edAI

Partaking in machine and deep learning research for Africa. Managed by @bonadossou and @ChrisEmezue

[Bonaventure F. P. Dossou @bonadossou]

🚨 Thrilled to share that two of my papers have been accepted to @aclmeeting SRW 2025! 🧠🌍
They tackle key challenges in low-resource learning, African NLP, and inclusive speech recognition, grounded in linguistic insight and practical efficiency. @McGill_NLP @Mila_Quebec

📝 1. Rethinking Full Finetuning from Pretraining Checkpoints in Active Learning for African Languages

⚡ The challenge: In active learning (AL), it is common to use Finetuning All (FA). This means retraining the model from its original checkpoint on the entire labeled dataset after every round.
❗ But this strategy is computationally expensive, especially for large models and African languages with limited data.
🔍 We investigate an alternative: Continual Finetuning (CF), where the model is updated only on newly acquired samples after each AL round.
📊 We analyze when CF is sufficient. It performs comparably or better when the language is included in pretraining or is typologically similar (for example, Bantu languages).
⚖️ For more distant or underrepresented languages, FA still offers greater stability.
✅ CF yields approximately 35% savings in memory, FLOPs, and training time. It is a compelling option for scalable multilingual AL.

🗣️ 2. Advancing African-Accented English Speech Recognition: Epistemic Uncertainty-Driven Data Selection for Generalizable ASR Models

🎤 The problem: ASR systems often fail on African-accented English due to data scarcity and underrepresentation.
🧪 We combine epistemic uncertainty, Bayesian active learning, and core-set selection to efficiently adapt state-of-the-art ASR models (Wav2Vec2, HuBERT, etc.).
📏 We also introduce U-WER, a new metric to track how model uncertainty evolves across difficult accents.
✅ Results: 27% WER improvement, 45% less data required, and stronger out-of-distribution generalization, with a model-agnostic, open-source pipeline.

🌍 Both papers contribute to building inclusive, efficient, and linguistically grounded AI systems for African languages and accents.

📩 Paper links and code will be shared soon.

[Lanfrica @lanfrica]

Happy 5th anniversary to our incredible partner, @AfricArxiv ! 🎉We applaud their remarkable achievements and the profound impact they've made in advancing scholarly research in Africa and beyond. Here is to many more years of collaboration and success! ❤️🌍#AfricArxiv

[Bonaventure F. P. Dossou @bonadossou]

Join us to celebrate Black excellence in AI. It’ll be in person at Mila, so spread the word around you.

We’ve amazing speakers, panel discussion, food, music, dance etc - in short it’ll be lots of fun and you don’t wanna miss that 😉

@black_in_ai @Mila_Quebec @mcgillu

[Dr Ignatius Ezeani @igezeani]

Looking forward to this exciting talk on "Advances in NLP for Africa" by the 'dynamic duo' @bonadossou and @ChrisEmezue from @MasakhaneNLP, @Mila_Quebec.

1pm UK time.

Do join us if you can.

[Naamii Nepal @naamii_nepal]

We are thrilled to have David Ifeoluwa Adelani in the 3rd Nepal Winter School in AI. David will present his research on NLP for Low-resourced Languages.
David is also an active member of Masakhane NLP for African Languages leading the Lacuna NER project.
@davlanade
@MasakhaneNLP

[Naamii Nepal @naamii_nepal]

We are thrilled to have Chris Emezue in the 3rd Nepal Winter School in AI, 2021.

https://t.co/jsPOq5zF7x

Looking forward to know more about his work on Language technologies involving low resourced African Languages.

