ONTOLOGY-BASED EMAIL CATEGORIZATION AND TASK
INFERENCE USING A LEXICON-ENHANCED ONTOLOGY
Prashant Gandhi and Roger Tagg
School of Computer and Information Science, University of South Australia
Keywords: Information overload, Personal information management, Task inference, Ontology, Lexicon.
Abstract: Today’s knowledge workers are increasingly faced with the problem of information overload as they use
current IT systems for performing daily tasks and activities. This paper focuses on one source of overload,
namely electronic mail. Email has evolved from being a basic communication tool to a resource used – and
misused – for a wide variety of purposes. One possible approach is to wean the user away from the
traditional, often cluttered, email inbox, toward an environment where sorted and prioritized lists of tasks
are presented. This entails categorizing email messages around personal work topics, whilst also identifying
implied tasks in messages that users need to act upon. A prototype email agent, based on the use of a
personal ontology and a lexicon, has been developed to test these concepts in practice. During the work, an
opportunistic user survey was undertaken to try to better understand the current task management practices
of knowledge workers and to aid in the identification of potential future improvements to our prototype.
1 INTRODUCTION
The problem of information overload is a multi-
faceted one, covering a wide range of technical and
social issues (Spira and Goldes, 2007). As a result,
research into easing it is complex and wide-ranging.
The work described in this paper is part of a
larger project named Virtual Private Secretary
(VPS). The theme in VPS is to apply IT in a way
much like a human secretary would organize the
work of her principal, i.e. through a knowledge of
the principal’s work structures and task types.
This paper focuses on the growing problem of
email overload that affects most knowledge workers
today. Though overload arises from a number of
different sources, it was considered imperative to
focus on email, since email is widely considered as
one of main contributors to information overload
incidents, 60% according to a recent study (Mulder
et al, 2006).
The current work, therefore, specifically
explores the use of ontology concepts and lexicons
for categorizing email messages around personal
work topics, whilst also inferring, from the emails,
any tasks that users need to act upon. An objective
here is to wean the user away from the traditional
(often cluttered) email inbox toward an environment
where sorted and prioritized lists of tasks can be
presented. A prototype email agent has been
developed to test these concepts in practice. In
addition, a user survey was undertaken to understand
the current task management practices of knowledge
workers and to aid in identifying future
improvements to the prototype system.
The remainder of the paper is organized as
follows. Section 2 gives more detailed motivation
for the work. Section 3 discusses the different
research approaches possible. Section 4 describes
the prototype itself. We conclude with a brief
evaluation, some reflections and some ideas for
future work.
2 MOTIVATION
Current opinion, e.g. (Hall, 2004; Dabbish et al,
2005; Spira and Goldes, 2007) suggests that there
are serious weaknesses in today’s commercial
personal information management (PIM) tools such
as Microsoft’s Outlook, IBM’s Lotus Notes
package, and Qualcomm’s Eudora Pro. Although
these tools typically provide spam-filtering
mechanisms and means for manually organizing
email messages into folders, and features for
creating rule-based filters, these are tedious and
cognitively demanding to use (Ducheneaut and
Bellotti, 2001).
102
Gandhi P. and Tagg R. (2009).
ONTOLOGY-BASED EMAIL CATEGORIZATION AND TASK INFERENCE USING A LEXICON-ENHANCED ONTOLOGY.
In Proceedings of the 11th International Conference on Enterprise Information Systems - Software Agents and Internet Computing, pages 101-107
DOI: 10.5220/0001996101010107
Copyright
c
 SciTePress

It is also often observed that these tools have
evolved from email clients into fuller PIMs, without
having been designed as such. In particular, the
integration of tasks with email is anything but
seamless.
At the same time, prototype systems developed
as a part of research initiatives have introduced
interesting strategies and techniques for handling
email workload. However, as of the time of writing,
none have led to widely adopted PIM improvements.
Our main research question is therefore as
follows: How can we develop a practical system that
can more accurately and reliably classify and
prioritize the task implications of email messages
around a user’s work activities in order to overcome
email overload?
