Ways to Improve N-Gram Language Models for OCR and Speech Recognition of Slavic Languages

The problems of n-gram models for the OCR and speech recognition for the Slavic languages are investigated. The paper proposes methods applicable for most Slavic languages. Two approaches are tested: filtering of the n-gram model and the alternative ways of carrying out the smoothing. The filtering relies on heuristics based on frequencies and morphological features of words. The smoothing uses classes based on morphological features in combinations with new discounting formula. The smoothing can also be combined with inner filtering. The numerical experiments for the Ukrainian language show that both approaches produce interesting results. However, smoothing is more promising while being more complex and requiring further investigation of development of proper classes based on morphological information in order to outperform standard smoothing techniques.