ConTask

Using Context-sensitive Assistance to Improve Task-oriented Knowledge Work

Jan Haas

, Heiko Maus

, Sven Schwarz

and Andreas Dengel

†

German Research Center for Artiﬁcial Intelligence (DFKI GmbH), Kaiserslautern, Germany

†

Computer Science Department, University of Kaiserslautern, Kaiserslautern, Germany

Keywords:

Task management, Proactive information delivery, Personal knowledge space, User observation, Agile task

modelling, Semantic Desktop.

Abstract:

The paper presents an approach to support knowledge-intensive tasks with a context-sensitive task manage-

ment system that is integrated into the user’s personal knowledge space represented in the Nepomuk Semantic

Desktop. The context-sensitive assistance is based on the combination of user observation, agile task mod-

elling, automatic task prediction, as well as elicitation and proactive delivery of relevant information items

from the knowledge worker’s personal knowledge space.

1 INTRODUCTION

Corporate work routines have changed vastly in re-

cent years, and today’s knowledge workers are con-

stantly presented with the challenge of negotiating

multiple tasks and projects simultaneously (Gonz

alez

and Mark, 2005). Each of these projects and tasks

stereotypically requires collaboration with different

teams, and dealing with a varied plethora of data, re-

sources, and technologies. In contrast to traditional,

static business processes, the task-oriented work of

clerks today is often highly fragmented. Each in-

terruption requires knowledge workers to mentally

re-orientate themselves, and permanent task switches

and disruptions are associated with a signiﬁcant over-

head cost (Mark et al., 2008). Especially for complex,

knowledge-intensive tasks requiring signiﬁcant quan-

tities of related documents and resources, reorienta-

tion entails a substantial cognitive overhead.

While an efﬁcient task execution requires the pro-

cessing of particularly task relevant information at

hand, today’s knowledge workers have to face the

known information overload. An increasing amount

of data is available, however, spread over various in-

formation sources such as email client, address book,

local and remote ﬁle systems, web browser, wikis,

and organizational structures. The efﬁcient and suc-

cessful processing of a task depends on the quality

of ﬁnding and selecting the most relevant information

for the task at hand. This represents a source of er-

rors in daily knowledge work. Important information

is not found, connections are overseen or relevant ex-

perts are not identiﬁed. The result consists of sub-

optimal problem solutions, unnecessary repetitions of

already accomplished tasks or wrong decisions.

As a solution to the outlined problems of the

knowledge worker, this paper proposes a context-

sensitive task management system named ConTask.

ConTask focuses on the areas of knowledge captur-

ing, knowledge reuse, and interruption recovery. By

tracking the user’s actions, the system provides au-

tomatic means to intelligently elicit task-speciﬁc rel-

evant information items and, thus, capture a task’s

context. This is used for proactively delivering such

context-speciﬁc, task-relevant knowledge to a user to

ensure a reuse of valuable task know-how. Thereby,

ConTask aims at the following goals:

• Automatically capture created/consulted informa-

tion objects and assign these to tasks to cap-

ture task know-how and to structure the personal

knowledge space in a task-centric way.

• Increase potential productivity for knowledge

workers and reduce resource allocation costs by

proactively providing relevant, task-related infor-

mation and resources.

• Enable and ensure task-speciﬁc know-how reuse.

• Facilitate reorientation back into an interrupted

task by reducing the cognitive and administrative

task switching overhead.

• Improve task-speciﬁc assistance by learning from

Haas J., Maus H., Schwarz S. and Dengel A. (2010).

ConTask - Using Context-sensitive Assistance to Improve Task-oriented Knowledge Work.

In Proceedings of the 12th International Conference on Enterprise Information Systems - Artiﬁcial Intelligence and Decision Support Systems, pages

30-39

DOI: 10.5220/0002895400300039

 SciTePress

knowledge workers’ feedback.

The paper is structured as follows: the next sec-

tions introduce the scientiﬁc background and give an

overview on related work. Section 4 explains the

main ingredients and concepts of ConTask. A sum-

mary and outlook on future work conclude the paper.

2 BACKGROUND

Various approaches identify business processes as

a means for structuring a company’s knowledge

(Abecker et al., 2002; Riss et al., 2005). As busi-

ness processes form the core operational sequences of

every company, their efﬁciency is critical for a com-

pany’s success. Knowledge workers are integrated

into crucial parts of the business processes, and the

quality of their (procedural) know-how decides be-

tween success and failure (for the company). They are

embedded in business processes where we are mainly

interested in supporting knowledge-intensive tasks for

know-how capture, provision, and reuse as well as as-

sistance in multitasking for supporting the knowledge

worker in his daily work directly.

Knowledge-intensive Tasks. Especially know-

ledge-intensive tasks entail the challenge of retriev-

ing, structuring and processing information, e.g.,

for judging a case or for making crucial decisions.

Knowledge-intensive tasks notably rely on pro-

cessing large quantities of relevant information and

coevally producing valuable knowledge to be reused

in similar situations (later on). Aiming at preserving

this valuable knowledge, tasks should be utilized to

structure a knowledge worker’s personal knowledge

space consisting of various resources like documents,

emails, and contact addresses, as well as real-life

concepts such as persons, projects, or topics.

