SEMANTIC APPROACH BASED MULTI-AGENT SYSTEM
FOR INFORMATION SEARCH ON THE WEB
University Center of Khenchela, Algeria
Biskra University, Algeria
Currentely most search systems developed for information retrieval are based on
vector representations as the space vector model, others use various statistical and /
or probabilistic approach. Generally, for a given user request, the documents
retrieved are not relevant, because the precision measurement is partial (presence
of noise), and furthermore other relevant documents will never be found, this
difficulty is an effect of a low recall measurement (presence of silence).
Our work is to propose a model whose objective is to improve the results towards
a user query, this will be done by acting on measures of precision and recall, for
this, first we use a multi agents system to reproduce the concepts of autonomy,
cooperation and communication, which are inherent to this type of search systems,
and secondly our approach will combine a syntactic search improved by the use of
semantics that provides the WordNet taxonomy with a semantic search engine
based domain ontology. The knowledge base (domain ontology) is used to
annotate documents by the concepts and the defined instances, therefore these
form a set of semantic index classes on which our search model is based.
Keywords: Semantic Web, Multi-agent System, Semantic Search, Ontology,
Semantic index, WordNet Taxonomy.
1 INTRODUCTION architecture based agents, we will deal mainly in our
approach with the problem of how to develop and
Today the web becomes an immense source of conduct a semantic search engine based ontology and
heterogeneous information that requires the using paradigm agent techniques. Nevertheless,
implementation of powerful tools for retrieving taking into account the lack and the use of
information quickly and to discover new knowledge. incompletes knowledge bases which limits the
Traditional techniques applied in many search exclusive use of models based ontology; we chose to
information models such as the vector space model retain the use of keyword based search to complete
suffer of the handicap of not being able to represent this lack while making some improvements based on
the semantic of documents contents, so that the the semantics offered by WordNet taxonomy.
indexing of a document produced at the end is a bag Section two describes some related works for
of independent terms without any semantic semantic search, next section defines elements used
relationship between them for representing domain knowledge and performing
The semantic information search is a complex semantic annotation process, in section four we'll
process, it has several stages, for example we have present the different multi-agents system components
semantic annotation, query processing, and stage of and describe their interactions, the last section
evaluation and classification of results. The includes a conclusion and some prospects for future
complexity comes from the nature of information improvements.
resources of semantic web that is not restricted to
multimedia objects; also other objects may be 2 RELATED WORKS
people, places, events and others exist.
Then, the semantic web doesn’t work only with In order to improve quality of models developed
the known hyperlink relationship; there are several for information search, many efforts have been
other types of relationships that link its different deployed to annotate documents with semantic
resources. We will introduce in the following an information. The related works to our approach
concerns: Each entry in WordNet is called "synset"; it is a
• Ontology based information search, which uses set of synonyms with the same meaning. For
reasoning mechanism and ontological query example, the words "car","auto" and "automobile"
languages to retrieve ontology instances (semantic are parts of the same synset, moreover, the same
layer) that annotate documents in the search domain. word can have several meanings, for the word "car"
• query expansion based information search, we find five possible senses. 
generally used with vector space models driven The synsets are connected at the top or bottom of
approaches. the hierarchy by different types of relationships,
Several work exist, in  is described an most relationships are hypernym / hyponym i.e. "Is-
architecture based agents and ontology which a" relationship and holonym / meronym relationships
introduces the interest of restricting research on the "Part-of", in our approach we use only hypernym/
web and the advantage of using software agents with hyponym relations and representations of names that
ontology. The architecture is detailed in several are commonly claimed to be the most representative
levels of sub systems layers. form for the semantics of a language, these names
Another semantic search system based agents is are extracted from documents and queries. Between
described in , it uses the concept of conceptual the words, several methods for calculating semantic
graph closely related to natural language, this type of similarity were tested on the taxonomy WordNet, we
knowledge’s representation provides the ability to find essentially two categories:
extract useful information by exploiting the logical • Methods based ontology structure.
relationships in the form of triplets (Relation, • Methods based information content of oncepts.
concept1, Concept2) between the terms in the An important property which we tried to take
documents collection. advantage of in our approach is that, most methods
In  we have a multi-agent system which of calculating semantic similarity try to assign higher
performs an intelligent search; it is based on a similarity to terms which are close together (in terms
semantic approach and a process of enrichment of of path length) and lower in the hierarchy (more
the ontological concepts by probabilistic notions. specific terms), than to terms which are equally close
The semantic approach is based semantic network together but higher in the hierarchy (more general
associated to domain knowledge; in the graph, edges terms).
