USING LIGHTWEIGHT KNOWLEDGE MODELLING TO
IMPROVE PROACTIVE INFORMATION DELIVERY
Oleg Rostanin, Heiko Maus
German Research Instutute for Artificial Intelligence, DFKI GmbH, Trippstadterstr., 122, Kaiserslautern, Germany
Takeshi Suzuki, Kaoru Maeda
Ricoh Co. Ltd, Yokohama, Japan
Keywords: Proactive information delivery, Lightweight knowledge modeling, Concept-based task tagging.
Abstract: The current work presents an integrated solution for task-centric proactive information delivery (PID) in
agile knowledge working (AKW) environments. The approach exploits a lightweight incremental modeling
of task relevant knowledge domains and process know-how using concept maps together with concept-
based task tagging to improve the quality of PID results. The feasibility of the described approach was
proved during the joint research project TaskNavigator conducted by Ricoh Co. Ltd and DFKI GmbH.
1 INTRODUCTION
During the last decade, a plenty of approaches for
intelligent user assistance in knowledge intensive
working environments were developed. Knowledge
intensive work consists of both strictly structured
processes that can be formally modeled and enacted
using workflow management systems (WFMS) and
agile processes (agile knowledge work, AKW) that
are highly dynamic that makes them difficult to be
formalized (Elst, L.v. et al, 2003).
Our current work concentrates on the support for
AKW, e.g., developing a software or writing a
project proposal. Although AKW is dynamic, it is
required to be managed to be successfully completed
in time. Task list management (TLM) tools are often
used for flexible time management and planning in
AKW environments. TLM tool is an ideal place for
intelligent assistance, e.g., proactive information
delivery (PID) that is required by a knowledge
worker coping with tasks. Generally, PID has two
main purposes, i) minimize information overload by
providing information adapted to the current task’s
needs, ii) diminish users’ risk of overlooking
important documents relevant to their tasks. We
distinguish light- and heavy-weight PID depending
on the needed modeling effort (Holz, H. et al, 2005).
Heavyweight approaches to learning-on-the-job
(Rostanin, O. et al, 2006) aim to educate users by
providing information according to users’
information needs and skill level. Such approaches
claim to ensure a precise information delivery. Their
major problem is a relatively large effort on process,
user and information modeling needed to introduce
these methods in an enterprise.
We propagate a lightweight PID approach in
combination with a TLM system TaskNavigator to
cope with requirements from AKW. The main goal
of our research is to find means to combine
advantages of light- and heavy- weight PID, i.e., low
modeling effort and high delivery precision.
In this paper we describe concepts of lightweight
PID and task tagging (section 2). Using concept
maps to improve PID is discussed in section 3.
Evaluation results of our approach are depicted in
section 4. Conclusions are provided in section 5.
2 PID IN TASKNAVIGATOR
TaskNavigator is a web-based TLM system
providing support for knowledge intensive business
processes (Holz, H. et al, 2006). By the mechanism
of task delegation, task comments and notification as
well as flexible task structure management implied
611
Rostanin O., Maus H., Suzuki T. and Maeda K. (2010).
USING LIGHTWEIGHT KNOWLEDGE MODELLING TO IMPROVE PROACTIVE INFORMATION DELIVERY.
In Proceedings of the 2nd International Conference on Agents and Artificial Intelligence - Artificial Intelligence, pages 611-614
DOI: 10.5220/0002729106110614
Copyright
c
 SciTePress

by work breakdown structure (WBS),
TaskNavigator becomes a powerful tool for work
coordination and collaboration in distributed teams.
2.1 Lightweight PID
The main idea of PID in TaskNavigator is to
proactively deliver task-relevant information e.g.,
documents, e-mails, web-history, wiki pages related
to the task without explicit user request. The
principle of lightweight PID is based on the
assumption that a task can be described sufficiently
by task title, description and comments as well as
documents attached to the task. The PID module of
TaskNavigator generates a keyword-based query
from the current task context represented by task
description and its attachments and sends request to
external information retrieval (IR) systems
automatically to get task-relevant information.
Results from the IR systems are sorted by their
relevance to the query and presented to the user.
The core advantage of lightweight PID is the low
level of human effort needed to make it work: the
user just types a task name in TaskNavigator to get
first PID results. A formal model of task information
needs is not required.
The main problem (P1) of lightweight PID is
that statistics-based query generation used in
lightweight PID can cause unsatisfactory quality of
generated queries or search results:
(P1.1), TF/IDF algorithm used for a query
generation has limitations, e.g. for the task
“Introduce new employee” the keyword “new” is
regarded as a stop word and removed from the
query, although it is essential in the given situation.
(P1.2), Compound search terms: Even if the
algorithm could identify the importance of the
keyword “new” for the task, the keyword does not
have sense as a query term without considering it in
the combination with the keyword “employee”.
