Putting Knowledge in Context to Facilitate Collaboration

Putting Knowledge in Context to Facilitate Collaboration


Brian Kettler, Jim Starz, Rachel Hingst, Jason Losco, and Mark Hoffman
Lockheed Martin Advanced Technology Laboratories
{bkettler, jstarz, rhingst, jlosco, moffman}@atl.lmco.com


ABSTRACT
and a decreasing supply of veteran analysts. Analysts lack

automated support from today’s tools to ensure
Today’s knowledge workers frequently suffer from
information they find remains accessible, exploitable, and
problems associated with information overload, changing
shareable over time. Analysts often maintain their own
missions, and workforce turnover. The Contrail project is
“shoeboxes” of information that they collect, organize, and
addressing how to improve personal and workgroup
maintain. These are often a mixture of paper and electronic
information management and collaboration in this
items (including documents, email messages, bookmarked
environment. Contrail leverages machine-understandable
web pages, contact information for human collaborators,
models of users’ contexts automatically built up by
etc.) and are organized according to the analyst’s own
observing users performing information-keeping actions.
worldview. Because these collections can be highly
We define context as those factors that bound and
idiosyncratic in content and organization, they may be
surround a problem. Contrail develops context models
difficult to share among analysts working the same or
using text analytics, semantic networks, and spreading
related tasks.
activation techniques to facilitate information retrieval

and collaboration. The former leverages a knowledge base
Recent work from the study of personal information
(“shoebox”)
of
information
items
automatically
management tools and techniques (e.g., [2]) has shed some
maintained by Contrail. Items in this shoebox are tagged
light on how knowledge workers collect, organize,
with their original context. This facilitates information
maintain, and exploit their personal collections. As
retrieval, recall, and sharing among collaborators.
suggested by work in context-aware computing, a rich,
Contrail has been focused on aiding intelligence analysts
machine-understandable notion of the analyst’s task
but is generally applicable to other knowledge workers.
context can bring increasing automation to reduce the
This paper describes the motivation for this work in
burden on knowledge analysts for these information
context-aware computing and ethnographic studies of web
management tasks, promote information sharing and
usage; the Contrail tools; and a preliminary experiment to
transfer within (and across organizations), and ultimately
evaluate their utility for intelligence analysts.
improve the timeliness and quality of analytic products.



This paper describes how a model of an analyst’s (or
KEYWORDS: Collaboration Enabling Technologies,
similar knowledge worker’s) context can be automatically
Information Infrastructure for Collaboration, Management
created and exploited to improve collaboration by
of Metadata for Collaboration, Collaborative Knowledge
intelligently tagging information items of interest (web
Management, Context-& Situation Based Collaboration,
pages, email, documents, etc). This includes “collaboration
Ontologies and Ethnographic Approaches
with one’s past self” in which users are able to re-find,

recall, and re-use information they have previously
captured, perhaps even a significant while ago. Context

tags also facilitate information transfer and sharing across
1. INTRODUCTION
users.


Intelligence analysts (and similar knowledge workers) face
Our Contrail tools for context-based information
the challenges of information overload and management
management and collaboration facilitate the capture,
common to other knowledge workers today. They also face
management, and sharing of information items. To
increasingly compressed timetables for delivering mission-
automatically build its model of a user’s context, Contrail
critical products to intelligence consumers, an increasing
semantically analyzes the information a user handles. This
scope of missions to support, need for interagency sharing,
goes beyond traditional keyword extraction by applying a

---

spreading activation algorithm over an ontology that is
information was found in the first place? Over time, how
represented using the Web Ontology Language (OWL).
can analysts be reminded of information previously
Contrail then uses model of a user’s context to
exploited and individuals with whom they have
automatically tag information and store it in a
collaborated.
semantically-indexed “shoebox.” These tags can then help

in the user retrieve this information when it is most
This problem is important because people invest
relevant future and to share this information in context
significant time and effort in accumulating personal
with other users.
collections of items (bookmarks, email, and files) (e.g.,

[2] [20]). Users frequently revisit (“refind”) previously
Section 2 motivates our approach by defining our view of
found items: e.g., for web pages, 80% of visits to web
context and describes related work. Section 3 presents the
pages are revisits [7]. Jones et al. [16] tell of a variety of
Contrail tools in some detail. Section 4 describes a
“keeping methods” (keeping actions) for web pages:
preliminary experiment to evaluate the utility of our tools
• Send email to self or others
by intelligence analysts. Section 5 presents future work.
• Print out the web page.

