Skip to Main Content
For general maximum entropy models (ME models), lots of features make the training cost large, while the feature selection problem is very complicated. A combined ME language model which is composed of some sub-models is provided. In this approach, a large feature set is partitioned into small ones, each of the sub-models is trained with small feature sets, and then combined with the linear interpolation approach to produce the final model. The experiment results show that the perplexities of the combined ME models are much lower than those of the general ME models using the same feature sets without feature selection, and the training cost of the combined ME model decreased significantly.