Zero-shot English-Assamese neural machine translation via pivot-based cross-lingual embedding alignment and transfer...

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Machine Translation

Background:

• Low-resource languages like Assamese present significant challenges for Neural Machine Translation (NMT) due to limited parallel corpora.
• Existing NMT models struggle with data scarcity, hindering effective translation development for underrepresented languages.

Purpose of the Study:

• To develop a novel zero-shot translation framework for English-to-Assamese translation.
• To leverage linguistic proximity between Assamese and Bengali to overcome data scarcity.
• To improve translation quality and efficiency for low-resource languages.

Main Methods:

• Utilized a multilingual transfer learning architecture (mBART) combined with cross-lingual embedding alignment (Procrustes optimization).
• Implemented subword tokenization, embedding projection, and dynamic vocabulary expansion to address out-of-vocabulary (OOV) challenges.
• Employed advanced language tagging mechanisms to reduce off-target translation errors.

Main Results:

• Achieved a BLEU score of 28.65 on an English-Assamese dataset, outperforming direct translation baselines by 7.53 points.
• Human evaluations demonstrated high adequacy (4.5/5) and fluency (4.6/5), with an 8.4% reduction in error rates.
• Demonstrated a 26.9% improvement in inference speed compared to baseline models.

Conclusions:

• Transfer learning and cross-lingual alignment are effective strategies for bridging resource gaps in NMT for linguistically proximate languages.
• The proposed framework successfully enables zero-shot translation for low-resource languages by exploiting shared semantic spaces.
• The methods developed significantly enhance translation quality, efficiency, and robustness for languages with scarce parallel data.