(P1.3), Verbose task descriptions can spoil
automatically generated query, e.g., for the task
“Create new DB for TouchMap weblog” with the
description “To install a new wordpress blog we
need a separate database on our mysql server” would
generate the query “create, db, touchmap, weblog,
wordpress, install, separate, database ...” that would
result in delivery of no or too many documents.
2.2 Task Tagging Improves PID
The objective of the TaskNavigator project was to
find an optimal solution that requires a minimally
possible modeling effort to achieve acceptable PID
results. Our claim here is that bottom up task
modeling realized by collaborative task tagging is
feasible and can improve PID results (C1).
Tagging is a wide-spread technology for
lightweight annotation of electronic resources by
manually or automatically assigning keywords to
them (Golder and Huberman, 2006). Considering
tasks in TaskNavigator as resources that are
annotated collaboratively by tags, we decompose C1
into the following sub-claims:
(C1.1) Task tags can be used as keywords to
refine a search query for task-related PID. Keywords
defined by users do not cause problems P1.1 and
P1.2 (if multi-word tags are allowed). The implicit
semantics behind task tags given by humans will
highlight the most important task aspects
suppressing the problem P1.3.
(C1.2) Provided the bag tagging model is used in
TaskNavigator, where different users can tag tasks
multiple times with the same tag, the popularity of
task-related tags can be used to specify weights of
single terms comprising a PID query. A weighted
query expresses the importance of each term thus
better specifying the task semantics (see P1.3).
(C1.3) Provided a list of tags of the parent task is
easy available in the current task details, the parent
task tags will ease the effort on current task tagging.
In order to implement this new vision on PID,
the process of the task-specific information delivery
will be extended as follows: i) Propose possible tags
to the user proactively; ii) User accepts/rejects tag
proposals or tags tasks manually (compound tags are
allowed); iii) In collaborative task management
environment, users can vote for or against task tags
assigned by themselves or by colleagues. iv) A new
PID query is generated by TaskNavigator
considering tags and tag votes as (compound) search
terms and their weights in the query.
Although task tagging can solve problems of
lightweight PID, there are severe problems going
along with tagging such as synonymy (P2.1),
homonymy (P2.2), polysemy (P2.3) - see (Goldman
06). In respect to the information retrieval, the
problem P2.1 (includes synonyms, misspelling,
different writing styles and different languages) is
the most critical. Provided the user tagged the task
with “digitalpaper”, documents containing “digital
paper” or “digitales Papier” (Ger.) will not be found
by the IR engine. The problem of homonymy can
emerge, for example, if the user tagged a task with
“SME” assuming “subject matter expert” but
received documents about “small and medium
enterprises”. The problem of polysemy is sometimes
difficult to recognize but it can spoil the IR results:
ICAART 2010 - 2nd International Conference on Agents and Artificial Intelligence
612

while expecting a description of the TouchMap
system, the user receives documents about the
TouchMap project which are related but not same.
3 TASK MODELING WITH
CONCEPT MAPS
A standard way of solving the problem of synonymy
and misspelling is to use mechanisms supporting
controlled vocabularies during the tagging process.
To solve homonymy and polysemy problems, more
sophisticated ontological modeling of the task-
relevant domains can be done. However, a sound
modeling of task context is practically impossible
for every task in TaskNavigator as they are mostly
ad-hoc in nature. Instead, we propose a method for
lightweight tasks modeling that realizes function of
the vocabulary control. The proposed solution is
based on the idea of concept maps.
LeCoOnt (http://lecoont.opendfki.de) is a web-
based tool for collaborative concept mapping
developed at DFKI (figure 1). It is aimed to combine
the graphical expressiveness and intuitiveness of
concept maps, a simple but well-defined information
model as well as vocabulary control to provide a
universal platform for lightweight knowledge
modeling using the concept map paradigm.
Figure 1: Task and domain modeling with LeCoOnt.
TaskNavigator integrates LeCoOnt as means to
control the vocabulary used for task tagging: the
LeCoOnt service realizes the auto-completion
function for manual tag input. The user can select
proposed concepts as task tags or create new tags.
Newly created tags are stored in the LeCoOnt
database as an unbound concept that can be later
used for domain modeling in LeCoOnt.
3.1 Concept-Based PID
By introducing the controlled vocabulary for task
tags, we are able to use it to identify concepts from
the knowledge base matching the current task and
thus not to rely on results of statistics-based
keyword extraction (P1.1-3). Figure 2 shows a
recommendation to add the concept “New
employee” as a tag for the task “Introduce John
Smith”. The user can tag the current task with
proposed concepts or attach concept information
items to the task. Figure 2 (left middle) illustrates
tags accepted by the user and attached to the task
“Introduce John Smith”.
