ADVISORY AGENTS IN THE SEMANTIC WEB

Ralf Bruns, Jürgen Dunkel

Department of Computer Science, University of Applied Sciences and Arts Hannover,

Ricklinger Stadtweg 120, D-30459 Hannover, Germany

Sascha Ossowski

AI Group, E.S.C.E.T.,Universidad Rey Juan Carlos Madrid,

Campus de Mostoles, Calle Tulipan s/n, E-28933 Madrid, Spain

Keywords: Semantic Web Application, Ontology, Agents, E-learning.

Abstract: In this paper, we describe the advances of the Semantic E-learning Agent project, whose obje

ctive is to

develop virtual student advisers that render support to university students in order to successfully organize

und perform their studies. The advisory agents are developed with novel concepts of the Semantic Web and

agent technology. The key concept is the semantic modeling of the domain knowledge by means of XML-

based ontology languages such as OWL. Software agents apply ontological and domain knowledge in order

to assist human users in their decision making processes. Agent technology enables the incorporation of

personal confidential data with public accessible knowledge sources of the Semantic Web in the same

inference process.

1 INTRODUCTION

E-learning has started to play a major role in the

learning and teaching activities at institutions of

higher education worldwide. The students perform

significant parts of their study activities

decentralized via the Internet. The main focus of

current E-learning systems is to provide an

appropriate technical infrastructure for content

engineering and information exchange.

The emerged individual ways of study are

ocation- and time-independent, consequently

requiring a permanently available and direct support

to answer questions and give advice. A recent

comparison of modern E-learning environments

(CCTT, 2004) revealed that intelligent advisory

agents are not applied so far in E-learning systems.

The objective of the Semant

ic E-learning Agent

(SEA) project (Dunkel, 2004) is to develop virtual

student advisers that render support to university

students, assisting them to successfully organize und

perform their studies. The experiences of human

course advisers show, that most students have

similar problems and questions. The advisory agents

should help to resolve these problems. Typical

questions concern the regulations of study (e.g. does

a student possess all requirements to participate in

an examination or a course?) or organizing student

mobility.

To achieve these goals, we propose a software

architecture

where virtual student advisers are

developed with novel concepts from Semantic Web

(Berners-Lee, 2001; Horrocks, 2002) and Intelligent

Agent (Wooldrige, 1995) technology. The basic idea

is to model the structure of our E-learning domain

by means of ontologies, and to represent it by means

of XML-based applied ontology languages.

Software agents apply the knowledge represented in

the ontologies during their intelligent decision

making process. We claim that this is a promising

approach because E-learning systems that

successfully support students in organizing their

studies are still to come. This paper reports on the

experiences gained from the development of an

advisory system that effectively integrates both,

Semantic Web and Intelligent Agent technology.

The first use case that has been implemented

reflects the counseling situation where a

student

intends to study a semester abroad within the

European Erasmus/ Socrates exchange program.

Together with the international coordinator, the

student has to choose the foreign university and the

Bruns R., Dunkel J. and Ossowski S. (2005).

ADVISORY AGENTS IN THE SEMANTIC WEB.

In Proceedings of the Seventh International Conference on Enterprise Information Systems, pages 90-96

DOI: 10.5220/0002546500900096

 SciTePress

foreign study program that matches best her/his

personal interests and her/his individual situation of

study. Subsequently, a study plan for the semester at

the host university must be determined that

corresponds to the home university syllabus. This

study plan constitutes the so-called Socrates

Learning Agreement.

The paper is structured as follows: In the next

section the employed knowledge representation

techniques and the developed knowledge models are

presented. The third section shows how automated

inference can be carried out on the knowledge

models. Subsequently, the software architecture of

the agent system is outlined. Finally, the last section

summarizes the most significant features of the

project and provides a brief outlook to future lines of

research.

2 KNOWLEDGE MODELING

The key concept of a semantic advisory system is

the semantic modeling of the domain knowledge

(e.g. university organization, degree requirements,

course descriptions, examination regulations) as well

as an individual user model, which reflects the

current situation of study (e.g. passed exams, current

courses). The fundamental structures of the available

domain knowledge as well as the basic facts (e.g.

offered courses) are defined in appropriate models.