Typically, knowledge-intensive activities are ex-

plorative and not completely known a priori (Elst

et al., 2003). As many parts of their execution might

not be predetermined, they can not be completely

modelled in advance. On the basis of this, we intro-

duce the concept of weakly-structured workﬂows con-

sisting of knowledge-intensive activities with speciﬁc

design decisions for applications supporting these

workﬂows. They mainly incorporate the two aspects

lazy and late modelling and the strong coupling of

modelling and execution of process-models. Lazy

modelling refers to an on demand reﬁnement of pro-

cess models initially only partially speciﬁed. This

pays off for weakly-structured workﬂows as details

of the execution of agile knowledge-intensive tasks

are not known in advance. This aspect is strongly

related to the coupling of modelling and execution

of process-models. Starting with a partial model,

weakly-structured workﬂows allow for dynamical re-

ﬁnement of the process model during its execution.

Our work resulted in the TaskNavigator (Holz

et al., 2006; Rostanin et al., 2009), a browser-based

workﬂow system. It supports weakly-structured

workﬂows through agile task management for teams,

proactive information delivery (PID) based on the

task context (mainly consisting of task name, descrip-

tion, and attached documents, i.e. text-based), pro-

cess know-how capture and re-use. Evaluations have

shown that a main drawback was the effort for users

to maintain their task, i.e., by uploading documents

to the task represented in the browser. The work pre-

sented here, is a continuation of the overall goal by

integrating this into user’s personal knowledge space

and allowing a much easier management of their tasks

embedded on their personal desktops.

Multitasking. Nowadays knowledge workers are

engaged in multiple tasks and projects in parallel

(e.g., (Gonz

alez and Mark, 2004)). Several stud-

ies have shown that task-oriented knowledge work is

highly fragmented (Czerwinski et al., 2004). Typi-

cally, knowledge workers spend only little time on

a certain task before switching to another. Task

switches and disruptions cause signiﬁcant overhead

costs (e.g., (Mark et al., 2008; Iqbal and Horvitz,

2007; Mark et al., 2005)). After an interrup-

tion, knowledge workers must reconstruct their task-

speciﬁc mental state. This encompasses, in amongst

other detail, memories around the task, including next

steps to take, required resources, critical factors and

deadlines.

In addition to the cognitive overhead, restructur-

ing of the desktop and physical work environment is

also often required. Resources such as documents,

websites, emails, and contact addresses must be relo-

cated and utilized. The retrieval of these various task-

speciﬁc resources represents a challenging and time

consuming problem. Furthermore, the cognitive chal-

lenge to remember all task-speciﬁc relevant informa-

tion can result in signiﬁcant difﬁculty for the success-

ful completion of a given task. Short term memory

loss concerning critical resources or other task-related

information can have major ramiﬁcations.

The frequent interruption of knowledge-intensive

tasks is often a contributing factor to workplace stress

and frustration. Studies show that constant inter-

ruptions in these situations lead to changes in work

rhythm, mental states and work strategies (Mark et al.,

2008). Often this results in attempts to compensate

lost time through an accelerated and therefore even

ConTask - Using Context-sensitive Assistance to Improve Task-oriented Knowledge Work

more stressful work pace. The higher workloads and

additional effort associated with frequently switching

tasks increase both pressure and frustration for to-

day’s knowledge workers.

3 RELATED WORK

Many different approaches focus on assisting knowl-

edge workers with knowledge-intensive tasks and in

multitasking work scenarios.

The TaskTracer system (Stumpf et al., 2005b)

consists of an user observation framework collecting

events from various ofﬁce applications. The system

utilizes the observed user actions and applies machine

learning algorithms for automatically predicting the

user’s current task and for associating accessed infor-

mation items like ﬁles or web pages with the elicited

task (Stumpf et al., 2005a). This enables a task-

speciﬁc provision of these information items.

The APOSDLE Project (Lokaiczyk et al., 2007)

represents a similar approach. The project integrates

task management, e-learning, knowledge manage-

ment and communication systems. APOSDLE uti-

lizes user observation to support the task-centric pro-

vision of suitable learning material and to associate

knowledge artefacts with corresponding tasks. The

user observation framework is realized by software

hooks on the operation system level. Observed user

actions are reported to the task predictor component, a

machine learning component that serves for task pre-

diction and task switch detection.

The OntoPIM (Lepouras et al., 2006) project sug-

gests an architecture for a task information system

including a monitoring system for user observation.

Observed user actions are interpreted by an inference

engine to elicit information relevant to the user’s tasks

and to proactively provide these items. This task-

speciﬁc PID is integrated into a selected target ap-

plication (e.g., into a web browser). Alternatively,

the windows context menu is enriched with task-

related information. Instead of integrating the task-

speciﬁc assistance into several applications, our ap-

proach aims at integrating the PID into a task manage-

ment system. This reduces the amount of applications

having to be adjusted and enables the agile task mod-

elling to directly use the captured and elicited task-

speciﬁc knowledge.

