A tagger-aided language model with a stack decoder

This contribution of this paper is to investigate the utility of exploiting words and predicted detailed semantic tags in the long history to enhance a standard trigram language model. The paper builds on earlier work in the field that also used words and tags in the long history, but offers a cleaner, and ultimately much more accurate system by integrating the application of these new features directly into the decoding algorithm. The features used in our models are derived using a set of complex questions about the tags and words in the history, written by a linguist. Maximum entropy modelling techniques are then used to combine these features with a standard trigram language model. We evaluate the technique in terms of word error rate, on Wall Street Journal test data.