A Language Model Using Products of Experts

Phil Cowans - University of Cambridge

Abstract

  Language modeling is essential for many tasks, including speech and handwriting recognition, data compression and predictive text entry. Traditionally, models such as prediction by partial match (PPM) are used. These blend predictions made by n-grams with differing context lengths. In this talk I will present a framework for creating PPM-like models using products of experts. This allows principled learning of the model parameters using Bayesian techniques, and also permits straightforward generalization of the model to include factors other than n-gram statistics.


Back to the Main Page

Charles Rosenberg
Last modified: Thu Jul 10 10:09:41 EDT 2003