ANALYZING WEB CHAT MESSAGES FOR
RECOMMENDING ITEMS FROM A DIGITAL LIBRARY
Stanley Loh
Catholic University of Pelotas – UCPEL, Brasil
Lutheran University of Brazil – ULBRA, Brasil
Ramiro Saldaña, Daniel Licthnow, Thyago Borges, Roberto Rodrigues,
Gabriel Simões, Leonardo Albernaz Amaral, Tiago Primo
Catholic University of Pelotas - UCPEL, Brasil
Keywords: text mining, recommendation, ontologies, chat, textual messages
Abstract: This work presents a recommender system that analyzes textual messages sent during a communication
session in a private Web chat, identifies the context of each message and recommends items from a Digital
Library. Recommendations are directly made to users in the chat screen and are decided by a software
system through a proactive paradigm, without any request of the users. A domain ontology, containing
concepts and a controlled vocabulary, is used to identify subjects in textual messages and to automatically
classify items of the Digital Library.
1 INTRODUCTION
According to NONAKA & TAKEUCHI (1995), the
majority of the organizational knowledge comes
from interactions between people. People tend to
reuse solutions from other persons in order to gain
productivity.
People communicate using synchronous
interactions (e.g., exchange of messages in a chat),
asynchronous interactions (e.g., electronic mailing
lists or forums), direct contact (e.g., two persons
talking) or indirect contact (when someone stores
knowledge and others can retrieve this knowledge in
a remote place or time).
Recommender systems can help in processes of
knowledge exchange and acquisition. A
recommender system is a software whose main goal
is to aid in the social process of indicating or
receiving indication about what options are better
suited in a special case for a certain individual
(RESNICK & VARIAN, 1997). The main goal is to
locate information sources related to a person’s
interest or need (MONTANER ET AL., 2002).
Recommendations are broadly used in electronic
commerce for suggesting products or providing
information about products and services, helping
people to decide in the shopping process
(LAWRENCE ET AL., 2001) (SCHAFER ET AL.,
2001).
Recommender systems are proactive devices,
because they can supply information without people
having to search, query or look for it. The offered
gain is that people do not need to request
information, but a software system decides what and
when to suggest. This kind of system is especially
useful when there are many options to choose and
users have little information about those options.
This work presents a recommender system to
support people when using a Web chat for
exchanging knowledge. Recommendations are made
by the system during the online discussion,
analyzing the context of the textual messages
exchanged by the chat participants. To decide what
41
Loh S., Saldaña R., Licthnow D., Borges T., Rodrigues R., Simões G., Albernaz Amaral L. and Primo T. (2004).
ANALYZING WEB CHAT MESSAGES FOR RECOMMENDING ITEMS FROM A DIGITAL LIBRARY.
In Proceedings of the Sixth International Conference on Enterprise Information Systems, pages 41-48
DOI: 10.5220/0002627900410048
Copyright
c
 SciTePress

to recommend, the system uses an ontology and a
Digital Library.
The paper is structured as follows. Section 2
discusses the proactive strategy used by
recommender systems to retrieve information.
Section 3 presents related works about
recommendations. Section 4 describes the proposed
system in details, including the functionality of each
component of the system. Finally, section 6 presents
concluding remarks and future work.
2 RECOMMENDATIONS: A
PROACTIVE STRATEGY FOR
INFORMATION RETRIEVAL
There are information retrieval systems to help
people to find information in digital libraries
(SPARCK-JONES & WILLET, 1997). Most of
these systems demand people to state an information
need as a query (in a formal language).
Taylor (apud OARD & MARCHIONINI, 1996)
defines 4 types of information need:
• visceral need: when the need is not consciously
perceived;
• conscious need: when the user perceives
his/her need and knows what he/she wants;
• formalized need: when the user expresses
his/her need in a formal way;
• compromising need: when the need is
represented in the system.
For using information retrieval systems, the user
has to formalize his/her information need
(formalized or compromising need). The problem is
that people are not able to specify what they need in
formal languages, because that is exactly what is
missing. Belkin and others (BELKIN ET AL., 1997)
define the information need as an Anomalous State
of Knowledge (ASK). Thus, any information need is
inherently hard to specify.