have weights that express the strength of semantic
relationship which the edge carries between the 3.2 Domain Ontology (Knowledge Base)
nodes (concepts). In computer science the ontology is a formal
Another interesting project developed at the declaration that associates the names of entities in
university George Mason , is interested in the universe of discourse (classes, relations and
searching information on heterogeneous databases functions) with document’s content understandable
(web), the research is guided by ontology and based by human and describing their meanings. This
multi agent system, this project uses a modular formalism also defines a set of axioms that constrain
conceptual model expressed in OWL, and allowing the interpretation of these terms
the integration of other ontologies and other The concept of ontology has long existed,
information resources. especially in philosophy, several definitions of
ontology have been made, the most used is the one
3 KNOWLEDGE REPRESENTATION given by Gruber “An ontology is an explicit
specification of a conceptualization“.
The system architecture includes the following The fundamental objective of the semantic web
components used for handling and representing the is to extend current interfaces oriented to human
domain knowledge: understanding in a format automatically interpretable
- WordNet Taxonomy by programs, this requires developing a rich and a
- Domain Ontology (Knowledge Base Annotation) standard scheme of knowledge representation, this
- Semantic Annotation schema is named "domain ontology".
An ontology as instrument for building
3.1 WordNet Taxonomy knowledge bases provides a controlled vocabulary to
WordNet is a lexical reference developed at formulate queries, representing knowledge
Princeton university offering two separate services: (concepts, relationships and functions), classify the
- A vocabulary describing the different meanings of content of the documents, and make expansions of
words. requests based on class hierarchy and rules on
- A concept hierarchy describing the semantic relationships. 
relationships between words. Bringing together the The ontology must be expressed in language
terms of natural language (English), about 160,000 enough expressive and carrying out reasoning
terms are organized in hierarchical taxonomies of mechanism, this understandable representation of the
names, verbs, adjectives and adverbs. knowledge will allow software agents ability to find
and handle domain entities. documents, these terms form the semantic index,
they are identified by URIs.
3.3 Semantic Annotation Process In our work we opted for OWL Lite as standard
The goal is to build a semantic index ontology specification because this language
according to domain ontology. The index process maintains a compromise between expressiveness to
can be divided into four steps:  formulate domain knowledge and ensure reasoning
1. Extraction of the index terms, a flat index is decidability (e.g. Jena …).
built, and then each term of this index is 3.3.1 Equivalent annotation class
associated with its weight calculated by a Annotation process and related techniques used
technique such as TF-IDF algorithm. such as the technique of generating a set of
2. We use the WordNet hierarchy for equivalents annotation classes are not part of the
identifying all the candidate concepts. objective of this work, we assume these classes have
3. Selecting candidate concept, this analysis been generated by using an appropriate inference
concerns its representation's degree for the mechanism applied to the knowledge base that
web pages contents; it is based on the specifies the domain ontology; nevertheless, we give
frequency of occurrence of the label's here an overview to clarify this concept.
concept and the relationships which the Let Ti: a term of the semantic index, and [Ti] its
candidate concept maintains with the other equivalent class, so we can have Ti1∈ [Ti], Ti2
concepts. ∈ [Ti] … (Fig.2)
4. Among these selected candidate concepts, We consider document “dj” and A, B, C three
we consider those that are specified in the strings in document “dj” semantically annotated by
ontology, i.e. that best represent the Ti1, Ti1, Ti2. From the semantic point of view these
semantic content of the pages, then those three strings are equal even with their different
concepts will form the terms of our syntax because A, B and C are semantically
semantic index. (Fig. 1) annotated with the same semantic index term Ti.
Figure 1: The indexing process Figure 2: Generating equivalent annotation classes
The semantic annotation allows agents who use 3.3.2 Weight mapping
a semantic search engine to decide intelligently about Weights of the annotations are used to evaluate
the relevance of the returned results, for these the relevance, the appropriateness and to implement
reasons the process of retrieving information a classification algorithm (ranking) of retrieved and
depends largely on the quality of formal semantic relevant documents.
