"Keeping it Real: From PowerLoom to KOJAK and Beyond"

02/11/05: 10:30 AM
11th Floor Large Conference Room
Host: Patrick Pantel, schedule

Abstract: The flagship products of our group developed over the course of many years are its Loom and PowerLoom knowledge representation reasoning (KR R) systems. While quite successful, there is bad news: "The world is big and messy." Real problems often present difficult challenges to traditional KR R systems due to noise, corruption, incompleteness, complexity and scale. In particular, the new area of link discovery which aims at finding hidden relations or linkages between entities in large amounts of low-leve data combines all these challenges in a single problem. In this talk I will present a high-level overview of our work on PowerLoom and our new suite of KOJAK link discovery tools that addresses some of these challenges. Starting from more traditional applications of KR R systems such as semantic interoperability which solely rely on deductive reasoning, I will describe how we have evolved PowerLoom's reasoning engine to support abductive reasoning such as query diagnosis in large, incomplete knowledge bases or partial pattern matching for plan and event recognition with large datasets. TO address issues of scale and dataset size, we use a combination of techniques such as resource-bounded inference, modeling of search control knowledge as well as tight integration with relational databases. For areas where logic-based inference is either not sufficient or not easily applicable, we use hybrid or purely statistical inference. For example, the KOJAK Group Finder combines logic-based reasoning and a statistical model to detect groups and comminities in low-level event data. The KOJAK Connection Finder uses a purely statistical model to find interesting entities based on a computed semantic profile. Finally, adapting a system such as KOJAK to different datasets in operational environments is in itself a difficult problem where we can apply a KR R system such as PowerLoom to formulate complex mappings between external and internal representations.

About Hans Chalupsky: Hans Chalupsky leads the Loom Knowledge Representation and Reasoning Group at the University of Southern California's Information Sciences Institute. He holds a Master's degree in computer science from the Vienna University of Technology in Austria (cum laude) and a Ph.D. in computer science from the State University of New York at Buffalo where he also held a Fulbright scholarship from 1987-1989. Dr. Chalupsky has 20 years of experience in the development and application of KR R systems such as RLL-1, the SNePS Semantic Network Processing System and PowerLoom. His research interests include knowledge representation and reasoning systems, ontology translation and maintenance, reasoning with partial and large-scale information, data mining and programming languages.

Last updated: Mon Jun 19 17:44:06 2006