Our methodology has been the development and
evaluation of a proof-of-concept prototype for
exploring the potential of using a personal ontology
(together with a set of lexical clues that indicate the
relevance of each ontology concept) for categorizing
email messages around user preferences and
identifying implied tasks from message content. We
then hope to understand the strengths and
weaknesses of this approach as a potential
component of next generation PIMs.
3 DIFFERENT APPROACHES TO
CATEGORIZING MESSAGES
The three main approaches appear to be machine
learning-based, ontology-based, and sender-assisted
techniques. Two other approaches are those based
on sender identity and those using threads.
3.1 Machine Learning Based
Categorization solutions based on machine learning
techniques have been dominant in the research
community (Sebastiani, 2002). This involves the
application of artificial intelligence (AI) theories to
build ‘intelligent’ software agents that can be trained
to make categorization decisions on a user’s behalf.
A learning-based classifier for a category can be
built through an inductive process where the system
observes the characteristics of a training set of
messages. Manual intervention is limited to deciding
whether or not messages have been sensibly
categorized.
The machine learning based approach relies
solely on ‘endogenous’ knowledge i.e. knowledge
gained only from the documents themselves. The
user does not stipulate the categorization scheme.
Example prototypes of this type include Maxims
(Maes, 1994), MailCat (Segal and Kephart, 1999),
and IEMS (Crawford et al, 2006). (Corston-Oliver et
al, 2004) aimed to automatically identify tasks in
email messages using machine-learning techniques.
Although they require lower maintenance,
classifiers suffer from the ‘slow start’ problem, since
they can only gradually build-up their competency
as more examples are provided, over time. This
problem becomes worse the more categories there
are, as compared with the simple spam/not spam
situation.
3.2 Ontology Based
The potential of leveraging ontology structures for
classifying documents has been attracting increased
research activity in recent years. For mail
categorization, an ontology would contain the
structure of a user’s or a groups work topics, task
types, priorities etc. This can be built and edited
using an ontology editor.
Some representative prototypes include ECPIA
(Li et al, 2006); CLIPS (Taghva et al, 2003), the
latter being a hybrid approach. An earlier prototype
within the VPS project, TaskMail (Punekar and
Tagg, 2005), also used an ontology, albeit a simple
flat one; hence this had number of shortcomings.
The key difference between an ontology and
machine learning-based system is that the former
relies on ‘exogenous’ information i.e. the category
structure specified within the ontology, which
someone has to create and maintain. However such
an explicit definition of rules by a user or group of
users could possibly lead to more accurate
categorization, since human judgment is being
indirectly leveraged. Additionally, the same
ontology could be re-used across multiple
applications (e.g. general document filing, personal
bookmarks, classifying events received from a
workflow system).
However, ontology alone is not enough. The
system needs a lexicon of words and phrases, which,
if they appear in a message, indicate that a category
applies to the message.
3.3 Sender Assisted
Since it is usually the sender of a message that wants
something done, it seems reasonable to expect him
or her to explicitly specify one or more categories or
tags for each document. This could be applied to
email by using an XML-style tagging protocol, or by
requiring senders to complete a recipient-dependent
ONTOLOGY-BASED EMAIL CATEGORIZATION AND TASK INFERENCE USING A LEXICON-ENHANCED
ONTOLOGY
103

pop-up form containing a set of drop down choices
that need to be made before the message can be sent.
The latter approach was the basis for another
earlier VPS prototype named NatMail (Tagg and
Mahalingam, 2005). Senders were required to
submit messages through a “Contact Me” web page.
The options in the drop down boxes were created via
a Wizard based on the recipient’s personal ontology.
The concept of sender-assistance might be
expected to incur resistance from senders, since it
demands a fundamental shift in work culture.
However we are all getting more used to filling in
web forms for airlines, banks, insurance companies
etc. Even some university professors ask their
students to do so.
The younger generation is already half way
there, since categorization is an essential part of
tools like Flickr and del.icio.us. In the long run the
approach might just need to gradually become part
of the work culture.
3.4 Other Classification Approaches
Another approach is the SimOverlap system of
(Dredze et al, 2006), who match people in an email
message with pre-defined activity participants. This
is a valid simplification in many work structures, but
not if the working roles are highly volatile.
