COILD
A machine translation parallel corpus and evaluation benchmark centered on Indian languages. It contains over 1.16 million human-translated and verified sentence pairs, covering 20 Indian language pairs across four major language families, accompanied by domain-oriented evaluation benchmarks.
Timeline (last 90 days)
COILD is a parallel corpus centered on Indian languages, containing over 1.16 million sentence pairs that have been human-translated and human-verified, covering 20 Indian language pairs.
The language pairs covered by COILD span four major language families: Indo-Aryan, Dravidian, Tibeto-Burman, and the Austroasiatic language family.
COILD is built entirely from original Indian-language source materials, with data drawn from licensed corpora and covering eight domains relevant to real-world applications.
COILD comes with a domain-centric evaluation benchmark containing 2,000 expert-verified sentences.
The evaluation benchmark supports multilingual and cross-lingual consistency evaluation across Indian language pairs.
The research team used COILD to conduct fine-tuning experiments on two multilingual neural machine translation models, IndicTrans2-Distilled and NLLB-200.
The fine-tuned models achieved consistent improvements across different language pairs, different domains, and both automatic evaluation metrics and human evaluation.
The COILD experiment also incorporates human evaluation, in which bilingual evaluators score translations for faithfulness and fluency.
All Facts (8)
COILD is a parallel corpus centered on Indian languages, containing over 1.16 million sentence pairs that have been human-translated and human-verified, covering 20 Indian language pairs.
50%UnverifiedThe language pairs covered by COILD span four major language families: Indo-Aryan, Dravidian, Tibeto-Burman, and the Austroasiatic language family.
50%UnverifiedCOILD is built entirely from original Indian-language source materials, with data drawn from licensed corpora and covering eight domains relevant to real-world applications.
50%UnverifiedCOILD comes with a domain-centric evaluation benchmark containing 2,000 expert-verified sentences.
50%UnverifiedThe evaluation benchmark supports multilingual and cross-lingual consistency evaluation across Indian language pairs.
50%UnverifiedThe research team used COILD to conduct fine-tuning experiments on two multilingual neural machine translation models, IndicTrans2-Distilled and NLLB-200.
50%UnverifiedThe fine-tuned models achieved consistent improvements across different language pairs, different domains, and both automatic evaluation metrics and human evaluation.
50%UnverifiedThe COILD experiment also incorporates human evaluation, in which bilingual evaluators score translations for faithfulness and fluency.
50%