IMPLEMENTATION ISSUES OF THE INFONORMA
MULTI-AGENT RECOMMENDER SYSTEM
Lucas Drumond, Rosario Girardi, D’jefferson Maranhão and Geraldo Abrantes
Department of Computer Science, Federal University of Maranhão, Avenida Dos Portugueses, São Luís, Brazil
Keywords: Recommender systems, Content based filtering, Collaborative filtering, Legal information systems, JADE.
Abstract: Recommender systems can help professionals of the legal area to deal with the growth and dynamism of
legal information sources. Infonorma is a multi-agent recommender system that recommends legal
normative instruments to users according to their particular interests using both content-based and
collaborative information filtering techniques. It has been modeled under the guidelines of the MAAEM
methodology. This paper discusses the main implementation issues of the Infonorma system.
1 INTRODUCTION
Recommender systems (Adomavicius and Tuzhilin,
2005) are a particular type of information filtering
applications that can be used as an approach to the
problem of storing, retrieving and structuring large
amounts of legal information to effectively satisfy
information needs of legal users.
Recommender systems perform tasks like
inferring user preferences and monitoring and
reasoning upon available information sources that
are inherently connected to the autonomous behavior
of software agents. Because of that, the multi-agent
paradigm is a reasonable approach for developing
recommender systems.
Infonorma is a multi-agent recommender system
for the legal domain that uses content based filtering
(Balabanovic and Shoham, 1997) for recommending
legal normative instruments to users according to
their particular interests. The knowledge about the
legal domain used by Infonorma is represented in
the ONTOJURIS ontology, which describes the
structure of a legal normative instrument as well as
the legal areas of Brazilian law. Ontologies
constitute the knowledge representation standard for
the Semantic Web (Antoniou and Van Harmelen,
2004)(Shadbolt and Berners-Lee, 2006) and the
ONTOJURIS ontology is written in OWL, according
to the W3C recommendation.
The Infonorma system was developed under the
guidelines of the Multi-Agent Application
Engineering Methodology (MAAEM) (Lindoso and
Girardi, 2006). The MAAEM methodology is
composed by three modeling phases: application
analysis, design and implementation. The analysis
and design of Infonorma are discussed in (Drumond
and Girardi, 2008). This paper introduces the
implementation phase of the Infonorma multi-agent
recommender system.
This paper is organized as follows. Section 2
presents an overview of the MAAEM methodology.
Section 3 introduces the Infonorma System with a
brief discussion of its specification and design.
Section 4 discusses the implementation phase of
Infonorma according to the guidelines of the
MAAEM methodology. Section 5 presents a brief
discussion on related work. Finally, section 6
concludes this paper with some remarks on further
work being conducted.
2 OVERVIEW OF MAAEM
METHODOLOGY
MAAEM is a methodology for requirement analysis,
design and implementation of multi-agent
applications through the reuse of software artifacts
such as domain models, multi-agent frameworks,
pattern systems and software agents. MAAEM also
supports the development of applications without
reuse, as is the case of the development of
Infonorma. The ONTORMAS (“ONTOlogy driven
tool for the Reuse of Multi-Agent Systems”)(Girardi
and Leite, 2008)(Leite and Girardi, 2008) ontology
128
Drumond L., Girardi R., Maranhão D. and Abrantes G. (2009).
IMPLEMENTATION ISSUES OF THE INFONORMA MULTI-AGENT RECOMMENDER SYSTEM.
In Proceedings of the 11th International Conference on Enterprise Information Systems - Software Agents and Internet Computing, pages 128-133
DOI: 10.5220/0002008701280133
Copyright
c
SciTePress
along with the Protégé platform is used as an
ontology-driven modeling tool and a storage
repository for products constructed on the Multi-
agent Domain Engineering and Multi-agent
Application Engineering processes. All the models
presented in this article were built using
ONTORMAS.