Another problem is that sometimes people do not
have a explicit knowledge (awareness) about what
they need (visceral need).
CHOUDHURY & SAMPLER (1997) identified
2 different processes for information acquisition:
reactive and proactive. In the first case, the user
knows what he/she needs and is able to specify what
is looking for. In the second case, the proactive one,
the user does not have a specific goal and the
purpose is to explore or monitor some situation, in
order to discover something new. The typical
characterization of this situation is when the user
asks “tell me something new or useful”. Proactive
strategies are suited for users with visceral needs.
An example of a proactive strategy is the work of
SWANSON & SMALHEISER (1997). They use
common words to relate texts and find analogies.
Their work has reached success discovering new
hypothesis in the medicine area.
Traditional information retrieval systems use the
reactive strategy. They consider that users have a
clear definition of what they are looking for and that
users are able to specify this using a formal language
(for example, keywords and Boolean operators).
Some recent researches are dealing with
proactive strategies for information retrieval through
the use of recommendations.
Recommender systems help users with conscious
or visceral needs. In the first case (conscious need),
recommendations help the user to find what he/she
wants without having to specify or formalize a
query. A software system automatically identifies
the user’s need, goal or interest and searches for
items that may be useful. This identification is made
by analyzing the user’s actions (as for example,
navigation across a web site).
In the second case (visceral need), a
recommender system may help the user by finding
useful and new items, without any intervention of
the user. Characteristics of the user (his/her profile)
or the user’s history (items bought, read, borrowed)
may be analyzed in order to identify the user’s
interesting areas or his/her background knowledge.
After that, the software system can look for new
items inside that context.
The recommender system presented in this paper
uses a proactive strategy for help users in these two
cases. Analyzing messages sent by the user, the
software system can identify the user’s interest (a
possible information need) and then select items
from a digital library to suggest to this user.
By other side, when one user sends a message,
all other users may receive recommendations,
because the system admits that the participants of
the discussion are also interested in the same subject.
Of course, the system first analyzes the user’s profile
for not recommending twice the same item or for not
recommending basic items for advanced users. This
way, the user, even without making any action, may
receive a suggestion.
3 RELATED WORK ABOUT
RECOMMENDATIONS
Recently, recommender systems are being used to
support knowledge acquisition. BRUSILOVSKY
(1996) discuss applications of adaptive hypermedia
systems (a kind of recommender systems) in
ICEIS 2004 - SOFTWARE AGENTS AND INTERNET COMPUTING
42

educational environments, to support students in
learning processes.
According to TERVEEN & HILL (2001), there
are 4 kinds of recommender systems. Content-based
systems use only customers’ preferences. Items to be
recommended are chosen from those similar to the
ones related to the customer, for example, products
in stock that are classified in the same section of the
products bought by the customer. Recommendation
support systems do not make automatically
recommendations but help people to produce
recommendations. Social data mining systems
discover preferences analyzing records from social
activities, like messages in newsgroups, citations in
scientific papers, usage logs of a system, peer-to-
peer services (like exchange of music and
documents), etc. Collaborative filtering systems do
not consider the content of the items but the
similarity among people and the items related to
them; the goal is to find people with similar
preferences and make cross-recommendations.
GroupLens system uses collaborative filters to
help people to find useful information (RESNICK
ET AL., 1994). The technique collects user’s
feedback to select new articles that can be
interesting to the user.
TERVEEN & HILL (2001) discuss the
PHOAKS system, which extracts addresses of Web
pages from messages in the Usenet newsgroup for
future recommendations.
Other system (proposed by Donath et al. apud
TERVEEN & HILL, 2001) analyzes messages in the
Usenet and other chats intending to later recommend
group of messages according to some attributes (for
example, presence of certain themes or discussions
with greater number of participations).
The TAPESTRY system allows people to
evaluate messages or news and it allows an user to
retrieve items based on content or in collaborative
evaluations (GOLDBERG ET AL., 1992).
Unfortunately, this system does not act in proactive
way, because users have to retrieve items using
queries.