The best known of other approaches is probably
that of detecting threads or ‘thrasks’, as in TaskVista
(Bellotti et al, 2003). In their TV-ACTA prototype,
(Bellotti et al, 2007) introduced a distinct strategy
based on the integration of ‘to-do’ lists with email.
Users can drag-and-drop email messages into a
system for creating to-dos, which can be then be
sorted according to properties such as deadline and
task type.
4 DESCRIPTION OF THE
PROTOTYPE
Regarding the technologies and systems to be used
to support the implementation of the prototype, it
was decided that:
• Personal ontologies would be represented in
the XML OWL (RDF) format using the
output of this university’s own EzOntoEdit
ontology editor (Einig et al, 2006).
• A tool named SnipCat (Srinivasan Kumaar,
2008) would also be used to insert lexical
clues into the ontology.
• The system should be developed in Java, and
the output task lists should be maintained in
a relational database, in our case Oracle.
Figure 1 outlines the steps involved. The process
can be broken down into three key phases, namely,
initiation, work topic categorization, and task
identification, as detailed below.
Figure 1: Prototype Process Overview.
4.1 Initiation Phase
We assume that the user has already created a
personal ontology; this is imported into the agent as
part of its internal database. When the agent is
started, it checks (at 5-minute intervals) whether the
ontology has changed from the previous session; if
there are any changes the complete new ontology is
imported.
4.2 Work Topic Categorization Phase
Having gained a copy of user’s latest ontology, the
agent then checks for new emails during the same 5-
minute intervals. If the user has not received any
new emails the process waits until the next 5-minute
interval. However, if new emails are detected the
message body is extracted for each email (identified
by a distinct email ID) and stored in individual text
ICEIS 2009 - International Conference on Enterprise Information Systems
104

files on a web server. Email attachments (if any) are
also stripped off and stored on the web server.
Treating each email message as a separate
document, the agent first tokenizes the contents (i.e.
From, To, Cc, Subject and Body) of each email,
removing punctuations.
Next, stop words present in the email message
are removed in order to reduce the search space.
This will hopefully improve classification
performance, in terms of speed and accuracy.
Having defined a feature set for a new email, the
agent then extracts the set of ontology clues from the
database. These include upper case and lower case
clue terms (stored separately) and phrases that were
provided by the user. The extracted clue phrases are
first compared with the content of the email message
(i.e. performing a free text search). The matching
phrases (if any) are identified and stored in a list.
The clue terms are then compared with the message
tokens and the matching terms (if any) are stored in
the same list. If no matching phrases or terms were
found at this stage, the email is labeled as
‘unidentifiable’ (i.e. the system was unable to
identify the work topic) and the task identification
phase commences.
However, if matching words or phrases were
found the agent proceeds to identify the number of
occurrences of each matched clue in the email. It
seems logical to assume that an email with 5
occurrences of the clue conference would have a
stronger association with the conferences work topic
than an email with only 1 occurrence of conference.
To achieve this, we build on the concept in our
ontology design of indication strength, which is the
subjective probability that presence of the clue’s
string indicates relevance to the ontology concept.
The weighted indication strength (WIS) for each
matched clue is calculated based on multipliers that
reflect the number of occurrences of the clue. The
heuristic multipliers we applied in our tests ranged
from 1.2 for 2 occurrences, to 1.66 for 10 or more.
Finally, the strengths of all the different clues
indicating the same ontology category are added. If
the total weighted indication strength for a category
exceeds the threshold value (we have set the default
at 1.0), the email is then labeled as belonging to that
work topic.
4.3 Task Inference Phase
The task inference phase now follows in a similar
but not identical fashion. The process begins with
the removal of old messages (typically previous
conversations between participants) within the body
of the email. This is more important in task
inference than with work topics, since the agent
needs to be prevented from erroneously inferring
tasks by scanning old messages in the thread.
In the ontology we have been using, we store a
number of message patterns that the user receives
regularly and which strongly indicate a particular
task type. If such a match is found, the email is
tagged with the appropriate task type (i.e.