A task management system embedded in the per-

sonal knowledge space along the vision we presented

in (Riss et al., 2005) is KASIMIR (Grebner et al.,

2008). It focuses on capturing, evolving, and pro-

viding process patterns from an organisational reposi-

tory. The tasks within the processes are enriched with

information items from the Nepomuk Semantic Desk-

top (see next section) by user interaction similar to

ConTask. Similar work on identifying process pat-

terns but coming from the process management side

and embracing task management is the Collaborative

Task Management (CTM) approach (Stoitsev et al.,

2008). The main difference of both approaches from

ConTask is the lack of user observation and PID our

system applies and their steps further to apply evolv-

ing process patterns in process management. We will

investigate the combination of these approaches in the

joint project ADiWa

started in 2009.

4 INGREDIENTS OF ConTask

ConTask is based on the following base components

to achieve a context-oriented personal task manage-

ment: a task management system, the Semantic Desk-

top, and a framework for retrieving user context.

Based on these components, ConTask supports with

proactive information delivery and agile task mod-

elling. It enables observing task management, provid-

ing a task prediction, and allowing relevance feedback

and learning. These components and concepts will be

detailed in the remainder of this section.

Task Management. As mentioned in Section 2, in

order to support knowledge workers in working with

knowledge-intensive tasks directly on their desktops,

a task management system called TaskPad was devel-

oped. TaskPad provides the possibility to work on

personal tasks and to synchronize tasks from different

sources such as TaskNavigator, where the tasks were

part of some agile workﬂows. TaskPad allows to ac-

cess, attach, and upload documents to TaskNavigator.

As both TaskNavigator and TaskPad are fully RDF/S

based, they rely on the ontology for weakly-structured

workﬂows developed in (Elst et al., 2003).

Apart from this, TaskPad provides the usual task

management functionality such as maintaining tasks,

taking notes, attaching URLs, notes, or documents,

and ﬁltering the task list

. Fig. 2 shows the task list as

Task Diary and the Task Editor.

Semantic Desktop. To represent and maintain the

user’s personal knowledge space we use and integrate

with the Nepomuk

Semantic Desktop which trans-

fers the idea of the Semantic Web to the user’s local

desktop (for a recent overview, see (Grimnes et al.,

http://www.adiwa.net

the ﬁlters are based on SPARQL queries

http://nepomuk.semanticdesktop.org/

ICEIS 2010 - 12th International Conference on Enterprise Information Systems

2009)). It serves to capture and represent the knowl-

edge worker’s personal mental models (Sauermann

et al., 2005). This personal knowledge space con-

sists of real world concepts such as persons, places,

projects or topics, as well as, the connections and re-

lations between them (see Fig. 1). Documents con-

tain information about these concepts and represent

the knowledge worker’s individual background, tasks

or personal interests. In this context, the Personal

Information Model Ontology (PIMO) serves to for-

malize and structure the personal knowledge space (a

PIMO excerpt can be seen on the upper left side with

PIMO classes).

It is the core of the Semantic Desk-

top and provides the possibility to associate real world

concepts with resources, such as documents, emails

or contact details, for personal information manage-

ment.

For example, Fig. 1 shows personal notes taken

during a meeting using a Semantic Wiki embedded

in the Nepomuk Semantic Desktop. Here, the meet-

ing itself is an instance of the PIMO class Meeting

together with linked concepts such as attendees, pre-

vious meetings or resources such as the calendar en-

try in MS Outlook. The wiki text shown is mixed

with concepts, such as the project it belongs to, top-

ics mentioned in the meeting, etc. Using a concept

within the text (e.g., by auto-completion) allows to

browse to the concept within the wiki text as well as

adds relations to the meeting instance automatically.

Furthermore, an ontology-based information extrac-

tion (iDocument, see (Adrian et al., 2009)) analyzes

text and provides proposals of concept which might

ﬁt for the current thing (lower left tab). Thus, dur-

ing everyday usage, the PIMO evolves with relevant

concepts of the user’s work. As the meeting is then

linked

User Context. The Context Service elicits the

user’s work context, which is a snapshot consisting

of contextual elements with relevance to the user’s

present goal or task (Schwarz, 2006; Schwarz, 2010).

Each contextual element corresponds to an entity

from the user’s PIMO. Contextual elements also con-

tain a value describing the certainty for actually be-

longing to the current context. As knowledge work-

ers regularly switch their tasks and, hence, their work

context, the Context Service maintains several so

called context threads. Each context thread represents

a context snapshot associated with a certain task and

contains all information items from the user’s PIMO

The PIMO is represented in RDF/S – the basic lan-

guage of the Semantic Web – and is able to include different

vocabularies resp. ontologies, i.e., it can be adapted to dif-

ferent domains.

that are relevant to that task. The Context Service ex-

pects explicit notiﬁcations about context switches to

correctly maintain context snapshots. A designated

API allows to inform the Context Service whenever

the user is switching to a different context thread.

