Exploring Historical Image Collections with Collaborative Faceted Classification

The US Government Photos and Graphics Collection include
some of the nation’s most precious historical documents.
However the current federation is not effective for
exploration. We propose an architecture that enables users
to collaboratively construct a faceted classifi cation for this
historical image collection, or any other large online multimedia
collections. We have implemented a prototype for the
American Political History multimedia collection from usa.gov,
with a collaborative faceted classifi cation interface. In addition,
the proposed architecture includes automated document
classifi cation and facet schema enrichment techniques.
Introduction
It is difficult to explore a large historical multimedia humanities
collection without a classifi cation scheme. Legacy items
often lack textual description or other forms of metadata,
which makes search very diffi cult. One common approach
is to have librarians classify the documents in the collection.
This approach is often time or cost prohibitive, especially
for large, growing collections. Furthermore, the librarian
approach cannot refl ect diverse and ever-changing needs and
perspectives of users. As Sir Tim Berners-Lee commented: “the
exciting thing [about Web] is serendipitous reuse of data: one
person puts data up there for one thing, and another person
uses it another way.” Recent social tagging systems such as del.
icio.us permit individuals to assign free-form keywords (tags)
to any documents in a collection. In other words, users can
contribute metadata. These tagging systems, however, suffer
from low quality of tags and lack of navigable structures.
The system we are developing improves access to a large
multimedia collection by supporting users collaboratively
build a faceted classifi cation. Such a collaborative approach
supports diverse and evolving user needs and perspectives.
Faceted classifi cation has been shown to be effective for
exploration and discovery in large collections [1]. Compared
to search, it allows for recognition of category names instead
of recalling of query keywords. Faceted classifi cation consists
of two components: the facet schema containing facets and
categories, and the association between each document and
the categories in the facet schema. Our system allows users to
collaboratively 1) evolve a schema with facets and categories,
and 2) to classify documents into this schema. Through users’
manual efforts and aided by the system’s automated efforts, a
faceted classifi cation evolves with the growing collection, the
expanding user base, and the shifting user interests.
Our fundamental belief is that a large, diverse group of people
(students, teachers, etc.) can do better than a small team
of librarians in classifying and enriching a large multimedia
collection.
Related Research
Our research builds upon popular wiki and social tagging
systems. Below we discuss several research projects closest
to ours in spirit.
The Flamenco project [1] has developed a good browsing
interface based on faceted classifi cation, and has gone through
extensive evaluation with digital humanities collections such as
the fi ne art images at the museums in San Francisco. Flamenco,
however, is a “read-only” system. The facet schema is predefi
ned, and the classifi cation is pre-loaded. Users will not be
able to change the way the documents are classifi ed.
The Facetag project [2] guides users’ tagging by presenting a
predetermined facet schema to users. While users participate
in classifying the documents, the predetermined facet schema
forces users to classify the documents from the system’s
perspective. The rigid schema is insuffi cient in supporting
diverse user perspectives.
A few recent projects [4, 7] attempt to create classifi cation
schemas from tags collected from social tagging systems. So far
these projects have generated only single hierarchies, instead
of multiple hierarchies as in faceted schemas. Also just as
any other data mining systems, these automatic classifi cation
approaches suffers from quality problems.
So far, no one has combined user efforts and automated
techniques to build a faceted classifi cation, both to build the
schema and to classify documents into it, in a collaborative and
interactive manner.
Architecture and Prototype
Implementation
The architecture of our system is shown in Figure 1. Users
can not only tag (assign free-form keywords to) documents
but also collaboratively build a faceted classifi cation in a wiki
fashion. Utilizing the metadata created by users’ tagging efforts
and harvested from other sources, the system help improve
the classifi cation. We focus on three novel features: 1) to allow
users collaboratively build and maintain a faceted classifi cation,
2) to systematically enrich the user-created facet schema, 3)
to automatically classify documents into the evolving facet
schema.
Figure 1. System Architecture
We have developed a Web-based interface that allows
users create and edit facets/categories similar to managing
directories in the Microsoft File Explorer. Simply by clicking
and dragging documents into faceted categories, users can
classify (or re-classify) historic documents. All the fi les and
documents are stored in a MySQL database. For automatic
classifi cation, we use a support vector machine method [5]
utilizing users’ manual classifi cation as training input. For
systematic facet enrichment, we are exploring methods that
create new faceted categories from free-form tags based on a
statistical co-occurrence model [6] and also WordNet [8].
