Adaptive Hybrid POS Cache based Semantic Language Model

This paper presents a language model as an improvement over the stochastic language model for developing a syntactic structure based on word dependencies in local and non local domain. The model copes with the issues of limited amount of training material and the exploitation of the linguistic constraints of the language. The proposed model is a dynamic probabilistic model which uses word dependencies based on their part of speech tags along with the tri-gram Model but also takes care of the influence of the word which are very far from the word being considered in a text and stores the word history in a dynamic cache for information mining using long distance dependency. The model based on second order Hidden Markov Model has been used and an improvement of 2% has been observed in the word error rate and 4% reduction in the perplexity when compared to the normal tri-gram model.