annotations defined by domain ontology. These annotation’s weights reflect the relevance
Documents in the domain of interest are of an instance for the semantic of the document
annotated by concepts and instances of concepts, this where it appears, our model is based on the
annotation has two relational properties that are frequency of occurrence of annotation instances in
instances annotations and documents annotated and each document and takes into account the principle
through which the concepts and documents are of generating an equivalent annotation class
linked. So, terms (class, concept, data type, object described above, so the adaptation of the Tf-Idf
property, data type property) defined in the ontology algorithm will consider the number of times an
are used as metadata to annotate the content of the annotation's label within an equivalent class appears
in a document, the formula is: system, it receives and transmits to the system the
user feedback and presents him the search results, the
freq x , d D (1) agent "Information-Research" collects in the area of
dx = * log
max freq y , d nx interest relevant information resources; it may deal
with several other subcontracts agents to accomplish
dx : weight of the instance "x" in document "d" this goal, while the "Domain-Ontology" agent
freqx,d : number of occurrences in "d" of keywords inspects and monitors the dynamic changes in
linked with instance "x" information resources contents, it extracts and stores
max y freqy,d : The frequency of occurrence of the in an RDF base document's links that are annotated
most repeated instance in the document "d" by concepts and instances of the specified ontology.
nx : The number of documents annotated by "x" In the management query unit (processing) is
D : Total number of documents. situated the "Query-Treatment" agent which
Based on works presented in ,  and to coordinates the activities of the system, it formulates
simplify the calculations we’ll only retain the and refines (prepare) the query to be submitted to the
importance of an instance in the document: agent "Information-Research".
max freq ,d
Let D = (d1, d2 ... dn) the collection of
documents in the search space.
([T1], [T2] ... [Tt]) the equivalents classes of the
semantic index terms.
freqxi,d : The total frequency of all elements of the
equivalent class [Ti], appearing in all the semantic
annotations of document "d".
max y freq yi ,d : Max (freq T1,d,freq T2,d…freq Tt,d)
Each document in space D will have a logical
view relative to the weight of its annotation
instances. So a document "dj" will be represented by
the vector weights, and we write:
Figure 3: System's architecture
dj ≅ ( w1 j , w2 j ,...wij ,..., wtj ) (3)
Similarly, for a given request, we get its logical
view to the whole equivalents classes ([T1], [T2] ...
q ≅ ( w1q , w2q ,...wiq ,..., wtq ) (4)
4 SYSTEM’S AGENTS
The main advantage of using paradigm agent is
to perfect the applications related to information
research. In particular, the objective is to design and
develop a multi-agent system which provide: 
1. Respond to user queries through a set of URLs
2. Automatically and simultaneously browse
multiple web sites to find concepts of interest.
3. Monitor changes on selected web sites.
4. Identify and extract useful information.
The architecture that we propose as shown in
Fig.3 is composed of three units:
• User Interface Unit
• Management Query Unit
• Research Information Unit Figure 4: Interaction in multi-agents system
Agent "User-Interface" is considered to be the -AUML: Sequence diagram-
door through which the query is entered in the
Finally returned results will be analyzed and provided by the agent "Evaluate-Ranking". In other
evaluated by the "Evaluate-Ranking" agent that to words, the keyword “Mi” with weight “Wi” is
determine their degree of relevance and decide to assigned a new weight “Wsi” calculated by the
accept or reject them. expression:
System’s architecture including agents is given
in Fig.3; the internal structure of each agent who Wsi=Sim* Wi. 0<= Sim <=1 (5)
composes the system will be detailed below.
All agents interact by sending messages in The keywords by Hyponym / hypernym:
FIPA-ACL protocols formatting, as showed in This task aims to assist the user to reformulate its
Fig.4 we have large type of information exchanged request by offering him choices, i.e. the synsets
between agents throughout their communications. excerpts from the WordNet hierarchy.
4.1 User-Interface Agent Another module that complements the first one
Interface agent resides on the desktop user; it prepares the same query based concepts; it is a
provides the interface to interact with the system. For semantic search which uses relationships between
a search session it records the user request in terms concepts as follow:
of keywords. An RDQL query will be generated from the
Possibly the user can define its Search domain keywords expressed in the original request, also this
and introduce various user preferences such as the may be done by the "User Interface" agent who in
favorite search engine (default Google), and a set of this case reaches the domain ontology and help the
variables defining thresholds calculations. Also this user to explicitly select classes and introduce the
agent presents the user the search results when they desired values of properties.
