Exploring Context’s Diversity to Improve Neural Language Model

Abstract

The neural language models (NLMs), such as long short term memery networks (LSTMs), have achieved great success over the years. However the NLMs usually only minimize a loss between the prediction results and the target words. In fact, the context has natural diversity, i.e. there are few words that could occur more than once in a certain length of word sequence. We report the natural diversity as context’s diversity in this paper. The context’s diversity, in our model, means there is a high probability that the target words predicted by any two contexts are different given a fixed input sequence. Namely the softmax results of any two contexts should be diverse. Based on this observation, we propose a new cross-entropy loss function which is used to calculate the cross-entropy loss of the softmax outputs for any two different given contexts. Adding the new cross-entropy loss, our approach could explicitly consider the context’s diversity, therefore improving the model’s sensitivity of prediction for every context. Based on two typical LSTM models, one is regularized by dropout while the other is not, the results of our experiment show its effectiveness on the benchmark dataset.