http://www.cs.umbc.edu/mariedj

USC Institute for Creative Technologies

People have remarkable abilities in recognizing, comparing, and reasoning about strategies in their daily lives despite their high level of abstraction. Strategies are often seen as the commonalities across analogous planning cases, but are more than simply abstract plans - they describe the reasoning processes of the agents in planning cases. As our best models of human analogical reasoning are premised on the alignment of structured representations, the mental representation of strategies must also include structured representations of the reasoning processes of people. This talk will describe a large-scale knowledge representation effort that was done in order to investigate the conceptual requirements of strategies and their underlying mental models. 372 strategies were collected and analyzed from 10 domains of human planning (business, object counting, education, government, artistic performance, personal relationships, science, warfare, and the anthropomorphic domains of animal behavior and immunology). The analysis of these strategies reveals that their component concepts (nearly one thousand terms) include large portions that have previously received little attention within the formal knowledge representation community. Many of these terms describe components a mental model of human reasoning (a folk psychology or theory of mind) that parallels much of the computational modeling work that has been done in Artificial Intelligence and Cognitive Psychology. By describing this mental model in some detail, this research offers new opportunities for the design of human-computer interactions and anthropomorphic agent architectures, and furthers the psychological investigation of social and introspective reasoning.

Bio:
Dr. Andrew S. Gordon is a Research Scientist at the University of Southern California Institute for Creative Technologies. Before arriving at USC, he held postdoctoral positions in the Psychology department of the University of California Los Angeles and the IBM T.J. Watson Research Center in Hawthorne, New York. He received his Ph.D. in Computer Science and B.A. in Cognitive Science from Northwestern University.

June 14, 2002

Talking To Truman - Exploring Intelligence As Communication

Dario Nardi

University of California Los Angeles

Truman is a scripting language, a virtual agent, and an experiment in social intelligence. Maybe you are a player character in a virtual world, or maybe you are interacting with a robot or an appliance - just speak into a microphone, let Truman "think", and then hear what the virtual non-player character, robot or appliance has to say about the time your neighbors, its feelings about you, or your apparent personality style. Truman is not Eliza or Alice, nor is it the Cyc Project. It is a natural language system that remembers information, conveys gossip, makes inferences and analogies, engages in group activities, and so on. How does it work? In keeping with a situated-action philosophy, dialog skills and reasoning skills are integrated, while the specifics of a Truman agent's knowledge are widely distributed in a context-sensitive subsumption architecture. The scripting language is a fully featured programming language designed to meet many of the needs of dialog and social reasoning. This use of a scripting language, instead of a formal architecture, allows a virtual agent to be organic and able to meet many of the often-arbitrary needs of natural language. This experiment has focused on handling what to hear, what to say, what to remember, what to reflect on, what to relate to others, and similar socially-relevant functions. Truman can also recall information from one interaction to the next with a user, share information between multiple users, learn new information from a trusted user, reflect on its own programming and progress, and engage in in-situ activities such as giving lectures or eating meals. Naturally, truman does not do everything. It does make mistakes. It's still growing. It isn't as smart as the average person. Rather, Truman is an on-going exploration of what happens when we try to embody a few everyday social processes in a machine.

Bio:
Dario Nardi is currently an assistant adjunct professor in UCLA's Program in Computing. He completed his Ph.D. in Systems Science in 1998 from SUNY Binghamton's Watson School of Engineering. He is also a research fellow with the Temperament Research Institute. Dario has authored and co-authored 4 books in organizational psychology, does research in artificial intelligence, web-based learning and undergraduate curriculum design, and runs live social-psychology simulations. More information can be found at www.darionardi.com.

June 18, 2002

A Machine Learning Approach to Improving Web Navigation

Corey Anderson

University of Washington

Most web sites today are designed one-size-fits-all: all visitors see the exact same pages regardless of interests, previous interactions, or, frequently, even browsing client (desktop PC or wireless PDA). But one size often does not fit all. Instead of presenting the same content, the web experience should be dynamic and personalized, adapting to visitors' preferences as evinced in previous interactions.

In this talk I will give an overview of our Proteus framework for personalizing the web experience. Proteus adapts the presentation of content and navigation between pages, and targets personalizations for individual visitors or groups of similarly behaving visitors. I'll detail one particular adaptation -- adding new navigational shortcuts -- and describe our MinPath algorithm for finding shortcuts efficiently. A challenge for MinPath and modeling web navigation is that training data for an entire site may be plentiful, but sparse for any individual page. This difficulty can be overcome, however, by noting that most large web sites have a rich underlying relational structure that can be exploited for generalization: pages can belong to different types (e.g., pages about laptop computers vs. pages about printers), with each type described by a different set of attributes (e.g., size of display vs. printing speed). We leverage this structure by developing Relational Markov Models (RMMs). States in an RMM belong to relations and are described by variables over hierarchically structured domains. Based on these hierarchies, the RMM defines sets of related states, learns transition probabilities between these sets, and uses shrinkage to estimate transitions between individual pages. I will present RMMs in detail and provide results showing that they outperform traditional Markov models for predicting web navigation by a substantial margin.