KOMOSINSKI ET AL. (2000) presents a system
that identifies terms in messages of a chat and gives
to the participants the definition of the terms during
the interaction.
The system presented in the current paper is
different from the others because it analyzes the
messages exchanged in the chat and identifies the
context of the discussion for then selecting items
from the digital library to recommend particularly to
each user.
4 DESCRIPTION OF THE SYSTEM
The goal of the present recommendation system is to
provide people with useful information during a
collaboration session. To do that, the system
analyzes textual messages sent by users when
interacting in a private Web chat, identifies
subjects/themes/concepts inside the messages and
recommends items cataloged in a digital library,
previously classified in the same subjects. The
Digital Library contains electronic documents.
A Text Mining module analyzes each message.
The words present in the message are compared
against the domain ontology. After, it passes the
concept identified to the Recommender module, that
looks in the library for items to suggest.
According to the classification of TERVEEN &
HILL (2001), the system is a content-based
recommender system because the context of the
messages is matched against the content of items in
the database. The system is also a social system
because it analyzes messages exchanged in a Web
chat. And the system uses collaborative filtering
because the database is created by people, especially
the digital library, where users of the community can
upload items into.
One difference of the proposed system from
others is that it is not necessary to store a profile for
a user to use the system and receive
recommendations. Messages sent by users are
enough for the system deciding what recommend.
In the next sections, each component of the
system is described in details. All the system was
implemented with PHP, Javascript and MySQL.
4.1 The Web Chat
The chat works like traditional ones in the Web. The
difference is that it is specially constructed for this
system and it is not open to non-registered users.
Thus, users have to be authenticated for using the
system. There is no limit for the number of persons
interacting at the same time.
At the moment, only one chat channel is allowed.
Thus a discussion session concerns all messages sent
during a day. In the future, this restriction will be
eliminated.
4.2 Message Analysis
The main component of the system is a Text Mining
Module. It works as a sniffer, examining each
message sent in the chat. This module is responsible
for identifying subjects in the messages. Subjects are
identified by comparing words present in the
ANALYZING WEB CHAT MESSAGES FOR RECOMMENDING ITEMS FROM A DIGITAL LIBRARY
43

message against terms defined in the ontology.
Generic terms like prepositions (called stopwords)
will be discarded. Each message is compared online
against all concepts in the ontology. The concepts
identified in the message represent user’s interests
and will be forwarded to the Recommender Module.
The text mining method employed in this system
(a kind of classification task) was first presented in
LOH ET AL. (2000). Instead of using Natural
Language Processing (NLP) to analyze syntax and
semantics, the method is based on probabilistic
techniques: subjects can be identified by cues. Using
a fuzzy reasoning about the cues found in a text, it is
possible to calculate the likelihood of a subject being
present in that text.
The algorithm is based on Rocchio’s and Bayes’
algorithms (ROCCHIO, 1966; RAGAS & KOSTER,
1998; LEWIS, 1998), since it uses a prototype-like
vector to represent texts and concepts. The method
evaluates the relationship between a text and a
concept of the ontology using a similarity function
that calculates the distance between the two vectors.
The vectors representing texts and concepts are
composed by a list of terms with a weight associated
to each term. In the case of texts, the weight
represents the relative frequency of the term in the
text (number of occurrences divided by the total
number of terms in the text). And the weight in the
concept vector represents the probability of the term
being present in a text of that subject. The next
section (the ontology) describes how concept
weights are defined.
The text mining method compares the vector
representing the text of a message against vectors
representing concepts in the ontology. The
comparison between the two vectors is done through
a fuzzy reasoning process, following (ZADEH,
1973) and (NAKANISHI et al., 1993). In the
comparison method, weights of common terms
(those present in both vectors) are multiplied. The
overall sum of these products, limited to 1, is the
degree of relation between the text and the concept,
meaning the relative probability of the concept
presence in the text or that the text holds the concept
with a specific degree of importance. The decision
concerning if a concept is present or not depends
then on the threshold used to cut off undesirable
degrees. This threshold is previously set by an
expert.