arrive, it can implement an intelligent behavior and The "Query-Treatment" agent interacts with the
learn from past experiences and user feedback on agent "Domain-Ontology" to run on the pattern of
earlier requests. domain ontology and instances of concepts specified
in OWL the RDQL query, the result is a set of
4.2 Query-Treatment Agent instances that strictly satisfy conditions of the RDQL
In our multi-agents system, this agent manages query. (Standard engine such as "Jena" is used to
the cooperative execution of the user request; it has execute RDQL queries). The execution is an
knowledge about each agent which includes the instantiation operation of the concepts of the
identification and roles that the agent can perform in ontology’s scheme OWL by values of variables used
its capabilities order. (Fig.5) According to their in the constructed query and the invocation of reason
various skills it allocates them tasks to achieve their such as Jena to infer the related knowledge.
common goal. Through interactions that the agent
maintains with the "Evaluate-Ranking" agent it 4.3 Information-Search Agent
performs various substitutions involving: The first research component of this agent is
based syntactic keywords and targets the area of
research (e.g. the web) through a traditional search
engine; however, to improve research results purely
syntactic we introduce a second component which
performs semantic search.
Both modules operate simultaneously, each one
receives input model adapted to query search mode
prepared by the agent "Query-Treatment" (keywords
to perform syntactic-semantic search and generated
instances of concepts derived from the execution of
RDQL query to perform a semantic search).
The agent can contract several other agents to
complete the research, choosing an agent for such
research can depend on the agent capability and the
nature of the information sought. (Fig.6)
4.3.1 Semantic-Syntactic search
Uses a syntactic search engine such as Google to
find in the area of interest documents that satisfy the
Figure 5: Query-Treatment agent structure submitted query. That is a search of purely syntactic
correspondence between the keywords in the query
The weight of keywords: the weights of and terms indexing the documents available in space
keywords are replaced by values calculated by a research.
heuristic evaluation of similarity; these values are
classes. By analogy with the space vector model,
semantic annotations are assigned weights reflecting
the importance of the annotation instance for the
document, therefore in RDQL queries; the variables
in the SELECT clauses are assigned weights
according to the principle of vector model.
The formula of cosine is used to calculate the
similarity document-query, so for a page "j" and a
request "q" we used the expression:
P j .q (7)
Sim ( Pj , q ) =
Pj . q
When the similarity is evaluated, it is compared
to a minimum threshold indicated by user variable
Rmin initially fixed.
Pj , q : weight vectors associated to Pj, q.
Figure 6: Information-Search agent structure |pj|,|q| : respectively “Pj” and “q” vectors norms
When Sim( Pj, q) > =Rmin, the page “Pj” is
4.3.2 Semantic search considered relevant, its link and the similarity's value
This module research in the RDF documents are returned to agent "Information-Search" for final
base, the RDF annotations that match tuples storage, in the case Sim( Pj, q) <Rmin the current page
instances recovered by the "Query-Treatment" agent. will be ignored, the process ends when all the pages
The module receives input instances which are the are crawled, at the end we will have obtained a set of
results of the RDQL query, then, documents whose all relevant pages, according to user feedback if the
links have been stored in the RDF database are number of relevant resources found is sufficient, an
analyzed and those annotated by these tuples algorithm for grading results is executed to present
instances are found, they are considered semantically the results in their degree of relevance; this algorithm
relevant. Then the agent records in a temporary file is implemented by the ranking component in the
the following details: Links of resources found and structure of this agent.
their evaluated similarities. If we want against include more resources
4.4 Evaluate-Ranking Agent (depends on user feedbacks), the "Evaluate-Ranking"
The "Information-Search" agent stores links of agent will explore the relationships between concepts
resources found in a temporary file to which the defined in the WordNet hierarchy to extract sets of
"Evaluate-Ranking" agent accesses, so it is a type of synonyms, hyponyms and hypernyms., Then the
memory that can be modeled by a blackboard. For expansion of the query will use the "synsets" in the
each entry, the "Evaluate-Ranking" agent download limits of depths set by the user, but generally when
page referenced by the link. Furthermore Keywords using an expansion with hypernym synsets the depth
that syntactically index the page or semantic index is set to "1" because the similarity tends to decrease
instances are assigned weights according to the when generalizing sense.
principle of vector model.
4.4.1 Evaluate module
Let Wij: weight of term "i" (keywords) in page j.
