ONTOLOGY-BASED FRAMEWORK FOR

DOCUMENT INDEXING

Djida Bahloul, Youssef Amghar, Pierre Maret

INSA de Lyon – LIRIS – FRE 2672. 7 avenue Jean Capelle. 69621, Villeurbanne – France.

ywords: Semantic indexing, Ontology, Knowledge modelling

Abstract: The work presented in this paper addresses a project of the Computer Centre CIRTIL who supported it. This

company wants to save and capitalize its knowledge and its know-how about the production activities,

especially concerning the technical incidents relating to software applications encountered during the

exploitation time. Indeed using a well accessing documents base, actors will be able to better solve

problems. Our purpose is to focus on ontology-based framework for indexing documents. The domain

ontology OntoCIRTIL has a structure which supports a semantic model based on semantic links and

inference mechanisms. In this paper, we present a new model called S

which permits to model knowledge

in upstream and index documents (or formalized knowledge) in downstream. To illustrate partial results,

this model is then applied to OntoCIRTIL.

1 INTRODUCTION

Professionals of library sciences (i.e. librarians)

accomplish indexing tasks in using indexing

resources as a controlled vocabulary (thesaurus).

The main characteristic of indexing resources

resides in the ability to offer a way to carry out a

relevant retrieval of document contents. Thus, it

contains descriptors that are the keywords describing

objects and their inter-relationships.

In this paper, we propose an indexing approach,

based on ontologies, and which is more efficient

than a simple use of taxonomy of concepts.

Therefore, we build a domain ontology which has a

significant capacity of expression thanks to the

possibility of introducing semantic links, structural

links and subsumption links. Compared to

descriptors, concepts of ontology, are used first, to

represent knowledge and then, to query the system

in order to access and retrieve these knowledge.

We discuss the role of ontologies on the

corporate memory building in section 2. Section 3

presents the environment of our work. Section 4

gives the description of our model called S

for

representing and indexing knowledge. This model is

applied to our domain ontology OntoCIRTIL.

2 ONTOLOGY, KNOWLEDGE

REPRESENTATION

Knowledge is a combination of information and

contexts required to perform a task by selecting,

interpreting and evaluating information according on

context of use (Weggeman, 1996). Ontologies are

used as a coherent support to describe and to share

knowledge. “Ontologies open the way to move from

a document oriented view of knowledge management

to a content-oriented view, where knowledge items

are interlinked, combined, and used.» (Staab & al.,

2001). Projects SHOE (Heflin & all, 2000) and

Ontobroker (Benjamins & all, 1998) use ontologies

to improve the searching abilities on the web. Both

systems are logical reasoning based on ontological

definitions. In SHOE ontologies are taxonomy

hierarchies queried by users. A web page on the

Internet can reference any ontology and exploit

definitions using tags. Ontologies are used more and

more in Knowledge Management System

development (Van Heijst & all, 1997). They

improve the knowledge engineering process.

2.1 Annotation with ontology

Ontologies can be used to improve information

retrieval through annotations of the resources

constituting a corporate memory (CM). Most recent

269

Bahloul D., Amghar Y. and Maret P. (2004).

ONTOLOGY-BASED FRAMEWORK FOR DOCUMENT INDEXING.

In Proceedings of the Sixth International Conference on Enterprise Information Systems, pages 269-274

DOI: 10.5220/0002606302690274

 SciTePress

approaches employ ontologies as domain-specific

vocabularies and concept structures according to

explicitly specified conceptualisations. This

approach extends traditional metadata technologies,

such as taxonomies and thesauri, by additional

relationships, axioms and general logical constraints

to allow reasoning. The CoMMA project (Gandon &

al, 2002) offers a solution to implement a CM based

on ontologies and on agent technology. It promotes a

wide vision of the document retrieval issue. The CM

is composed of heterogeneous evolving documents,

structured using semantic annotations expressed

with concepts and relationships provided by a shared

ontology. The approach for ontology-based

knowledge management (Staab & al., 2001) includes

a tool suite and a methodology for developing

knowledge management systems. OntoAnnotate

allows users to create objects and describe them with

their attributes and relationships.

2.2 Ontology as an indexing resource

Thesaurus and ontology provide common properties

such as the organization of terminology for covering

of a broad range of terminology used in a particular

domain, the use of hierarchical structure (terms are

grouped into categories and subcategories).

According to (Saadani & al, 2000) the more

important differences is the informality and

ambiguity of relationships in a thesaurus. Ontology

introduces a host of structural and conceptual

relationships including superclass/subclass/instance

relationships, property values time relationships, and

others depending on the representation language.

