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

Area of Science:

• Computational Linguistics
• Natural Language Processing
• Machine Translation

Background:

• Statistical machine translation (SMT) has benefited from linguistic features, but their application in neural machine translation (NMT) for low-resource languages remains underexplored.
• Low-resource languages like Thai and Myanmar present unique challenges for NMT systems due to limited data.
• Previous NMT models have not fully leveraged linguistic information for these specific language pairs.

Purpose of the Study:

• To propose and evaluate transformer-based NMT models incorporating linguistic features for Thai-Myanmar, Myanmar-English, and Thai-English translation.
• To investigate the impact of adding part-of-speech (POS) or universal part-of-speech (UPOS) tags to NMT models.
• To compare the performance of novel multi-source and shared-multi-source transformer models against baseline NMT architectures.

Main Methods:

• Development of transformer, multi-source transformer, and shared-multi-source transformer models.
• Integration of POS/UPOS tags into source, target, or both sides of the input data.
• Utilizing a standard transformer model and an Edit-Based Transformer with Repositioning (EDITOR) model as baselines for comparison.

Main Results:

• The incorporation of linguistic features demonstrably enhanced the performance of transformer-based NMT models for low-resource language pairs.
• Shared-multi-source transformer models augmented with linguistic features outperformed both baseline transformer and EDITOR models.
• Significant improvements were observed in Bilingual Evaluation Understudy (BLEU) and character n-gram F-score (chrF) metrics.

Conclusions:

• Linguistic features are crucial for improving NMT performance, especially in low-resource scenarios.
• The proposed shared-multi-source transformer architecture effectively utilizes linguistic information for enhanced translation quality.
• This research provides a viable approach for advancing machine translation capabilities for under-resourced languages.