MMTAfrica: Multilingual Machine Translation for African Languages

1 Introduction

• MMTAfrica focuses on multilingual translation among six African languages plus English and French to address African-language under-representation.The selected African languages span widely and least spoken languages in Africa.Source: §1 · p. 1
• BT&REC combines a reconstruction objective with backtranslation to exploit African-language monolingual data in multilingual translation.Source: §1 · p. 2
• MMTAfrica reports significant spBLEU gains over the M2M MMT benchmark on the FLORES Test Set.Source: §1 · p. 2
• The work creates the MMTAfrica Test Set to provide a representative evaluation resource and reports benchmark results with it.Source: §1 · p. 2

2.1 Multilingual Machine Translation (MMT)

• Multilingual NMT uses one model across directions, enabling scalability, transfer among related languages, low-resource improvements, and zero-shot translation.Source: §2.1 · p. 2
• Increasing a multilingual decoder vocabulary is computationally costly because training and inference time scale linearly with its output-layer size.Source: §2.1 · p. 2
• Massively multilingual NMT can show poor zero-shot performance, spurious correlations, and off-target translation into the wrong language.Source: §2.1 · p. 2
• Prior research motivates backtranslation and monolingual-data self-supervision for improving low-resource and zero-shot multilingual translation.Source: §2.1 · p. 2

3 Data Methodology

• The study gathers parallel data from JW300, Tatoeba, KDE4, MultiParaCrawl, and African-language resources, using language-specific preprocessing and sequence-length limits.The corpus contains 5,424,578 training samples and 4,000 development samples across directions.Source: §3 · p. 3
• Monolingual African-language data are collected from diverse sources to address severe imbalance and limited-domain coverage in parallel corpora.The parallel data are especially uneven for Fon-related directions and mostly cover biblical or internet domains.Source: §3 · p. 3

3.1 Data Set Types in our Work

• The FLORES Test Set contains 1,012 parallel sentences and is used for most directions except those involving Fon and Kinyarwanda.Source: §3.1 · p. 3
• The MMTAfrica Test Set samples equal numbers of sentences from each parallel-data domain to broaden evaluation coverage.It is small but includes samples from many domains and existing test sets.Source: §3.1 · p. 3
• Table 2 documents parallel-sample counts by direction and highlights the largest and smallest datasets, including greater resources for Swahili–English than Fon–French.Source: Table 2 · p. 4

4 Model and Training Setup

• The model represents translation as text-to-text generation by prepending a target-language tag to each source sentence.For source sentence x_i in language X, the input is <Ytag> x_i and the target is its translation y_i.Source: §4 · p. 4

4.1 Model Setup

• mT5-base provides the multilingual encoder-decoder foundation, but computational constraints limit the model to 580M parameters.The authors use the multilingual T5 architecture and fine-tune its base variant rather than larger versions.Source: §4.1 · p. 4

4.2 Training Setup

• The final multilingual setup covers all directions among eight languages except English–French, preserving focus on African-language translation.The language set is English, French, Igbo, Fon, Swahili, Kinyarwanda, Xhosa, and Yoruba; English–French is omitted because its high-resource data could overshadow other directions.Source: §4.2 · p. 5

4.2.1 Our Contributions

• Backtranslation samples fresh monolingual sentences for each language and generates candidate translations with random sampling to broaden training exposure.The method varies sampled monolingual sentences across backtranslation rounds and uses two generated predictions before randomly selecting one.Source: §4.2.1 · p. 5
• Reconstruction corrupts monolingual sentences through random swaps and deletion, then trains the model to recover the original sentence under its language tag.The corruption uses two swaps and deletion probability 0.2, integrating a denoising objective into fine-tuning.Source: §4.2.1 · p. 5

5.0.1 Baseline

• The baseline compares three fine-tuning settings—BASE, BT, and BT&REC—to assess how backtranslation and reconstruction affect African-language translation.Experiments use Fon and Igbo to contrast a language absent from versus included in initial multilingual pretraining, across 10 directions.Source: §5.0.1 · pp. 5–6
• BT&REC jointly samples 500:50 monolingual sentences for baseline training and 100:50 for final experiments, using different sentences for each objective.The authors leave the effect of alternative ratios for future ablation studies.Source: §5.0.1 · p. 6

5.0.2 MMTAfrica

• MMTAfrica fine-tunes across all language directions involving eight languages except English–French, using the proposed BT&REC setting.The final experiment builds on the baseline configuration and adapts it for the broader multilingual system.Source: §5.0.2 · p. 6
• Backtranslation was computationally expensive: generating roughly 3,000 translations per round with 100 sentences per language required parallelization and reducing the count to 50 and then 10.The process was parallelized with Python multiprocessing to mitigate runtime.Source: §5.0.2 · p. 6
• Gradient accumulation increased the effective batch computation from 64 to 4096 sentences, while PyTorch DataParallel distributed training across GPUs.These choices addressed computational-resource constraints and the reported stability of large-batch updates.Source: §5.0.2 · p. 6

6.1 Baseline Results and Insights

• BT improves over BASE in most baseline directions, except eng→ibo is relatively unchanged and fra→ibo is worse.When the target is Fon, BT significantly outperforms both BASE and BT&REC.Source: §6.1 · pp. 6–7
• BT&REC gives the best results for Igbo targets, including cases where BT underperforms BASE, such as fra→ibo and eng→ibo.The authors attribute this to Igbo being represented in mT5 pretraining, allowing the decoder to use reconstruction more effectively.Source: §6.1 · p. 7
• Backtranslation improves cross-lingual mapping for low-resource African languages, although its computational cost motivates parallelization and reducing the number of backtranslated sentences.These propositions are drawn from the Fon–Igbo baseline analysis.Source: §6.1 · p. 7
• For African languages included in mT5 pretraining, BT&REC performs best, whereas for unseen languages reconstruction can initially add noise and harm cross-lingual learning.For Kinyarwanda and Fon, the authors propose reconstruction-only finetuning before introducing BT&REC; their final ratio was 2:1 for backtranslation and reconstruction.Source: §6.1 · p. 7

6.2 MMTAfrica Results and Insights

• +0.58 to +19.46 spBLEU: MMTAfrica improves over the M2M MMT benchmark on all reported FLORES language pairs except swa→eng, which is 2.76 spBLEU lower.The gains range from +0.58 in swa→fra to +19.46 in fra→xho.Source: §6.2 · p. 7
• MMTAfrica Test Set results show that BT&REC improves spBLEU on a majority of tasks, while translation into Yoruba generally has the lowest scores.The Fon→Yoruba direction is an exception, achieving the highest score among Fon-target directions; the authors relate this to shared linguistic characteristics and propose Yoruba-to-Fon transfer for future work.Source: §6.2 · pp. 7–8

• MMTAfrica is introduced as a multilingual machine translation model for six African languages.Source: §7 · p. 8
• The paper presents a new reconstruction objective and analyses intended to provide insights for future research.Source: §7 · p. 8