In our system, the structural part of the

knowledge base is modeled by means of ontologies,

which formally define domain entities and the

relations among them. For this purpose, we apply

Semantic Web technology based on XML. We have

chosen the W3C standard ontology language OWL

(Web Ontology Language) (W3C-OWL, 2004) to

model the knowledge required in the advisory

system. Software agents use this information as the

basis for their reasoning and negotiation. Due to the

standardization of these technologies, knowledge

models can easily be shared and reused via the

Internet. Thus, the developed ontologies can serve as

standardized and open interfaces for the

interoperability of different E-learning systems.

2.1 Ontologies

In order to implement the counseling situation of the

Erasmus/ Socrates exchange program, information is

necessary about the possible exchange universities

and their offered degree programs. In addition,

further information about the living conditions of a

particular university city and its urban infrastructure

may influence the decision.

Several interrelated ontologies have been

developed for our advisory agents: Two central

ontologies describe the organizational structure of a

university and the offered courses in a semester. To

facilitate the comparison of different study places

and course contents, two subordinated ontologies are

used. The individual study situation of a specific

student is represented by a separate ontology.

Dividing the knowledge base of the advisory

system in several different ontologies is crucial to

yield a coherent scope of each ontology and to

facilitate reusing existing ontologies (Noy, 2001). In

the following, we describe the responsibilities of the

employed ontologies in some more details.

• University Ontology

The university ontology is the core knowledge

base of the SEA project. It models the essential

parts of the organizational structure of a

particular university and the departments with

the different programs of study. Its main domain

concepts are: university, department, degree

program, offered degrees.

The following example shows an excerpt of an

instance of the university ontology.

<uni:DegreeProgram

rdf:ID="FHH_Master_CS">

...

<uni:numberOfStudents rdf:datatype=

"http://.../XMLSchema#int">

547

</uni:numberOfStudents>

<uni:hasContent rdf:resource=

"http://../subject.owl#softwareEng"/>

<uni:hasContent rdf:resource=

"http://../subject.owl#compGraph"/>

...

</uni:DegreeProgram>

At first, a degree program instance with id

FHH_Master_CS is created. The property

numberOfStudents specifies how many

students are enrolled and has the XML schema

data type int. The property

hasContent

describes the contents of the degree program

and refers to a computer science instance of the

subject area ontology specified by the URI.

• Course Ontology

The course ontology models the courses per

semester for a degree program. This information

changes from semester to semester and can only

be provided by the responsible department.

Several properties describe an individual course,

e.g. course name, teaching language, number of

credit points, keywords describing the course

content, and the semester when the course takes

place. This knowledge will be used in the

ADVISORY AGENTS IN THE SEMANTIC WEB

second step of our sample use case when open

courses of the home syllabus are matched with

courses at the exchange university.

Each university participating in the Socrates

program should build its own instance of these

ontologies. Additionally, for our counseling scenario

we need further information that is provided by two

additional ontologies.

• Regional Ontology

The regional ontology models the relevant

properties of a study place, e.g. in which

country, state, and region it is located, number

of inhabitants, which infrastructure is available

(e.g. airport, station, theatre). Each study place

is represented by an instance of this ontology,

thus allowing a comparison due to the students

living preferences. It is expected that for many

cities this information will be available on the

Semantic Web in the near future.

• Subject Area Ontology

To find an appropriate study plan at the

exchange university, home and foreign courses

must be compared based on their contents. A

simplified taxonomy is modeled in the subject

area ontology, e.g. one instance for computer

science, one instance for mechanical

engineering, and so forth.

These two ontologies define some transitive

properties that are used for inference and reduce the

number of facts significantly. An example for

transitivity is the property

isLocatedIn of the

regional ontology.

<owl:TransitiveProperty

rdf:ID="isLocatedIn">

<rdfs:domain rdf:resource="#region"/>

<rdfs:range rdf:resource="#region"/>

</owl:TransitiveProperty>

For example, from the two facts, that Hannover

is located in Lower Saxony, and that Lower Saxony

is located in Germany, it can be concluded that

Hannover is located in Germany. In a similar way a

hierarchy of subtopics is modelled in the subject

area ontology.