• Save the web page as a file.
2. CONTEXT: WORKING DEFINITION
• Paste URL into a document.
AND RELATED WORK
• Add hyperlink to a web site.

• Do nothing (and enter URL directly later, search
Context-aware computing (CAC) applications tailor their
for or access from another web site).
behavior to the tasks and circumstances at hand and
• Bookmark the page.
provide the user increased automation such as
• Write down the URL on paper.
proactive/semi-autonomous
retrieval,
delivery,
and

presentation of relevant information from a variety of
Each method has its own advantages and disadvantages
sources (including human collaborators). For example,
[16]. For example, the “do nothing” methods require little
much of the CAC literature is from the ubiquitous
effort when something is found but can tax a user’s
computing community, where location of mobile users and
memory at refinding time. Bookmarks are less taxing to
computers is a key element of context (e.g., [6]).
refind but take greater effort to create, maintain, and find

in long lists or folder hierarchies. These keeping methods
Modeled after [3], we broadly construe “context” as those
vary in their ability to capture the context of an item: i.e.,
factors that bound and surround a problem. These factors
why the item was deemed important and saved in the first
include a worker’s job function, experience, skill level,
place—the “chain of reasoning”, search path, etc. that led
training, and other personal factors. Also included are
to the item. Bookmarks in lists and folders provide little
organizations’ policies, politics, and culture, and the work
context as the user only sees a name (which is typically
processes used (e.g., information available, procedures
unchanged by users from the default—the web page’s Title
followed, collaboration occurring, etc.). Often these
tag). Pasting a URL in an email or document, however,
contextual factors are on the periphery and not the
may be a way of keeping more context: i.e., if a user’s
conscious focus of an analyst’s deliberations: e.g., the time
comments or text pasted from the web page are also
pressures on an analyst, the analyst’s biases/pre-
included.
dispositions, etc. (e.g., [9]). The impact of context on

intelligence analysis has been well argued based on
Our hypothesis is that context (as accumulated by prior
cognitive and ethnographic studies (e.g., [11]).
and additional Contrail tools) can be used to improve

refinding, reusing, and sharing of items in one’s
Our current focus is exploiting context to aid analysts in
collections (“shoebox”). This, in turn, leads to the research
collecting, organizing, retrieving, and sharing personal
and development of capabilities including context capture,
information. Personal information management (PIM) is
context management, and context exploitation.
the “everyday process carried out by an individual as he or

she gathers, handles, and organizes information” [19]. A
Our work has benefited from related work in the research
lesser studied aspect of PIM is “keeping found things
community and the commercial world. The information
found” (KFTF) [17].1 For example, once found, how are
retrieval community has used user context to find
things organized for recall and reuse later on? What can be
information more relevant to the user’s task at hand.2
done to avoid the need to repeat the process by which the
Several commercial applications have utilized a user’s


1 Keeping Found Things Found® is a project at the Univ. of Washington,
2 e.g., The ACM SIGIR 2005 Information Retrieval in Context (IRiX)
http://kftf.ischool.washington.edu/
workshop: http://irix.umiacs.umd.edu/

