Unsupervised model adaptation using information-theoretic criterion

Abstract

In this paper we propose a novel general framework for unsupervised model adaptation. Our method is based on entropy which has been used previously as a regularizer in semi-supervised learning. This technique includes another term which measures the stability of posteriors w.r.t model parameters, in addition to conditional entropy. The idea is to use parameters which result in both low conditional entropy and also stable decision rules. As an application, we demonstrate how this framework can be used for adjusting language model interpolation weight for speech recognition task to adapt from Broadcast news data to MIT lecture data. We show how the new technique can obtain comparable performance to completely supervised estimation of interpolation parameters. © 2010 Association for Computational Linguistics.