Abstract

In language modeling, it is difficult to incorporate entity relationshipsfrom a knowledge-base. One solution is to use a reranker trained with globalfeatures, in which global features are derived from n-best lists. However,training such a reranker requires manually annotated n-best lists, which isexpensive to obtain. We propose a method based on the contrastive estimationmethod~\cite{smith2005contrastive} that alleviates the need for such data.Experiments in the music domain demonstrate that global features, as well asfeatures extracted from an external knowledge-base, can be incorporated intoour reranker. Our final model achieves a 0.44 absolute word error rateimprovement on the blind test data.

 

Quick Read (beta)