The labels of attached concepts together with
their alternative labels will be used by the PID
engine to generate new queries. A simple query
expansion realized by using concept alternative
labels will ease the problem of synonymy (P2.1).
Furthermore, relations of the concept used for
tagging to other concepts in the knowledge base can
be exploited to disambiguate meanings of keywords
presented by tags and filter delivered documents.
3.2 Conceptual Task Modeling
Whereas the task tagging represents a bottom-up
approach to task modeling, the LeCoOnt tool can be
used as means to lightweight top-down task
modeling. In figure 1 an informal process model
“Introduce a new employee” created in LeCoOnt is
shown. Having attached the concept “Introduce a
new employee” as a task tag, a TaskNavigator user
can decompose the task into subtasks according to
the task model defined in the concept map. Created
subtasks will be automatically tagged by
corresponding concepts from the concept map and
inherit information items attached to the concepts.
4 EVALUATION
In order to show the feasibility of the approach, a
case study was conducted at the DFKI that lasted for
3 months. Totally, 11 subjects took part at the
experiment: 4 students, 9 researchers and 2
consultants. During the case study, users created 376
tasks as well as attached 624 documents and 164
comments to their tasks. We classified users in two
groups: 7 users those who used TaskNavigator for
part of their work and initiated 97% of the tasks; and
ii) the rest with rather short usage period small
number of own created tasks. The type of tasks
conducted with TaskNavigator ranged from personal
tasks such as workshop preparation or writing
publications to project tasks such as project
organization or customer relations.
Over the case study period, 458 tags were added
to tasks. During task tagging, 70 new concepts were
USING LIGHTWEIGHT KNOWLEDGE MODELLING TO IMPROVE PROACTIVE INFORMATION DELIVERY
613

Figure 2: Concept-based PID.
created. Considering both numbers of tasks and
given tags, each task got enriched description by 1.2
tags in average.
Over 80% of tags were reused by some means,
which means a number of tags being used in the
system is fairly maintained to reduce risks
introduced with tagging. Over half (54%) of the tags
were automatically provided by the system. Finally
24% of the tags were proposed by the concept-based
PID and added to tasks by users. For the controlled
tasks, the subjects compared the query generated
from the task’s textual context to the query
generated from the concepts attached to the tasks.
Once tags were available, usually the tag-based
query terms were rated better. The overall
impression of the subjects was, that both,
lightweight and tag-based PID compliment each
other, therefore, they should be used in combination.
5 CONCLUSIONS
The uniqueness of the TaskNavigator approach of
concept-based PID is in using lightweight concept
maps instead of formal ontologies to describe
knowledge domains and support task tagging.
According to our case study, a bearable user effort
spent for task tagging, either manual or supported by
the system, allows to improve results of PID as well
as to develop the corporative knowledge base.
As a feasibility test with real users showed, both
lightweight and extended PID approaches
complement each other and should be used together.
Whereas the concept based PID solves many
problems of lightweight one, lightweight PID can
help to solve the problem of a system cold start
specific to tag-based PID: if there are few concepts
available in the knowledge base, lightweight PID
keyword proposals can be used to initialize it.
Some conceptual aspects could not be tackled in
the project’s time frame: e.g., the PID engine used in
this work considers neither different user skill and
knowledge levels. Another critical issue is a
seamless integration into the user’s workspace.
REFERENCES
Elst, L.v., Aschoff, F.-R., Bernardi, A., Maus, H. and
Schwarz, S. (2003). Weakly-structured workflows for
knowledge-intensive tasks: An experimental
evaluation. IEEE WETICE Workshop on Knowledge
Management for Distributed Agile Processes
(KMDAP03). IEEE Comp. Press.
Golder, S.A. and Huberman, B.A. (2006). Usage patterns
of collaborative tagging systems. Journal of
Information Science, 32, 198-208.
Url: http://jis.sagepub.com/cgi/reprint/32/2/198
Holz, H., Maus, H., Bernardi, A. and Rostanin, O. (2005).
From Lightweight, Proactive Information Delivery to
Business Process-Oriented Knowledge Management.
Journal of Universal Knowledge Management, 2,
2005, pp. 101-127.
Holz, H., Rostanin, O., Dengel, A., Suzuki, T., Maeda, K.,
Kanasaki, K. (2006). Task-Based Process Know-how
Reuse and Proactive Information Delivery in
TaskNavigator. In Proc of 15
th
ACM Conference on
Information and Knowledge Management CIKM '06.
Arlington, USA.
Rostanin, O., Holz, H., Ullrich, C. and Song, Sh. (2006).
Project TEAL: Add Adaptive e-Learning to your
Workflows. In Tochtermann, K., Maurer, H. (Eds.)
Proc. of 6
th
International Conference on Knowledge
Management, I-KNOW'06. Graz, Austria.
ICAART 2010 - 2nd International Conference on Agents and Artificial Intelligence
614