In this scenario, the User Observation Hub

(UOH) serves as a technical means to automatically

observe the user’s actions (see also Fig. 3). This is

used for gathering evidence about relevant concepts

belonging to the user’s current context. The UOH in-

cludes an extensive user action ontology formalizing

all types of so called native operations (NOPs) that

are observable during daily knowledge work. The on-

tology comprises operations such as browsing a web-

site, adding a bookmark, reading an email or access-

ing a ﬁle in the ﬁle system. To gather these user ac-

tions, the user observation framework provides a set

of installable observers, which report the correspond-

ing user actions to the UOH. The observer framework

includes plugins for Mozilla Firefox and Thunderbird

and a Windows ﬁle system observer. Observed user

actions are gathered in the UOH and distributed to

registered listeners. The Context Service is such a

designated user observation listener receiving notiﬁ-

cations about observed user actions.

As outlined, the allocation and retrieval of relevant

information items represents a time consuming chal-

lenge for today’s knowledge workers. Thus, ConTask

aims at increasing potential productivity for knowl-

edge workers and reducing allocation costs by proac-

tively providing relevant, task-related information and

resources.

Proactive Information Delivery. The user’s PIMO

represents a range of concepts and resources the

knowledge worker deals with during daily work.

Therefore, elements in the worker’s PIMO are taken

to serve as items for the proactive information deliv-

ery (PID). Aiming at a task-centric work support, the

information items are provided in a task-centric man-

ner. Thus, the PID structures the personal knowledge

space in a task-oriented way.

To automatically elicit task-speciﬁc relevant infor-

mation items from the PIMO, ConTask utilizes the

Context Service. Based on the user’s interactions

with desktop applications, such as browsing a web-

site or writing an email, the Context Service elicits

task-speciﬁc relevant information items using tech-

niques of machine learning, entity recognition, and

document similarity (for details see (Schwarz et al.,

2008)). In addition, a History Service provides task-

speciﬁc records of all accessed information items.

http://usercontext.opendfki.de/wiki/UserObservationHub

ConTask - Using Context-sensitive Assistance to Improve Task-oriented Knowledge Work

Figure 1: Nepomuk Semantic Desktop: semantically-enriched meeting notes (some text obscured) with PIMO relations

(among them the task accessible in ConTask as shown in Fig. 2).

Being registered as a UOH listener, the History Ser-

vice maintains a detailed task-speciﬁc access history

of PIMO concepts and resources (based on the ob-

served user actions).

Fig. 2 depicts a screenshot of ConTask, where

the so called PID Sidebar proactively provides task-

centric access to relevant information items. The PID

Sidebar is part of the Task Editor which allows for

easy consultation and modiﬁcation of a task’s prop-

erties. Below the task’s name, status, and time con-

straints, the so called task attachments represent the

information items explicitly associated with the task.

PID Sidebar and Task Editor are located on the right-

hand side of the screenshot. ConTask provides further

interaction possibilities on both PID items and task

attachments, such as viewing an item speciﬁc access

history, opening resources with the associated appli-

cation or viewing resp. editing PIMO elements with

the PIMO Editor.

While the Task Editor contains manually attached

information items, the PID Sidebar shows additional,

potentially relevant information items. The History

Service is used to deliver directly accessed items, and

the Context Service is used to propose automatically

reasoned items.

PID Categories. The PID Sidebar provides infor-

mation items in the following three categories:

• Elicitation For Task. This category contains the

most relevant information items from the user’s

PIMO with respect to the task. It comprises both

directly observed and also elicited items, that have

not been directly accessed by the user.

• History For Task. This section contains all PIMO

elements for which a user action has been ob-

served during the execution of this task. This

category does not comprise any concluded or

elicited elements, but only directly observed el-

ements such as directly accessed ﬁles or websites.

The items here stem from the History Service.

Items already shown in the Elicitation For Task

are omitted.

• History. This category contains the most fre-

quently accessed PIMO elements that are not con-

tained in the ﬁrst two categories, but that have

been directly accessed by the user since a certain

point of time. These items also stem from the His-

tory Service.

Elicitation For Task provides the most relevant

elicited information items computed by the Context

Service’s machine learning techniques. The PID

Sidebar supports the user by proposing potentially rel-

evant items which have not been explicitly associated

with the task at hand. The effectiveness of these al-

gorithms is a critical factor for the acceptance of this

service. Low quality proposals will merely distract

and annoy the user while good quality proposals have

ICEIS 2010 - 12th International Conference on Enterprise Information Systems

Figure 2: ConTask: Task Diary with open tasks; Task Editor with the current active task and the PID sidebar with proposed

concepts from the PIMO and recently accessed resources (the context menu opens the PIMO editor in Fig. 1).

the potential to increase the user’s work performance

and foster a successful task execution. Particularly for

newly created tasks, there is not enough explicit in-

formation available that can be used to automatically

learn what this task is about. As a consequence, the

Context Service is not able to determine good propos-

als. For these cases, the PID Sidebar’s category His-

tory For Task contains all information items the user

has accessed while working on the task. According to

the user observation, these items have been recently

touched by the user while the task at hand was se-

lected as the current task.

The problem is, we can not expect the user to

make every task and every task switch explicit—

consider interruptions or phone calls, for example.

