Word Perplexity Reduction

We propose a distribution-based pruning of n-gram backoff language models. Instead of the conventional approach of pruning n-grams that are infrequent in training data, we prune n-grams that are likely …...