Generally, ontology contains far more relationships,

which are formally defined and unambiguous. The

ontology can reason about the meaning of concepts

by comparing logical concept structures. An

example is given in (Desmontils & al, 2002) in

indexing web sites with a terminology-based

ontology. When concept C2 satisfies the requirement

of being a specialization of concept C1, then C2 can

automatically be classified below C1. This gives rise

to query processing and searching which is not

possible with a thesaurus.

2.3 Domain ontology: OntoCIRTIL

The domain we consider addresses the call centre

activity of CIRTIL, a Computer Centre which has

the following missions: i) management of the

Information System of the Covering Branch; ii)

setting in production and exploitation of software

applications; iii) realization and/or contribution to

the national and regional projects; iv) technical and

functional help to customers. A first work allowed

us to build OntoCIRTIL that intends to model

technical and functional incidents, actors who treat

these incidents, applications concerned by these

incidents and the entities characterising these

incidents, applications and actors. Main purposes of

OntoCIRTIL are to represent knowledge, to permit

indexing documents and consequently feeding CM

and to reuse and share capitalized knowledge.

Several approaches to ontologies development

have been proposed. To build OntoCIRTIL, we were

inspired by methodologies proposed in (Uschold &

all, 1996), (Guarino, 1995) and (Fernández-López &

al., 1999). We chose a middle-out approach. Indeed,

Uschold argues that this approach is most effective

when the basic concepts in a domain are identified

first (e.g., Employee), and later generalised (Person)

and/or specialised (Secretary).

INCIDENT

using

Incident_name (string)

Incident_state (integer)

severeDegree (integer)

frequency (integer)

DOCUMENT

RESOURCE

COMMENT

CARD

INCIDENT

EVENT

APPLICATION

SOLVING

RESOURCE

concerns

described_in

EMPLOYEE

CUSTOMER

MONITORING

TOOL

Is-a link

Semantic link

attribute

treated_by

reported_by

rised_by

ACTOR

necessitates

INCIDENT

FAMILY

composition link

PROCESS

TIMES

occurs

impacts

TECHNICAL

INCIDENT

FUNCTINAL

INCIDENT

Figure 1: OntoCIRTIL SNF.

ICEIS 2004 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS

270

OntoCIRTIL is presented in semantic network

formalism (SNF). A semantic network is basically a

directed graph with nodes representing concepts

from the discourse area and edges expressing

relationships between these concepts. A concept is

defined in a single way. Knowledge relating to a

concept is factorized on the level of the node. This

knowledge is expressed by attributes (property), and

by relationships. For example, attributes of the node

"Incident" are: Incident Name, Incident State,

Gravity Degree and Frequency. Relationships linked

to "Incident" are "Incident concerns Product" or

"Incident treated-by Employee".

3 USER’S ENVIRONMENT

Our work contributes to the elaboration of a CM

environment. This environment is directed towards

several user types: experts of field, actors (end-users

of knowledge) and administrators. It allows indexing

and integration of knowledge in the CM as well as

querying of heterogeneous information sources.

These tasks are classified through two processes

(capitalization and restitution) which are related with

several databases. Hereafter, we describe these tasks

and define the role of databases. In this paper, only

dark components of Figure 2 are detailed.

The capitalization process consists of the

capture, the treatment and the integration of

knowledge in the CM. Knowledge is made explicit

and reusable. Each actor can participate to this

process according to access rights. Main tasks are:

Creation: Achieved directly by actors and can be

result of the drafting of business documents,

technical reports, actors’ experiences, etc.

Knowledge must be created or converted regarding

the conventions of the company.

Dematerialization: Knowledge dematerialization

is an operation that consists of translating knowledge

issued from documents or actors experience into a

computational form. This allows a huge amount of

knowledge to be stored and retrieved.

Formalization: XML and RDF(S) are used for

describing syntax and semantics of semi-structured

information sources. RDF provides a simple data

model for representing formal semantics of

information, i.e. meta-information. RDF Schema

defines a simple ontology modelling language on top

of RDF that can be used to define vocabulary and

structure of meta information.

Indexing: Indexing techniques allow to create

and to describe objects with their attributes and

relationships. Objects are knowledge items found on

web pages, in spreadsheets, or in text documents.

This task updates the knowledge database (§3.3).

The Restitution process consists of extracting

data, documents or document fragments for end-

users. Hyper-navigation and interfaces contribute to

this process. Two modes of restitution are proposed:

navigation through taxonomy of ontology and

interrogation based on a search engine.