The method is based on the relevancy index
proposed in (RILOFF & LEHNERT, 1994) whose
definition is "a collection of features that, together,
reliably predict a relevant event description". Some
terms may indicate the presence of a subject with a
degree of certainty. Therefore the fuzzy reasoning
process must evaluate the likelihood of a concept
being present in a text, analyzing the strength of its
indicators. The process is like an abductive
reasoning. According to GULLA (1997), in a
deduction, if “A Æ B” and “A is truth" then we can
infer “B is truth”. In abduction, if “A Æ B” and “B
is truth” then “A is a probable cause for B being
truth”. This means that if words describing a concept
c
appear in a text, there is a high probability of that
the concept c
will be present in that text.
When two or more concepts are identified in the
same message, the degrees of relationship between
the message and the concepts are used to form a
ranking. Only the top concept in the ranking is
considered. If two concepts are identified with the
same degree, but one is “father” of the second in the
ontology, only the more specific concept (son) is
considered.
New terms, used in the messages but not present
in the ontology, are stored for future analysis.
4.3 The Ontology
A domain ontology is a description of “things” that
exist or can exist in a domain (SOWA, 2002). It is a
formal and explicit definition of concepts (classes or
categories) and their attributes and relations (NOY
& McGUINNESS, 2002).
A domain ontology is similar to a thesaurus. In
fact, a thesaurus is a kind of ontology, but this will
not be discussed in this paper (see GILCHRIST,
2003, for a detailed discussion). According to
FOSKET (1997), thesaurus is a device to control
terms in texts. Thesauri provide knowledge maps,
representing concepts or ideas of the application
domain and indicating relations among them. A
thesaurus also defines terms used to describe a
concept.
In the proposed system, the ontology is
implemented as a set of concepts in a hierarchical
structure (a root node, fathers and sons). Each
concept has associated a list of terms and their
respective weights that help to identify the concept
in texts (messages or electronic documents).
Weights are used to state the relative importance or
the probability of the term for identifying the
concept in a text. The relation between concepts and
terms is many-to-many, that is, a term may be
present in more than one concept and a concept may
be described by many terms.
The ontology is used to identify subjects in
textual messages and to automatically classify items
of the digital library.
A software tool is used to manage and configure
the ontology, including functions to visualize the
structure of concepts and the list of terms, to insert a
new concept and its respective list of terms, to
insert/remove terms and to modify the weights. A
ICEIS 2004 - SOFTWARE AGENTS AND INTERNET COMPUTING
44

group of administrators is responsible for creating
and updating the ontology. New terms, found by the
Text Mining module, may be added as a new
concept, or inside an existing concept or the new
term may be added to the stopword list.
Currently, the system uses a domain ontology for
Computer Science, but others can be used. For this
purpose, the domain ontology has a root node called
“ontology”. Under this node, other ontologies may
be aggregated.
Concepts and the hierarchy were based on the
ACM classification for Computer Science. Software
tools supported people in identifying terms for each
concept. Terms and weights were defined using the
supervised learning strategy (for machine learning):
experts selected texts about a concept
(approximately 100 per concept) and a software tool
identified the most important terms for each class.
The texts where extracted from the Research Index
database (www.researchindex.com). A probability
measure was used to define the weight of each term
in a concept. Stopwords (prepositions and others)
were removed.
Furthermore, experts reviewed the ontology
eliminating terms present in many concepts and
adding word variations with the same weight as the
principal. This last task was important since texts
were written in English and Portuguese. So, the
terms used in the ontology come from these two
languages.
4.4 The Digital Library
The Digital Library used by the recommender
system is a repository of electronic documents.
The inclusion (upload) of items in the Digital
Library is responsibility of authorized people and
can be made offline in a specific module.
A module for upload of items from non-
authorized people is being developed, so that these
items can be approved or rejected later by referees.
The classification of the electronic documents is
made automatically by software tools, using the
same text mining method used in the Text Mining
Module. A difference is that a document may be
related to more than one concept. A threshold is
used to determine which concepts can be related to
one document. Thus, the relation between concepts
and documents in the Digital Library is many-to-
many. The relationship degree is also stored.
