. Typically, an

ontology is incomplete if it does not include all

relevant concepts and their lexical representations.

Furthermore, partitions are incompletely defined if

knowledge about disjointness or exhaustiveness of a

partition is omitted.

Redundancy Error. It is a type of error that occurs

when redefining expressions that were already

explicitly defined or that can be inferred using other

definitions.

Inconsistency Error. This kind of error can be

classified in circularity errors, semantic

inconsistency errors, and partition errors.

 Circularity Errors. A circularity error is

identified, if a defined class in an ontology is a

specialization or generalization of itself. For

example, the concept Woman is a subclass of the

concept Person which is a subclass of the

concept Woman.

 Semantic Inconsistency Errors. It refers to an

incorrect semantic classification. For example,

the concept Car is a subclass of the concept

Person.

 Partition Errors. A class partition error occurs, if

a class is defined as a common subclass of

several classes of a disjoint partition. For

example, the concept Dog is a subclass of the

concepts Pet_Animal and Wild_Animal which

are disjoint subclasses of the concept Animal.

In the current stage, we are interested in kinds of

errors that can be automatically detected (i.e.,

without human expert involvement). Redundancy

verification has already been dealt with in the

second stage of this process. As for the completeness

assessment, it could not be done at this early stage of

domain ontology development. Therefore, only

inconsistency errors, particularly those of circularity

and partition types, are addressed in the present

work. After the check of the resulting core, we

proceed to the update of the concepts and relations

lists as well as the ontology core.

4 IMPLEMENTATION DETAILS

The methodology for core domain ontology

generation from LMF-standardized dictionaries is

implemented by a Java-based tool that enables users

to automatically build the core structures formalized

in OWL-DL, a sublanguage of OWL (Dean and

Schreiber, 2004). Indeed, an OWL-DL formalized

ontology can be interpreted according to description

logics, and DL-based reasoning software (e.g.,

RacerPro or Pellet) can be applied to check its

consistency or draw inferences from it. To take

advantage of this, we have decided to incorporate

the Pellet reasoner into our system. Indeed, it is an

open-source Java-based OWL-DL reasoner tool

(Sirin et al., 2007). Its consistency checking ensures

that an ontology does not contain any contradictory

facts. After the loading of the built OWL file, Pellet

determines if the ontology is actually consistent by

calling the isConsistent()method, whereby its

boolean return shall decide whether the addition

operation could be performed in the resulting core.

5 EXPERIMENTATION

AND EVALUATION

The assessment of the high performance of the

developed system as well as the good quality of the

obtained ontologies is shown through the experiment

carried out on the Arabic dictionary. The latter‟s

standard model (Baccar et al., 2008) and

experimental version has been worked out by our

research team. This dictionary is covering various

domains, of which animals, plants, astronomy and

sports are but a few. Besides, thanks to LMF meta-

model, our dictionary would certainly be an

extendable resource that could be incremented with

entries and lexical properties, extracted from other

sources (e.g., Arabic lexicons, text corpora).

As far as the evaluation of the obtained results is

concerned, we can obviously see that besides the

fully automated level, many important benefits are

noticeable in the proposed approach. Indeed, we can

firstly point out that all concepts and relations

represented in the core domain ontology are relevant

to the considered domain. In fact, the LMF-

standardized dictionaries are undeniably widely-

accepted and commonly-referenced resources;

thereby they simplify the task of labeling concepts

and relationships. Moreover, there is no ambiguity

insofar as we check the duplication of core ontology

elements before their construction. In addition, no

ICSOFT 2011 - 6th International Conference on Software and Data Technologies

400

inferred knowledge is explicitly represented. Finally,

there are no consistency errors since we managed to

check the coherence of the generated ontology with

a specialized tool. Furthermore, a series of statistical

studies were conducted on various domains toward

the comparison of the obtained core ontologies with

the corresponding handcrafted expected domain

ontologies. We found out that about 80% of all

concepts and 30% of all relations can be deduced

and formalized without human expert involvement.

6 CONCLUSIONS AND FUTURE

WORK

The main contribution of the current research work

is to propose a novel approach for the domain

ontology generation starting from an LMF

standardized dictionaries (ISO-24613). Firstly, it

consists in building an ontology core. Secondly, the

constructed core will be further enriched with

additional knowledge included in the text available

in the dictionary itself. The originality of this

approach lies in the use of a unique, finely-

structured source and rich in lexical as well as

conceptual knowledge.

Both qualitative and quantitative evaluations

have shown that the constructed core elements stand

for basic structures of a good quality, prone to be

further fleshed out with the additional information.