In contrast to these ontologies, which model

public accessible information, the user ontology

serves as the knowledge model of a particular user,

e.g. student or faculty member and, consequently,

contains confidential information.

• User Ontology

The major classes of this ontology are

Student

and

Faculty. Relevant information of a student

are, e.g. login name, student ID, current

semester, passed/failed courses etc. Every

student owns her/his own instance file of this

ontology, reflecting her/his individual progress

of study. This information allows the adviser to

give a personalized advice considering the

individual situation of a student.

Note that the different ontologies are not

isolated, but related to each other. So, e.g. a student

instance of the user ontology is related to a course

instance of the university ontology via the property

isEnrolledIn. Figure 1 shows the entire structure

of the ontologies with the interrelating properties

and some of their classes.

University

Department

DegreeProgram

Region

Student

Course

Subject

Area

Subject Area

Ontology

Course

Ontology

Regional

Ontology

User

Ontology

University

Ontology

isLocatedIn

isEnrolledIn

comprises

hasContent

University

Department

DegreeProgram

Region

Student

Course

Subject

Area

Subject Area

Ontology

Course

Ontology

Regional

Ontology

User

Ontology

University

Ontology

isLocatedIn

isEnrolledIn

comprises

hasContent

Figure 1: Sketch of ontology structure

In a Semantic Web infrastructure the knowledge

is spread over the Internet in form of different OWL-

files. We can distinguish two types: OWL schemas

and OWL instances. In our advisory system there are

five different OWL schema files, each containing

just one of the described ontologies. To prevent

inconsistencies, OWL schema files are located only

once on a central web server.

However, the OWL instance files are created and

maintained locally. It is crucial, that the OWL

instances conform to the language specification

defined in the OWL schemas and refer also to other

instances.

2.2 Ontology Development

The previous section described the knowledge base,

i.e. the ontologies and their corresponding facts,

from a logical point of view. To make the

knowledge usable for the advisory agents, internally

or via the internet, they must be defined in a formal

ontology language suitable for reasoning. For this

purpose, we applied the W3C standard ontology

language OWL (Web Ontology Language) (W3C-

OWL, 2004) based on XML and RDF/ RDF Schema

(W3C-RDF, 2004). The expressiveness of OWL-DL

was sufficient to model our domain knowledge.

Only a few shortcomings of OWL came up, which

we resolved within the inference engine, as describe

in the next section.

To develop complex ontologies an adequate tool

support is indispensable. OWL is intended for the

ICEIS 2005 - SOFTWARE AGENTS AND INTERNET COMPUTING

usage of software programs and cumbersome for

humans, as the short OWL example in the previous

section illustrates. In our project we used the well-

known Protégé Version 2.1 with the OWL Plugin

(Protégé, 2004) for ontology development. Except

some smaller technical problems we made good

experiences with this tool. It allowed to specify

ontologies with a graphical user interface and to

generate the corresponding OWL files, avoiding a

potentially error prone “manual” OWL coding.

Furthermore, facts in form of OWL instances were

created on base of these ontologies.

3 INFERENCE

The semantic advisory agents should act similar to

human advisers according to their knowledge

modeled in the ontologies. This is achieved by using

the rule-based inference engine JESS (Java Expert

System Shell) (Friedman-Hill, 2004) to carry out the

automated inferences entailed by the semantics of

OWL. JESS provides a convenient way to integrate

reasoning capabilities into Java programs. With the

JESS language complex rules, facts and queries can

be specified.

3.1 OWL Transformation

To make use of the knowledge modeled in an

ontology, the OWL semantics must be mapped into

facts and rules of an inference engine. Because JESS

does not provide an interface to import an OWL

ontology, we employed the tool OWL Engine to

load OWL ontologies and OWL instances into a

JESS knowledge base (OWL Engine, 2004), which

provides an XSLT-based transformation process.

The OWL inference engine consists of three

different parts. One file contains JESS rules

describing the OWL meta model, i.e. the OWL built-

in rules. Two XSLT stylesheets transform files with

OWL schemata or with OWL instances into JESS

assertions.