For the specification of a design solution, roles
are assigned to agents structured and organized into
a particular multi-agent architectural solution
according to non-functional requirements. The
interactions between the agents must be modeled as
well as the knowledge shared by them.
Application Analysis is performed through the
following modeling tasks: concept modeling, goal
modeling, role modeling, role interaction modeling
and user interface prototyping.
Application Design approaches the architectural
and detailed design, defining a solution to the
requirements specified in the analysis phase.
Application Implementation approaches the
mapping of design models to agents, behaviors and
communicative acts, concepts involved in the JADE
framework (Bellifemine et al., 2003), which is the
adopted implementation platform. The modeling
tasks and respective products of the implementation
phase of the MAAEM methodology are shown in
Table 1.
The information source used by Infonorma is a
legislative repository, a repository composed by
normative instruments. Such instruments have a type
(the nature of the instrument e.g. Ordinary Law,
Complementary Law, etc.) and legal branches in
which they are classified. The users specify their
interests in terms of such types and legal branches.
Figure 1 shows the goal model of Infonorma, in
which the general goal of providing legal normative
recommendations is achieved by three specific
goals, namely Model legal users, Content based
filter new legal information and Deliver
recommendations. The fulfillment of these specific
goals requires the exercise of some responsibilities
that are assigned to the agents in the system.
3 OVERVIEW OF INFONORMA
SPECIFICATION AND DESIGN
Infonorma is a system that provides its users with
personalized recommendations of legal normative
instruments. Legal normative instruments are
documents that compose the Brazilian jurisprudence.
Their structure is represented in the
ONTOJURIS domain ontology.
Each legal user has a profile composed by his/her
own interests and identification and represented by a
user model as an instance of ONTOJURIS.
The Infonorma system is composed by two kinds
of agents: one Filter agent, in charge of the
Legislative repository monitoring, Information items
classification into legal branches and Content based
similarity analysis responsibilities and one User
Modeler agent for each user, in charge of the
Explicit profile acquisition, User model creation and
Filtered information delivery responsibilities.
For a detailed description of the Infonorma
specification and design, please refer to (Drumond
and Girardi, 2008).
Table 1: Modeling tasks and products of the
Implementation phase of the MAAEM methodology.
Phases Tasks Products
Application Implementation
Mapping from
design to
Implementation
Agents and from
Responsibilities to
Behaviors
Model of
Agents and
Behaviors
Application Implementation
Model
Mapping of
Agents Interaction
and
Communication
Acts
Model of
Agent
Communicativ
e Acts
Agent
Implementation
Executable Software
Agents
4 IMPLEMENTATION OF
INFONORMA
As stated earlier, the Infonorma system was
implemented using the JADE platform. The
interactions of the user with the system take place
through a web site. The actions of the user in the
web site are recorded in a database monitored by the
User modeler agents. This section discusses the
main implementation issues of the agents of the
Infonorma system.
4.1 Mapping from Design to
Implementation Agents and from
Responsibilities to Behaviors
In this phase, the design agents are mapped to
implementation agents. As the implementation
IMPLEMENTATION ISSUES OF THE INFONORMA MULTI-AGENT RECOMMENDER SYSTEM
129
Figure 1: Goal Model of Infonorma.
platform adopted in Infonorma is the JADE
framework, each implementation agent is a JADE
agent. The responsibilities identified initially in the
Goal model are mapped to JADE behaviors and the
activities identified in Activity model are mapped to
JADE behavior methods.
The User Modeler agent is responsible for the
Explicit profile acquisition, Filtered information
delivery and User model creation responsibilities.
They were mapped, respectively, to the
AcquireUserProfiles, DeliverFilteredInformation
and CreateUserModels behaviors. The
AcquireUserProfiles behavior has one main method,
the validateProfile one, which should verify if new
profiles were specified and, in this case, fire the
CreateUserModels behavior. The
AcquireUserProfiles behavior was specified as a
One Shot behavior, since it is performed only once
for each new user model specified.