---

desktop (open windows, stored folders and files, etc) as a
finds relevant items (web pages, email messages, etc.)
source of context to augment or replace a user’s explicit
through web browsing, web searches, reading email, etc.
query for relevant information. These include Autonomy’s
Through instrumentation and logging services, Contrail is
Intelligent Data Operating Layer® search tools,
notified of these “keeping actions,” such as the
blinkx.com’s video search tool, MediaRiver’s (formerly
bookmarking of a web page using Internet Explorer (IE)
Intellext’s) Watson® contextual search tool, Google
(Section 3.1). Contrail then performs a semantic analysis
Desktop® and others. Other tools have mined context
of each kept information item’s content using text
from the user’s past search history (web pages visited,
analytics and other methods. Using the results of this
queries) (e.g., [4]). While Contrail mines content that a
analysis, Contrail updates its model of the analyst’s
user has collected to build a (machine-exploitable) model
context (Section 3.2) and stores a copy of the kept item in
of her context, it applies a level of semantic analysis to
her Semantic Shoebox. The item is automatically
augment the mined content with information from an
annotated with content, metadata, and context tags.
ontologically-grounded knowledge base (Section 3.2).
Metadata tags are derived from the keeping action itself
Contrail is also focused currently on using context to tag,
(e.g., the URL and name of a web bookmark). Context tags
store, re-find, and share information (Section 3.3) versus
are derived from Contrail’s current model for the analyst.
query augmentation or relevance feedback techniques.
The analyst does not have to do any additional work

beyond her original action such as bookmarking the web
Current desktop tools such as web browsers provide little
page using IE.
support for capturing the context of items and helping the

user to recall that context later. The previously mentioned
As the analyst continues to do research, Contrail uses its
KFTF project work, leveraged by Contrail, has mostly
model of her current context to find and recommend
focused on studies of how users do this and less on tool
relevant past kept items in her Semantic Shoebox (Section
development. Microsoft’s Stuff I’ve Seen project provides
3.2). It does this by matching the current context to the
prototype tools for searching multiple collections of items
context tags that were originally stored with the item when
(e.g., web, email, etc.) simultaneously [10] [12].
it was originally kept. The analyst can view the prior
Microsoft’s MyLifeBits project has focused on capturing
context of a recommended item to recall why, when, and
and searching a lifelong log of information (including
how it was kept. This contextual knowledge can help the
audio/visual sensor data, electronic documents, etc.) [13].
analyst in reusing the item in the current context. As an
This log provides a personal memory and context to draw
analyst reviews items recommended by Contrail, she may
upon. Contrail’s end goals are similar to those of these
save them via keeping actions, thus repeating the cycle
projects but the underlying technology approaches are
shown in Figure 1.
different.


For a hypothetical use example, imagine a Middle East
3. THE CONTRAIL TOOLS
terrorism analyst Jane Jackson is tasked with evaluating

Palestinian terrorists ability to attack the continental U.S.
Figure 1 shows the high-level concept of operations for the
With Contrail running in the background, she begins some
Contrail tools. As an analyst does her research online, she
on-line research for recent finished intelligence and


relevant reports. As she is researching she bookmarks a
couple of relevant products using her browser, and she
copies and pastes several key pieces of information from
different sources into her working report. These “keeping
actions” are automatically captured by Contrails as she
works, updating her current context to reflect her current
interests. Jane likes to keep her Contrail Sidebar display up
as she works—this display recommends documents and
Contrail
recom-
potential collaborators based on her current context. She
mends
context-
notices that an analyst named John Smith is recommended,
relevant
sources
based on a collaborative exchange they had many months
before on Palestinian issues. Jane remembered that Smith
Behind the scenes,
Contrail captures
had since transferred to another position. She decides to
context
search Smith and other analysts’ Contrail shoeboxes to see
what work they have determined to be relevant. She sees

that Smith has captured several interesting articles on neo-
Figure 1. High-level Concept of Operations
Nazi/Palestinian cooperation. This was an angle she had
for Contrail Tools

---

not considered. She is able to open the articles and see not
3.1. Detecting Information Keeping Actions
only the text of the article, but what additional sources