A major advantage of the XSLT stylesheets

approach is that the stylesheets can be easily

adjusted to individual requirements. In our project

we expanded the transformation rules for the

owl:transitiveProperty and the owl:UnionOf

OWL constructs.

3.2 Ontology Reasoning

Mainly, the advisory agents reason on the basis of

the OWL knowledge model loaded into the JESS

knowledge base. For our sample use case, the

semantic expressiveness of OWL is nearly

sufficient. But to express more complex expert

knowledge, e.g. complex examination regulations,

domain-specific rules must be developed. Inference

engines such as JESS provide their own languages to

specify complex rules for developing rule-based

systems. A simple example for a domain-specific

rule, out of the scope of OWL, is a JESS rule that

categories cities according to their size.

The data modeled in OWL is usually domain

specific, but independent of a certain application.

The OWL properties define rules, which represent

the general structure of the knowledge. They are

mainly data-oriented, usage-independent and

applicable to different applications. Additional rules

specified in an inference engine are process-

oriented; they specify the reasoning capabilities of

an advisory system and are tailored to a specific use

case.

4 AGENT ARCHITECTURE

The software architecture of an advisory system

should reflect the situation of a real counseling

interview. In our use case, a student intends to study

abroad for one semester and consults the

international coordinator of the department to get

advice. Together they first look for an appropriate

exchange university and then for a study plan, which

fits best with the course program at the home

institute.

All students are characterized by their personal

situation and intents; the international coordinators

give their advice on base of a profound knowledge

of the study regulations and the different exchange

programs.

Multi-agent systems provide a software

paradigm that fits well to the described situation

(Woolridge, 2002). The advisory system can be

viewed in terms of autonomous agents of two

different types: student agents and international

coordinator agents. These agents interact to find an

exchange university and a suitable study plan. Multi-

agent technology provides the right level of

abstraction to model a negotiation process between

independent partners (Jennings, 2000; Kraus, 1997)

and, consequently, is well-suited for our purposes.

4.1 Agent Structure and Semantic

Web

Figure 2 outlines the internal structure of the

advisory system with two different types of agents:

the student agent and the international coordinator

ADVISORY AGENTS IN THE SEMANTIC WEB

Adviser Server

coordinator

agent

OWL instance

coordinator data

Student Server

student

agent

OWL instance

student data

University Server

OWL instance

university +

course data

University Server

OWL instance

university +

course data

Schema Server

OWL

schemas

OWL instance

subject area

City Server

OWL instance

regional data

City Server

OWL instance

regional data

reference

Adviser Server

coordinator

agent

OWL instance

coordinator data

Student Server

student

agent

OWL instance

student data

University Server

OWL instance

university +

course data

University Server

OWL instance

university +

course data

Schema Server

OWL

schemas

OWL instance

subject area

City Server

OWL instance

regional data

City Server

OWL instance

regional data

reference

agent. The two agent types are conceptually

identical: both reason on a knowledge base using

JESS as inference engine.

Student Agent

Coordinator Agent

university/course

ontologies

and facts

user ontology

and facts

JESS rules

behaviors

JESS engine

Student Agent

Coordinator Agent

university/course

ontologies

and facts

user ontology

and facts

JESS rules

behaviors

JESS engine

Figure 2: Agent structure

Each agent loads its individual knowledge base

dynamically according to the actual counseling

situation. As described above, the knowledge is

specified in OWL-files and spread over the Internet.

To build up its knowledge base each agent has to

process the following steps:

1. According to the status of the interview, the

agent determines the required information for

the actual counseling context.

2. If the information is publicly available, the

agent locates the corresponding OWL files in

the Internet.

3. It downloads the OWL instances, transforms

them and imports them into JESS using OWL

Engine.

The international coordinator agent requires

information about all exchange universities and their

course contents. To gain this knowledge, it can

dynamically expand its knowledge base by

accessing the locally stored OWL-files of the

universities registered in the advisory system.

Beyond the information the coordinator agent

collects in the Internet, it can hold some private

knowledge. For example, it may know about all

exchange agreements of its university or the

utilization of the courses in its department.

