Statistical Versus Neural Machine Translations for Khmer Braille

Abstract

For individuals with visual impairments, reading Braille text is crucial for acquiring information. However, the scarcity of available text in the Khmer Braille script presents a significant challenge. In this paper, we assess statistical and neural machine translation models (SMT versus NMT) trained on our developing Khmer-Braille corpus, which is of limited size (20K sentences). We employed phrase-based statistical machine translation (PBSMT) and Operation Sequence Model (OSM) for the SMT, and Sequence-to-Sequence (Seq2Seq) and Transformer architectures for NMT. Our experiments reveal that SMT models achieve significantly higher BLEU scores and lower word error rate (WER) compared to NMT models.