SADIM: AN AID SYSTEM FOR MANAGEMENT ENGINEERING

DIAGNOSIS USING KNOWLEDGE EXTRACTION AND

MATCHING TECHNIQUES

Jamel Kolsi

LOGIQ – LARIS Laboratory, Sfax- TUNISIA

Lamia Hadrich Belguith, Mansour Mrabet, Abdelmajid Ben Hamadou

Faculty of Management and Economic Sciences, LARIS Laboratory, Sfax- TUNISIA

Keywords: Engineering management, dysfunction, term extraction, knowledge acquisition, witness sentences, key

ideas, socio-economic diagnosis.

Abstract: This paper describes an aid system of management engineering diagnosis "Système d’Aide au Diagnostic

d’Ingénierie de Management " SADIM, the aim of which is to detect the dysfunctions related to the

enterprise management. This system allows the acquisition of knowledge based on textual data (given in

French) related to the diagnosis, the matching and the assignment of witness sentences to the key ideas that

correspond to them. SADIM can also serve as a part of a decision aid system as it includes carrying out

diagnosis which can helps experts and socio-economic management consultants to take decisions that would

make enterprises reach the required standards through council interventions.

1 INTRODUCTION

Looking for a better industrial productivity is the

major concern of the organization in a context

where priority is given to the decision making,

reduction of the production cycle time, flexibility for

risk facing, best quality, etc.

One of the means used to reach these objectives

is the application of the management engineering

process. The latter is applied in enterprises and

organizations in four steps namely the socio-

economic diagnosis, socio-economic innovation

project, implementation and result evaluation (Savall

and Zardet, 1989).

Our assignment is located at the level of

diagnosis. The data relative to this diagnosis are

formed of witness sentences fitting key ideas. These

data can hide useful knowledge, dependences or

inter-relations.

The socio-economic expert conducts semi-

directive interviews with executives, mastery agents,

workers… These interviews will represent the

dysfunctions in the form of witness sentences.

Given the important number of collected witness

sentences, the expert finds it hard to synthesize these

sentences into key ideas. This synthesis can be made

easier if the expert starts from a basis of key ideas

that she/he has collected through several diagnoses.

The automatic tools proposed in this domain are

characterized by the non-automation synthesis of

witness sentences into key ideas. This is the case of

SEGESE system (SocioEconomic Management

Expert System) (Savall and Zardet, 2004). This

system presents a problem of key idea redundancy.

The problem is related to the significance of the key

ideas rather than to the way they are formulated.

This situation is due to a difficulty that the expert

meets in the research of the dysfunction key ideas

that correspond to the witness sentences. This

situation incites the expert to insert other key ideas.

By adopting an extraction approach and an

automatic manipulation of textual data relative to the

management engineering diagnosis and in order to

solve the problems of SEGESE, we propose a

system baptized SADIM.

In what follows, we present a brief overview of

previous works on knowledge extraction, then we

propose our method of aid for management

engineering diagnosis and finally we expose an

assessment of our method.

331

Kolsi J., Hadrich Belguith L., Mrabet M. and Ben Hamadou A. (2006).

SADIM: AN AID SYSTEM FOR MANAGEMENT ENGINEERING DIAGNOSIS USING KNOWLEDGE EXTRACTION AND MATCHING TECHNIQUES.

In Proceedings of the Eighth International Conference on Enterprise Information Systems - AIDSS, pages 331-334

DOI: 10.5220/0002488003310334

 SciTePress

2 A BRIEF OVERVIEW OF

WORKS RELATED TO

KNOWLEDGE EXTRACTION

The work related to automatic knowledge extraction

could be classified in two basic categories: the

methods of terminological extraction and the

methods of knowledge acquisition.

2.1 Terminological Extraction

Methods

We could distinguish three main approaches of

automatic terminological extraction : the structural

approaches, the non structural approaches and the

mixed approaches (Chevallet, 2003).

The structural approaches use two kinds of

techniques: some use the syntactic and lexical rules

and often require grammars (TERMINO tool of

(David & al, 1990)), the others adopt surface

analysis and terms contexts (LEXTER tool of

(Bourigault, 1994)) and often use the recognition

syntactic patterns.

The non-structural approaches use some

statistical and quantitative methods (MANTEX tool

of (Oueslati, 1999)).

The mixed approaches (SORT tool of (Daille,

2002)) use the two previously described approaches.

The outcoming terms of these terminological

extraction approaches can be used by tools of

knowledge acquirement. We describe some of them

in what follows.

2.2 Knowledge Acquisition Methods

Knowledge acquisition methods use different

techniques such as:

• Techniques of relation acquirement between

terms (TERM tool of (Oueslati, 1999)).

• Techniques using rules (SEEK tool of (Jouis,

1995)).

• Techniques using lexico-semantic patterns

(Oueslati, 1999).