The student agent is characterized by its

individual study situation, which can be described

by the study year, the attended lectures, and the

passed exams. Of course, this information is

confidential and, therefore, it is represented in a

personal OWL instance file, which is protected

against unauthorized access.

Figure 3 depicts the distribution of the

knowledge sources in the Semantic Web. The

coordinator agent and the student agent reside on

different servers, where their private knowledge is

stored in corresponding OWL instance files.

Figure 3: Distributed knowledge sources in the

Semantic Web

Furthermore, each agent can have more

sophisticated reasoning capabilities expressed by

some further JESS rules, as explained in subsection

3.2.

4.2 Agent Interaction and

Negotiation

In a real counseling situation a problem is resolved

by a communication and negotiation process, which

is characterized by an information exchange among

the different dialog partners. In a multi-agent system

the communication between the agents reflects this

negotiation process between clients and advisers.

Depending on how the consultancy is developing,

different information is exchanged between the

agents. The agent behaviors implement the

negotiation protocol determining the rules that

govern the interaction (Jennings, 2000; Ossowski,

2002). The following use case scenario outlines how

the agents interact.

1. A student starts his/her personal student agent

(SA) to search for a suitable exchange semester,

and logs in.

2. The SA loads the OWL user ontology and the

OWL instance data representing the students

specific study situation into its JESS knowledge

base.

3. The student can enter some preferences

regarding the exchange university (e.g. the

subject of study, the teaching language, desired

location). The SA extracts the specified

parameters and queries further personal data

(e.g. the aimed degree) from the knowledge

base. Then the SA sends a request to the

international coordinator agent (ICA).

ICEIS 2005 - SOFTWARE AGENTS AND INTERNET COMPUTING

4. The ICA collects instance data about all

universities registered in the system as well as

about the study places and loads them in its

knowledge base. During its initialization, the

ICA has already loaded all ontology schema

files.

5. Then the ICA reasons on the knowledge base,

aggregates the results and sends a ranked list of

appropriate foreign degree programs to the SA.

6. The SA receives the result and presents it to the

student, who chooses his/her favorite exchange

university and degree program. The student’s

decision and further user instance data are send

to the ICA (e.g. the study program based on the

passed exams).

7. The ICA accesses the OWL course instance

data of the selected foreign degree program via

the Internet, and loads it into its knowledge

base. Usually the courses information is

maintained in each exchange university

separately. On the basis of the expanded

knowledge base the ICA suggests the foreign

courses that are fitting best to the study program

of the home university, see figure 4.

Figure 4: Study plan proposal of the coordinator agent

8. The SA receives the results from the ICA and

the student chooses manually the desired course

plan out of the different suggested options.

Finally, the SA generates a formal document,

called Socrates Learning Agreement,

determining the personalized exchange study

plan.

The student agent protects the confidential

information of its human owner. Similar to a real

consultancy situation it only reveals sensitive private

data, if it is indispensable for finding a solution. The

knowledge of the student agent is rather restricted; it

mainly knows the personal situation of its owner.

The international coordinator agent has a much

broader knowledge, which it dynamically expands in

the Semantic Web.

In the current implementation the international

coordinator agent has no own intentions, i.e. it

leaves all decisions about the exchange program to

the user agent, who delegates them to its human

user. Of course, the behavior of both agents could

implement personal desires and intentions. For

example, the coordinator agent could present only a

selection of possible exchange universities,

depending on the exchange agreements or the

number of applicants.

4.3 Implementation Issues

Powerful agent development frameworks facilitate

the development of multi-agent systems. The

semantic advisory agents are developed with JADE

(Java Agent Development Framework) (Bellifemine,

2002), which complies with the FIPA (Foundation

of Intelligent Physical Agents) standards (FIPA,

2003). JADE includes two main components: a

FIPA-compliant agent platform and a framework to

develop Java agents. The core part of the FIPA

architecture is a standard for agent communication,

i.e. its ACL (Agent Communication Language). The

interaction between the student and the international

coordinator agent is based on the exchange of ACL

messages.

To avoid that a user has to install the student

agent on her/his computer, we chose a web

architecture: the user agent resides on a central

server and has a web interface implemented with

JavaServer Pages.