The task management tool does not know whether the

user is actually working on the currently selected task.

As a consequence, one can not automatically attach

recently touched items to the task. Rather ConTask

proposes the items allowing users to attach them with

a simple drag&drop gesture. The section Task Predic-

tion will explain the how ConTask tries to keep track

of the user’s current task nevertheless.

For example consider the following scenario. The

user accesses a website that is associated with the

current task, but the Context Service does not deter-

mine this website as relevant, as it has to be further

stimulated by being re-accessed by the user. In this

case, the website would be provided in the History

For Task section. If the relevance value of the web-

site increased based on further stimulation, the web-

site would be included into the Elicitation For Task

category and would therefore be removed from the

History For Task.

Manual task switches sometimes have fuzzy

boundaries (Stumpf et al., 2005a). For instance if a

knowledge worker is just reading a web page that is

related to his current task but that leads him to switch

to a new one: should the resource be associated with

the old or with the new task? Additionally, if a knowl-

edge worker explicitly consults a related task (and

hence switches to the other task) for reusing there-

stored know-how, the History For Task for the orig-

inating task will not contain the consulted material.

A solution for this case provides the History category

at the bottom of a task’s PID Sidebar. It contains all

information items which have been recently accessed

by the user, but which have neither been associated

with, nor elicited for, nor accessed during the task at

hand. This overview over all recently accessed items

allows an easy reuse of know-how across different

tasks (via simple drag&drop a resource can be added

as a task attachment).

Agile Task Modelling. Based on the proactively

provided PIMO elements, ConTask enables agile, on-

the-ﬂy task modelling. As knowledge-intensive tasks

can not be completely designed or modelled before

their execution, agile, lazy modelling was chosen to

allow for task reﬁnement during the execution pro-

cess. Via context menu or drag&drop, new items, e.g.,

proposed items from the PID Sidebar, can be easily

added as task attachments. This enables knowledge

workers to associate information objects to tasks and

to thereby classify work knowledge to independent

task information units. The result is a task-centric

structuring of the user’s personal knowledge space.

This aligns with suggestions in (Abecker et al., 2002;

ConTask - Using Context-sensitive Assistance to Improve Task-oriented Knowledge Work

Holz et al., 2005), where minimal analysis and initial

modelling overhead are identiﬁed as one of the key

requirements for successful business process-oriented

knowledge management.

Furthermore, task-oriented structuring of the

user’s personal knowledge space enables intuitive and

direct process know-how reuse. For example, while

working on a report document for project x, the

knowledge worker may remember an already com-

pleted similar report for project y. In case a relevant

information item of project y is actually relevant for

project x, ConTask allows to sight, attach, and reuse

these items from one task to another with a few clicks.

By explicitly attaching reused items to tasks, Con-

Task supports a light-weight capturing of task-speciﬁc

knowledge and, hence, provides the basis for the inte-

gration of these tasks in organizational workﬂow sys-

tems such as TaskNavigator.

The explicitly attached and conserved task-

speciﬁc items as well as the task-speciﬁc history also

facilitate rapid reorientation when switching back to

an interrupted task. ConTask reduces the mentioned

cognitive and administrative overhead consisting of

remembering and reallocating task-speciﬁc relevant

information items. By double clicking attached or re-

cent documents and resources, task-speciﬁc working

states can be easily reconstructed and the task can be

resumed without much delay.

As the PID Sidebar merely proposes potentially

relevant items, ConTask provides the possibility of re-

jecting non-relevant or unsuitable suggestions. The

system remembers these decisions and does not pro-

vide rejected information items again. Both kinds

of feedback, acceptance and rejection, represent ev-

idences for adapting the context of a task and serve

as relevance feedback for the task-speciﬁc PID. That

way, user feedback enables automatic learning and

system improvement (Abecker et al., 2002; Riss et al.,

2005; Holz et al., 2006).

Observing Task Management. During the execu-

tion of a task, the User Observation Hub (UOH)

observes the user’s task-speciﬁc behavior. Besides

tracking actions inside ofﬁce applications, ConTask

also observes the user’s explicit task management in-

teractions such as reuse/open, drag&drop, and reject

operations within the Task Editor and the PID Side-

bar. Automatic, unobtrusive learning is applied to im-

prove the task-speciﬁc PID. The observed user events

are utilized for the following two goals:

• Relevance feedback on proactively provided in-

formation items.

• Automatic task switch detection based on the

user’s interaction with the system.

To integrate explicit task management operations

into the user observation framework, the user action

ontology has been enriched with NOPs for task oper-

ations. These additional NOPs resemble a task oper-

ations ontology capable of representing user interac-

tion with an agile task management tool. Observable

actions of the ConTask system which are also passed

to the UOH are switching to some task or giving rel-

evance feedback on PID items, for example. The task

operations ontology is designed to represent NOPs

from different task management applications. It com-

prises a minimal set of operations that are necessary

for the purpose of relevance feedback on proactively

provided information items and automatic task switch

detection. It contains operations such as task creation,

attribute modiﬁcation such as by adding a task de-

scription or attachments. Further operations inform

about accessing attachments or interacting with the

PID sidebar. A Task Observation Service inside Con-

Task describes the user’s task actions according to the

NOP ontology and report these to the UOH.

