YodiV3: NLP for Togolese Languages with Eyaa-Tom Dataset and the Lom Metric

Bakoubolo Essowe Justin, Kodjo François Xegbe, Catherine Nana Nyaah Essuman, Afola Kossi Mawouéna SamuelACL 2025

TL;DR

This paper introduces YodiV3, a comprehensive NLP pipeline for 10 Togolese languages, including machine translation, speech recognition, and text-to-speech. They also present Eyaa-Tom, a new multi-domain parallel corpus, and the Lom metric, a framework to assess language technology readiness, highlighting significant disparities in resource availability.

What Problem Does This Paper Address?

Natural Language Processing (NLP) has made incredible strides, but its benefits are not evenly distributed. Many languages, particularly those in Africa, are severely under-represented in terms of available data and NLP tools. This "digital divide" limits information access and technological inclusion for speakers of these languages. While large tech companies are expanding their multilingual models, they often struggle with domain-specific content and languages not explicitly included in their training data. This paper tackles this gap by focusing on Togolese languages, aiming to build a functional NLP pipeline for them.

Methods

The core of this work is the YodiV3 system, a multi-faceted NLP pipeline. It leverages existing large, pre-trained multilingual models like Meta AI's NLLB (for translation) and MMS (for speech) as a starting point. However, YodiV3 goes beyond simply using these off-the-shelf models. They adapt and fine-tune these models on their newly created Togolese dataset, Eyaa-Tom, with a focus on domain-specific content (religious, healthcare, financial). For languages not covered by NLLB/MMS, they employ data augmentation and smaller neural models.

Key components of YodiV3 include:

  • Machine Translation (MT): An encoder-decoder neural translation model capable of translating between Togolese languages and French/English.
  • Automatic Speech Recognition (ASR): Explores both standard autoregressive and faster non-autoregressive Transformer models.
  • Text-to-Speech (TTS): Includes models for Ewè and Kabyè, and also a novel "Togolese-accented" French/English voice cloning approach for code-switching scenarios.

A novel contribution is the Lom metric, a qualitative scoring framework designed to quantify the "AI-readiness" of a language. It aggregates the availability of eight foundational resources: a major religious text corpus (WOG), a digital dictionary, NMT, ASR, TTS, spoken language identification (SLID), text language identification (TLID), and OCR. This provides a clear, albeit simplified, overview of a language's digital resource landscape.

Datasets

The primary new dataset introduced is Eyaa-Tom. This multi-domain parallel corpus includes text and audio for 10 Togolese languages: Ewè, Kabyè, Adja, Tem (Kotokoli), Moba, Lamba (Togo), Konkomba, Mina (Gen), Bassar, and Nawdm. The data spans religious texts, healthcare, and financial domains, with translations to French and English. A significant portion of this data was collected through community contributions via the Umbaji platform and fieldwork. They also collaborated with Mozilla Common Voice to incorporate validated voice samples.

Key Insights and Ideas

  • Leveraging Pre-trained Models is Crucial: The paper strongly reinforces the idea that large, pre-trained multilingual models are essential for jump-starting NLP in low-resource settings. They significantly reduce the data and compute needed.
  • Domain-Specific Data Matters Immensely: Even with large pre-trained models, fine-tuning on domain-specific data yields substantial improvements. For example, Ewè-to-French translation on health-related sentences saw a dramatic increase in BLEU score when trained on a health subset compared to a model trained only on religious texts.
  • The Lom Metric is a Useful Communication Tool: While qualitative, the Lom metric provides a clear and easily understandable way to communicate the state of NLP resources for a language, highlighting disparities and guiding future development efforts.
  • Community Involvement is Key: The success of Eyaa-Tom relies heavily on community contributions, emphasizing the importance of participatory approaches for low-resource language NLP.

Key Results

  • YodiV3 demonstrates improved translation quality and speech capabilities for low-resource Togolese languages by integrating domain-specific data and pre-trained models.
  • The Lom metric reveals significant disparities in resource availability among the ten Togolese languages, with Ewè and Kabyè having the highest scores (7/8) and languages like Mina and Adja scoring much lower (3/8).
  • Fine-tuning pre-trained NLLB models greatly stabilized training and yielded higher translation quality for these low-resource languages.
  • The paper highlights the effectiveness of their approach by showing improved performance on tasks like Named Entity Recognition (NER) for Togolese names.

Limitations and Future Work

The authors acknowledge that the YodiV3 models, while promising, are still in their early stages. The ASR component is currently most accurate for Ewè and Kabyè, with baseline models for others needing further improvement. TTS is also primarily developed for Ewè and Kabyè, with plans to extend it to other languages requiring more recording efforts.

Future work includes:

  • Integrating additional datasets and models from projects like Masakhane and updates from NLLB/MMS.
  • Expanding the Eyaa-Tom corpus to cover more Togolese languages.
  • Exploring semi-supervised and active learning techniques to maximize the use of limited data.
  • Refining the Lom metric, potentially by weighting categories.
  • Deploying YodiV3's translation and TTS capabilities via an API for real-world applications (e.g., rural clinics, community radio) to gather feedback.
  • Conducting human evaluations with native speakers to assess translation quality and user acceptance of TTS and ASR.

Summary updated

BibTeX
@inproceedings{justin2025yodiv,
  author = {Bakoubolo Essowe Justin and Kodjo François Xegbe and Catherine Nana Nyaah Essuman and Afola Kossi Mawouéna Samuel},
  title = {{YodiV3: NLP for Togolese Languages with Eyaa-Tom Dataset and the Lom Metric}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2025},
}
Figure