Process of restitution (Inference Engine)

- Retrieval

- Querying

- Browsing & displaying

Documents

database

Data

warehouse

Knowledge

Base

Process of capitalization (Indexing Engine)

- Creation (e.g. drafting)

- Dematerialization (e.g. tools of

electronic management of documents)

- Formalisation (e.g. XML standards:

XLL, XSL, RDF, …)

- Indexing and classification

Semantic

models

Interface

models

indexing

Production

databases

Documents realted to

processes activities

Information Sources

- Documents (web

pages, document,

diagram, …);

- Databases;

- Quality Procedures;

Actors' company

- Experts (know-how,

experiences, …)

- Actors' daily tasks

Users

Decision-maker,

Employees, …

matching

Figure 2: Corporate Memory Environment dedicated to Actors' Company.

ONTOLOGY-BASED FRAMEWORK FOR DOCUMENT INDEXING

271

Incident

Family

Event

Actor

Resolution

Is-a

Cause

Component-of

necessitate

reported-by

treated-by

Date

Hour

Occur-at

INCIDENT

Figure 3: Semantic interpretations of concepts

Retrieval: Main function allowing users to locate

knowledge, search relevant documents. This module

is the kernel of inference engine.

Navigation: Actors may explore the CM without

asking a particular question. They exploit the

ontology for navigation purposes. Also, the ontology

allows deriving additional links and descriptions.

Querying: Is a treatment of queries formulated

by end-users through a high level language. The

query language has to be able to query

heterogeneous databases. For instance, we can use

XQuery or OQL to query XML-based CM.

Browsing & displaying: Concerns documents

(‘fragments) as results of queries and presentation to

end-user. User profiles and auto-adaptive interfaces

are used for improving access to knowledge.

Roles of databases: A CM is built around many

databases. Each database contains data, models and

rules that ensure the storage and the treatment of

knowledge. We briefly present these databases.

Semantic model: It describes how knowledge is

related semantically into the CM. It includes

relationships models, properties corresponding to the

semantic of each relationship (type of relation,

properties of combination with the other relations in

the same context) and inferences models. Also, it

permits to ensure the coherence supported by

required conditions of database management (rules

of priority, rules of incompatibility). The semantic

model is considered at metalevel which allows

cognitive representations of semantics primitives.

Knowledge base: It contains representations

(formal) of concepts and relationships of the domain

(here: the management of the technical hitches). It

contains also the metadata of documentary

descriptions (Dublin Core).

Documents base: This database contains

structured documents supported by XML standard.

The basic structure of an XML-document is given

by the hierarchically nested elements, thus it is a

natural approach to model it as a tree-graph.

Data warehouse: A data warehouse is a central

repository for all or significant parts of the data that

an enterprise's various business systems collect. It is

enriches from one or more production databases.

Interface models: Several types of models may

be useful to represent interfaces and communication

setting process. These models can help users to

navigate between the nodes of the ontology. A

variety of models have been proposed to help

retrieve knowledge.

4 THE MODEL S

This paper focuses on capitalization process and

requirements for a CM. For that we propose an

approach to model knowledge aiming to enhance

performances of retrieval systems. Before presenting

the model S

, we briefly expose some motivation

which guided us to elaborate such a model.

Relationships of ontology define and enrich the

semantic between the concepts. In certain cases, the

type of relationship can change the semantic. Thus,

two concepts C1 and C2 linked by two distinct

relationships R1 and R2 produce two semantics.

Let us consider two concepts “Incident” and

“Actor” and two relationships “reported-by” and

“treated-by”. Semantic can be variable according to

the relationship employed between each of these two

concepts. Reported-by(Incident,Actor) means that a

given incident has been notified by an actor during

his task; and Treated_by(Incident, Actor) means that

an actor has resolved the incident.

In addition, several types of relationships offer

various viewpoints on one concept. In fact, figure 3

shows that the concept "Incident" has many

significances, defined through relationships. Such

explanations are necessary to define all variety

semantic of knowledge considered by end-users.

To elaborate such indexing model, we build a

domain ontology allowing modelling knowledge

according to various facets and viewpoints. Our

domain comprises the range of knowledge

associated with processes of activity domain of

CIRTIL Company. In this section, we present the

principles and the components of model S

The OntoCIRTIL aims first to represent domain

knowledge and then, to be used like a resource of

indexing. To achieve this goal, we take advantage of

conceptual relationships when we built OntoCIRTIL.