5 CONCLUSION

In this paper, we described how Semantic Web and

Agent Technology can be integrated to build an

intelligent advisory system for an E-learning

environment. Our goal is to create and deploy

semantic advisory agents capable of supporting

university students in successfully organizing and

performing their studies.

Due to the use of Semantic Web languages the

developed knowledge models can easily be used in

distributed systems and shared among software

agents via the Internet. In dependence of the state of

the consulting interview, agents acquire dynamically

useful knowledge from distributed sources in the

Semantic Web and integrate it in their personal

knowledge base. Agent technology enables the

incorporation of personal confidential data with

ADVISORY AGENTS IN THE SEMANTIC WEB

publicly accessible knowledge sources of the

Semantic Web.

The major difficulty encountered was the

integration of the different concepts – on the one

hand the knowledge bases written in RDF and OWL,

on the other hand the inference engine JESS and the

agent environment JADE. We implemented a

prototype system, where the agents were able to

reason upon the knowledge base in the desired

manner. Our experiences show that the employed

technologies are mature and well-suited for the

implementation of advisory systems.

In our future work, we will implement more use

cases for the Semantic E-learning Agent project. For

example, advisers should be able to announce new

opportunities for students who are looking for

suitable thesis subjects and to answer questions

regarding the regulations of study.

REFERENCES

Bellifemine, F, Giovanni, C., Trucco, T. and Rimassa, G.,

2002. JADE Programmers’s Guide. Retrieved August

25, 2004, from http://jade.tilab.com/

doc/programmersguide.pdf

Berners-Lee, T., Hendler, J. and Lassila, O., 2001. The

Semantic Web. Scientific American, Vol. 5.

CCTT - Center for Curriculum, Transfer and Technology,

2004, Retrieved August 25, 2004, from

http://www.edutools.info/course/compare/all.jsp

Dunkel, J., Bruns, R. and Ossowski, S., 2004. Semantic E-

Learning Agents. In Proc. of the Sixth International

Conference on Enterprise Information Systems ICEIS

2004 (Porto, Portugal), 271-278.

FIPA - Foundation of Intelligent Physical Agents, 2003.

Retrieved August 25, 2004, from http://www.fipa.org

Friedman-Hill, E., 2000. JESS, The rule engine for the

Java platform. Retrieved August 25, 2004, from

http://herzberg.ca.sandia.gov/jess/

Horrocks, I. and Hendler, J. (eds.), 2002. The Semantic

Web, First International Semantic Web Conference.

Springer.

Jennings, N.R., Parsons, S., Sierra, C. and Faratin, P.,

2000. Automated Negotiation. In Proc. 5

Int. Conf.

On Practical Application and Intelligent Agents and

Multi-Agent Systems (Manchester, UK), 23-30.

Kraus, S., 1997. Negotiation and cooperation in multi-

agent environments. Artificial Intelligence, 94(1-2),

79-98.

Noy, N.F. and McGuiness, D.L., 2001. Ontology

Development 101, A Guide to Creating your first

Ontology. Stanford Knowledge Systems Laboratory

Technical Report KSL-01-05.

Ossowski, S.; Omicini, A., 2002. Coordination Knowledge

Engineering. The Knowledge Engineering Review 17

(4), 309−316

OWL Inference Engine, Retrieved from

http://mycampus.sadehlab. cs.cmu.edu/public_pages/

OWLEngine.html

The Protégé Project. Retrieved August 25, 2004, from

http://protege.stanford.edu/

Wooldridge, M. and Jennings, N., 1995. Intelligent Agents

- Theory and Practice. Knowledge Engineering Review

10 (2), 115–152.

Woolridge, M., 2002. An Introduction to Multiagent

Systems. Wiley.

W3C – The World Wide Web Consortium. RDF Primer –

W3C Recommendation 10 February 2004.

Retrieved August 25, 2004, from

http://www.w3.org/TR/rdf-primer/

W3C – The World Wide Web Consortium. OWL Web

Ontology Language Reference – W3C

Recommendation 10 February 2004.

Retrieved August 25, 2004,

from http://www.w3.org/TR/owl-ref/

ICEIS 2005 - SOFTWARE AGENTS AND INTERNET COMPUTING