Translation-Augmented Multilingual Summarization for Low-Resource Languages

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

A translation-augmented approach addresses the challenge of automatic text summarization for low-resource languages, where labeled training data is scarce. This method systematically translates high-quality English summarization corpora into target low-resource languages using NLLB-200. The resulting parallel data trains and evaluates sequence-to-sequence models. Experiments across Swahili, Hausa, and Afrikaans compared monolingual fine-tuning (MONO), cross-lingual transfer (XLT), and joint multilingual training (TAMT) on mBART-large-50. Monolingual fine-tuning achieved the best performance for Swahili (ROUGE-L 13.9) and Afrikaans (ROUGE-L 15.7), surpassing the Lead-3 baseline. Cross-lingual transfer remained strongest for Hausa (ROUGE-L 14.5). Native language token availability in mBART-50 is a critical determinant of fine-tuning performance, influencing when the expected TAMT > MONO > XLT ordering breaks down. The dataset, code, and evaluation infrastructure are released to support future research.

Key takeaway

For NLP engineers extending summarization capabilities to low-resource languages, your choice of fine-tuning strategy should account for the target language's token support within your base model. While monolingual fine-tuning excels for some languages, cross-lingual transfer may be superior for others. You should evaluate native language token availability in models like mBART-50 to inform your approach, potentially avoiding suboptimal performance from a one-size-fits-all strategy.

Key insights

A translation-augmented approach addresses data scarcity for low-resource language summarization by leveraging English corpora.

Principles

• Native language token availability critically impacts fine-tuning performance.
• The expected performance hierarchy (TAMT > MONO > XLT) can vary by language.

Method

Systematically translate high-quality English summarization corpora into low-resource languages using NLLB-200. Use this parallel data to train and evaluate sequence-to-sequence models.

In practice

• Utilize NLLB-200 for translating English summarization corpora.
• Consider monolingual fine-tuning for languages like Swahili and Afrikaans.
• Evaluate cross-lingual transfer for languages similar to Hausa.

Topics

• Multilingual Summarization
• Low-Resource Languages
• NLLB-200
• mBART-large-50
• Sequence-to-Sequence Models
• ROUGE-L

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• Understanding LLM Behavior in Multi-Target Cross-Lingual Summarization — LLMs perform cross-lingual summarization and translation jointly in later layers, not sequentially.
• Modular Monolingual Adaptation using Pretrained Language Models — Modular PLM adaptation, freezing embeddings while tuning the rest, improves low-resource language performance.
• Multilingual Coreference Resolution via Cycle-Consistent Machine Translation — Coreference resolution in low-resource languages can be significantly improved by cycle-consistent machine translation data augmentation.
• Evaluating Cross-lingual Knowledge Consistency in Code-Mixed vis-a-vis Indian Languages using IndicKLAR — Code-mixing inputs significantly improves large language model knowledge recall in Indian languages, nearly matching English performance.
• Merging Methods for Multilingual Knowledge Editing for Large Language Models: An Empirical Odyssey — Shared covariance in vector merging is crucial for effective multilingual knowledge editing in LLMs.
• A NATO-accredited organization published a report concerning language-based disparities in LLM performance (Published April 25, 2026) — LLM performance varies significantly by language, but adaptation strategies can bridge the gap.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.