MKWI 2010 – Conceptual Analysis and Ontological Modelling in Information Systems

27
Assisting the Discovery and Reuse of Document-
based Knowledge using Semantic Metadata
Hinnerk Brügmann
Department of Information Systems, University of Erlangen-Nuremberg
Lange Gasse 20, 90403 Nuremberg, Germany
hinnerk.bruegmann@wiso.uni-erlangen.de

1 Introduction
Due to the high importance computer-mediated communication and IT systems
have in the execution of today’s business processes, enterprises are producing a
rapidly growing amount of information as by-product of their business activities
(Zadok et al. 2004, p. 2f). Lyman and Varian (2003, p. 4) estimate the annual
growth of newly created information to be 30%.
According to studies of Merrill Lynch and Ferris Research only 20% - 40% of
this information is stored in a structured, semantically described form in electronic
databases or structured business applications (e.g. SAP R/3). The other 60% - 80%
are contained in the unstructured form of electronic documents1 (e.g. Microsoft
Office files, emails, images, multimedia files, or web based content). This is even
more dramatic as the annual growth rate of newly created unstructured content is
considerably higher than that of structured information. (Ferris Research 2008;
Blumberg and Atre 2003)
The unstructured nature of the bulk of existing as well as newly generated in-
formation is cause for a whole host of inefficiencies and problems. In a single-user
desktop environment unstructured documents are spread across local folder hie-
rarchies and contained as attachments in email messages. This leads to potential
redundancy and an increased required effort to locate document-based knowledge.
Such problems multiply in a multi-user enterprise environment with numerous
competing repositories, document management systems, shared network drives
and local file systems for each distinctive user. On one hand a person’s ability to
overlook the available number of information sources diminishes as the amount of
documents in numerous different repositories grows. On the other hand a multi-

1 In this context the term document is understood as relating to electronic files containing unstruc-
tured information.

Hinnerk Brügmann

28
user environment with local file system repositories lacks the central coordinating
instance of a single-user environment (where such would be the individual user
himself).
As a result individual knowledge workers in such multi-user enterprise envi-
ronments spend a substantial part of their working hours looking for the correct
and most up-to-date information needed in workflow steps or tasks (Guther 2007,
p. 25).
On an enterprise wide level there is one scenario we want to point out as an
exemplary motivational use case which is usually referred to as E-Discovery in the
context of legal litigation (Runyon 2007, p. 3f). Lawyers and internal staff of large
institutional clients face the problem of how to efficiently conduct searches for
relevant documents in large heterogeneous electronic data sets, for the purpose of
responding to litigation demands. According to Baron and Thomson lawyers typi-
cally overestimate their true rate of recall, i.e. how well their searches for docu-
ments have uncovered all relevant evidence (Baron and Thompson 2007, p. 141).
For this reason the analyst firm Forrester Research expects that companies will
heavily invest in technology and software products to assist in the discovery
process. But even though they estimate spendings to grow from $1.4 billion in
2006 to more than $4.8 billion in 2011 the methods employed by today’s profes-
sional E-Discovery tools still require a high amount of additional reviewing to
validate the results (Murphy 2006, p. 3).
While investments in advanced Enterprise Content Management (ECM) solu-
tions are made these often fail to meet expectations to bring order into the docu-
ment chaos as business units and individual users tend to circumvent seemingly
cumbersome ECM restrictions on document handling (Gilbert et al. 2006, p. 3f).
On top of that a working Information Lifecycle Management, which could at least
partly alleviate the problem of information overload and redundancy by deleting
outdated or redundant documents, is missing in many enterprises.
This paper illustrates an approach to assist in the enterprise wide lifecycle man-
agement and discovery of electronic documents by interrelating unstructured doc-
uments and correlating those documents to known business entities (employees,
partners, projects, products or processes) in a querieable way.
2 Related Work
Previous research approaches to establish and make use of relationships among
documents can be separated into two groups. The first group makes use of the
content of documents to detect similarity or references in documents, while the
second group of work relies on analyzing user activity involving documents usage
to infer document interrelations.

