A hybrid language model based on a recurrent neural network and probabilistic topic modeling

Resumo

A language model based on features extracted from a recurrent neural network language model and semantic embedding of the left context of the current word based on probabilistic semantic analysis (PLSA) is developed. To calculate such embedding, the context is considered as a document. The effect of vanishing gradients in a recurrent neural network is reduced by this method. The experiment has shown that adding topic-based features reduces perplexity by 10%.