Novel topic n-gram count LM incorporating document-based topic distributions and n-gram counts

In this paper, we introduce a novel topic n-gram count language model (NTNCLM) using topic probabilities of training documents and document-based n-gram counts. The topic probabilities for the documents are computed by averaging the topic probabilities of words seen in the documents. The topic probabilities of documents are multiplied by the document-based n-gram counts. The products are then summed-up for all the training documents. The results are used as the counts of the respective topics to create the NTNCLMs. The NTNCLMs are adapted by using the topic probabilities of a development test set that are computed as above. We compare our approach with a recently proposed TNCLM [1], where the long-range information outside of the n-gram events is not encountered. Our approach yields significant perplexity and word error rate (WER) reductions over the other approach using the Wall Street Journal (WSJ) corpus.