MKWI 2010 – Conceptual Analysis and Ontological Modelling in Information Systems

29
2.1 Content-based Relation Building
Implicit structure and similarity in document content can be used to define and
measure the relationship between two documents.
The project Stuff I’ve Seen (Dumais et al. 2003) relies on document metadata
and in part document content to create a context-snapshot of the time when a
document was accessed. A user might then at a later time query the Stuff I’ve Seen
data-store with arbitrary keywords to be presented with a list of documents he
accessed earlier. The project has a strong emphasis on the user-interface part to
allow for convenient querying those relations during a users day-to-day work in
information-heavy workflow tasks. This interaction allows the user to quickly re-
fine their query based on whatever contextual knowledge he can remember. A
second research project, Haystack (Adar et al. 1999) consolidates all accessed in-
formation in its own internal information-store and tries to recognize the occur-
rence of named entities (e.g. persons) in texts. In the Haystack data model, a typi-
cal application file is shredded into many individual information objects of various
types that are connected through application-specific relationships. A main prere-
quisite for such an approach is the need to give each information object a unique
identifying name. This way relations of documents to those entities and, indirectly,
to other documents are supposed to be detected. (Karger and Jones 2006, p. 82)
While techniques like Named Entity Recognition in electronic documents are
already quite advanced, more sophisticated parsing of electronic document content
is still problematic. One example are the current limitations of Natural Language
Processing (NLP) techniques which are not sufficiently solved to work resource-
efficiently with an acceptable margin of error or require an undue amount of ma-
nual effort to adjust and train the NLP algorithms (Castell et al. 2007, p. 14f).
Another limitation to content based approaches stems from the fact that certain
documents (e.g. multimedia files) contain content in proprietary formats which
complicate the content extraction or make it altogether impossible.
2.2 Activity-based Relation Building
A completely different view of document relationship discovery is less focused on
the documents themselves. Instead, the activities of a user around a document or
section are deemed the critical information for discovering document relationships.
Reading, editing, copying, pasting, sending email attachments or downloading files,
indeed any action a user can take with a document, are used to discover key
relationships. These approaches are mainly based on two assumptions: (a) the user
switches between different activities are detectable, and (b) each activity is associa-
ted with a set of resources relevant to that activity. Resources can be documents,
photographs, podcasts, email messages, web pages, RSS feeds, social bookmarks,
chats and so on. Activity management systems help with task switching and re-

Hinnerk Brügmann

30
source (re-)discovery by providing a context for organizing and accessing related
resources.
In TaskTracer (Dragunov et al. 2005) the user indicates when he begins a task
and when the task is complete. TaskTracer then monitors documents and user
activity to learn relevant folders and file locations, specific files manipulated, and a
range of application settings relevant to the task. Once the user has done this the
first time, on subsequent engagements with the task, TaskTracer will identify the
active task and reset the state of relevant applications and documents. Another
project, ActivityExplorer (Millen et al. 2005), takes a slightly different approach. In
ActivityExplorer the user specifies the boundary of tasks by explicitly indicating
the set of documents that are part of the task. In essence, this model has the user
explicitly indicate how documents are related; there is no automatic relationship
discovery. Explicit articulation of activity in ActivityExplorer, combined with tag-
ging, has been combined in a search interface to exploit the user specified relation-
ships. Relying on “tasks” as the principle means of document relationship discov-
ery requires identifying the connections of one piece of information in one applica-
tion to “task” related information in the same or other applications. Often a single
task or workflow will require the use of multiple, differing applications, resulting in
different interaction techniques and different representations to support the user.
Usage-Tracking has been deployed for a variety of purposes under a common
rubric of looking for the user’s intention. For instance, implicit feedback systems
attempt to infer user intent based on observable behavior, such as which docu-
ments she does and does not select for viewing, and how long he views them
(Oard and Kim 2001, p. 1f).
3 Our Approach
All of the above mentioned approaches operate in the domain of a single-user
desktop environment. As mentioned in section 1 the potential redundancy of doc-
uments as well as the corresponding complexity to unravel the information-chaos
rises exponentially when one looks at a multi-user environment. To the best of our
knowledge no research project has applied Activity-based Relation Building to the
domain of a multi-user environment as of yet.
Due to the described limitations of Content-based Relation Building our ap-
proach is set to avoid the ambiguities that arise when dealing with natural language
texts or visual images in multimedia files in an automated way. Instead we are try-
ing to utilize the capabilities of humans and computer systems where they fit best:
The cognitive ability of humans to understand the deeper meaning of unstructured
information and the analytical capabilities of computer systems to deal with large
quantities of structured data. At the same time the whole process should be as
transparent to the knowledge worker as possible to avoid additional workload on
his part.