Smith reviewed before and after and what other sources he
Keeping actions (described above) trigger Contrail to
reviewed with similar context and context terms. This rich
update its context model and augment its Semantic
information set provides her with another critical potential
Shoebox. Contrail’s Keeping Action Detector receives user
threat to the U.S. via neo-Nazi groups working with
events from instrumentation such as Battelle’s Glass Box
Palestinian terrorists that she may not have found
software, which captures events ranging from low level
otherwise.
ones such as keystrokes and mouseclicks to higher level

ones such as web queries [8]. The Glass Box works with
As the above example shows, Contrail supports the sharing
MS Windows™ and a number of desktop applications,
and retrieval of kept items from other analyst’s shoeboxes.
such as MS Word™ and MS Internet Explorer™. Contrail
The contextual knowledge appended to these items by
could also leverage other user-activity monitoring software.
Contrail helps one analyst quickly understand the potential
As captured events are often low level, Contrail does some
relevance and pedigree of an item retrieved from another
aggregation using software agents. For example, a “cut”
analyst’s shoebox. This supports rapid knowledge transfer.
(or “copy”) event from a web page is matched with a
With the current trend in the intelligence community of
corresponding “paste” event to aggregate an (implicit)
analysts switching jobs frequently, this system provides an
bookmark event.
extremely valuable tool to help a newcomer exploit the

information/knowledge captured by predecessors. This is
Contrail has been used to detect and process the following
in contrast, for example, to current methods of sharing
keeping actions:
bookmarks (e.g., via email or social bookmarking tools
• A user explicitly bookmarks a web page using
such as del.icio.us) where minimal contextual information
Internet Explorer, the Contrail plug-in to Mozilla
is provided with a shared bookmark. For example,
Firefox, or social bookmarking tools such as
del.icio.us supports tagging of bookmarks that enables
del.icio.us
some sharing of context but requires additional effort by an
• A user implicitly bookmarks a web page: e.g., a
analyst when bookmarking a web page. Also one analyst’s
user cuts text from a web page in a web browser
tags may not hold the same meaning for another analyst
and pastes it into an MS Word document.
who is viewing them.
• A user receives an email that contains a web page

URL from a specified user, including himself.
Figure 2 shows the Contrail tools and services. They are
• A user explicitly captures a web page, document,
discussed in detail in the sections that follow.
or email message using Contrail’s Desktop Sidebar

tool: e.g., a user drags a document onto the sidebar.

Except for those actions using Contrail’s Firefox™ plug-in
or the Contrail Desktop Sidebar tool, no additional user
effort is required by Contrail to capture these keeping
actions.

3.2. Tagging and Storing Found Items

When a keeping action is detected, Contrail’s Context
Aggregator (CA) performs the following actions
automatically:
1. The CA semantically analyzes the item (e.g., web
page, document, email, text that was cut/pasted)
associated with a keeping action by applying a
COTS text analytics tool to extract entities.
Contrail supports plug-and-play of different entity
extractors. We have experimented with Inxight’s
Thingfinder™ and Lockheed Martin’s Aerotext™
products. Entities are matched to concepts and

individuals (instances) stored in Contrail’s
Figure 2. Contrail Tools and Services (dashed boxes
Reference Database, a repository organized using
are external to Contrail)
domain ontologies described using the industry

---

standard Web Ontology Language (OWL) [14].
supports this (e.g., [1], [21]). A simple example is the
We call the matched concepts and individuals
mnemonic strategy of remembering a new person’s name
“context terms.”
by also noting several facts about the person and using
2. The CA stores extracted entities as content tags
those to more richly index the name. If a person notices
with the item in the Semantic Shoebox. Like the
and commits many details about a party to memory (e.g.,
Reference Database, the Semantic Shoebox is
the venue, people attending, atmosphere, etc.), that party
organized using OWL ontologies. This repository
event is likely to be recalled in more detail in more future
leverages the open source Sesame RDF/OWL
situations. The details stored in memory function as
triple store and additional services from our
retrieval cues. This is the premise behind case-based
Semantic Object Web (SOW) framework. The
reasoning systems in which episodes (“cases”) are stored
SOW framework provides a number of data
in great detail for recall and reuse in future situations [18].
markup/ingest tools to populate the repository with

