Thai-English Neural Machine Translation Method with Local and Global Syllable Feature

Neural Machine Translation (NMT) has superseded Statistical Machine Translation (SMT) owing to the advent of deep learning in natural language processing. However, enhancing the quality of low-resource NMT involves more nuanced strategies beyond data-dependent methods, especially for languages like Thai that lack substantial parallel corpus data for Thai-English translation. Addressing this problem, we introduce a Thai-English NMT methodology that amalgamates local and global syllabic features. Our approach employs a CNN-based module for extracting local syllable characteristics and capturing interdependencies among Thai syllable characters. Concurrently, a global syllable feature extraction mechanism is used to preserve sentence-level syllable traits. The integration of these features significantly augments Thai syllable context representation. In experiments conducted on the Thai-English corpus from the AL T dataset, our method improved the BLEU score by 2.11 points above the baseline, achieving a score of 12.99. This research holds potential implications for the translation of other syllable-based, low-resource languages, as well.