MKWI 2010 – Conceptual Analysis and Ontological Modelling in Information Systems

31
To this end we propose to integrate the different types of available document-
related metadata, in particular static granular metadata (section 3.1) and dynamic
contextual metadata (section 3.2) of electronic documents. As will be shown it
becomes necessary to implement an initial seeding of external entity information to
act as relationship-anchors (section 3.3).
3.1 Static Granular Document-Metadata
One special peculiarity of unstructured information in electronic documents is that
those documents can be analyzed at different levels of granularity. Depending on
the particular document type one chunk of information may be contained inside
one sentence, span multiple paragraphs, or make up the whole document. Also
multiple information chunks may rest inside one document instance. This becomes
even truer when one considers an image or multimedia file with the different in-
formation chunks contained in sections such as foreground, background or
specific subparts of the image.
Due to their multi-informational nature electronic documents are, in contrast
to structured data, usually not automatically filed into single, one-dimensional tax-
onomies. Rather this classification falls to the human minds of knowledge workers
who perform the cognitive task to classify these documents into one or more cate-
gories (Kwasnik 1989). To give an example a person might manually save an in-
coming email attachment into his local folder hierarchy. He may also store the
attachment in multiple folders using symbolic links or creating redundant copies if
the document contains information bits relating to different topics.
The granularity of many electronic documents is in itself additional meta-
information that would not be present had the contained information, piece by
piece, been extracted, structured and stored in a database system. A human user
performed a cognitive effort when assembling multiple chunks of information into
one document. This allows the assumption that those pieces of information are
connected to each other semantically. A similar line of thought can be drawn re-
garding the filing and categorizing a human person does when organizing his per-
sonal archive, email management application or file system.
The, from a knowledge management point of view, at first sight complicating
aspects of unstructured information in an electronic document may actually prove
to be a possible point of leverage. They can, as in the saying the whole is more
than the sum of its parts, provide additional meta-information about semantic
relations the document may have regarding contained information or regarding
semantic relations the document may have to other documents.
3.2 Dynamic Contextual Document-Metadata
In addition to the granular aspect of electronic documents further meta-
information can be extracted from the context in which knowledge workers access

Hinnerk Brügmann

32
and modify those documents (see section 2.2) (Barreau 1995, p. 329ff; Shen et al.
2005, p. 2f). In this paper the context of document access is understood as the sum
of all actions taken by a user as well as the workplace environment shortly before,
during, and after access of an electronic document.
The individual elements compromising such context can be separated into two
dimensions: time and scope. The dimension of time is split into context elements
which happened before, during, and after the document access. The dimension of
scope differentiates among user actions and workplace environment. Of course the
workplace environment and especially changes to it are always results of user ac-
tions. Therefore we restrict the definition of a user action to some input from the
user which directly relates to a document, not to an application in itself. In contrast
the workplace environment component in the context of a document access con-
sists of all opened desktop and web applications as well as all content and docu-
ments visible in these applications. Additionally spontaneous communication by
short email or instant messaging, which may not contain much meaningful infor-
mation in itself, can act as glue by appearing in the same context as two or more
other actions, linking them together.
To give a simple business scenario example: A user in the sales department
copies some textual content from document price-list.doc into a new document sales-
offer.doc and saves it in the local file system folder customer-alpha.

scope
user action workplace environment
time
before
(WE0) opened document
price-list.doc
during
(UA1) copying text from price-
list.doc into new document sales-
offer.doc
(WE1) open documents
price-list.doc and sales-
offer.doc
after
(UA2) saving document sales-
offer.doc into folder customer-
alpha on the local file system
(WE2) opened document
price-list.doc
Figure 1: Exemplary context of electronic document access

Figure 1 shows the context of the action UA1. The user action (in this case
copying) allows the assumption (not certain knowledge) of an existing semantic
relation between the documents price-list.doc and sales-offer.doc. The strength of this
assumption as well as the type of semantic relation depend on heuristics based on