Task Prediction. As outlined, Context and History

Service form the basis of the task-speciﬁc PID. Both

services observe the user’s desktop activity and record

or elicit relevant information items corresponding to

a certain context thread. However, both services rely

on explicit information about the currently processed

task. If users are stressed or get interrupted, they will

not use the task management tool for every tiny task

deviation. Hence, not every task switch is technically

observed. To compensate for this, ConTask contains

a Task Elicitation Service realizing a task prediction.

The Task Elicitation Service maps the tasks of the

task management tool to context threads maintained

by the Context Service. The observed user actions

are treated as evidences to predict and update the cur-

rently active context thread. If the Context Service

detects a context switch, a corresponding task switch

is also proposed to the user. And vice versa: If the

user switches a task, the Context Service is informed

to switch to the corresponding context thread, too.

The detection of the currently processed task is

based on the observed user interaction with ConTask,

such as attaching relevant information items to a task.

To realize this, the task operations ontology is divided

into the following two categories:

• Operations with a Direct Implication on a Task

Switch. These operations are interpreted as strong

indication for the fact that the user works on a

certain task. Based on that, any of these user ac-

tions immediately leads to the conclusion of a task

switch. They comprise all write access operations

on a task, such as adding an attachment or editing

ICEIS 2010 - 12th International Conference on Enterprise Information Systems

the task’s description, as well as opening the Task

Editor window.

• Operations with a Weak Implication on a Task

Switch. These operations only lead to a task

switch conclusion if the following condition holds

true: No other operation with either weak or di-

rect implication on a task switch occurs within a

certain amount of time t. Operations with weak

implications consist of the selection of a task in

the task list and the event that occurs if the Task

Editor becomes the active window.

The reasoning for the timeout value, associated with

the selection of a task, is that knowledge workers

sometimes browse their task list by clicking on each

single entry. This serves to get an overview on their

tasks and to determine which tasks are most critical

at the moment. To avoid that each selection during

browsing triggers a task switch and therefore context

thread switch, selections are only interpreted as task

switches after the timeout.

A similar reason explains the timeout value corre-

sponding to the focus gain event of the Task Editor.

As several Task Editor windows might be open at the

same time, the user might quickly switch focus be-

tween two or more Task Editors to compare the cor-

responding tasks. The timeout value avoids that every

focus change directly determines a task switch.

The automatic task elicitation serves to precisely

determine which user actions and corresponding in-

formation objects occurred in the context of which

task. This aims at increasing the Context and History

Service’s effectiveness. As the assistance is realized

in an unobtrusive way, the user just needs to work

and interact with ConTask, without having to deal

with explicitly telling the system to perform a task

switch. Since knowledge workers frequently perform

task switches and face interruptions of their current

task, automatic switch detection reduces interaction

overhead and still guarantees that information items

are associated with the correct task.

In addition to this, the Context Service also

provides the possibility of automatically detecting

whether the last user action(s) match better with dif-

ferent context thread(s) than the current one. In this

case, the service notiﬁes interested listeners about a

potential switch. ConTask utilizes these switch de-

tection capabilities. However, a task switch can not

be performed automatically and, hence, the user is

consulted via a speciﬁc popup listing potential task

switches. The window slides up from the bottom of

the screen and only remains visible for a short amount

of time. If the user does not interact within this time-

frame, the popup disappears.

Figure 3: ConTask components and interrelations.

Relevance Feedback and Learning. The Task

Elicitation Service also utilizes the observed interac-

tions with ConTask for the transmission of feedback

to the Context Service. Actions such as the assign-

ment of a PID item to a task are interpreted as pos-

itive feedback. They increase the relevance value of

the item in the corresponding context thread.

The relevance feedback serves to increase the

Context Service’s effectiveness. The cycle of proac-

tively providing contextual information to the user

and transmitting user feedback to the Context Service

leads to a better synchronization of context thread

and task: Relevant information items from the con-

text thread are proposed in the PID Sidebar. Some

of them may be added to the task by the user (via

drag&drop from the sidebar to the task attachments).

In that case, the resulting feedback leads to increased

relevance values of these items in the context thread.

Fig. 3 shows how user observation data is utilized in

ConTask to realize a knowledge improvement cycle.

As an extension to the mentioned user observa-

tion providers, such as Mozilla Firefox, ConTask uti-

lizes the Task Observation Service to become observ-

able. Interested listeners, such as the Task Elicitation

Service, receive the transmitted task events from the

User Observation Hub. On the basis of the received

task NOPs, the Task Elicitation Service transmits rel-

evance feedback to the Context Service and switches

both Context and History Service to the thread be-

longing to the current task. The History and Context

Service gain their data from the UOH and complete

the loop by providing information items for the PID

within ConTask.