As for each new user profile, only one user
model must be created, the CreateUserModels
behavior also is One Shot. For each new profile
acquired, the computeAffinities method computes the
legal affinities, i.e. the level of interest, that the legal
user has with the legal branches. After that, the data
acquired in the profile and the computed affinities
are recorded as an instance of the User Model class
in the ONTOJURIS ontology, in the context of the
saveLegalUserModel method.
The DeliverFilteredInformation is a Cyclic
behavior because of the evaluateProposal method.
This method constantly checks the message queue of
the agent waiting for similarity analysis proposals.
For each proposal, this method checks whether the
user is interested in the type of the proposed
information item. When a proposal is accepted and
the filtered information items are sent by the Filter
agent, the DeliverFilteredInformation behavior
should produce the personalized recommendations
(constructRecommendationMessage method) and
make them available to the user
(deliverRecommendation method).
The Filter agent, responsible for the Legislative
repository monitoring, Information items
classification and Content-based similarity analysis
responsibilities, has the
MonitorLegislativeRepository,
ClassifyInformationItems and
FilterInformationItems behaviors respectively. Each
time the MonitorLegislativeRepository behavior is
performed, it must check if there are the new
information items (monitorInformationSource
method). This is done through the following query.
Query that checks whether there are new instruments of a
certain type in the legislative repository.
ICEIS 2009 - International Conference on Enterprise Information Systems
130
When new items are detected they are identified
and the current state of the legislative repository is
recorded (extractNewItems method). It may seem
natural to specify this behavior as a cyclic one, but,
as the legislative repository does not change
constantly, this behavior can be executed between
certain time intervals. Thus it was specified as
Ticker. As each information item must be classified
just one time, the ClassifyInformationItems behavior
was specified as a One Shot behavior. The
classification process determines the legal affinity
level of the normative instrument with each legal
branch in the ONTOJURIS ontology. It starts
counting the relevant keywords that are found in the
normative instrument (preprocessing method). Then,
the computeLegalAffinities method computes the
level of affinity with each legal branch and the
normalizeAffinities method normalizes the computed
values, so that they range from 0 to 1.
computeLegalAffinities Method from the
ClassifyInformationItems behavior.
Finally, the proposeSimilarityAnalysis method
sends to every User Interface agent a proposal with
the type of the current normative instrument.
When the information items are classified, they
are filtered by the FilterInformationItems behavior, a
One Shot behavior, since it is executed only once for
each new information item. The similarityAnalysis
method computes the similarity analysis based on
the closeness between the user model and the
information item. This closeness is measured by a
function that measures the distance between them.
For a more detailed description of these functions,
please refer to (Drumond and Girardi, 2008). The
methods that compute both the closeness and the
distance between normative instruments are the ones
that follow.
Methods for computing the closeness and the distance
between information items and user models.
4.2 Mapping from Agent Interaction to
Communicative Acts
Figure 2 shows the Model of Agent Communicative
Acts of Infonorma. This model describes how
knowledge of the multiagent society is exchanged by
the agents through messages sent from one to
another. The messages from the Agent Interaction
Model are mapped to FIPA-ACL messages.
First of all, when new information items are
available, the Filter agent sends a
PROPOSE(normativeInstrumentType) message to
each User Modeler agent in context of the
ProposeSimilarityAnalysis method. This message
means that Filter agent is proposing to perform the
similarity analysis of a normative instrument, which
type is specified in the message, with each user
model. The User Modeler agent evaluates the
proposal and checks whether its respective user has
IMPLEMENTATION ISSUES OF THE INFONORMA MULTI-AGENT RECOMMENDER SYSTEM
131
Figure 2: Model of Agent Communicative Acts of Infonorma.
interest in the type of the instrument through the
evaluateProposal method. If so, it accepts the
proposal sending an
ACCEPT_PROPOSAL(ONTOJURIS user model
instance) message which contains an instance of an
ONTOJURIS user model. Otherwise the User
Modeler agent just ignores the message.
