About FAI...

The Forum for Artificial Intelligence meets every other Friday at 3pm to discuss topics in artificial intelligence. After the formal talk, we continue our conversation at the Crown and Anchor. All are welcome to attend.

Please send questions or comments to Kenneth Stanley or Tal Tversky.

Full Schedule: 2/82/153/14/12

February 8th

PAI 3.14

Improving Machine Learning Approaches to Noun Phrase Coreference Resolution

Prof. Claire Cardie
Cornell University

This talk will first introduce noun phrase coreference resolution, one of the critical problems that currently limit performance for many practical natural language processing tasks. We then present a machine learning-based solution to noun phrase coreference that extends earlier work in the area and produces the best empirical results to date --- for both learning- and knowledge-based approaches to the problem --- on two standard coreference data sets. Performance gains accrue from two very different sources of change: first, we propose and evaluate three extra-linguistic modifications to the machine learning framework; second, we more than triple the number of linguistic knowledge sources made available to the learning algorithm. We conclude with a discussion of why we view these seemingly promising results as ultimately disappointing, and identify a key area of research where progress in noun phrase coreference resolution is likely to be made.

February 15th

ACES 2.402

Semantics, Pragmatics and Rhetorical Structure

Prof. Nicholas Asher
UT Department of Philosophy

This talk will introduce issues in discourse interpretation. I'll trace a short history of semantics from Montague Grammar to Dynamic Semantics and beyond. I'll concentrate on the architecture of a theory of discourse interpretation that exploits rhetorical function, and I'll give some applications, time permitting, to dialogue and quantification.

March 1st

ACES 2.402

Using Ideal Observer Analysis to Understanding Human Spatial Navigation

Prof. Brian J. Stankiewicz
University of Texas at Austin Department of Psychology
Center for Perceptual Systems

Humans possess a remarkable ability to navigate through complex environments with relative ease and accuracy. As such, they provide us with both an existence proof that a robust spatial navigation system can be built along with a working system that can be reversed engineered. In this talk I will describe a series of studies investigating human spatial navigation performance in complex indoor environments. The studies will investigate the effect of increasing layout size on spatial navigation performance along with the effect of reducing visual information on navigation performance. Because we are manipulating the information available to the subject in these experiments it is important to determine what changes in performance are due to limitations in human processing (e. g., memory, strategy, etc.) and what changes are simply due to task demands (e.g., increased uncertainty). To make this differentiation, I will also describe an ideal navigator model of indoor spatial navigation that uses principles from Partially Observable Markov Decision Process to provide an upper limit on navigation performance in each of these studies. Human performance will be compared against the ideal observer's performance to provide an navigation efficiency measure.

April 12th

ACES 2.402

Semi-supervised Clustering

Sugato Basu and Mikhail Y. Bilenko
Department of Computer Sciences
University of Texas at Austin

Clustering is a type of unsupervised learning that involves partitioning a set of unlabeled objects into meaningful groups. In semi-supervised clustering, some labeled data is used along with the unlabeled data to obtain a better clustering. We propose to incorporate initial supervision into clustering in two ways: (1) Learning meaningful distance metrics from labeled data, and (2) Initializing and constraining the clustering algorithm with labeled data. We present some initial results on each of the two methods and propose a unified framework which can be applied to several domains, including text and biological data.

Past Schedules

Fall 2001

Spring 2001

Fall 2000

Spring 2000

fai (fA) n. Archaic. [Middle English]: Faith.