MKWI 2010 – Conceptual Analysis and Ontological Modelling in Information Systems

33
typical usage patterns of document access. Additionally the metadata gained from
observing user action UA1 might be influenced by the corresponding workplace
environment WE1. So could for example a window-setup which shows both do-
cuments next to each other without overlap strengthen the assumption of an
existing relationship. Part of the context of UA1 is the second user action UA2.
This in turn can be used to establish a new (or strengthen an existing) semantic
relation of sales-offer.doc and (to a lesser extend) price-list.doc to other documents
contained in folder customer-alpha.
From a modelling perspective UA2 has its own context, which would then
contain UA1 as an action that happened shortly before UA2.
3.3 External Business Entity Metadata
So far only one abstract entity-type was mentioned, document. Semantic relations
among documents can be deduced by analyzing both the taxonomic placement in
file hierarchies as well as the context of document access and modification. Both
scenarios build upon the cognitive effort a knowledge worker put into those tasks.
For the resulting network of relations to have any use in the exemplary use
case of E-Discovery a link to the business domain must be established. For this
purpose we propose the insertion of external entity-types (e.g. product, process,
role or person). These external entities, when linked to documents by semantic
relations, can then be part of the context of a document access. The above de-
scribed inferring of additional relations by analyzing the context of document
access can in that case allow new assumptions to be made regarding semantic rela-
tions between external entities and documents. If we know document price-list.doc in
the example shown in Figure 1 has a semantic relation to some specific product
the assumption can be made that document sales-offer.doc also has a (however light
or strong) relation to the same product.
The question is than how to initially link these external entities to documents.
Fortunately certain relations are, in a usual enterprise environment, already present
in machine readable ways. For example an Active Directory installation in combi-
nation with an Identity and Access Management (IAM) solution relates person
records to product- or process-related roles. Also at least the centralized Docu-
ment Management solutions often require users to tag uploaded documents or to
enter metadata to place a documents main topic into a predefined taxonomy. Last-
ly collaborative project workspaces inherently infer a project-relation to the con-
tained documents.
4 Aggregation and Analysis of Semantic Metadata
Section 3 mentioned the uncertainty which comes with every assumption made on
semantic relations among entities. After all, without explicitly asking the specific

Hinnerk Brügmann

34
user an automated system cannot know for sure if the exemplary documents price-
list.doc and sales-offer.doc actually have any sort of semantic relation. The research
hypothesis underlying this paper is that, similar to the approach of collaborative
filtering, with a high number of data sets the amount of white noise of false positi-
ves can be reduced below an acceptable margin of error. The use case of this app-
roach being a collaborative enterprise environment - with a high number of users
and multiple occurrences of the same document in various repositories – has the
secondary effect that the amount of both reliable2 taxonomical as well as unreliable
contextual and inferred (meta-) data might be immense. Before dynamic contextual
information as described in section 3 can be utilized the basic methods of
aggregating and analyzing semantic metadata in general need to be evaluated.
Section 4.1 describes the prototypical “ConSense” system to construct and ag-
gregate basic semantic relations among documents in a multi-user scenario. Section
4.2 builds upon this prototype and shows how the immense amount of generated
semantic (meta-) data can be handled. To evaluate the technical viability of the
described approach the test case is limited to static document-based metadata. The
processing of dynamic contextual metadata will have to be shown in a following
paper.
4.1 Overview of the ConSense Prototype
To aggregate and analyze the assumptions made on semantic relations we
implemented a distributed prototypical application “ConSense”.
The business domain in this scenario is the ecosystem of knowledge workers in
the enterprise accessing and modifying electronic documents as well as collaborat-
ing in teams and exchanging email messages.
Client-side plug-ins on the knowledge workers workstations extract static me-
tadata from documents as well as detect and record the context of user activities
related to document access and modification. Additionally sensors obtain the con-
text of document-related user actions in the enterprises central document man-
agement applications and collaborative project workplace solutions. To ensure
interoperability among different subsystems ontologies representing the set of
concepts specific to the enterprise domain and the possible relationships among
those concepts are used. The main ontology classes document, process, product,
role and person can further be customized depending on industry-specific peculiar-
ity. A good part of the Information-Lifecycle-specific abstract classes, as well as
their possible relationships are imported from the FOAF and Dublin Core ontolo-
gies. The sensor plugins validate gathered context input against the common do-
main ontologies and persist it in the form of semantic networks in local Resource
Description Framework (RDF) triple-stores. In the next step the sensor-plugins

2 In this context the term reliable is used to differentiate first-order metadata from assumptions
containing unreliable second-order (inferred) metadata.