Each ACCEPT_PROPOSAL(ONTOJURIS user
model instance) message received by the Filter agent
is answered with an INFORM_REF(Filtered
information) message which content is a triple
<ONTOJURIS user model, normative instrument,
similarity> representing the result of similarity
analysis, performed by the similarityAnalysis
method, between the normative instrument which
type was in the
PROPOSE(normativeInstrumentType) message and
user model specified in the answer to the first
message. This way, the User Modeler agent is able
to produce the personalized recommendations
(constructRecommendationMessage method) and
deliver them to the user.
The JADE platform provides a way to visualize
the messages exchanged by the agents using the
Sniffer agent.
5 RELATED WORK
Previous work on Infonorma has already been
published. The specification of the system is
discussed in (Drumond, Girardi and Leite, 2007)
while its architectural design is introduced in
(Drumond, Girardi and Leite, 2007b). Both the
specification and design of the system were revised
and updated and are presented with more details in
(Drumond and Girardi, 2008), which also discusses
the ONTOJURIS ontology and the detailed design of
the system.
Work has been done on the usage of artificial
intelligence techniques to improve the effectiveness
of recommender systems. Some of these approaches
use ontologies, like the one in (Middleton et al.,
2002) and (Middleton et al., 2004) and machine
learning, as in the PersoNews system (Banos et al.,
2006).
The ROSA system (Girardi and Ibrahim, 1995)
uses semantic cases to represent natural language
queries and natural language descriptions of
software components in an information retrieval
system. Infonorma uses this approach adapted to
information filtering. Besides that, ROSA
represented semantic cases with frames while
Infonorma considers the semantic cases as instances
of an ontology.
There has been much work done in the domain
of Artificial Intelligence and Law. Pinkwart et. alii
(Pinkwart et al., 2006) developed a tutor system that
uses collaborative filtering to identify weak points in
students legal arguments. The development and
usage of legal ontologies to represent and access
legal information has been approached in
(Kralingen, 1995), (Tiscornia, 2001), (Valente,
1995) and (Visser, 1995). Benjamins et. alii
(Benjamins et al., 2005) introduce an overview of
the application of Semantic Web technologies to the
legal domain. A formal criteria for describing legal
information on different levels is presented in
ICEIS 2009 - International Conference on Enterprise Information Systems
132
(Tiscornia, 2001) in the context of the Italian project
Norme in rete (Law in the net).
6 CONCLUSIONS
This work discussed the implementation of the
Infonorma multi-agent recommender system. This
system is composed by two kinds of agents: the User
Interface and the Filter agents which perform
content based filtering and explicit user modeling for
recommending legal information items.
Infonorma was developed under the guidelines of
MAAEM, a methodology for the development of
multi-agent systems, through the instantiation of
ONTORMAS, an ontology-driven tool which guides
the specification and design of multi-agent domain
models and applications.
We are currently working on the evaluation of
the effectiveness of the Infonorma system, as well as
its extension with collaborative and hybrid filtering
approaches and implicit profile acquisition. To this
extent, a semantic portal where users can navigate
through legal information and visualize the
recommendations generated by Infonorma and other
legal information systems is under development.
REFERENCES
Adomavicius G., Tuzhilin A.: Toward the next generation
of recommender systems: a survey of the state-of-the-
art and possible extensions. IEEE Trans Knowl Data
Eng. 17(6), 734--749 (2005)
Antoniou G., Van Harmelen F.: Semantic web primer,
MIT Press (2004)
Balabanovic, M., Shoham, Y.: Fab: content-based,
collaborative recommendation. Commun ACM 40(3),
pp. 66--72 (1997)
Banos, E., Katakis, I., Bassiliades, N., Tsoumakas, G.,
Vlahavas, I.: PersoNews: A Personalized News Reader
Enhanced by Machine Learning and Semantic
Filtering. In: Proc. 5th Int. Conf. on Ontologies,
DataBases, and Applications of Semantics. LNCS,
vol. 4275, pp. 975--982. Springer-Verlag, Montpellier,
France (2006)
Bellifemine, F., Caire, G., Poggi, A., Rimassa, G.: JADE a
white paper. Exp vol. 3 n. 3, http://
www.jade.tilab.com/ (2003)
Benjamins, V., Casanovas, P., Breuker, J., Gangemi, A.:
Law and the semantic web, an introduction. Lect
Notes Comput Sci 3369, pp. 1--17 (2005)
Drumond, L., Girardi, R.: A Multi-agent Legal
Recommender System. J. of AI and Law. 16, 175-207.