ForInfoOnly, Reply etc.) and the task identification
process is brought to an end. This is because these
message pattern clues always have an indication
strength of 1.0.
However, if a matching message pattern was not
found, the list of remaining clue strings (words or
phrases) that indicate each task type, together with
their respective locations and strengths, are retrieved
from the database. Thereafter, the clue strings are
matched one task type at a time, upper case and
lower case being taken into account.
4.4 Presentation of the Task Lists
To present the results of the above email
classification activities, we have designed a
prioritized “to do list” style interface. The idea is to
encourage users to start by viewing only the
messages that imply high priority tasks, and to group
these by work topic; they can then view the less
important ones later. Hopefully, this could help
negate feelings of email overload. The interface
consists of mailbox-like panels; a screenshot is
shown in Figure 2 below.
Any one of three panels can be chosen. The
default is ‘High Priority Tasks’, which includes
those emails mapped to the task types DefiniteTask,
Reply, AppointmentInvitation, VoteApprove, and
ConditionalTask. The ‘Low Priority Tasks’ panel
includes those mapped ForInfoOnly, Questionnaire,
ConfInvite, and PrivateCommunication. Two
buttons displayed at the bottom of both of the above
panels facilitate user interaction with the interface.
The ‘Email Inbox’ panel, which displays all
emails, was included since it was found, through the
survey discussed below, that users would feel more
comfortable with the system if they could switch
back and forth between the ‘new’ interface and the
traditional inbox view, without having to go back to
the old email client. This relates to the issue of trust
that arises when using a software agent.
ONTOLOGY-BASED EMAIL CATEGORIZATION AND TASK INFERENCE USING A LEXICON-ENHANCED
ONTOLOGY
105

Figure 2: Example of a Task List from the Prototype.
5 EVALUATION
The system was built successfully as planned.
Evaluation was in three parts.
• a first version was tested on one academic
and then demonstrated at a project fair;
• a questionnaire survey was carried out on
40 people attending the project fair
• a revised version with improved
performance was built and tested.
In the first version, categorization into work
topics was quite successful, with an average
precision of 80%. However the task type
categorization was disappointing; the average
precision was under 50%. Many messages that
should have indicated definite tasks were only
graded as low priority, or the agent failed to put
them into any task category.
The survey showed positive responses from the
majority of respondents, who included a mix of
students, academics and people outside academia.
They expressed interest in the idea of moving to a
more structured, task list style interface. However,
interest is a long way removed from changing one’s
everyday computing habits. Issues raised included
ontology change management, the need to still refer
to the inbox on a regular basis, and the lack of a
feature for managing task deadlines and reminders.
In the tests of the second version, performance
did improve, but the task inference was still not
adequate for use in a real world environment. Our
assessment was that the agent’s inability to
recognize deadline dates and times was a
contributing factor, and some level of sender
assistance (e.g. by tagging deadlines) might also be
needed to make a significant difference.
6 CONCLUSIONS AND FUTURE
WORK
The main contribution of this work has been to
demonstrate the concept of an ontology-driven email
categorization agent. In regards to system
performance, the results produced so far have been
positive. The strong results achieved in classifying
emails around work topics, is particularly
encouraging.
Admittedly, limited user testing of the prototype
system has been undertaken in this work due to the
time constraints and the need to acquire ethics
approval.
Future initiatives would need to focus on not
only testing the system with a larger user base and
sample data sets but also over a longer period of
time. However, having said that, the user testing
undertaken so far has been useful in terms of
gauging system performance, user attitudes and
acceptance, as well as for establishing future
research direction.
ICEIS 2009 - International Conference on Enterprise Information Systems
106

This work does not entirely rule out use of a
machine learning approach, and acknowledges that a
combined ontology and machine learning-based
model might be the way forward in the future,
especially for overcoming the challenges of implicit
task identification.
ACKNOWLEDGEMENTS
We would like to thank Harshad Lalwani for his
technical inputs and help in developing the VPS
email agent prototype.
REFERENCES
Bellotti, V., Ducheneaut, N., Howard, M., & Smith, I.