When inserting a item in the base, the user have
to designate whether the item is basic, intermediate
or advanced. This information will be used later by
the recommender module to avoid suggesting basic
items to experienced users.
4.5 Recommendations
The goal of the Recommender Module is to offer
electronic documents, stored in the Digital Library,
to the chat participants. The module uses a content-
based technique, where only items classified in the
concept identified in the message are recommended.
The action of this module starts when it receives
a concept from the Text Mining Module. Then, it
searches the Digital Library for items classified in
the same concept. Each time the Text Mining
Module identifies a concept in a message, it sends
this concept to the Recommender Module. Similarly,
the searches happen online, that is, immediately
when the Recommender Module receives a concept.
Since the discussion in the chat is synchronous,
recommendations should not interrupt the users. So,
indications are given in a separate frame and not
inside the chat window. Recommendations are
particular of each user. Thus, each user receives a
different list of suggestions in the screen.
The Recommender Module uses a Profile Base
for registering information about each user as
demographic characteristics and his/her history in
the system, like items read from, uploaded to or
downloaded from the Digital Library.
The system also uses a punctuation schema to
identify interesting areas for each user and to
determine his/her degree of knowledge about
subjects (represented by concepts of the ontology).
Each operation of the user inside the system gives
him/her some points in the profile. Operations that
punctuate are: download and upload of items in the
Digital Library, opening a PDF file, participating in
a discussion, sending messages and the authorship of
electronic documents stored in the Digital Library.
The knowledge degree is used for not
recommending basic items for advanced users and to
determine authorities in each subject.
Additionally, the system must not recommend:
a) the same item twice in the same section;
b) items already associated to the user
(downloaded, uploaded or read);
c) items marked by the user to not be
recommend (button “never show me again”).
As the recommendation list may increase without
limit, there are options for the user to reduce the
overload in the recommendation frame. For
example, there is an option to eliminate a item from
the session list (“don’t show me again in this
session”) and other option to set in the user profile
that the item should not be recommended again
(“never show me again”).
Furthermore, the user can access details of the
item being recommended as title, authors, keywords,
abstract and the related concepts with the degree of
ANALYZING WEB CHAT MESSAGES FOR RECOMMENDING ITEMS FROM A DIGITAL LIBRARY
45

relationship (this information is stored in the digital
library).
Figure 3 shows a snapshot of the system in a real
situation. There is an area where the nicknames of
the participants of the chat appear (users), an area
where the messages can be viewed (discussion), an
area where the recommendations appear individually
for each user (recommendations) and another area
for the user writing the messages (message).
The prototype system is accessible at
http://gpsi.ucpel.tche.br.
5 CONCLUSIONS
This work presented a recommendation system to
support knowledge exchange and acquisition in a
Web chat discussion. The system allows proactive
information retrieval because people receive
suggestions of electronic documents without having
to search the digital library
The main advantage of the system is to free the
user of the burden to search for information during
the online discussion. Users do not have to choose
attributes or requirements from a menu of options, in
order to retrieve items of a database. The system
decides when and what information to recommend
to the user. This proactive approach is useful for
non-experienced users that receive hits about what to
read in a specific subject. With the system, user’s
information needs are discovered naturally during
the conversation.
This system points to a new kind of support for
users in online discussions. Furthermore, Digital
Libraries get a new facet with the additional help of
recommender systems. Feigenbaum (apud DAVIES,
1989) compares the current libraries with future
ones. The existing ones consist of a warehouse of
passive objects. In the other side, libraries of the
future will be a collection of active documents,
helping people to discover new knowledge through
providing connections and associations previously
unknown, analogies and new concepts, without
people having to state clearly their information
needs.
At the moment, the ontology only concerns a
small set of the Computer Science area. It is possible
to extend this ontology or even to create and use
other domain ontologies with concepts from other
subject areas. Similarly, the current digital library
only has items related to Computer Science. For
other communities using the system, the digital
library has to be populated with items from other
areas.
Some experiments were carried out with the
recommendation system. Results reveal a promising
strategy, reducing the charge for the user find useful
information, especially when dealing with great
volumes of documents in Digital Libraries.