Springer (2008)
Drumond, L., Girardi, R., Leite, A.: A Case Study on the
Application of the MAAEM Methodology for the
Specification Modeling of Recommender Systems in
the Legal Domain. In Proceedings of the 9th
International Conference on Enterprise Information
Systems, pp. 155--160. INSTICC, Funchal, Portugal
(2007)
Drumond, L., Girardi, R., Leite, A.: Architectural Design
of a Multi-Agent Recommender System for the Legal
Domain. In Proceedings of the 11th International
Conference on Artificial Intelligence and Law, pp.
183--188. ACM Press, New York (2007b)
Girardi, R., Ibrahim, B.: Using English to Retrieve
Software. J Syst Software, Special Issue on Software
Reusability 30(3), pp. 249--270 (1995)
Kralingen, R.: Frame-based conceptual models of statute
law, Computer/Law Series 16 (1995)
Girardi, R., Leite, A.: A knowledge-based tool for multi-
agent domain engineering. Knowledge-Based
Systems, vol. 21, n. 7, pp. 604—611 (2008)
Leite, A., Girardi, R.: ONTORMAS: Uma ferramenta
dirigida por ontologias para a Engenharia de Domínio
e de Aplicações Multiagente. The XI Iberoamerican
Workshop on Requirements Engineering and Software
Environment (2008)
Lindoso, A., Girardi, R.: The SRAMO technique for
analysis and reuse of requirements in multi-agent
application engineering. IX workshop on requirements
engineering, Cadernos do IME, UERJ Press, 20, pp.
41--50. Rio de Janeiro (2006)
Middleton, S., Alani, H., Shadbolt, N., De Roure, D.:
Exploiting synergy between ontologies and
recommender systems. In Proceedings of the WWW
International Workshop on the Semantic Web, vol. 55
of CEUR Workshop Proceedings, Maui, USA (2002)
Middleton, S., Shadbolt, N., De Roure, D.: Ontological
User Profiling in Recommender Systems. ACM
Transactions on Information Systems, 22, pp. 54--88
(2004)
Pinkwart, N., Aleven, V., Ashley, K., Lynch, C.: Using
collaborative filtering in an intelligent tutoring system
for legal argumentation. In: Weibelzahl, S., Cristea, A.
(eds.) Proceedings of workshops held at the 4th
international conference on adaptive hypermedia and
adaptive web-based systems. Lecture notes in learning
and teaching, Dublin, Ireland, pp. 542—551 (2006)
Shadbolt, N., Hall, W., Berners-Lee, T.: The semantic web
revisited. Intelligent Syst 21(3), p. 96–101 (2006)
Tiscornia, D.: Ontology-driven access to legal
information. DEXA 12th international workshop on
database and expert systems applications, pp. 792
(2001)
Valente, A.: Legal knowledge engineering: a modelling
approach. University of Amsterdam, the Netherlands,
IOS Press, Amsterdam, The Netherlands (1995)
Visser, P.: Knowledge specification for multiple legal
tasks; a case study of the interaction problem in the
legal domain, computer/law series, n. 17, Kluwer Law
International, The Hague, The Netherlands (1995)
IMPLEMENTATION ISSUES OF THE INFONORMA MULTI-AGENT RECOMMENDER SYSTEM
133