Developing an English-Efik Corpus and Machine Translation System for Digitization Inclusion

1. Problem Statement

The paper addresses the critical gap in Natural Language Processing (NLP) for low-resource languages, specifically Efik, an indigenous language spoken in Southern Nigeria and parts of Cameroon.

• Data Scarcity: Unlike high-resource languages (e.g., Swahili, Yoruba), Efik lacks sufficient parallel corpora (aligned sentence pairs) required to train effective Neural Machine Translation (NMT) systems.
• Technological Exclusion: Existing translation tools fail to capture cultural nuances, idiomatic expressions, and complex morphological structures of Efik, leading to poor translation quality.
• Language Decline: The paper highlights the sociolinguistic decline of Efik due to a lack of digital support and the prioritization of English, necessitating technological interventions for revitalization.
• Limitations of Previous Work: Prior attempts were either rule-based (limited by small dictionaries) or relied on domain-specific data (e.g., Bible verses), failing to generalize to daily conversation and diverse domains.

2. Methodology

The study employs a rigorous pipeline involving community-driven data curation and state-of-the-art transfer learning.

A. Dataset Creation (The Efik Corpus)

• Source: A community-curated parallel corpus of 13,865 high-quality English–Efik sentence pairs.
• Data Collection:
  • Sources included controlled Large Language Model (LLM) generation, manual writing by bilingual linguists, and digitization from printed dictionaries.
  • Domains: The dataset is diverse, covering daily conversation (52%), family/relationships, agriculture, environment, religion, health, and education.
• Annotation Process:
  • Six native speakers and two trained linguists translated English sentences into Efik.
  • Quality Control: Translations were independently verified by at least two other annotators. Disagreements were resolved by a linguist reviewer.
  • Format: Stored as UTF-8 encoded JSONL for compatibility with standard NLP toolkits (HuggingFace, OpenNMT).
• Split: The data was split into Training (10,398), Validation (1,387), and Testing (2,080) sets.

B. Model Training

The authors utilized a transfer learning approach, fine-tuning two pre-trained multilingual models:

1. mT5-small: A text-to-text Transformer pre-trained on 101 languages.
2. NLLB-200 (No Language Left Behind): A sequence-to-sequence model covering over 200 languages.

• Hyperparameters: Models were trained using the Hugging Face Transformers library with specific tuning for low-resource settings (e.g., mT5 used a learning rate of 1e-4 for 20 epochs; NLLB used 2e-5 for 6 epochs with FP16).

C. Evaluation

• Automatic Metrics:
  • BLEU: Measures n-gram overlap.
  • chrF (Character n-gram F-score): Crucial for Efik due to its morphological complexity and tonal nature.
• Qualitative Analysis: Human evaluation of translation outputs to assess semantic fidelity, fluency, handling of named entities, and cultural nuances.

3. Key Contributions

1. First Diverse Parallel Corpus: The creation of the first large-scale, multi-domain English–Efik parallel corpus (13,865 pairs) that moves beyond religious texts to include daily life, agriculture, and education.
2. Benchmarking State-of-the-Art: The first comprehensive evaluation of fine-tuned mT5 and NLLB-200 specifically for the English–Efik language pair.
3. Reproducible Pipeline: The paper provides a fully documented, reproducible pipeline for data cleaning, annotation, and model fine-tuning, serving as a blueprint for other low-resource African languages.
4. Community-Centric Approach: Demonstrates the efficacy of combining native speaker expertise with computational linguistics to build high-quality datasets.

4. Results

The study compared zero-shot performance against fine-tuned models.

• Zero-Shot Performance: Both models performed poorly without fine-tuning (BLEU < 0.22), confirming that pre-trained models cannot handle Efik without specific adaptation.
• Fine-Tuned Performance:
  • NLLB-200 significantly outperformed mT5 in both directions.
  • English → Efik: NLLB achieved a BLEU of 26.64 and chrF of 51.04.
  • Efik → English: NLLB achieved a BLEU of 31.21 and chrF of 47.92.
  • mT5 achieved lower scores (BLEU ~14–18), indicating struggles with fluency and semantic accuracy.
• Qualitative Findings:
  • NLLB demonstrated superior handling of cultural context, named entities (e.g., "William Otu"), and complex sentence structures.
  • mT5 frequently omitted repeated phrases, introduced spurious content, and failed to translate specific cultural terms accurately.
  • Both models still struggled with highly idiomatic expressions and older vocabulary, highlighting the limits of current data volume.

5. Significance and Conclusion

• Feasibility of Low-Resource MT: The study proves that high-quality machine translation for low-resource languages is achievable with relatively small, high-quality, community-curated datasets and modern multilingual models.
• Cultural Preservation: By enabling accurate translation, the system supports the digitization and revitalization of the Efik language, countering its decline.
• Inclusive NLP: The work advocates for "culturally grounded evaluation" and inclusive data practices, moving away from purely statistical metrics to ensure translations respect linguistic and cultural nuances.
• Future Directions: The authors suggest expanding the dataset size, applying phonological normalization for dialectal variations, and utilizing techniques like back-translation and crowdsourced annotation to further improve model robustness.

In summary, this paper establishes a foundational resource and a high-performance baseline for English–Efik translation, demonstrating that NLLB-200 is currently the superior architecture for this specific low-resource pair, while emphasizing that data quality and community involvement are as critical as model architecture.