The Task Elicitation Service only depends on the

task operations ontology and the assumptions on task-

oriented work that are utilized for the purpose of task

ConTask - Using Context-sensitive Assistance to Improve Task-oriented Knowledge Work

prediction. The History and Context Service only

depend on the NOP ontology formalizing the user’s

desktop activity. Thus, any task management appli-

cation that is compatible with the task operations on-

tology can be integrated into the created scenario re-

alizing the knowledge improvement cycle. Necessary

integration steps involve the extension of the system’s

user interface with observation calls to the Task Ob-

servation Service. If the task management application

includes a PID component and allows for agile task

modelling, the Context and History Service could be

utilized to proactively provide relevant, task-speciﬁc

PIMO elements to the user.

Feasibility Study. The current proof-of-concept

implementation of ConTask delivers early indications

that the system actually has the potential to assist the

user while keeping the additional work at a minimum.

In the long run, a robust context identiﬁcation is es-

sential for keeping the assistance scalable: The con-

text identiﬁcation algorithm is expected to estimate

the correct context in most of the cases. We created

a ground truth by tagging a large set of observed user

operations, manually assigning user actions to “con-

texts”. A ten-fold cross validation on this ground

truth data shows that 78% of the operations are iden-

tiﬁed correctly, 9% of the guesses were incorrect, and

13% of the cases were not identiﬁed at all. Striving

for a best-effort strategy, a relatively high number of

unidentiﬁed cases (13%) is not considered harmful for

the user’s actual work. An amount of 9% incorrect

context guesses is not very high, but this is a critical

value as false identiﬁcations may lead to false con-

text switch proposals and, hence, to disruptions of the

user. One cause for the false identiﬁcations is that the

user observation software does not recognize some of

the user operations. Additionally, users mentally sep-

arated some contexts which were technically identi-

cal. Additional sensors providing evidences for ad-

ditional contextual elements will reduce these prob-

lems. Hence, we will continuously enhance the user

observation and context elicitation technology.

5 SUMMARY AND OUTLOOK

This paper addressed challenges in today’s knowledge

work: continuously increasing quantities of infor-

mation, knowledge intensive tasks, and highly frag-

mented multitasking work scenarios.

The context-sensitive task management system

ConTask was designed to address these challenges

and alleviate the knowledge worker’s job. ConTask

is integrated into the Semantic Desktop and com-

bines task management with context-speciﬁc assis-

tance. The assistance is based on the combination of

user observation, automatic elicitation and proactive

information delivery of relevant information items

from the user’s PIMO. ConTask enables agile task

modelling for deﬁning tasks on the ﬂy and striving

for a task centric structuring of the personal knowl-

edge space. Observation of the user’s interaction with

ConTask is utilized for relevance feedback and auto-

matic task prediction to increase the precision of the

PID while keeping the required task management to a

minimum. Thereby, the system realizes a knowledge

improvement and learning cycle.

As ConTask is only capable of detecting task

switches to already existing tasks, a possible improve-

ment would be an algorithm for detecting the user is

working on a new task which is not yet reﬂected in

the system. Similarly, making proposals for reﬁning a

task into subtasks based on the observations (e.g., by

detecting that involved resources of a task can be sep-

arated in two topic clusters but within the task). This

would signiﬁcantly support agile task modelling.

We currently expand the observed area to the

physical desktop of a user by using a digital camera

and applying image recognition algorithms to recog-

nize user actions with paper documents on the desk

(Dellmuth et al., 2009). So far, recognizable actions

are placing, removing, and moving a paper document

on the desk as well as arranging a pile all enriching

the user context.

For getting better suggestions on PID, we cur-

rently investigate to embed the ontology-based infor-

mation extraction system iDocument to extract PIMO

entities in observed text snippets as well as to contex-

tualize with the PIMO as background knowledge as

it is done in the Nepomuk Semantic Desktop (Adrian

et al., 2009) and the TaskNavigator in (Rostanin et al.,

2009).

In the ADiWa project we will investigate the usage

of the PIMO-based user context for to realise dynamic

business processes.

ACKNOWLEDGEMENTS

This work has been partly funded by the German Fed-

eral Ministry of Education and Research (BMBF) in

the ADiWa project (01IA08006).

REFERENCES

Abecker, A., Hinkelmann, K., Maus, H., and Mller, H.-J.,

editors (2002). Geschftsprozessorientiertes Wissens-

ICEIS 2010 - 12th International Conference on Enterprise Information Systems

management. xpert.press. Springer.

Adrian, B., Klinkigt, M., Maus, H., and Dengel, A.

(2009). Using idocument for document categorization

in nepomuk social semantic desktop. In Pellegrini,

T., editor, i-Semantics: Proceedings of International

Conference on Semantic Systems 2009, JUCS.

Czerwinski, M., Horvitz, E., and Wilhite, S. (2004). A diary

study of task switching and interruptions. In CHI’04:

SIGCHI conference on Human factors in computing

systems, pages 175–182. ACM.

Dellmuth, S., Maus, H., and Dengel, A. (2009). Support-

ing knowledge work by observing paper-based activ-

ities on the physical desktop. In 3rd Int. Workshop

on Camera Based Document Analysis and Recogni-

tion (CBDAR’09). Proceedings.

