Communications of the ACM

"Flamenco" allows users to move through large information spaces in a flexible manner without feeling lost.

We plan to augment the system in several ways, including incorporating thesaurus term matching and using techniques from adaptive user interface research.

There are a number of issues associated with the creation of metadata itself that are not addressed in this article. The most pressing problem is how to decide which descriptors are correct or at least the most appropriate for a collection of information. Another problem relates to how to assign metadata descriptors to items that currently do not have metadata assigned. Many researchers are addressing these issues, and the field of automated text categorization is making great strides. Additionally, many important collections with hand-assigned hierarchical metadata already exist.

We illustrate the interface using an architectural image database containing about 40,000 photographs of landscapes and buildings from a wide variety of historical periods, styles, and geographic regions (see Figure 1). The images are classified under about 16,000 hierarchical metadata terms, which we manually reorganized into nine facets: people, locations, structure types, materials, periods, styles, view types, concepts, and building names.

We use a brief scenario to demonstrate how the interface works. Imagine a user named Claire who has a beach house she plans to renovate, with the goal of bringing more natural light into the living room. Before she meets with the architect, she browses through the architectural image collection to gather a few ideas.

She begins at the starting page (see A in Figure 1). This page shows an overview of available topics, each hyperlinked to the equivalent of a query on the corresponding metadata term, and each link showing how many items have been assigned that topic label. To help her in this search, the starting page also includes three sample images from each facet. To begin her search, Claire may either click one of these links or issue a keyword search.

Claire opts to begin by clicking the hyperlink "interior views" in the View Types facet and has arrived at what we call the "matrix view" (see B). There is a column of metadata on the left and the images in the current result set on the right. The matrix shows query previews for all of the metadata terms assigned to the images in the current result set. These previews are updated as constraints are added or removed. The caption under each image gives the name of the building, the location, and the architect.

Claire's eye is drawn to one image showing an interior flooded with daylight. She clicks on this image to see a more detailed view (C). After reading the metadata categories assigned to the image, Claire clicks on the term "windows" found under the Structure Types facet. This refines her query because it conjoins the metadata term "windows" with the current query. Doing this creates a new matrix page (see D).

Now the query, consisting of metadata from the two selected facets (View Types and Structure Types), is shown at the top of the screen in the form of hyperlinked history trails (or "breadcrumbs"). The images are grouped according to subcategories of the "windows" metadata category; up to four sample items are shown in each subcategory. Note that the interface allows the user to navigate multiple hierarchies simultaneously.

To further refine her search, Claire can select terms from other facets by clicking in the matrix on the left or by selecting a subcategory on the right. The results set can be broadened (expanded) to include more items by selecting a general category within the breadcrumb or by clicking the X to remove a category constraint. Assume that Claire clicks on the "openings" category, just above "windows" in the breadcrumb, to relax the Structure Types constraint. This brings her to E.

Clicking on an image within the "skylights" subcategory brings her to the image detail (F) where she sees several other helpful terms: "daylight," "beams," and "beach houses." This page allows Claire to make lateral moves, shifting to associated categories that were not part of the original query. We have found this facility is important for promoting shifts to areas of the collection that users had not considered previously.

The interface makes a keyword search facility available at all points in the interface. The scope of the search is by default the current result set, although users can also choose to search through the entire collection.

Thus Claire might have begun her exploration by running a keyword search on the word "light." In this case, a list of all matching metadata terms appears above the result set, with search terms highlighted as shown in G. Selecting the metadata term "skylights" in the list converts the keyword constraint into a category constraint as shown in H. This, in turn, could lead her to the image detail in F.

In some situations, there are too many subcategories or keyword matches to fit on the page. When this occurs, an alphabetized list is presented on a page of its own so the user can make a selection. The links labeled "more"—visible in the matrix view—can also take the user to listing pages of this type.

Returning now to the discussion of usability guidelines for search interfaces, notice this interface supports six out of Shneiderman et al.'s eight design desiderata. It is consistent and it constantly gives feedback on the query state. Reversal of actions is supported by various methods for going back, canceling query terms, and starting over. The system maintains the query state entirely in the URL so the browser's back button and bookmark features work correctly. These browser features help to keep the user in control and also serve as important memory aids. The use of query previews emphasizes recognition over recall, which also reduces short-term memory load, and helps provide an information scent of where to go next. The ever-present search box provides a shortcut for the user who has a specific goal in mind.