Haystack: A User Interface for Creating, Browsing, and

Organizing Arbitrary Semistructured Information

David R. Karger Dennis Quan
MIT CSAIL IBM T. J. Watson Research Center
200 Technology Sq., Cambridge, MA 02139 USA 1 Rogers Street, Cambridge, MA 02142 USA
karger@theory.lcs.mit.edu dennisq@us.ibm.com
ABSTRACT possible to modify software as needed to integrate specific
Much past HCI research has examined the usability con- domains on an ad hoc basis, one might hope to solve the
cerns of information management software for specific do- problem once and for all by designing a system in which all
mains such as object-oriented software design, e-mail, and information is naturally integrated from the start. Systems,
the Web. We believe that many of the results uncovered by such as Lifestreams [7], Presto [2], and Semantic Regions
these studies are applicable across multiple domains but [8] show that many personal organization and navigation
that more broadly-scoped experiments require a system that tasks occur most usefully at the attribute level, regardless of
can integrate multiple data sources. Haystack is a general- application. Arguably, these systems owe their flexibility to
purpose information management environment designed to the generality of their attribute-value pair-based data mod-
attack this very problem. Haystack’s user interface, which els. In other words, the data model is what enabled these
incorporates capabilities from previous research such as systems to push beyond the restrictions of a single domain
context-specific visualization paradigms and attribute-based to explore cross-domain information interaction.
categorization, is built upon a highly expressive semistruc-
tured data model and data integration capabilities. In our Our Haystack project applies this idea by layering a system
demonstration we show how combination of a direct- atop a fully general semantic network data model. All in-
manipulation-based UI paradigm and an expressive, feder- formation—both data and metadata—is represented in a
ated data model can begin to address many of the informa- labeled directed graph [6]. But unlike some other rich-data-
tion management problems plaguing general desktop com- model systems, we aim to move beyond low-level inter-
puting today and can serve as a basis for further, yet unex- faces to the data model such as graph navigation (e.g., Is-
plored, crossover information interaction experiments. aViz [14]) or frame-based navigation (e.g., Protégé [10]).
Instead, objects in our data model are rendered to the screen
Author Keywords by components called views (as in model-view-controller).
Views, semistructured data, direct manipulation, RDF. Various views (e.g., an icon-sized view, a paragraph-sized
view, a full-screen view, a one line summary view, etc.) are
ACM Classification Keywords defined for each type of object in the data model. Views of
H.5.2. Information interfaces and presentation (e.g., HCI): a complex object are defined recursively in terms of views
User Interfaces. of other objects in the data model to which the base object
is “connected.” Once rendered, the view serves as a proxy
INTRODUCTION for the object—views can be right-clicked to access context
The usability problems inherent in programs that help users menus and dragged and dropped to effect operations [5].
to interact with their information have been explored for To populate the model, a multi-agent environment incorpo-
many domains; good examples include Whittaker and rates information from different sources (e-mail, the file
Sidner’s study into e-mail overload [4], Storey et al.’s visu- system, MP3 headers, etc.) [6]. Shared ontologies unify
alizations for object-oriented program design [13], and metadata (such as title and creation date) that recur in many
Abrams et al.’s study of Web bookmark usage [3]. Many separate domains. Furthermore, higher level concepts such
proposed systems addressing these usability issues have in as collections and contexts tie objects in our data model
common the notion that cross-domain approaches are key. together and represent generalizations of common themes
For example, Bellotti et al.’s Taskmaster system demon- discussed in the HCI literature that address problems of
strated that e-mail and task management need to be pre- organization and contextualization, respectively. These two
sented to the user within a single framework [1]. While it is concepts are described below.
Copyright is held by the author/owner(s). COLLECTIONS: CROSS-DOMAIN ORGANIZATION
CHI 2004, April 24–29, 2004, Vienna, Austria.
ACM 1-58113-703-6/04/0004. Although not often thought of as being the same type, items
ranging from photo albums and bookmark hierarchies to to-
do lists and e-mail inboxes are all collections. Considering
that users more naturally group objects by purpose rather