Freq _ t i
Wij = (6)
max( freq _ t j ) j =1, n
n: the number of keywords
User's query is also represented by the weight
vector Q = (w1q, w2q, wiq ,..., wnq), where wiq is
the weight of the keyword “i” in query Q, a keyword
may be a keyword's synonym, its hyponym or
The semantic annotation is based ontology, the
ontology defines the concept's terms used as
metadata to form the semantic index, and thus, these
terms are identified by URIs and may be equivalent Figure 7: Evaluate-Ranking agent structure
4.4.2 Calculation of similarities according to dynamic web changes (i.e. on the fly).
We based our approach on user feedback in the Moreover the agent uses a standard engine (e.g. Jena
choice of terms that will be used to extend the query, and racer) to infer knowledges formalized in the
in (Fig.8) we give a formula to review similarities ontology and executes the RDQL query, then results
and update the weight of keywords. This formula is (instances) obtained would be communicated to the
improved from those presented in , our choice to "Query-Treatment” agent.
use these forms is justified by opportunity for The other task of the agent is to browse the web (area
considering the structure of the ontology through two of interest) at regular intervals to detect documents
parameters: annotated RDF consistent with the specified domain
The length of the path linking concepts Ca and Cb ontology, the document's links found will be stored
The depth of concepts Ca, Cb in the hierarchy in the database documents annotated RDF. Agent
can therefore take into account the dynamics of
information on the web in independently and
D*n Hhypernyms 5 CASE STUDY
Min.( Da ,Db )+1
Sim (Ca, Cb)
2 We chose to experiment and implement our
(8) model the tourism domain where tourist sites on the
web are annotated by their owners by RDF triples
Cb,Ca and instances defined in the domain ontology used.
1 Hyponyms Tourism activity has several domains (transport,
D*n entertainment, sports, scientific conferences, etc.),
Max Da ,Db )
( but to simplify the analysis, we will limit study to
2 hotel domain for which we associate a domain
ontology named “hotel”.
1 Cb,Ca Synonyms
0 Cb ,Ca Not linked 5.1 Ontology UML Specification
The ontology ("hotel") is specified in the
language OWL-Lite; this ontology is associated with
Figure 8: Forms of calculating similarities a UML diagram specifying the classes (concepts),
the properties and relationships between concepts
D: depth hierarchy; n: minimum path length between and some instances of concepts. (Fig. 9)
concepts Ca and Cb (number of arcs).
Da, Db: concept’s depths.
Terms of synonymous synsets are assumed to
have a similarity of "1" i.e. they are identical.
For hyponyms / hypernym synsets and
considering the user choice, a similarity value is sent
with the term to "Query-Treatment" agent to
recompose query and restart a new syntactic search.),
the process is repeated until all sets constructed
would be entirely explored.
4.4.3 Ranking results
It is desirable to present to the user the obtained
results ordered by their relevance measure. The
agent accesses the temporary database of relevant
documents, for each one it estimate its similarity
with submitted query. The final similarity is
calculated by an expression of type:
SimF=a*Simsyn+b*simsem ; a ∈[0,1] ; b=1-a (9) Figure 9: UML diagram.Oontology “hotel”
Documents returned having a high similarity are An inference engine applied to the ontology
those with: a # 0 and b # 0. schema and defined instances, will infer knowledge
other than those explicitly declared, inference is a
4.5 Domain-Ontology Agent mechanism that is based on the expressiveness of the
Attached to the domain ontology, the main goal language (OWL-Lite) and its formal semantics based
of this agent is to maintain the ontology closely and on description logics, especially this concerns
restrictions on classes, on the properties among <class>3</class>
classes and a set of defined axioms on classes, for </Hotel>
example, we specify that a 5 star rated hotel must </rdf:RDF>
have as service “guided-visits” by the class: guided- The execution of code associated with this
visits=((hotel) ∩ (> = 5 rated.star)). model produced the following results:
Type: Chelia is
5.2 Inference Models
The integration of the Jena API in our model http://mydomain/ontology/infohotel/chelia rdf:type
will allow it deriving additional RDF assertions http://mydomain/ontology/infohotel/hotel
included in the OWL knowledge base; this
mechanism supports the languages RDF /RDFS and Type: Chelia is
OWL and uses an inference model which has two http://mydomain/ontology/infohotel/chelia rdf:type
• The schema of the model
• The instances of the model 6 CONCLUSION
The example below is an illustration of an
inference model used by inference engine RDFS. The proposed semantic research model based
Inference is performed by the transitive relation on multi-agent system and using domain ontology
properties which defines 'room service “as a sub illustrate the concept of cooperative resolution of
property” of the property “hotel service”. distributed problems, the process combines a search
5.2.1 Model’s schema engine based ontology with a traditional search-
based keyword which include relations of synonymy
<?xml version="1.0"?> and hyponymy provided by the WordNet taxonomy.
