Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AfriNLLB: Efficient Translation Models for African Languages

Yasmin Moslem, Aman Kassahun Wassie, Amanuel Gizachew Abebe

TL;DR

AfriNLLB tackles the scarcity of efficient machine translation for African languages by compressing the NLLB-200 baseline through iterative decoder-layer pruning, multi-stage fine-tuning, and sequence-level knowledge distillation. The approach is supported by curated, high-quality multilingual data and a robust filtering pipeline, with validation on Flores200NLLB2022. The authors release two model variants (Transformers and CTranslate2) and all training data to enable replication and deployment in resource-limited settings. Results show meaningful speedups while preserving translation quality, highlighting practical impact for African-language MT and paving the way for broader language coverage and deployment in constrained environments.

Abstract

In this work, we present AfriNLLB, a series of lightweight models for efficient translation from and into African languages. AfriNLLB supports 15 language pairs (30 translation directions), including Swahili, Hausa, Yoruba, Amharic, Somali, Zulu, Lingala, Afrikaans, Wolof, and Egyptian Arabic, as well as other African Union official languages such as Arabic (MSA), French, Portuguese, and Spanish. Our training data covers bidirectional translation between English and 13 languages, and between French and two languages (Lingala and Wolof). AfriNLLB models are based on NLLB-200 600M, which we compress using iterative layer pruning and quantization. We fine-tune the pruned models on parallel corpora we curated for African languages, employing knowledge distillation from a larger teacher model. Our work aims at enabling efficient deployment of translation models for African languages in resource-constrained settings. Our evaluation results demonstrate that AfriNLLB models achieve performance comparable to the baseline while being significantly faster. We release two versions of the AfriNLLB models, a Transformers version that allows further fine-tuning and a CTranslate2 version for efficient inference. Moreover, we release all the training data that we used for fine-tuning the baseline and pruned models to facilitate further research.

AfriNLLB: Efficient Translation Models for African Languages

TL;DR

AfriNLLB tackles the scarcity of efficient machine translation for African languages by compressing the NLLB-200 baseline through iterative decoder-layer pruning, multi-stage fine-tuning, and sequence-level knowledge distillation. The approach is supported by curated, high-quality multilingual data and a robust filtering pipeline, with validation on Flores200NLLB2022. The authors release two model variants (Transformers and CTranslate2) and all training data to enable replication and deployment in resource-limited settings. Results show meaningful speedups while preserving translation quality, highlighting practical impact for African-language MT and paving the way for broader language coverage and deployment in constrained environments.

Abstract

In this work, we present AfriNLLB, a series of lightweight models for efficient translation from and into African languages. AfriNLLB supports 15 language pairs (30 translation directions), including Swahili, Hausa, Yoruba, Amharic, Somali, Zulu, Lingala, Afrikaans, Wolof, and Egyptian Arabic, as well as other African Union official languages such as Arabic (MSA), French, Portuguese, and Spanish. Our training data covers bidirectional translation between English and 13 languages, and between French and two languages (Lingala and Wolof). AfriNLLB models are based on NLLB-200 600M, which we compress using iterative layer pruning and quantization. We fine-tune the pruned models on parallel corpora we curated for African languages, employing knowledge distillation from a larger teacher model. Our work aims at enabling efficient deployment of translation models for African languages in resource-constrained settings. Our evaluation results demonstrate that AfriNLLB models achieve performance comparable to the baseline while being significantly faster. We release two versions of the AfriNLLB models, a Transformers version that allows further fine-tuning and a CTranslate2 version for efficient inference. Moreover, we release all the training data that we used for fine-tuning the baseline and pruned models to facilitate further research.
Paper Structure (17 sections, 3 figures, 6 tables)

This paper contains 17 sections, 3 figures, 6 tables.

Table of Contents

  1. AfriNLLB: Background & Motivation
  2. Data
  3. Data Sources
  4. Data Processing
  5. Rule-based filtering
  6. Language detection
  7. Semantic filtering
  8. Quality estimation
  9. Validation and Test Data
  10. Methodology
  11. Layer importance evaluation:
  12. Layer pruning:
  13. Fine-tuning:
  14. Knowledge distillation:
  15. Evaluation and Results
  16. ...and 2 more sections

Figures (3)

  • Figure 1: Quality-Efficiency Comparison. The iterative-pruned models demonstrate a superior balance of speed and quality compared to the middle-pruned variants. The 548M models include 12 encoder layers and 8 decoder layers (i.e. 4 decoder layers are pruned), while the 498M models include 8 encoder layers and 8 decoder layers (i.e. 8 layers are pruned, 4 from the encoder and 4 from the decoder). The chart reports the average chrF++ scores across all language pairs before and after fine-tuning the pruned models.
  • Figure 2: Translation performance (chrF++) from English/French to African languages.
  • Figure 3: Translation performance (chrF++) from African languages to English/French.