Language-Model Investigations Related To Broadcast News
Dietrich Klakow, Xavier L. Aubert, Peter Beyerlein, Reinhold Haeb‐Umbach, Meinhard Ullrich, Andreas Wendemuth, Patricia Wilcox · 1998
In this paper we present some experiments that have been performed while developing language models for the PHILIPS Broadcast News system. Three main issues will be discussed: construction of phrases, adaptation of remote corpora to this task, and the combination of the di#erent models. Also, perplexities on the 1997 evaluation data are reported. 1. Introduction Two main lines have been pursued to improve the system: construction of phrases and adaptation. The #rst topic is the combination of words to phrases. It is questionable whether words are really the best basic units for the estimation of stochastic language models - grouping frequentword sequences to phrases can improve language models. This issue has already been raised elsewhere #1,2#. Tests done on Wallstreet Journal #WSJ# and Broadcast News #BN# revealed that bigram perplexity #PP# can be reduced by up to 29#. In tests on WSJ we also observed reductions in word error rate #WER# of about 10#, which is partly due to langua...