@ChrisEmezue

[Naamii Nepal @naamii_nepal]

We are thrilled to have Bonaventure Dossou in the 3rd Nepal Winter School in AI, 2021.
https://t.co/jsPOq5i3IX

Looking forward to know more about his work on Language technologies involving low resourced African Languages. More about him here https://t.co/ECTS4hNSLV

@bonadossou

[WiNLP @WiNLPWorkshop]

We will be back at #EMNLP2022 🎊🎉

[Paul Rayson @perayson]

Thanks to both @ChrisEmezue and @bonadossou for the great talk in the UCREL CRS series on "Advances in NLP for African Languages" ... watch on YouTube, details below ⬇️⬇️⬇️

[CoLLAs 2026 @CoLLAs_Conf]

We are thrilled to announce a new focused #MachineLearning conference: International Conference on Lifelong Learning Agents (CoLLA) https://t.co/ueLgjFzrTR 1/

[edAI @edAIOfficial]

Congratulations to @bonadossou for his paper accepted at the Affective Behavior Analysis workshop at the International Conference on Computer Vision @ICCV_2021.

We are proud to see you keeping the standards
@black_in_ai @blackinxnetwork @BWiSNetwork @BlackWomenInAI @DiverseInAI

[Yenoukoume Gbenou @yeno_gbenou]

Delighted to share the acceptance of our paper "FSER: Deep Convolutional Neural Networks for Speech Emotion Recognition" with my partner @bonadossou, at the Affective Behavior Analysis in-the-wild (ABAW) workshop at International Conference on Computer Vision @ICCV_2021

[edAI @edAIOfficial]

Another great mind of African NLP celebrating his birthday today. Thanks for your active work, inspiration and motivation.

Happy warm and loving birthday to our beloved @ChrisEmezue. The sky is the limit 🎂

[Bay Area NLP @BayAreaNLP]

Join us on May 7th to see the dynamic duo @bonadossou & @ChrisEmezue speak about Advances in #NLProc for African Languages!
https://t.co/dV6bAjK4L1

[edAI @edAIOfficial]

What a milestone by @ChrisEmezue and @bonadossou.

Kudos!!!

[Bonaventure F. P. Dossou @bonadossou]

1/ We (@ChrisEmezue and I) are delighted to officially introduce you, the FFRTranslate: the first Neural Machine Translation engine from Fon to French and vice-versa.

Official website: https://t.co/CZ6J2sJYtm

[edAI @edAIOfficial]

@bonadossou and @ChrisEmezue keep amazing us.

Great paper, really relevant for ASR in African Languages.

Let's hope it draws inspiration to more discoveries in the field.

[Bonaventure F. P. Dossou @bonadossou]

Okwugbé, first ASR initiative for Fon and Igbo is now available. Such a nice work with @ChrisEmezue.

We open-source codes and data to encourage similar works on low-resourced African Languages 🤗.

Fon: https://t.co/vdtyd3NQkR
Igbo: https://t.co/TWtrjxZQTq

[edAI @edAIOfficial]

Happy birthday to our beloved member @bonadossou. Keep shining and inspiring. We wish you the best, many more years and research for the humanity in general and particularly for the African NLP!

[Bonaventure F. P. Dossou @bonadossou]

I wanted to share subsequently few notes about my accepted accepted. This one is with @M___Sabry. After describing how word embedding works, with full explanatory guidelines, we created Word2vec and Poincaré embeddings for Fon and Nobiin.

[Bonaventure F. P. Dossou @bonadossou]

This one w/@ChrisEmezue is about the importance of involving human in the data collection and cleaning process. We performed phrase-based tokenization on the datasets w/- natives interference in the cleaning process, and 👇🏾

[Bonaventure F. P. Dossou @bonadossou]

We present a state-of-art ASR model for Fon, as well as benchmark ASR model results for Igbo. Our linguistic analyses provide valuable insights and guidance into creation ASR models for other African low-resourced languages, as well as guide future NLP research for Fon and Igbo.

[Bonaventure F. P. Dossou @bonadossou]

Okwugbé that means speak the language, in combined Igbo and Fon. Using Fon and Igbo as our case study, we conduct a comprehensive linguistic analysis of each language and describe the creation of End-to-End, deep neural network-based speech recognition
models for both languages.