knowledge about specific instances (people,
The current context is the CNG nodes (terms) with the
organizations, places).
highest activation levels. Node activation levels change
3. The CA augments context terms using the
over time as the content of documents handled by the user
spreading activation algorithm (described below)
changes.
Thus
the
context
CNG
represents
which updates the current context model for the
Contrail’sestimate of topics of interest to the user. The
user. The augmented set are stored as context tags
most active nodes in the CNG become the context tags for
with the item in the Semantic Shoebox.
a kept item. The current context model is currently the 20
4. The CA stores metadata tags with the item
most active nodes and their associated activation levels.
depending on the type of keeping action. For an

explicit bookmarking action using IE, for example,
Thus Contrail builds the current context model “bottom
the bookmark’s name and folder name are stored.
up” based on the dynamic actions of the user. Additional
The type, date/time, and user are also stored for
research is planned to investigate how to integrate “top
every keeping action. A user can supply optional
down” and less dynamic information into the current
metadata (e.g., a description) for kept items, either
context model such as the user’s identity, long-term
at keeping time or later (using the Contrail
interests, current resources (including collaborators), etc.
Refinder).
Each kept item is associated with a snapshot of the user’s

context taken at the time the keeping action was performed.
Figure 3 summarizes the information (tags, etc.) stored in
Taken in aggregate, these snapshots form the basis of a
the Semantic Shoebox that constitute the context of a kept
Context Log, which can later be mined for trends and
item. This information provides rich retrieval cues for the
monitored by collaborators. Contrail also has an
item. When people commit something to long term
experimental graphical user interface by which a user
memory, the more richly indexed that thing is, the more
optionally can edit his or her current context model: e.g.,
likely it is to be successfully remembered in the future.
add or remove a context term or change the activation for a
Much research from the psychology of long-term memory
term.



3.3. Refinding, Sharing, and Collaboration

Contrail’s current context model, context log, and semantic
shoebox for each user can be exploited in a number of
ways by Contrail and other applications, via Contrail’s
web service-oriented interface. Analysts can directly
access knowledge from Contrail through two (client-side)
applications, the Contrail Desktop Sidebar and Contrail
Refinder. The Sidebar presents kept items from the
analyst’s shoebox that are most relevant to the current
context. These are determined by comparing the current
context to context tags associated with kept items and
returning the best matches. If these relevant kept items
have an associated person (e.g., the sender of an email

message), that person is also listed as a recommended
Figure 3. Elements of Context Associated
collaborator. As described in Section 3.1, the Desktop
with a Found (and Kept) Item
Sidebar can also be used to manually capture an item (e.g.,

---

document, email, etc) and insert it into the Semantic
4. PRELIMINARY EVALUATION
Shoebox, which also updates the current context model.


Our team has completed an experiment with the National
The Contrail Refinder tool, shown in Figure 4, presents a
Institute of Science and Technology (NIST) to evaluate the
more comprehensive view of a Semantic Shoebox and
Contrail tools using intelligence analysts. The goals of the
displays a variety of information (textually and
NIST study were to determine (1) whether the Refinder
graphically) associated with a kept item including its
implementation in Contrail was able to support more rapid,
metadata, content, and context tags. A user may perform a
more accurate citation assignments than the Google®
one-button search to display those items most relevant to
baseline; (2) whether analysts could validate the content
the current context. He can also search for items associated
and weightings of the items in the Contrail system’s
with specific keywords or tags.
context representation; and (3) whether they could

appreciate any enhancement in the Contrail context over
From an item, a user can drilldown to view the item itself
time.
(e.g., web page), context at capture time, content tags, and

metadata. The context at capture time includes both the
The NIST experiment subjects were military intelligence
context tags and also the item’s temporal context. The
analysts of varying backgrounds. NIST facilitated
temporal context display shows those items kept just
experiment sessions (without involvement by Lockheed
before and just after a particular item.
Martin). Their primary tasks were: (1) Refinder