We expect to at the end create rich and valuable

semantic resources that are suitable for NLP tasks.

The next challenges deal with how to exploit the

wealth of information in the handled dictionary and

preserve in the same time the good quality of

yielding ontologies. Indeed, although systematic

organization provided by LMF structure, much

implicit information still needs to be analyzed

toward digging out more ontological knowledge.

That is why, ongoing work deals with the

investigation on words bearing other relationship to

the dictionary entry. We also plan to support the

enrichment mechanism with rules maintaining the

coherence of domain ontologies throughout their

construction process.

REFERENCES

Almeida, M. B., 2009. A proposal to evaluate ontology

content. Applied Ontology, 245-265.

Aussenac-Gilles, N., Despres, S., Szulman, S., 2008. The

TERMINAE Method and Platform for Ontology

Engineering from texts. Bridging the Gap between

Text and Knowledge. IOS Press, 199–223.

Aussenac-Gilles, N., Kamel, M., 2009. Ontology Learning

by Analyzing XML Document Structure and Content.

In KEOD’09, 159-165.

Baccar, Ben Amar, F., Khemakhem, Gargouri, B., Haddar,

K., Ben Hamadou, A., 2008. LMF standardized model

for the editorial electronic dictionaries of Arabic.

NLPCS’2008, Barcelona, Spain, 64-73.

Baccar, Ben Amar, F., Gargouri, B., Ben Hamadou, A.,

2010. Towards Generation of Domain Ontology from

LMF Standardized Dictionaries.SEKE 2010, Redwood

City, San Francisco Bay, USA, 515-520.

Buitelaar, P., Cimiano, P., Haase, P., Sintek, M., 2009.

Towards Linguistically Grounded Ontologies.

ESWC2009, Heraklion, Greece.

Chrisment C., Haemmerlé, O., Hernandez N., Mothe J.,

2008. Méthodologie de transformation d'un thesaurus

en une ontologie de domaine. Revue d'Intelligence

Artificielle 22(1): 7-37.

P. Cimiano, A. Mädche, S. Staab, J. Völker, 2009.

Ontology Learning. In: S. Staab & R. Studer.

Handbook on Ontologies. Springer.

Dean, M., Schreiber, G., 2004. OWL Web Ontology

Language reference. W3C recommendation, W3C.

Francopoulo, G., George, M., 2008. Language Resource

Management-Lexical Markup Framework (LMF).

Technical report, ISO/TC37/SC4 (N330 Rev.16).

Gómez-Pérez, A. 2004, Ontology evaluation. In Staab, S.,

Studer, R. (eds.), International Handbooks on

Information Systems.

Hirst, G., 2009. Ontology and the Lexicon. In: S. Staab &

R. Studer. Handbook on Ontologies. Springer.

ISO 24613. Lexical Markup Framework (LMF) revision

16. ISO FDIS 24613:2008.

Jannink, J., 1999. Thesaurus entry extraction from an on-

line dictionary. In Proceedings of Fusion ’99.

Kurematsu, M., Iwade, T., Nakaya, N., Yamaguchi, T.,

2004. DODDLE II: A Domain Ontology Development

Environment Using a MRD and Text Corpus.

IEICE(E) E87-D(4) 908-916.

Ma, Y., Audibert, L., Nazarenko, A., 2010. Formal

Description of Resources for Ontology-based

Semantic Annotation. In LREC 2010, Malta.

Montiel-Ponsoda, E., Peters, W., Auguado, de Cea, G.,

Espinoza, M., Gómez Pérez, A., Sini, M., 2008.

Multilingual and localization support for ontologies.

Technical report, D2.4.2 Neon Project Deliverable.

Na, H.-S., Choi, O.-H., Lim, J.-E., 2006. A Method for

Building Domain Ontologies based on the

Transformation of UML Models. (SERA'06), 332-338.

Pazienza, M. T., Stellato, A., 2006. Exploiting Linguistic

Resources for building linguistically motivated

ontologies in the Semantic Web. (OntoLex2006).

Rigau, G., Rodríguez, H., Agirre, E., 1998. Building

accurate semantic taxonomies from monolingual

MRDs. COLING-ACL’98, Montreal, Canada.

Sirin E., Parsia B., Cuenca Grau B., Kalyanpur A., Katz

Y., 2007. Pellet: A practical OWL DL reasoner.

Journal of Web Semantics, 5(2):51-53.

DOMAIN ONTOLOGY GENERATION USING LMF STANDARDIZED DICTIONARY STRUCTURE

401