Bio:
Corin Anderson is a final-year Ph.D. student at the University of Washington. His research interests cover artificial intelligence, machine learning, intelligent user interfaces, data mining, and applications of these to challenging problems. Outside of research, Corin occupies himself with running, cycling, photography, and contributing to open source software.

August 2, 2002

Searching Spoken Word Collections

Douglas W. Oard

University of Maryland
on sabbatical at USC Information Sciences Institute

Spoken word collections promise access to unique and compelling content, and most of the needed technology to realize that promise is now in place. Decreasing storage costs, increasing network capacity, and easy availability of software to exchange digital audio make possible physical access to spoken word collections at a previously unimaginable scale. Effective support for intellectual access -- the problem of finding what you are looking for -- is much more challenging, however. In this talk I will review the work that has been done on this problem at the Text Retrieval Conferences and the Topic Detection and Tracking evaluations, and I will present some early results from a user study comparing present manual and automated approaches to indexing spoken word collections. I will then describe a unique resource, a collection of 116,000 hours of oral history interviews recorded in 32 languages in 67 countries, and explain how we are leveraging an unprecedented manual indexing effort to develop the ability to index similar materials automatically.

Bio:
Doug Oard is an Associate Professor at the University of Maryland, with a joint appointment in the College of Information Studies and the Institute for Advanced Computer Studies. He will be on sabbatical in the Natural Language Group at USC-ISI through August, 2003. He holds a Ph.D. in Electrical Engineering from the University of Maryland, and his research interests center around the use of emerging technologies to support information seeking by end users. Dr. Oard's recent work has focused on cross-language information retrieval, retrieval from audio, data mining from text, and the exchange of ratings by networked users. Additional information is available at http://www.glue.umd.edu/~oard/.

September 13, 2002

Text Mining with Information Extraction

Raymond Mooney

University of Texas at Austin

Information extraction (IE) is a form of shallow text understanding that locates specific pieces of data in natural language documents. An IE system is therefore capable of transforming a corpus of unstructured texts or web pages into a structured database. Our previous work has focused on using machine learning methods to automatically construct IE systems from training sets of manually annotated documents. Our current research focuses on a form of text mining that extracts a database from a document corpus using a learned IE system and then mines this database for interesting patterns using rule induction. The noise and variation in automatically extracted text requires rule mining methods that allow "soft" matching to the data based on textual similarity. We have developed two methods for inducing "soft matching" rules from textual data, one based on integrating rule induction and nearest-neighbor learning and another based on modifying association rule mining. Results on several extracted datasets will be presented.

Bio:
Raymond J. Mooney is a Professor in the Department of Computer Sciences at the University of Texas at Austin. He received his Ph.D. in 1988 from the University of Illinois at Urbana/Champaign. He is an author of over 80 published papers in artificial intelligence, primarily in the area of machine learning, and a former editor for the Machine Learning journal. His recent research has focused on text mining, learning for natural-language processing, information extraction, text categorization, recommender systems, relational learning, inductive logic programming, and semi-supervised learning. Additional information is available on the World Wide Web at http://www.cs.utexas.edu/users/mooney.

September 23, 2002

Social Agents in Digital Cities

Toru Ishida

Kyoto University

The research community tackling agents and multiagent systems has studied and developed various agents. Typically, personal agents belong to humans and help their users to operate complex computer/communication systems. In the area of multiagent systems, though computational agents are designed to interact (both collaborate and compete) with each other, interaction among humans and computational agents has not been studied intensively. In this talk, however, I focus on social agents that can be members of a human community: social agents support human-human interactions, while personal agents support human-computer interactions. To understand the nature of social agents, we need a research platform to play with them. This talk proposes a scenario description language called Q for designing interactions among a large number of agents, and its application to digital cities and crisis management simulations. We started a five years project to develop social agents for supporting social interactions in digital cities.

Bio:
Toru Ishida received the B.E., M.Eng. and D.Eng. from Kyoto University, Kyoto, Japan, in 1976, 1978 and 1989, respectively. He is an IEEE fellow from Jan. 2002. He is currently a professor of Department of Social Informatics, Kyoto University, Kyoto, Japan. From 1978 to 1993, he was a research scientist of NTT Laboratories. He has been working on parallel, distributed and multiagent production systems from 1983. He proposed parallel rule firing, and extended it to distributed rule firing. Organizational self-design was then introduced into distributed production systems for increasing adaptiveness. From 1990, he started working on realtime search for learning autonomous agents. Again, organizational adaptation becomes a central issue in controlling multiple problem solving agents. He started working on community computing in 1995. He is leading digital cities and intercultural collaboration projects in Kyoto.

September 27, 2002

Congestion Games and Emergent Coordination in Non-Stationary Environments

Aram Galstyan