• Techniques using templates (PALKA tool of

(Kim & Moldavan, 1993))

3 PROPOSITION OF AN AID

METHOD FOR THE

DIAGNOSIS

To elaborate our method, we have exploited the

principles of morphological analysis technique of

treatment (Daille & al, 2002):

• Abréviations: "Sté, SGBD, MEO,PC,..".

• Inflectional paradigms: I work, she

works,…the lemma is work.

• Derivational paradigms: "nation, nationalisé,

nationaliser" the root is "nation".

• Suffixations: assembler/assemblage

exécuter/exécution".

• Préfixations: "faire/défaire, faire/refaire".

• Compound noun: "mise en oeuvre, mise à

niveau", company manager, resource

management, business structure, company

structure.

We have also used some techniques of ontology to

define semantic relations as (Amarnath, 2003):

• "Sorte de", (kind of): join heteronyms to

hyponyms (computer material / printer).

• "Partie de", (Part of): join an element to a

whole (diagnosis/ management engineering).

• "Action/Objet", (Action/Object) :

(crisis/economy).

• "Objet/propriété",(Object/property) :

(inflation/rate).

• "Objet/procédé", (Object/process) :

(enterprise/state to rank).

• "Relation causale", (causal Relation) :

(dysfonctionnement/deficit).

4 SADIM SYSTEM

In order to test our method, we developed an aid

system for management engineering diagnosis

(SADIM) to create and update a data of knowledge

basis which essentially includes witness sentences

and key ideas.

This system involves five steps: pre-treatment of

witness sentences and key ideas, extraction of the

simple terms and the compound terms in

witness

sentences, validation of the simple terms and the

compound terms of the key ideas, matching the key

ideas with the witness sentences, classification of

key ideas with reference to each witness sentence

and linking the witness sentence to key idea. (see

the following figure)(Kolsi & al, 2005).

ICEIS 2006 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS

332

Figure 1: SADIM General architecture.

Step 1: Pre-treatment of witness sentences and

key ideas

In this step, the system performs the following tasks:

• Deleting the separators: The system replaces all

the following separators (:; . ? ! ( ) [] {} =" * +...)

except the hyphen (-) by a space since it can exist

at the level of compound terms. The objective is

to have only one separator (the white space).

• Deleting the empty words: The empty words do

not have a semantic content; their function is to

structure the speech by a correct syntactic form

that is the case of tool words. Thus we delete all

empty words in witness sentences and in key

ideas, with the exception of ("de, en, d’,à ) since

the latter exist in some compound nouns.

• Deleting the multiple spaces: To unify the word

separators by only one space, we replace all the

existing multiple spaces in the witness sentences

and in the key ideas by only one space.

Step 2: Extraction of simple terms and compound

terms of witness sentences

This step consists in the extraction of simple terms

and compound terms of witness sentences by using a

Key Words Dictionary (KWD). This dictionary

contains 950 words; each of them is represented

with all its derivatives, its synonyms and the words

of the same class.

In this step, the system loads all the words of the

witness sentence in a table and detect all possible

compound terms (i.e. composed of three words or

two words) and simple terms (i.e. composed of one

word).

Stage 3: Validation of the simple terms and

compound terms of the key ideas

While seizing the key ideas, the expert introduces

for each key idea, its corresponding simple and

compound terms. In this stage, the system does the

following treatments for each key idea:

 The extraction of the simple and compound

terms (the same way the treatment is carried out

in step 2).

 Validation of the previously mentioned terms.

Stage 4: matching of the key ideas with each of

the witness sentence

This stage consists in the elaboration of a statistical

table that includes some statistical data on the

similarity, the synonymy and the adherence to the

same class between words of a witness sentences

and those of key ideas. These statistical data are

going to be used to carry out the matching between

each of the witness sentences and the key ideas.

This treatment can be classified as follows:

• Similarity treatment: The proposed system

calculates the number of similar words in both

witness sentences and key ideas taking into

account the morphological variations. The

results of this treatment will be stored in a

statistical table.

• Synonymy treatment: This treatment provides as

a result the number of synonymous words in

witness sentences and key ideas. The results of

this treatment will be stored in a statistical table.

• Adherence to the same class treatment: This

treatment provides in the same way as an

outcome the number of terms of the same class

between the witness sentences and the key ideas.

The results of this treatment will be stored in a

statistical table.

• Matching of witness sentences with the key

ideas.

Starting from the statistical table of the

similarity, the synonymy and the adherence to the

same class, the system does a matching between

each of the witness sentences and the key ideas to

provide as a result a table that contains a list of

candidate key ideas that correspond to each of the

witness sentences.

Stage 5: Ranking of the key ideas and fitting of

witness sentences to the key ideas.