Capability—compare versus baseline tools, and evaluate
This provides the user a way to recall additional items that
use of Contrail versus current tools (e.g. Google®) for
might be relevant to their current needs. Viewing the
information (re)finding and (2) Context Assessment—
context under which an item was captured provides a way
Quality of Context Model: evaluate ability of Contrail
to help recall how, when, and why it was originally kept
context model to track and represent user context over time.
and thus how it might be useful in the current context.
Ideally, the experiment would enable a measure of the

refinding/reminding utility of Contrail for a single user
A user’s search via the Refinder can also access items
once a significant time period (e.g., one year) had elapsed,
stored in other users’ shoeboxes. A user can also export
but this was infeasible for the NIST study. Throughout the
items (and their associated context, content, and metadata
study, NIST captured quantitative as well as subjective
tags) to a file to later import into another shoebox.
inputs from subjects.




Figure 4. Contrail Refinder (Item Browser, Item General Details, and Item Source Details Screens)

---

The experiment yielded very interesting feedback that has
context aggregation algorithms. Contrail is focused on re-
helped guide the direction of our Contrail tools. NIST’s
finding of documents, but we plan to add pro-active,
final report showed quantitative results including:
context-aware web search capabilities to find new
• Seventy-five percent of the 20 terms in the current
information and collaborators. We will also focus on the
context (concepts & individuals) were judged
group context of a team of analysts and how that relates to
fitting by analysts (Figure 5).
its members’ individual contexts. This will enable tools for
• Of those items judged fitting, they were generally
collaboration among group members and oversight by
ranked higher by Contrail (Figure 6). Comparing
group managers (i.e., allowing managers to identify
each analyst’s data based on the value at the
problems like circular reporting or groupthink early in an
baseline (measurement period 0), NIST found that
analytic effort). Finally, the collected context over time of
three of the four analysts encountered richer
an individual or group should prove to be a wealth of
contexts in terms of fractional weights of ‘good’
information to mine for preferences, trends, lessons
items.
learned, and auditing purposes. We plan to investigate the

use of machine learning techniques in this area.

6. CONCLUSION

Contrail has shown the promise of context in enabling the
automated capture, organization, retrieval, and sharing of
information. This provides analysts with the benefits of
tailored automation including reductions in effort and
gains in efficiency. Context can facilitate collaboration,
allowing analysts to leverage the “network effect” to meet
today’s new and varied intelligence challenges. We have
developed a novel approach that combines machine-

understandable context representations from semantic web
Figure 5. Number of Context Terms that Fit by
technologies, context-aware computing methods, and text
Subject and Measurement Period; Each Line
analytics. The utility of Contrail has been demonstrated in
Represents a User
an experiment conducted by NIST. We have also begun to

apply Contrail’s techniques in applications outside
intelligence analysis to support other kinds of knowledge
workers.

ACKNOWLEDGMENTS

This work was funded by the Air Force Research
Laboratory. The experiment described in this paper was
supported by Dr. Emile Morse and her colleagues at the
National Institute of Standards (NIST).

REFERENCES


[1] Barsalou,
Lawrence
W.,
COGNITIVE
Figure 6. Fractional weight of Fit Items to Total
PSYCHOLOLOGY: AN OVERVIEW FOR COGNITIVE
Weight by Subject and Measurement Period; Each
SCIENTISTS. Lawrence Erlbaum, 1992.
Line Represents a User3


[2] Boardman, R. and M. A. Sasse, “Stuff goes into the
5. FUTURE WORK
computer and doesn't come out: a cross-tool study of
personal information management,” In Proceedings of the

SIGCHI Conference on Human Factors in Computing
Our research agenda includes the enrichment of Contrail’s
Systems, CHI '04. ACM Press, New York, Vienna,
context model and the improvement of the underlying
Austria, April 24-29, 2004, pp. 583-590.


3 Due to system problems, the information could only be collected
subjectively for users cast1 and cast2.

---

[3] Brézillon, P., “Context in Problem Solving: A Survey.
[11] Heuer, R., PSYCHOLOGY OF INTELLIGENCE
Knowledge Engineering Review,” Vol. 14, No. 1, 1999,
ANALYSIS. History Staff, Center for the Study of
pp. 47-80.
Intelligence, CIA, 1999.