Note that the architecture has an open design so that it can
be integrated with existing websites or content management
systems. As such the system can be readily deployed to enrich
existing digital humanity collections.
We have deployed a prototype on the American Political
History (APH) sub-collection (http://teachpol.tcnj.edu/
amer_pol_hist) of the US Government Photos and Graphics
Collection, a federated collection with millions of images
(http://www.usa.gov/Topics/Graphics.shtml). The APH
collection currently contains over 500 images, many of which
are among the nation’s most valuable historical documents.
On the usa.gov site, users can explore this collection only
by two ways: either by era, such as 18th century and 19th
century, or by special topics, such as “presidents” (Figure 2).
There are only four special topics manually maintained by the
collection administrator, which do not cover most items in
the collection. This collection is poor with metadata and tools,
which is common to many digital humanity collections that
contain legacy items that have little pre-existing metadata, or
lack resources for maintenance.
Figure 2. American Political History Collection at usa.gov
The prototype focused on the collaborative classifi cation
interface. After deploying our prototype, the collection has
been collaboratively classifi ed into categories along several
facets. To prove the openness of system architecture, the
prototype has been integrated with different existing systems.
(Figure 3)
The system integrated with a Flamenco-like Interface
The system integrated with Joomla!, a popular
content management system
Figure 3. Multi-facet Browsing
As users explore the system (such as by exploring faceted
categories or through a keyword search), besides each item
there is a “classify” button which leads to the classifi cation
interface. The classifi cation interface shows the currently
assigned categories in various facets for the selected item. It
allows user to drag and drop an item into a new category. At
this level user can also add or remove categories from a facet,
or add or remove a facet.
Faceted Classifi cation button on the bottom of the screen (the button
to the right links to a social tagging system, del.icio.us)
The classifi cation interface. Users can create/edit facets
and categories, and drag items into categories
Figure 4. Classifi cation Interface
Evaluation and Future Steps
Initial evaluation results in a controlled environment show
great promise. The prototype was tested by university students
interested in American political history. The collection was
collaboratively categorized into facets such as Artifact (map,
photo, etc.), Location, Year, and Topics (Buildings, Presidents,
etc.) The prototype is found to be more effective than the
original website in supporting user’s retrieval tasks, in terms
of both recall and precision. At this time, our prototype does
not have all the necessary support to be deployed on public
Internet for a large number of users. For this we need to work
on the concept of hardening a newly added category or facet.
The key idea behind hardening is to accept a new category or
facet only after reinforcement from multiple users. In absence
of hardening support our system will be overwhelmed by
the number of new facets and categories. We are also
exploring automated document classifi cation and facet schema
enrichment techniques. We believe that collaborative faceted
classifi cation can improve access to many digital humanities
collections.
Acknowledgements
This project is supported in part by the United States National
Science Foundation, Award No. 0713290.
References
[1] Hearst, M.A., Clustering versus Faceted Categories for
Information Exploration. Communications of the ACM, 2006,
49(4).
[2] Quintarelli, E., L. Rosati, and Resmini, A. Facetag: Integrating
Bottom-up and Top-down Classifi cation in a Social Tagging System.
EuroIA 2006, Berlin.
[3] Wu, H. and M.D. Gordon, Collaborative fi ling in a
document repository. SIGIR 2004: p. 518-519
[4] Heymann, P. and Garcia-Molina, H., Collaborative Creation
of Communal Hierarchical Taxonomies in Social Tagging Systems.
Stanford Technical Report InfoLab 2006-10, 2006.
[5] Joachims, T. Text categorization with support vector
machines. In Proceedings of 10th European Conference on
Machine Learning, pages 137–142, April 1998.
[6] Sanderson, M. and B. Croft, Deriving concept hierarchies
from text. SIGIR 1999: p. 206-213.
[7] Schmitz and Patrick, Inducing Ontology from Flickr Tags.
Workshop in Collaborative Web Tagging, 2006.
[8] WordNet: An Electronic Lexical Database. Christiane
Fellbaum (editor). 1998. The MIT Press, Cambridge, MA.