Starting from the result table of the previous

treatment and to facilitate the user's choice, the

system provides a grading scale of key ideas that

correspond to each witness sentence. This could be

illustrated as follows:

• If the key idea and the witness sentence share

one common term, add 3 points to the score of

the key idea;

• If the key idea and the witness sentence share

one synonymous term, add 2 points to the score

of the key idea;

Step 5 : Ranking of the KI by a WS and

affectation of WS to the KI

Step 1 : Pre-treatment of the WS and KI

Step 3 : Validation of the simple terms and

composed terms of the KI

Step 2: Extraction of the simple terms and the

composed terms of the WS

Step 4 : Matching of the KI with each of the WS

Pre-treated W S

Pre-treated KI

WS terms

KI terms

Ki by WS

Empty Word Dictionary

Key Ideas (KI) Witness Sentences (WS)

Key Words Dictionary (KWD)

WS affected by KI

SADIM: AN AID SYSTEM FOR MANAGEMENT ENGINEERING DIAGNOSIS USING KNOWLEDGE

EXTRACTION AND MATCHING TECHNIQUES

333

• If the key idea and the witness sentence share

one same class term, add 1 point to the score of

the key idea;

SADIM displays the key ideas corresponding to

each witness sentence on the basis of the already

calculated score. Depending to the user's choice the

system matches each witness sentence to a chosen

key idea.

5 ASSESMENT OF SADIM

The first assessment of SADIM is based on a test

corpus that contains 990 witness sentences (WS) and

390 key ideas (KI). The corpus contains sentences of

two types:

- Type 1: Presence of a Common term between KI

terms and those of WS.

- Type 2: Absence of common terms between terms

of WS and the KI but there can be some semantic

ties between these terms.

In this assessment we determine the recall and

the precision measures that are extensively used in

the domain of information research. We are going to

adapt these measures to our diagnosis method in the

following way:

(Number of KI correctly generated with SADIM to WS/ number of KI

correctly generated with the expert to W

(Number of KI correctly generated with SADIM to WS/ Number of KI

correctly generated with SADIM to WS

)

Results of type 1 sentences

For this type of sentences, the results are as follows:

Recall = 94% Precision = 90%

Results of type 2 sentences

For this type of sentences the results are as follows:

Recall = 64% Precision = 50%

According to the previous results we can identify a

general recall and precision of the order:

General recall = 79% General precision = 70%

6 CONCLUSION

In this paper we started with a presentation of the

concept of diagnosis of the management

engineering, then we gave a brief overview of the

methods of knowledge extraction from the textual

data.

In a latter step we exposed a method of

knowledge extraction that permits to solve

insufficiencies of the SEGESE tool. This method has

led to the emergence of the SADIM system.

We finally made an experimentation of SADIM

in order to give evidence to our method contribution.

As perspectives we intend to spread the

application of our approach into other domains and

to integrate the training and ontological techniques

in SADIM.

REFERENCES

Amarnath, G. and al., 2003. “Towards a formalization of

disease-specific ontologies for neuroinformatics,”

Neural Netwoks, Special issue, No. 16, pp. 1277-1292.

Bourigault, D., 1994. “Terminology extraction software

for the aid of text-based acquisition knewledge,”

Thesis deposited in the school of Higher Studies

Sciences Paris, 1994.

Chevallet, J., 2003. “Language processing for Knowledge

extraction for information research,” Third scientific

days of young research GEI03.

David S., Plante P., 1990. Termino version 1.0. Rapport de

recherche du Centre d'Analyse de Textes

parOrdinateurs, Université du Québec à Montréa

Daille B., 2002. “Mixed approach for terminology

automatic extraction: lexical statistics and linguistic

filters,” Doctorate Thesis, University of Paris VII.

Daille B., Fabre C. and Sebillot P,., 2002. “Applications of

computational Morphology: many morphologies,” P.

BOUCHER, 2002. ISBN 1-57473-125-4.

Jouis, C. 1995. “SEEK, Knowledge acquisition software

using linguistic background without using knowledge

on the external world,” Days concerns acquisition,

validation and training, Grenoble..

Kim, J. and Moldovan D., 1993. “Palka: a system for

lexical knowledge acquisition,” in proceeding of the

second international conference on information and

knowledge management, pages 124-131, Washington,

United States..

Kolsi J. , Belguith L., and Abdelmajid B., 2005.

“Conception and development aid system of

management engineering diagnosis,” Third scientific

days of young research GEI05,.

Oueslati, R., 1999.“Corpus-based knowledge acquisition

aid,” Doctorate Thesis, University Louis

PasteuStrasbourg.

Savall H. and Zardet V., 1989. “Mastering the hidden

coasts and performances: periodically negociable

activity contract,” preface: Marc andré Lanselle,

Edition Economica, edition 2, 351 P.

Savall H. and Zardet V., 2004. “Research in management

science: Qualimetric, seeing complex object,” preface:

BOJE David (USA), Edition Economica, 432p.

∑

WSexpert to e with thgenerated correctly KI ofNumber

WS toSADIM with generated correctly KI ofNumber

Recall

Precision

∑

WS toSADIM with generated correctly KI ofNumber

ICEIS 2006 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS

334