(2003) Taking Email to Task: The Design and
Evaluation of a Task Management Centered Email
Tool. Proceedings of the computer-human interaction
conference (CHI) 2003, CHI Letters 5 (1). 297-304.
Bellotti, V., Thornton, J., Chin, A., Schiano, D., & Good,
N. (2007). TV-ACTA: Embedding an Activity-
Centered Interface for Task Management in Email.
CEAS, 2007.
Corston-Oliver, S., Ringger, E., Gamon, M., & Campbell,
R. (2004). Task-focused Summarization of email. Text
Summarization Branches Out: Proceedings of the
ACL-04 Workshop, 43–50.
Crawford, E., Kay, J., & McCreath, E. (2006). IEMS - the
Intelligent Email Sorter. Proceedings of the
International Conference on Machine Learning, 25-32
Dabbish, L.A., Kraut, R.E., Fussell, S., & Kiesler, S.
(2005) Understanding Email Use: Predicting Action
on a Message. Proceedings of the SIGCHI conference
on Human factors in computing systems, 691-700.
Dredze, M., Lau, T., & Kushmerick, N. (2006).
Automatically Classifying emails into Activities.
Proceedings of the 11th International Conference on
Intelligent User Interfaces, 70-77.
Ducheneaut, N., & Bellotti, V. (2001). E-mail as Habitat:
an exploration of embedded personal information
management. Interactions 8, 5, 30-38.
Einig, M, Tagg, R. & Peters, G (2006), Managing the
Knowledge Needed to Support an Electronic Personal
Assistant, in Proceedings of the ICEIS conference,
Paphos, Cyprus.
Hall, S. (2004). A Better Microsoft Outlook . Blog,
http://nwvc.blogs.com/northwest_vc/2004/02/a_better
_micros.html (Viewed 27 November 2008).
Li, W., Zhong, N., & Liu, C. (2006). ECPIA: An Email-
Centric Personal Intelligent Assistant. Lecture notes in
computer science, 4062, 502.
Maes, P. (1994). Agents that Reduce Work and
Information Overload. Communications of the ACM,
37, 31-40.
Mulder, I., de Poot, H., Verwij, C., Janssen, R., & Bijlsma,
M. (2006). An Information Overload Study.
Proceedings of the CHI Conference SIG of Australia
on Design: Activities, Artefacts and Environments,
245-252.
Punekar & Tagg, R. (2005). Ontology Assisted Pre-
processing of Incoming email: Working paper,
University of South Australia.
Sebastiani, F. (2002). Machine Learning in Automated
Text Categorization. ACM Computing Surveys 34 1, 1-
47.
Segal, R., & Kephart, J. (1999). MailCat: an intelligent
assistant for organizing e-mail. Proceedings of the
third annual conference on Autonomous Agents, 276-
282.
Spira, B., & Goldes, M. (2007). “Information Overload:
We have met the enemy and he is us”, Basex Inc.,
http://freya.basex.com/web/webdownloads.nsf/e67dc0
f5617d6e9c85256a99005ea0e7/ff7810b594932040852
570b9008022ce/$FILE/BasexInformationOverloadWh
itePaper.pdf (Viewed 18 May, 2008)
Srinivasan Kumaar, R. (2008), Helping Users to Create
and Maintain a Personal Knowledge Base or Ontology
to Support Task-oriented Work Management,
Masterate Thesis, University of South Australia.
Tagg, R. (2007), Task Integration for Knowledge
Workers, in ICEIS Workshop on Computer Supported
Activity Coordination, Madeira, Portugal, May 2007.
Tagg, R. & Mahalingam, N. (2005). Improving Customer
Service and Reducing Administrative Overload with
Sender-assisted Message Categorization: Working
paper, University of South Australia.
Taghva, K., Borsack, J., Coombs, J., Condit, A., Lumos,
S. & Nartker, T. (2003). Ontology-based Classificat-
ion of email. Proceedings of ITCC 2003 Conference
on Information Technology: Coding, 194-198.
ONTOLOGY-BASED EMAIL CATEGORIZATION AND TASK INFERENCE USING A LEXICON-ENHANCED
ONTOLOGY
107