Result Details
Language modeling of Czech using neural networks
The work concentrates on language modeling of Czech using neural networks
language modeling
Language models are used in many systems involving natural language processing, like speech and handwriting recognition. The most widely used techniques are based on backoff n-grams. However, it is commonly believed that this approach is insufficient. One of the best improvements over back-off language models has been achieved by using neural networks that project words onto a continuous space. This work concentrates on comparison of standard 4-gram language model with modified Kneser-Ney smoothing and neural network, both trained on spoken corpora with 1M words. Significant improvements in perplexity are reported.
@inproceedings{BUT25344,
author="Tomáš {Mikolov}",
title="Language modeling of Czech using neural networks",
booktitle="Proc. 13th Conference STUDENT EEICT 2007",
year="2007",
pages="1--3",
publisher="Faculty of Electrical Engineering and Communication BUT",
address="Brno",
isbn="9788021434103",
url="http://www.fit.vutbr.cz/research/groups/speech/publi/2007/mikolov_eeict_2007.pdf"
}