Abstracts for Marti A. Hearst --SIMS

The EECS Research Summary for 2003

BioText - Infrastructure for Mining of Biological Text

The BioText project's main goal is to provide an intelligent information extraction and retrieval system for use in biomedical and genomics research. The system would enable fast and flexible access to text-based information needed by biological scientists, and would also provide an efficent, modular infrastructure for NLP scientists developing text-mining and text-analysis algorithms [1-3].

We are working on the design and implementation of the system’s infrastructure. Our main interest is in extending object relational databases to support the special requirement of information extraction from biomedical text. Current plans are for the system to include:

• Efficient storage representation and access methods for semantically annotated text. The system should be able to efficiently store and access different layers of annotation of the same text, including part-of-speech tagging, resolving definitions of acronyms [3], and mapping to a domain-specific ontology [2]. It should also provide an API for updating and retrieving annotations, and for registering NLP algorithms into the system.
• Extended query semantics. Combining relational query language and keyword based search and support for the use of synonyms and definition substitution as part of the query language and access methods. The system should support queries like "find ligands that bind to a given receptor;" and "find all documents that reference a specific gene and its synonyms."
• Ranking results based on their relevance to the query. Relevance metrics should incorporate support and confidence information, time of publication, number of references, and other tunable parameters. The challenge here is to integrate the process of ranking and result discovery.

[1]

M. Hearst, "Untangling Text Data Mining," Proc. ACL Mtg. Assoc. Computational Linguistics, University of Maryland, June 1999.

[2]

B. Rosario, M. Hearst, and C. Fillmore, "The Descent of Hierarchy, and Selection in Relational Semantics," Proc. Assoc. Computational Linguistics, July 2002.

[3]

A. Schwartz and M. Hearst, "A Simple Algorithm for Identifying Abbreviation Definitions in Biomedical Text," Proc. Pacific Symp. Biocomputing, Kauai, HI, January 2003.

Flexible Searching Using Faceted, Hierarchical Metadata

The FLAMENCO project (FLexible information Access using MEtadata in Novel Combinations) is exploring a new method for searching large online collections. Our goal is to design a search interface that supports task-oriented search and offers users a "browsing the shelves" experience. The Flamenco interface: integrates search and browsing; uses query previews to guide users; dynamically presents metadata to organize the search results and suggest next steps; and offers multiple methods for expanding and refining a search.

We have conducted a series of usability studies, which have helped refine the interface design and demonstrated that users strongly prefer the metadata-based approach over a baseline, Google-style image search interface.

Our methods and tools may be applied to any collection of images or text documents that have been classified using faceted metadata. So far we have applied this approach to: architecture images, plant/animal images, fine arts images, Epinions product information, a portion of MEDLINE medical texts, and to OCR'd tobacco industry documents.

The FLAMENCO group is currently developing methods to optimize the system's speed, offer users search history and personalization features, and enable the addition and classification of new items in a collection.

Figure 1: The Flamenco image browser. Categories available within the current result set are displayed on the left. This image set is constrained by location and structure type, and grouped by style.

Figure 2: Search results for "garden." The top portion of the screen is devoted to presenting the categories that produced the result set displayed on the lower portion of the screen. Users have the option of disambiguating their text query, or exploring all the results that contain their search term.

Figure 3: View of an individual item, with contextualized links for expanding the query in several conceptual directions, to see all "entrance spaces" or all "opera houses," or, more broadly, all buildings related to the "performing arts."

[1]

M. Hearst, J. English, R. Sinha, K. Swearingen, and P. Yee, "Finding the Flow in Website Search," Communications of the ACM, Vol. 45, No. 9, September 2002.

[2]

K.-P. Yee, K. Swearingen, K. Li, and M. Hearst, "Faceted Metadata for Image Search and Browsing" (submitted).

[3]

M. Hearst, "Next Generation Web Search: Setting Our Sites," IEEE Data Engineering Bulletin, Vol. 23, No. 3, September 2000.

[4]

A. Elliott, "Flamenco Image Browser: Using Metadata to Improve Image Search During Architectural Design," Proc. ACM CHI Conf. Companion, Seattle, WA, April 2001.

Peephole Displays: Handheld Computers as Virtual Windows

The small size of handheld computers makes them conveniently mobile, but limits the amount of information that can be shown on their screens. This work introduces "peephole displays," an interaction metaphor in which the handheld computer is a movable window on a larger virtual workspace anchored to the user’s physical reference frame. Peephole displays enable new forms of two-handed interaction for simultaneously navigating and manipulating information, including the ability to create and edit objects larger than the screen and the ability to drag and drop in three dimensions. I developed four iterations of the peephole hardware, and built and tested several peephole-augmented applications including a drawing program, a map viewer, and a calendar. A user study of 24 participants shows that the peephole technique can be more effective than current methods for navigating large information spaces on handheld computers.

Figure 1: Viewing a map. These images were made by blending two photographs taken from the same viewpoint. The position of the device is tracked and the display scrolls to produce the illusion of a movable view on a large street map floating in space. Notice how Gravier St., visible in both views, maintains a fixed position with respect to the outside world.

Figure 2: Note-taking on a large workspace. By using both hands together, a user can continue to write beyond the bounds of the screen.

Figure 3: Drawing on a large workspace. By using both hands together, a user can draw a figure larger than the screen in a single, natural stroke.

Figure 4: Viewing a calendar. Conventional PDA calendar programs have a day view, a week view, and a month view; the peephole calendar combines the strengths of all three into a single modeless view. This combined view is shown on the lower plane and an overview of the entire year is shown on an upper plane. The user can lift the display to switch to the year view and thereby navigate to a different month.

Figure 5: 3D drag-and-drop. The clipboard plane is located in space above the drawing plane. To transfer items to and from the clipboard, the user simply picks an item and lifts or lowers the display. Unlike conventional clipboards, this clipboard can contain more than one item and permits its contents to be viewed.

User Interaction Design for Secure Systems

The security of any system that is configured or operated by human beings depends on the information conveyed by the user interface, the decisions of the users, and the interpretation of their actions. This work establishes some starting points for reasoning about security from a user-centered perspective: it proposes to model systems in terms of actors and actions, and introduces the concept of the subjective actor-ability state. Ten principles for secure interaction design are identified; examples of real-world problems illustrate and justify the principles.

The results of this work come from discussing design challenges and user experiences at length with designers and users of software intended to be secure. After much debate and several iterations of refinement, we have formed the following set of design principles:

1. Path of least resistance: the most natural way to do any task should also be the most secure way.
2. Appropriate boundaries: the interface should expose, and the system should enforce, distinctions between objects and between actions along boundaries that matter to the user.
3. Explicit authorization: a user's authorities must only be provided to other actors as a result of an explicit user action that is understood to imply granting.
4. Visibility: the interface should allow the user to easily review any active actors and authority relationships that would affect security-relevant decisions.
5. Revocability: the interface should allow the user to easily revoke authorities that the user has granted, wherever revocation is possible.
6. Expected ability: the interface must not give the user the impression that it is possible to do something that cannot actually be done.
7. Trusted path: the interface must provide an unspoofable and faithful communication channel between the user and any entity trusted to manipulate authorities on the user's behalf.
8. Identifiability: the interface should enforce that distinct objects and distinct actions have unspoofably identifiable and distinguishable representations.
9. Expressiveness: the interface should provide enough expressive power to (a) describe a safe security policy without undue diffiulty and (b) allow users to express security policies in terms that fit their goals.
10. Clarity: the effect of any security-relevant action must be clearly apparent to the user before the action is taken.