<!DOCTYPE rdf:RDF [ <!ENTITY hotelerie The semantic search uses as support an RDQL
'http://mydomain/ontology/infohotel/'> query generated from query keywords, then an
<!ENTITY rdf 'http://www.w3.org/1999/02/22-rdf- inference engine such as "Jena" will use the ontology
syntaxns#'> scheme to retrieve defined instances in
<!ENTITY rdfs 'http://www.w3.org/2000/01/rdfschema#'> correspondence with keywords in the query, these
instances will be sought in the RDF database and
<rdf:RDF xmlns:rdf="&rdf;" xmlns:rdfs="&rdfs;"
return the documents that they annotate.
xmlns:xsd="&xsd;" As prospects for research in this area and in
xml:base="http://mydomain/ontology/infohotel/" relation with our model, we propose to enrich the
xmlns="&hotelerie;"> knowledge base agents with techniques for
<rdf:Description rdf:about="&hotelerie;room-service"> formulation query including explicit rules and policy
<rdfs:subPropertyOf decision, this will allow the "Query-Treatment"
rdf:resource="&hotelerie;hotelservice"/> agent to optimize the request in an intelligent way, it
</rdf:Description> is true that over the query is well-defined, better
relevant results are obtained.
Also, to take advantage of new technologies
</rdf:Description> applied to artificial intelligence systems, we intend to
<rdf:Description rdf:about="&hotelerie;classement"> couple the agent "Query-Treatment" with a system of
<rdfs:range rdf:resource="&xsd;integer" /> reasoning from cases (CBR), this will enable and
</rdf:Description> perfect the search process by reasoning from cases
</rdf:RDF> already resolved and stored in the CBR data base.
5.2.2 Model’s instances 7 REFERENCES
 G. Varelas, E. Voutsakis, P. Raftopoulou
<?xml version="1.0"?> “Semantic similarity methods in WordNet and
<!DOCTYPE rdf:RDF [ <!ENTITY hotelerie their application to information retrieval on the
Web”, WIDM’05, November 5, 2005, Germany.
<!ENTITY rdf 'http://www.w3.org/1999/02/22-rdf-
Copyright 2005 ACM 1-59593-194-5/05/0011.
<!ENTITY rdfs 'http://www.w3.org/2000/01/rdfschema#'>  T. Osman, D. Thakker, G. Schaefer, P. Lakin,
<!ENTITY xsd 'http://www.w3.org/2001/XMLSchema#'> “An integrative semantic framework for image
]> annotation and retrieval” International
<rdf:RDF xmlns:rdf="&rdf;" xmlns:rdfs="&rdfs;" Conference on Web Intelligence 0-7695-3026-
xmlns:xsd="&xsd;" 5/07, 2007 IEEE/WIC/ACM
xml:base="http://mydomain/ontology/infohotel/"  S. Jun-feng et al.
xmlns="&hotelerie;"> “Ontology-based information retrieval model for
the semantic Web” Changsha ,China 2005
<room-service rdf:resource="&hotelerie;internet" />
 P. Castells, M. Fernandez, D. Vallet,
“An adaptation of the vector-space model for
ontology-based information retrieval”
IEEE transactions on knowledge & Engineering,
Vol. 19, No. 2, February 2007 1041-4347/07
 E. Toch & A. Gal
“A semantic approach to approximate service
retrieval” ACM Transactions on Internet
Technology, Vol. 8, No. 1, Article 2, Nov.2007.
 B.Espinasse ; S.Fournier ; F.Freitas
“AGATHE : Une architecture générique à base
d’agents et d’ontologies pour la collecte
d’information sur domaines restreints du web”
 Tanveer J Siddiqui ; U.S tiwary
“Integration notion of agency and semantics in
information retrieval :An intelligent multi agent
model” University of Allahabad IEEE 2005
 Carmine Cesarano et al.
“An Intelligent search agent system of semantic
information retrieval on the Internet” ACM 2003
 Larry Kershberg et al.
“Knowledge Sifter: Agent based ontology-
driven search over heterogeneous databases
using semantic web services”
George Mason University USA 2004
 E. Desmontils & C. Jacquin
"Indexing a Web Site with a Terminology
IRIN, Université de Nantes-IOS Press, 2002
 Kwang Mong Sim
“Towards Holistic Web-based Information
Retrieval: An Agent-based Approach”
IEEE/WIC International Conference on Web
Intelligence (WI’03) IEEE 2003