Although the ontological relationships are

numerous, it is difficult to give an exhaustive list of

it. This can be explained by the fact that relations in

ontologies are not standardized like those in

thesaurus. Nevertheless, the model proposed in this

paper, called S

, is based on three views i.e. three

spaces for organizing knowledge fragments

ICEIS 2004 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS

272

represented by concept of domain ontology:

semantic, structural and subsumption spaces.

Semantic space: This space gives a view

allowing discovering concepts across semantic links.

This space is domain dependent because

relationships used are defined and interpreted by

users. Our environment offers primitives to define

such relationships such as: create, delete, and

rename semantic links. The graphs path algorithms

make possible to navigate in this space in order to

discover concepts: to displace from a concept to

another according to the semantic link defined by

the expert. This space embeds semantic links.

Structural space: This space takes in account

structural dimension of the knowledge concepts and

links with document fragments. These fragments are

represented by concepts and organized according to

aggregation links. Besides the graphs path

algorithms, it is interesting to define a zooming

function for discovering various levels of the

selected concepts. This function allows exploring

deeply the various parts of one concept. This space

allows representing structural links.

Subsumption space: This space makes it

possible to organize the definitions of concepts by

using subsumption links. This space increases

semantic space by using a new dimension which

allows improving concepts description in term of

related definitions. The organization of knowledge

in this space allows actors to retrieve concepts (and

related documents) with definitions (which concerns

ancestors in the subsumption graph).

4.1 Description of link types of S

The concept is the core of our model: concepts are

objects (or fragments) of knowledge. Each concept

is linked with others concepts by at least a type of

links (Semantic, Structual, Subsumption link). Each

type of relationships gives a particular semantic

between two concepts. To model these relationships,

our approach is based on manual linguistic analysis.

We applied a same method to define the concepts, to

choose the representative links, i.e. we proposed

several lists which were modified then validated by

actors according to the ontological commitment.

Semantic links. The names links are related to

the usual language (natural language) of the

community for which ontology is available. Several

approaches are used for naming links: verbs or

prepositions (Sherratt & all, 1990); verbs or nouns

(Heeren & all, 1993); verbs or logical connectors

(Malone & all, 1984). This facilitates their use in

particular to retrieve knowledge in the CM because,

these relationships are considered as keywords.

Semantic relationships express clearly "evident" and

no ambiguous knowledge. Two types of semantic

links are illustrated in the following examples.

Example 1: Impact link (Incident, Process)

In this representation the type of semantic link is

called Logical Link. Generally, the definition of

logical link relates to a description of a true logic in

the real world. The link between the two concepts:

Incident and Process allows informing the actor (i.e.

technician) to take new disposition. Instantiation

example is: a server failure as an incident, which

impacts the process of exploitation of software

located on workstations. Examples of logical links

used in our ontology are: cause, necessitate,

implicate, etc.

Example 2: Occur link (Incident, Data-Hour)

Another type of semantic link is a temporal link.

In ontology, the concept of time can be even defined

at concept level. However, it can be better clarified

by a link between two concepts. It is the case of link

"to occur" established between Incident and Date-

Hour. This representation indicates that any incident

can occur at a given moment. Other links such as:

treated-by, written, organized, etc. compose the

relationships employed in our domain ontology.

Note that relationships allow the bi-directional and

opposite (inverse) semantic expression.

Structural links They express a strong property

between the whole and the parts, as well as

subordination between the parts and the whole. The

parts can be created after the composite itself, but

once created, they "lives" and "dies" with him (i.e.

they share its duration life). Parts can also be

explicitly withdrawn before the death of the

composite. Composition can be recursive.

Subsumption links. They are largely used and

are considered as the foundation of ontology,

because all ontologies are presented in taxonomy.

Subsumption allows information about concepts to

be associated with their most general concept, and it

allows information to filter down to more specific

concepts in the taxonomy via inheritance.

In addition our indexing model proposes a

mechanism which allows deducting new knowledge

in order to enrich a semantic. In some cases,

knowledge is also defined with rules: Chief Project

is a person who manages project. These rules permit

to express implicit useful knowledge of information

retrieval. We write constraints that control semantic

relationships between concepts in order to support

reasoning. For example, if “Incident” necessitates

“Resolution” and if this “Incident” is treated-by

“Employee” then a link is inferred: “Employee”

proposes a solution.

