Improving neural machine translation with POS-tag features for low-resource language pairs

• DOI: 10.1016/j.heliyon.2022.e10375
• Publication Date: 2022-08-22T16:23:59Z
• Authors: Zar Zar Hlaing; Ye Kyaw Thu; Thepchai Supnithi; Ponrudee Netisopakul
• Keywords: H1-99; Science (General); Neural machine translation; 02 engineering and technology; Linguistic features; Social sciences (General); Q1-390; Universal part-of-speech; 0202 electrical engineering, electronic engineering, information engineering; Part-of-speech; Transformer architecture; Research Article

Integrating linguistic features has been widely utilized in statistical machine translation (SMT) systems, resulting improved quality. However, for low-resource languages such as Thai and Myanmar, the integration of neural (NMT) systems yet to be implemented. In this study, we propose transformer-based NMT models (transformer, multi-source transformer, shared-multi-source transformer models) using two-way Thai-to-Myanmar, Myanmar-to-English, Thai-to-English. Linguistic part-of-speech (POS) tags or universal (UPOS) are added each word on either source target side, both sides, proposed conducted. The take two inputs (i.e., string data with POS tags) produce tags. A model that utilizes only vectors was used first baseline comparison models. second model, an Edit-Based Transformer Repositioning (EDITOR) also compare our addition model. findings experiments show adding enhances performance a language pairs. Moreover, best results were yielded more significant Bilingual Evaluation Understudy (BLEU) scores character n-gram F-score (chrF) than EDITOR