1
than by type (though these may sometimes coincide) [1], the ticket. We also demonstrate visualizing objects such as
the real distinction between these different forms of collec- Charles’s flight, research papers, and instant messaging
tion is the style of presentation. In Haystack, all collections conversations.
are treated the same. Any collection can contain multiple
types of objects at once. Views let users vary the collec- FUTURE WORK
tion’s presentation depending on task, regardless of the con- The Haystack system is Open Source, downloadable from
tained objects’ types. our Web site at http://haystack.lcs.mit.edu/. While still
somewhat rough from a usability perspective, the current
The “directory” is a collection commonly used for generic prototype illustrates the base functionality upon which we
organization. Existing implementations have conflated the intend to improve the user experience and research the long
two distinct roles such collections play in many file sys- term effects of unified information management.
tems: storage and organization [2]. While restricting files to
being stored in one directory at a time may be logical for ACKNOWLEDGMENTS
storage efficiency, such a restriction inhibits the use of di- This research was supported by MIT Project Oxygen and
rectories (as well as bookmark folders, e-mail folders, and the MIT-NTT collaboration.
other strict hierarchies) as effective instruments for organi-
zation. In Haystack, objects can easily be associated with REFERENCES
multiple collections. Early user studies, both within Hay- 1. Bellotti, V., Ducheneaut, N., Howard, M., and Smith, I. Tak-
stack and in the context of the Web browser, suggest that ing Email to Task: The Design and Evaluation of a Task Man-
users benefit significantly from this added cross-domain agement Centered Email Tool. Proc. CHI 2003.
flexibility [11, 12]. 2. Dourish, P., Edwards, W.K., et al. Extending Document Man-
agement Systems with User-Specific Active Properties. ACM
CONTEXTS: CROSS-DOMAIN FUNCTIONALITY Transactions on Information Systems 18 (2), 140–170.
One particular useful role that applications play today is to 3. Abrams, D., Baecker, R., and Chignell, M. Information Ar-
aggregate a collection of commonly-used views and opera- chiving with Bookmarks: Personal Web Space Construction
tions in a single location. This is extremely useful until the and Organization. Proc. CHI 1998.
views and operations needed by a user diverge from the 4. Whittaker, S. and Sidner, C. Email Overload: Exploring Per-
predefined aggregation by application. Instead of locking-in sonal Information Management of Email. Proc. CHI 1996.
a specific aggregation, Haystack offers contexts, which are 5. Quan, D., Karger, D., and Huynh, D. RDF Authoring Envi-
lightweight assemblies of views and operations. Rather ronments for End Users. Proc. Semantic Web Foundations and
than“switching” applications to indicate the current context, Application Technologies 2003.
users can specify the contexts that are active. A pane ap- 6. Quan, D., Huynh, D., and Karger, D. Haystack: A Platform for
pears on the left hand side of the screen for each active con- Authoring End User Semantic Web Applications. Proc. Int’l
text. This is done in analogy to people’s desks, where an Semantic Web Conf. 2003.
active context may have an associated folder or paperwork 7. Freeman, E. and Gelernter, D. Lifestreams: A Storage Model
placed on the desk. Context-specific commands and infor- for Personal Data. SIGMOD Record 25 (1).
mation are then shown on screen in the left hand pane.
8. Kang, H. and Shneiderman, B. MediaFinder: An Interface for
Grouping commands and objects by context rather by ap- Dynamic Personal Media Management with Semantic Re-
plication allows functionality from multiple applications gions. Proc. CHI 2003.
(e.g., schedule appointment from the calendar program, add
9. Yee, K., Swearingen, K., Li, K., and Hearst, M. Faceted Meta-
to expense account from the expense tracker) to become
data for Image Search and Browsing. Proc. CHI 2003.
available for use within any given user-centric context (e.g.,
travel planning). 10. Eriksson, H., Fergerson, R., Shahar, Y., and Musen, M. Auto-
matic Generation of Ontology Editors. Proc. 12th Banff
Knowledge Acquisition Workshop 1999.
EXAMPLE SCENARIO
11. Quan, D. (2003). Designing End User Information Environ-
Our demonstration illustrates a scenario in which a Charles, ments Built on Semistructured Data Models. Doctoral Disser-
a biologist, is preparing to attend the fictitious HIV 2004 tation.
conference in San Francisco with his colleagues, John and
12. Quan, D., Bakshi, K, Huynh, D., and Karger, D. User Inter-
Mary. Charles starts off by reading e-mail messages from faces for Supporting Multiple Categorization. Proc.
his colleagues within Haystack and using Haystack’s col- INTERACT 2003.
lection tools to organize lists of to-do’s (e.g., book ticket
13. Storey, M., Best, C., Michaud, J., Rayside, D., Litoiu, M., and
using John’s itinerary), talks to see during the conference,
Musen, M. SHriMP Views: An Interactive Environment for
and authors of related work he should get in contact with. Information Visualization and Navigation. Proc. CHI 2003.
Charles then books an airline ticket from within Haystack
14. Pietriga, E. IsaViz. http://www.w3.org/2001/11/IsaViz/.
by dragging and dropping objects such as John’s itinerary
and an icon view representing the person who will pay for