Elst, L. v., Aschoff, F.-R., Bernardi, A., Maus, H., and

Schwarz, S. (2003). Weakly-structured workﬂows

for knowledge-intensive tasks: An experimental eval-

uation. In IEEE WETICE Workshop on Knowledge

Management for Distributed Agile Processes (KM-

DAP’03). IEEE Computer Press.

Gonz

alez, V. M. and Mark, G. (2004). “Constant, constant,

multi-tasking craziness”: managing multiple working

spheres. In CHI’04: SIGCHI conference on Human

factors in computing systems, pages 113–120. ACM.

Gonz

alez, V. M. and Mark, G. (2005). Managing currents of

work: multi-tasking among multiple collaborations.

In ECSCW’05: 9th European Conference on Com-

puter Supported Cooperative Work, pages 143–162.

Grebner, O., Ong, E., and Riss, U. (2008). Kasimir - work

process embedded task management leveraging the

semantic desktop. In Multikonferenz Wirtschaftsinfor-

matik, pages 1715–1726.

Grimnes, G. A., Adrian, B., Schwarz, S., Maus, H., Schu-

macher, K., and Sauermann, L. (2009). Semantic

desktop for the end-user. i-com, 8(3):25–32.

Holz, H., Maus, H., Bernardi, A., and Rostanin, O. (2005).

From Lightweight, Proactive Information Delivery

to Business Process-Oriented Knowledge Manage-

ment. Journal of Universal Knowledge Management,

0(2):101–127.

Holz, H., Rostanin, O., Dengel, A., Suzuki, T., Maeda,

K., and Kanasaki, K. (2006). Task-based process

know-how reuse and proactive information delivery in

TaskNavigator. In CIKM’06. ACM Conference on In-

formation and Knowledge Management.

Iqbal, S. T. and Horvitz, E. (2007). Disruption and recov-

ery of computing tasks: ﬁeld study, analysis, and di-

rections. In CHI’07: SIGCHI conference on Human

factors in computing systems, pages 677–686. ACM.

Lepouras, G., Dix, A., and Katifori, A. (2006). OntoPIM:

From personal information management to task infor-

mation management. In SIGIR’06 Personal Informa-

tion Management Workshop.

Lokaiczyk, R., Faatz, A., Beckhaus, A., and Goertz,

M. (2007). Modeling and Using Context, volume

4635/2007 of LNCS, chapter Enhancing Just-in-Time

E-Learning Through Machine Learning on Desktop

Context Sensors, pages 330–341. Springer.

Mark, G., Gonzalez, V. M., and Harris, J. (2005). No

task left behind?: examining the nature of fragmented

work. In CHI’05: SIGCHI conference on Human fac-

tors in computing systems, pages 321–330. ACM.

Mark, G., Gudith, D., and Klocke, U. (2008). The cost of

interrupted work: more speed and stress. In CHI’08:

SIGCHI conference on Human factors in computing

systems, pages 107–110. ACM.

Riss, U., Rickayzen, A., Maus, H., and van der Aalst, W.

(2005). Challenges for Business Process and Task

Management. Journal of Universal Knowledge Man-

agement, 0(2):77–100.

Rostanin, O., Maus, H., Zhang, Y., Suzuki, T., and Maeda,

K. (2009). Lightweight conceptual modeling and

concept-based tagging for proactive information de-

livery. Ricoh technology report 2009, no. 35, Ricoh

Co Ltd., Japan.

Sauermann, L., Bernardi, A., and Dengel, A. (2005).

Overview and Outlook on the Semantic Desktop. In

1st Workshop on The Semantic Desktop at ISWC’05,

volume 175 of CEUR Proceedings, pages 1–19.

Schwarz, S. (2006). A context model for personal knowl-

edge management applications. In Modeling and Re-

trieval of Context, Second Int. Workshop, MRC 2005,

volume 3946 of LNCS, pages 18–33. Springer.

Schwarz, S. (2010). Context-Awareness and Context-

Sensitive Interfaces for Knowledge Work Support.

PhD thesis, University of Kaiserslautern.

Schwarz, S., Kiesel, M., and van Elst, L. (2008). Adapting

the multi-desktop paradigm towards a multi-context

interface. In HCP-2008 Proc., Part II, MRC 2008

– 5th Int. Workshop on Modelling and Reasoning in

Context, pages 63–74. TELECOM Bretagne.

Stoitsev, T., Scheidl, S., Flentge, F., and Mhlhuser, M.

(2008). From personal task management to end-user

driven business process modeling. In Business Pro-

cess Management, volume 5240 of LNCS. Springer.

Stumpf, S., Bao, X., Dragunov, A., Dietterich, T. G., Her-

locker, J., and et al. (2005a). Predicting user tasks: I

know what you’re doing! 20th National Conference

on Artiﬁcial Intelligence (AAAI-05).

Stumpf, S., Bao, X., Dragunov, A., Dietterich, T. G., Her-

locker, J., Johnsrude, K., Li, L., and Shen, J. (2005b).

The TaskTracer system. 20th National Conference on

Artiﬁcial Intelligence (AAAI-05).

ConTask - Using Context-sensitive Assistance to Improve Task-oriented Knowledge Work