[4] Capra, R., Perez-Quiones, M.A., and Ramakrishnan, N.,
[12] http://research.microsoft.com/adapt/sis/
“WebContext: Remote Access to Shared Context,” In the

Proceedings of the Workshop on Perceptive User
[13] http://research.microsoft.com/barc/mediapresence/
Interfaces, November 2001.
MyLifeBits.aspx


[5] Ceglowski, M., A. Coburn, and J. Cuadrado, “Semantic
[14] http://www.w3.org/2004/OWL/
Search of Unstructured Data Using Contextual Network

graphs,” Preliminary Whitepaper, National Institute for
[15] Jones, W., H. Bruce, and S. Dumais, “Keeping Found
Technology and Liberal Education, Middlebury College,
Things Found on the Web,” In the Proceedings of the
Middlebury, Vermont, 2003.
Tenth international Conference on information and

Knowledge Management, ACM Press, New York, H.
[6] Cheverst, K., K. Mitchell, and N. Davies, “Exploring
Paques, L. Liu, and D. Grossman, Eds., Atlanta, GA,
Context-aware
Information
Push,”
Personal and
October 5-10, 2001, pp. 119-126.
Ubiquitous Computing, Springer-Verlag, Vol. 6, Issue 4,

2002, pp. 276-281.
[16] Jones, W., S. Dumais, and H. Bruce, “Once Found, What

Then?: A Study of Keeping Behaviors in Personal Sse of
[7] Cockburn, A. and B. McKenzie, “What do Web Users Do?
Web Information,” In the Proceedings of the 65th ASIST
An Empirical Analysis of Web Use,” International Journal
Annual Meeting, Toms, E.G. (ed.), Philadelphia, PA,
Human-Computer Studies Academic Press, Inc., Vol. 54,
November 18-21, 2002, pp. 391-402.
Issue 6, 2001, pp. 903-922.


[17] Jones, W., H. Bruce, and S. Dumais, “How Do People Get
[8] Cowley, P., J. Haack, R. Littlefield, and E. Hampson,
Back to Information on the Web? How Can They Do It
“Glass Box: Capturing, Archiving, and Retrieving
Better?” In the Proceedings of the 9th IFIP TC13
Workstation Activities,” In Proceedings of the Third ACM
International Conference on Human-Computer Interaction,
Workshop on Continuous Archival and Retrieval of
Zurich, Switzerland, September 1-5, 2003.
Personal Experiences, ACM, New York, Santa Barbara,

CA, October 28-28, 2006), pp. 13-18.
[18] Kolodner, Janet L. CASE-BASED REASONING, Morgan

Kaufmann Publishers, San Mateo, CA, 1993.
[9] Devlin, K., “Confronting context effects in intelligence

analysis:
How
can
mathematics
help?”
2005,
[19] Lansdale, M., “The Psychology of Personal Information
http://www.stanford.edu/~kdevlin/papers.html.
Management,” Applied Ergonomics, PubMed, Vol. 19, No.

1, pp. 55–66.
[10] Dumais, S., E. Cutrell, J.J. Cadiz, G. Jancke, R. Sarin, and

D.C. Robbins, “Stuff I've Seen: A System for Personal
[20] MacKay, B., M. Kellar, and C. Watters, “An Evaluation of
Information Retrieval and Re-use,” In the Proceedings of
Landmarks for Re-finding Information on the Web,” In the
the 26th Annual international ACM SIGIR Conference on
Proceedings of the Human Factors in Computing Systems
Research and Development in Information Retrieval, ACM
Conference, ACM, Portland, OR, 2005, pp. 1609-1612.
Press, New York, Toronto, Canada, July 2- August 01,

2003, pp. 72-79.
[21] Williams, Michael D., and James D. Hollan, “The Process

of Retrieval from Very Long-Term Memory,” Cognitive
Science, Vol. 5, 1981, pp. 87-119.

---