MKWI 2010 – Conceptual Analysis and Ontological Modelling in Information Systems

35
add further statements based on the initially gathered static metadata by using heu-
ristics based on typical usage patterns of document access. These statements con-
tain assumption on the similarity or implicit correlation of either document-pairs
or relations of documents to seeded business entities.
Document-related metadata with relevance beyond that of the local user (that
is higher-level metadata relating to the domain-specific business entities) is then
submitted to a central semantic information repository which is also implemented
as an RDF triple-store. To reduce the amount of necessary network communica-
tion between a client and the central semantic store the transmission of RDF
triples is selective. Even though both subsystems reference the same ontologies -
meaning they have the same understanding of document and real-world entities
and their possible interrelations – different subsets of the ontologies are in actual
use. So will the central semantic store only consider aggregated and higher-level
predicates which will actually be the target of subsequent discovery queries (regard-
ing domain specific-questions) as input from the sensor plugins. This allows for
baseline metadata to remain in the local RDF-stores on the clients while only high-
er-level metadata (the common ontology-subset) is transferred and made available
to other users. Additionally the central semantic store – upon receiving sensor data
- condenses this new data by merging statements with the same subject resource
into the existing network. This is for example used to aggregate the reification-
statements on the confidence of inferred interrelations from the client sensors.
The central semantic store, representing a virtualized view of assumptions on
real-world relations among business-entities and documents, can in turn be queried
by a user or knowledge manager to visualize and cluster relationship types accord-
ing to his task-specific discovery needs. In the current implementation a
SPARQL/Update endpoint is exposed as a query interface (a future prototype
should allow for more sophisticated access and come with domain-specific pre-
configured queries).
For example the litigation manager of the E-Discovery scenario could query
for and detect documents having a relation to a product line being the subject in a
legal low-product-quality complaint by a client. The query parameters could then
be narrowed to documents having additional relations to the companies Quality
Management process. Alternatively a rule- or heuristic based electronic service can
directly access the semantic aggregation layer and query or manipulate the semantic
network.
4.2 Coping with very large sets of inferred Metadata
To test the scalability of the used approach we ran an initial test of a prototypical
client-side sensor-plugin on four test- clients. External entities were seeded in the
RDF-encoded form of a corporate directory and a list of corporate products and
services.

Hinnerk Brügmann

36
The sensor-plugins read and locally stored all static meta-attributes of documents3
contained in the users “My Documents” folders. Extracted meta- attributes were
then matched to a prepared set of RDF predicates referencing a subset of the
FOAF and Dublin Core ontologies. The average number of documents on these
test clients was in the range of 10,000. Including meta-attributes specific to Micro-
soft Office documents on average 50 RDF-triples per document were generated.
This resulted on average in 500,000 triples containing reliable information.
Comparing MD5 hashes of document content the overlap of reoccurring docu-
ments across clients lay at 8% of the total amount of detected documents. In the
next step the sensor-plugins added further statements based on document name
and location in the local file system. For example the titles of document- as well as
folder-pairs were compared using the Smith-Waterman-Gotoh-Algorithm. When
the similarity exceeded a minimum-threshold a statement containing the predicate
“similarContent” connecting the two documents or folders in focus was added. In
a similar way links to the business domain were created by matching properties of
the initially seeded business entities to meta-attributes of documents and their
containing folder taxonomies (Named Entity Recognition). Additionally these
meta-statements containing unreliable assumptions were each referenced with a
reification-statement of predicate-type “confidence” proportional to the numeric
result of the respectively applied similarity-algorithm.
In the test setup this raised the total number of statements in the local RDF-
store of each client by 400,000 to 900,0004. Seeing that an average RDF-statement
in this test consisted of ~290 bytes (with dereferenced namespaces ~460 bytes)
such an amount of RDF-triples is beyond an acceptable technical boundary both
regarding transmission of data to a central semantic store as well as regarding the
performance of further inferencing in the resulting semantic network.
Combining the aggregation-methods described in section 4.1 the number of
RDF-statements persisting from the local RDF stores into the central semantic
store could be reduced by 95%. In the test setup this resulted in a store of 160,000
statements of which 18,000 specifically described semantic relations of the 800
identified common documents on all four clients.
Decoupling the semantic repositories of clients and the central store has two
further beneficial side effects: (1) It allows for system-resilience regarding outage or
unavailability of individual sensors. (2) The sensors can filter data to be transmitted
against predefined black- or white-lists and prevent too personal predicates (“ha-
sRead”, “visitedWebsite”) from leaving the boundaries of the client desktop.

3 In this test Microsoft Office files, multimedia files, pdf as well as postscript files, and text-files
(including program source code) were considered.
4 Further tests involving a dramatically higher document population (as could be expected in an
enterprise document repository) of 200,000 documents resulted in the expected 10 million reliable
statements and an additional 20 million unreliable assumptions using the same algorithms. The num-
ber of possible document-interrelations rose exponentially.