However, yet some poor recommendations are
being generated. This is due in great part to the
ontology. We detected some problems like, for
example, the use of generic words (like “software”)
appearing with high weight in different concepts;
lack of word variations and lack of specific concepts
(some concepts are too much broad, including
different items).
Quality of the recommendations can be improved
by developing better the ontology. For example, we
plan to use a stemming algorithm to treat word
variations. We are also investigating how to
automatically divide concepts in sub-concepts in
order to accommodate the specificity of the sub-
groups. A software tool is being developed to
identify ambiguous terms (that appear in many
concepts) and to automatically correct their weights.
Furthermore, recommendation quality is also
dependent on the quality of the digital library. For
example, few documents may cause poor
recommendations independently of the methods
used. We are studying an automatic software that
will search the Web for electronic documents under
concepts present in the ontology. After finding the
documents, other software will extract document
information as authors, title, abstract and keywords,
besides the existing automatic classification of the
document in the ontology concepts.
Regarding performance, it is possible to say that
the recommendations are made very quickly (in
milliseconds).
As a great number of recommendations may be
sent to users, the system allows the user to set a
threshold and only the items related to the concept
with a degree above this threshold will be presented
to the user.
The punctuation schema is still being tested. We
have to determine how many points will be given for
each operation. An ongoing implementation will
determine initial points for the user profile according
to his/her publications as specified in the curriculum
vitae.
Other future works include the implementation
of other recommendation techniques (like
collaborative filtering) and techniques to minimize
information overload due to the great number of
items recommended. A special study is being
conducted to use relevance feedback to narrow the
list of recommendations. Users would read some
items of the list and rate them, and the system would
use this information to eliminate items from the list
or to reorder the items in a new ranking.
An interesting future research is to use context in
interpreting each message. Currently, each message
ICEIS 2004 - SOFTWARE AGENTS AND INTERNET COMPUTING
46

is analyzed independently of others. This can lead to
mistakes in the subject identification. For example,
if one message has the words “neural nets” and the
next one has “learning”, the system should
recognize that “learning” is related to “machine
learning”, since it is assumed that the discussion
would not deviate from the context. However, in the
current state, the system may identify “learning” in a
context like “Computers in Education”. We are
studying a solution that considers the context of a
discussion, that is, the system will assume that
discussions do not move far from one subject to
other. Using the hierarchy of concepts, it is possible
to identify the distance of each movement from one
message to another, and this will be used to
disambiguate a message: when two or more subjects
are identified in a message, the nearest one must be
used.
ACKNOWLEDGEMENTS
This work is partially supported by CNPq, an entity
of the Brazilian government for scientific and
technological development.
REFERENCES
N. J. BELKIN; R. N. ODDY; H. M. BROOKS, ASK for
information retrieval: part I - background and theory,
in: K. SPARCK-JONES and P. WILLET, Peter, eds.,
Readings in Information Retrieval (Morgan
Kaufmann, San Francisco, 1997).
P. BRUSILOVSKY, Methods and techniques of adaptive
hypermedia, User Modeling and User Adapted
Interaction 6 (2-3) (1996) 87-129.
V. CHOUDHURY and J. L. SAMPLER, Information
specificity and environmental scanning: an economic
perspective, MIS Quarterly 21 (1) (1997) 25-50.
R. DAVIES, The creation of new knowledge by
information retrieval and classification, Journal of
Documentation 45 (4) (1989) 273-301.
D. J. FOSKET, Theory of clumps, in: K. SPARCK-
JONES and P. WILLET, eds., Readings in
Information Retrieval (Morgan Kaufmann, San
Francisco, 1997).
A. GILCHRIST, Thesauri, taxonomies and ontologies – an
etymological note, Journal of Documentation 59 (1)
(2003) 7-18.
D. GOLDBERG et al., Using collaborative filtering to
weave an information tapestry, Communications of the
ACM 35 (12) (1992) 61-70.
J. A. GULLA et al., An abductive, linguistic approach to
model retrieval, Data & Knowledge Engineering,
23(1) (1997) 17-31.
