Modelling Retrieval Models in a Probabilistic Relational Algebra with a new Operator: The Relational Bayes

Thomas Roelleke (Queen Mary University of London)

CSIRO ICT

DATE: 2008-04-09
TIME: 14:00:00 - 15:00:00
LOCATION: CSIRO Seminar Room, S206, CSIT Building (Building 108)
CONTACT: JavaScript must be enabled to display this email address.

ABSTRACT:
The work on probabilistic DB technology led to results that feed into DB+IR technology. The talk will browse research on probabilistic DB and reasoning including Cavallo/Pitarelli:VLDB:87 (theory of probabilistic DB), Fuhr/Roelleke:TOIS:97, Chaudhuri...Weikum:04/06 (probabilistic ranking of tuples), Dalvi/Suciu:04/05 (efficient processing of safe expressions), and our recent contribution, the relational Bayes.

The relational Bayes is a new probabilistic relational operator. Traditional database technology is based on five operators. Probabilistic extensions based on those five only captured probability aggregation, but not estimation. The Bayes operator embeds probability estimation conceptually into the probabilistic relational paradigm.

Through the relational Bayes, IR models such as tf-idf, binary-independent retrieval, and language modelling can be expressed in probabilistic logical models. This will be illustrated in examples and a system demo. The outlook addresses optimisation, design and verification of probabilistic logical programs, and applications such as RSS retrieval.
BIO:
Thomas Roelleke is a researcher and lecturer at Queen Mary University of London (QMUL). Thomas started his IT career at Nixdorf Computer as product manager for Unix/DB/4GL technology. While studying Computer Science & Engineering in Dortmund, he consulted Nixdorf in Europe as Unix/4GL/C tutor. From 1994-1999, he was a researcher/lecturer at the University of Dortmund, and after his PhD, he was appointed as strategic IT consultant at comdirect bank Germany, and he continued his research as a research fellow at Queen Mary University London.

Thomas' research contributions include - a probabilistic relational algebra (PRA, TOIS 97), - a probabilistic object-oriented logic (POOL, SIGIR 96/98, chapter in 2002 book Intelligent Exploration of the Web, PhD thesis "POOL: A Probabilistic Object- Oriented Logic", Shaker Verlag 99), - the probabilistic inference engine HySpirit (EDBT 98, FQAS 98, BTW 97), - the probability of being informative (SIGIR 03), - an idf formulation of the probabilistic retrieval model (SIGIR 05), - a general matrix framework for information retrieval (IP&M 06), - a parallel derivation of probabilistic retrieval models (SIGIR 06), and - Probabilistic SQL and the relational Bayes (TREC 05, VLDB Journal 08).

Thomas was a co-organiser of the DB+IR workshop at SIGIR 04, a member of the panel "DB and IR: Rethinking the great divide" at SIGMOD 05, and he is the founder of a spin-out to exploit innovative DB+IR technology for information management. He serves as a reviewer for numerous conferences and journals in the field.