Obtaining Parallel Sentences in Low-Resource Language Pairs with Minimal Supervision.

Area of Science:

• Computational Linguistics
• Natural Language Processing
• Machine Learning

Background:

• Machine translation performance heavily depends on the availability of parallel sentences.
• Low-resource languages often lack sufficient parallel data, hindering translation system development.
• Recent machine learning advances enable cross-lingual representation learning from nonparallel data.

Purpose of the Study:

• To develop a novel methodology for obtaining parallel sentences with minimal supervision for low-resource languages.
• To leverage a small bilingual seed lexicon to bootstrap the process.
• To construct and evaluate a machine translation system using the harvested parallel data.

Main Methods:

• Establishing cross-lingual semantic mapping using a seed lexicon.
• Constructing a deep learning classifier to extract bilingual parallel sentences.
• Applying the methodology to Uyghur-Chinese language pair for data harvesting.

Main Results:

• Successfully harvested large-scale, high-accuracy parallel sentences for the Uyghur-Chinese low-resource language pair.
• Demonstrated the effectiveness of the proposed methodology in improving machine translation system performance.
• Validated the approach's ability to generate valuable parallel corpora from limited initial resources.

Conclusions:

• The proposed method offers an effective solution for parallel sentence acquisition in low-resource scenarios.
• Minimal supervision, via a seed lexicon, is sufficient to build robust machine translation resources.
• This work opens new avenues for low-resource language machine translation research and development.