L. J. KOMOSINSKI et al., The usage of agents to support
dialog mediation between students via Internet, in:
Proc. RIBIE’2000 5th Iberoamerican Conference on
Informatics in Education, Viña del Mar, Chile,
December 2000. (in portuguese)
R. D. LAWRENCE et al., Personalization of supermarket
product recommendations, Journal of Data Mining and
Knowledge Discovery 5 (1/2) (2001) 11-32.
D. D. LEWIS, Naive (bayes) at forty: the independence
assumption in information retrieval, in: Proc.
European Conference on Machine Learning, Lecture
Notes in Computer Science, v.1398 (Springer, Berlin,
1998) 4-15.
S. LOH; L. K. WIVES; J. P. M. OLIVEIRA, Concept-
based knowledge discovery in texts extracted from the
Web, ACM SIGKDD Explorations 2 (1) (2000) 29-39.
M. MONTANER; B. LOPEZ; J. L. de la ROSA,
Improving case representation and case base
maintenance in recommender agents, in: Proc. 6th
European Conference on Case Based Reasoning
(2002).
H. NAKANISHI; I.B. TURKSEN; M. SUGENO, A
review and comparison of six reasoning methods,
Fuzzy Sets and Systems 57 (3) (1993) 257-294.
I. NONAKA and T. TAKEUCHI, The knowledge-creating
company: how japanese companies create the
dynamics of innovation. (Oxford University Press,
Cambridge, 1995).
N. F. NOY and D. L. McGUINNESS, Ontology
Development 101: a guide to creating your first
ontology (2002) Available at
http://protege.stanford.edu/publications/
D. W. OARD and G. MARCHIONINI, A conceptual
framework for text filtering. University of Maryland,
Technical Report EE-TR-96-25 (1996). Available at:
<http://www.ee.umd.edu/medlab/filter/>. Access in
June 1998.
H. RAGAS and C.H.A. KOSTER, Four text classification
algorithms compared on a Dutch corpus, in: Proc.
SIGIR’98 International ACM-SIGIR Conference on
Research and Development in Information Retrieval
(ACM Press, Washington, 1998) 369-370.
P. RESNICK et al., GroupLens: an open architecture for
collaborative filtering of Netnews, in: Proc.
Conference on Computer Supported Coooperative
Work (1994) 175-186.
P. RESNICK and H. VARIAN, Recommender systems,
Communications of the ACM 40 (3) (1997) 56-58.
E. RILOFF and W. LEHNERT, Information extraction as
a basis for high-precision text classification, ACM
Transactions on Information Systems 12 (3) (1994)
296-333.
J. J. ROCCHIO, Document retrieval systems -
optimization and evaluation, Ph.D. Thesis, Harvard
ANALYZING WEB CHAT MESSAGES FOR RECOMMENDING ITEMS FROM A DIGITAL LIBRARY
47

Computation Laboratory, Harvard University, Report
ISR-10 to National Science Foundation (1966).
J. B. SCHAFER et al., E-commerce recommendation
applications, Journal of Data Mining and Knowledge
Discovery 5 (1/2) (2001) 115-153.
J. F. SOWA, Building, sharing, and merging ontologies
(2002). Available at http://www.jfsowa.com/ontology
K. SPARCK-JONES and P. WILLET, Eds., Readings in
Information Retrieval, San Francisco, Morgan
Kaufmann (1997).
D. R. SWANSON and N. R. SMALHEISER, An
interactive system for finding complementary
literatures: a stimulus to scientific discovery, Artificial
Intelligence 91(2), (1997) 183-203.
L. TERVEEN and W. HILL, Beyond recommender
systems: helping people help each other, in: J.
CARROLL, ed., Human computer interaction in the
new millennium (Addison-Wesley, 2001).
L. A. ZADEH, Outline of a new approach to the analysis
of complex systems and decision processes, IEEE
Transactions on Systems, Man and Cybernetics SMC-
3 (1) (1973) 28-44.
Figure 1: A snapshot of the system in a real situation
ICEIS 2004 - SOFTWARE AGENTS AND INTERNET COMPUTING
48