Optimizing Isarn-Thai SMT: Word-Based Alignment Adjustment through Linguistic Knowledge and Semantic Similarity

This research presents an Isarn dialect to Thai Statistical Machine Translation (SMT) model. Initially, word-based alignment is trained on a bilingual parallel corpus consisting of 51,810 sentence pairs. An Isarn-Thai dictionary containing 12,850 word-pairs and 40 transformation rules are employed to refine the word alignment process. Cosine similarity is used to evaluate the closeness between Thai words in alignment pairs and their corresponding translations in the dictionary, enhancing alignment accuracy. The alignment adjustments are achieved through three main strategies: dictionary-based locks, transformation rules, and word-form matching. A phrase-based translation model is then constructed. The effectiveness of the proposed model is tested against a set of 2,380 sentence pairs and evaluated using the BLEU score, which reached a value of 92.34. The results demonstrate that combining these three alignment strategies significantly improves translation performance, producing more accurate translations compared to standard alignment methods.