A STUDY OF DATA AUGMENTATION AND ACCURACY IMPROVEMENT IN MACHINE TRANSLATION FOR VIETNAMESE SIGN LANGUAGE

Abstract

Sign languages are independent languages of deaf communities. The translation from normal languages (i.e., Vietnamese Language - VL) as long as other sign languages to Vietnamese sign language (VSL) is a meaningful task that breaks down communication barriers and improves the quality of life for the deaf community. In this paper, we experimented with and proposed several methods for building and improving models for the VL to VSL translation task. We presented a data augmentation method to improve the performance of our neural machine translation models. Using an initial dataset of 10k bilingual sentence pairs, we were able to obtain a new dataset of 60k sentence pairs with a perplexity score no more than 1.5 times that of the original dataset. Experiments on the original dataset showed that rule-based models achieved the highest BLEU score of 68.02 among the translation models. However, with the augmented dataset, the Transformer model achieved the best performance with a BLEU score of 89.23, which is significantly better than that of other conventional approach methods.