ONTOLOGY-BASED FRAMEWORK FOR DOCUMENT INDEXING

273

4.2 Indexing tool

We developed a tool for hitches management called

MaTIP (Management of the Technical Incident

Project). The objective is to capitalize data,

information and knowledge allowing the

identification, the management and the anticipation

of the dysfunction and technical anomalies at the

exploitation time of applications. This Knowledge

Base contains OntoCIRITIL. It groups the concepts

and relationships identified during the

conceptualisation process. The employees can

modify, enrich and validate the ontology. One of the

objectives of our contribution is to integrate the

indexing operation into the daily activities of the

actors. To achieve this goal, we take into account

that users hardly change their practices.

KnowIndexe is a simple application that actors

can use easily to index or to retrieve formalized

knowledge from the CM. The indexing technique is

achieved through the ontology considered as an

indexing resource. The indexing mechanism

comprise three steps: Selection of knowledge

(document or fragment) in the usual environment of

actors; Selection of representative' concepts in the

ontology describing the selected knowledge;

Indexing in generating a correspondence between

concepts and knowledge.

5 CONCLUSION

In this paper we have presented a semantic model

based on ontology of domain intended to index

technical documents in the context of the CM. We

presented first an environment dedicated to actor's

company, as a framework for CM development. We

outlined the particularities of the domain ontology

built for this objective. Then, we presented the

model S

and its components. We explained that the

ontological relationships allow a strong semantic. In

this context, we proposed three link types. The first

experimentation applied to a project of CM permits

first, to expose real needs and then to test and

validate our approach with the KnowIndex indexing

tool. The interest of our contribution is to develop an

indexing model which exploits the ontological

relationships. The application of the model to a

small corpus showed that the approach is time-

consuming in particular when the ontology must be

built. Nevertheless, the implementation of the

structural space gave good results for users.

REFERENCES

Benjamins, V. R. and Fensel, D. The ontological

engineering initiative (KA)2. In Guarino, N., editor,

Formal Ontology in Information Systems, pages 287–

301. IOS Press.

Desmontils.E, Jacquin.C. "Indexing a Web Site with a

Terminology Oriented Ontology", The emerging

semantic web, I.F. Cruz, S. Decker, J. Euzenat and D.

L. McGuinness Ed., IOS Press, Amsterdam, pp. 181-

197, 2002, ISBN 1-58603-255-0.

Fernández, M.; Gómez-Pérez, A.; Pazos, J.; Pazos, A.

Building a Chemical Ontology using methontology

and the Ontology Design. IEEE Intelligent Systems

and their applications. #4 (1):37-45. 1999.

Gandon.F, Dieng-Kuntz.R, Corby.O, Giboin.A. Semantic

Web and Multi-Agents Approach to Corporate

Memory Management. 17th IFIP World Computer

Congress IIP, Eds Musen M., Neumann B., Studer R.,

p. 103-115, August 25-30, 2002, Montréal, Canada.

Guarino, Nicola. 1995. Formal Ontology, Conceptual

Analysis and Knowledge Representation, I. J. of

Human-Computer Studies, 43, 625–640.

Heeren, E. and Collis, B. Design considerations for

telecommunications-supported cooperative learning

environments: concept mapping as a telecooperation

suppport tool. J. of educational multimedia and

hypermedia, 4(2), 1993, 107-127.

Heflin, J. and Hendler, J. Semantic interoperability on the

web. In Extreme Markup Languages 2000.

Malone, J. and Dekkers, J. The concept map as an aid to

instruction in science and mathematics. School science

and mathematics, 84(3), 1984, 220-232.

Saadani.L, Bertrand-Gastaldy.S. Conceptual Maps and

Thesauri: A Comparison of Two Models of

Representation from Different Disciplinary Traditions

CAIS 2000. Canadian Association for Information

Science Proceedings of the 28th Annual Conference.

Sherratt, C. S. and Schlabach, M. L. The application of

concept mapping in reference and information

services. RQ, 30, 1990, 60-69.

Staab S., Schnurr H.-P., Studer R., Sure Y.(2001):

Knowledge Processes and Ontologies. IEEE

Intelligent Systems. 16(1), Special Issue on

Knowledge Management, January/February 2001.

Uschold M, Gruninger M. Ontologies: principles, methods

and applications. Knowledge Engineering Review

1996;11(2).

Van Heijst, G., Schreiber, A.T., Wielinga, B.J. Using

explicit ontologies in KBS development. I. J. of

Human-Computer Studies, 46 (2/3), 1997, pp 183-292.

Weggeman, M. Knowledge management: The modus

operandi for a learning organization. In J.F.

Schreinemakers, editor, Knowledge Management—

Organization, Competence and Methodology, pages

175–187. Ergon Verlag,W¨urzburg, D, 1996.

ICEIS 2004 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS

274