MKWI 2010 – Conceptual Analysis and Ontological Modelling in Information Systems

37
5 Conclusions and Outlook
The described approach is intended to utilize the cognitive acts knowledge workers
perform during their everyday work, by analyzing the context of those acts and
aggregating assumptions on document interrelations using semantic network
techniques. In the test setup heuristics were used to condense the initially very
large number of resulting semantic statements. This lowered the amount of data
which needs to be transferred to a queriable central semantic store to an acceptable
amount making further inference in the network technically viable.
In future research the underlying hypothesis of the feasibility to extract seman-
tic relations from work-context will need further evaluation. Do humans in PC-
based workplace environments really behave in such a way that the extracted as-
sumptions on interrelations hold truth?
Also the mentioned heuristics to detect document-interrelations will have to be
further tested and improved to reduce the number of false-positives. Here domain-
specific heuristics might prove to be valuable. Lastly legislative aspects, especially
privacy concerns of employees, have to be considered. Here a combination of
white- and/or blacklisting named entities might be feasible, to specifically include
business process relevant documents only or to exclude documents and communi-
cation of sensible parties from the context readings.
References
Adar E, Kargar D, Stein LA (1999) Haystack: per-user information environments.
In Proc. CIKM'99, pp. 413-422.
Baron JR, Thompson P (2007) The search problem posed by large heterogeneous
data sets in litigation: possible future approaches to research. ICAIL '07:
Proceedings of the 11th international conference on Artificial intelligence and
law, pp. 141-147.
Barreau D (1995) Context as a factor in personal information management
systems. Journal of the American Society for Information Science 5(46), pp.
327-339.
Blumberg R, Atre S (2003) The problem with unstructured data. DM Review
Magazine, http://www.dmreview.com/issues/20030201/6287-1.html, last
accessed on 200909-05.
Castell P, Fernández M, Vallet D (2007) An adaptation of the vector-space model
for ontology-based information retrieval. IEEE Transactions on Knowledge
and Data Engineering, vol. 19, no. 2, pp. 261-272.

Hinnerk Brügmann

38
Dragunov AN, Dietterich TG, Johnsrude K, McLaughlin M., Li L, Herlocker, J
(2005) TaskTracer: a desktop environment to support multi-tasking
knowledge workers. In Proc. IUI’05. ACM Press.
Dumais ST, Cutrell E, Cadiz J, Jancke G, Sarin R, Robbins DC (2003) Stuff I've
Seen: a system for personal information retrieval and re-use. In Proc.
SIGIR'03, pp. 72-79.
Ferris Research (2008) Industry statistics. http://www.ferris.com/research-
library/industry-statistics/. last accessed on 2009-07-20.
Gilbert MR, Shegda KM, Logan D, Chin K, Bell T, Latham L, Knox RE, Lundy J
(2006) Key issues for enterprise content management. Gartner Research.
Guther M (2007) Duet - eine Software verknüpft Microsoft Office mit SAP-
Geschäftsanwendungen. IM - Information Management & Consulting.
Karger DR, Jones W (2006). Data unification in personal information
management. Commun. ACM 49 (1), pp. 77-82.
Kwasnik BH (1989) The influence of context on classification behavior. Doctoral
thesis, Rutgers University.
Lyman A, Varian B (2003) How much information? Techreport, University of
California,Berkeley.
Millen DR, Muller MJ, Geyer W, Wilcox E, Brownholtz B (2005) Understanding
users and usage patterns: Patterns of media use in an activity-centric
collaborative environment. In Proc. CHI 2005, ACM Press.
Murphy B (2006) Believe It - eDiscovery Technology Spending To Top $4.8
Billion By 2011. Forrester Research.
Oard DW, Kim J (2001) Modeling information content using observable behavior.
In Proceedings of the 64 Annual Meeting of the American Society for
Information Science and Technology.
Runyon B (2007) Data classification is a vital first step in information life cycle
management. Gartner Research.
Shen X, Tan B, Zhai C (2005) UCAIR: Capturing and exploiting context for
personalized search. Techreport, University of Illinois at Urbana Champaign.
Zadok E, Osborn J, Shater A, Wright C, Muniswamy-Reddy K, Nieh J (2004)
Reducing storage management costs via informed user-based policies. 12th
NASA Goddard, 21st IEEE Conference on Mass Storage Systems and
Technologies (MSST 2004)