💬 NLP

Automatic Correction of Writing Anomalies in Hausa Texts

This paper addresses the challenge of writing anomalies in Hausa texts by creating a large-scale parallel dataset of over 400,000 noisy-clean sentence pairs and demonstrating that finetuned transformer models, particularly M2M100, can effectively correct these errors to significantly improve downstream NLP tasks.

Original authors: Ahmad Mustapha Wali, Sergiu Nisioi

Published 2026-05-06

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ahmad Mustapha Wali, Sergiu Nisioi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the Hausa language as a vibrant, bustling marketplace in West Africa, spoken by about 80 million people. For decades, this marketplace has been thriving orally. But recently, people have started bringing their goods to a digital marketplace (social media, texting, and the internet). The problem? In this digital rush, the "goods" (the text) are getting messy.

People are typing too fast, mixing up letters, or forgetting to put spaces between words. It's like someone trying to write a recipe but accidentally swapping "salt" for "sugar" or running all the words together into one giant, unbreakable block. While a human can easily guess what they meant, computers get completely confused. This messiness stops helpful tools like translation apps or search engines from working properly for Hausa speakers.

This paper is about building a digital janitor to clean up this mess automatically.

The Problem: The "Typos" of the Digital Age

The authors noticed that Hausa text online is full of specific types of errors:

Character Swaps: Hausa has special letters with "hooks" (like ɓ, ɗ, ƙ, ƴ) that look like standard English letters but sound different. People often type the plain English versions (b, d, k, y) instead. It's like writing "cat" when you meant "bat"—a small change that completely alters the meaning.
Spacing Chaos: Sometimes spaces are deleted (making "go home" look like "gohome"), and sometimes they are added where they don't belong.

The authors point out a classic "chicken-and-egg" problem: To teach a computer to fix these errors, you need a huge library of "messy" sentences paired with their "clean" versions. But no one had ever collected this for Hausa.

The Solution: The "Synthetic Noise" Factory

Since they couldn't find enough real-world messy text to train on, the authors decided to manufacture the mess.

The Clean Library: They gathered over 400,000 clean, high-quality Hausa sentences from sources like Wikipedia and official documents.
The Noise Machine: They built a computer program that intentionally "ruined" these clean sentences. It randomly swapped letters, deleted spaces, and duplicated characters, mimicking the way humans actually make mistakes on social media.
The Calibration: They didn't just break the text randomly; they studied real Twitter posts to ensure their "fake" mistakes looked and felt exactly like the "real" mistakes. They created a massive dataset of 400,000 pairs: Messy Sentence $\rightarrow$ Clean Sentence.

The Training: Teaching the Computers

With this new dataset, they trained several different types of AI models (think of them as different students taking a test). They asked these models: "Here is a messy sentence; please rewrite it correctly."

They tested various "students," including:

M2M100: A multilingual model designed to translate between 100 languages.
AfriTeVA: A model specifically trained on African languages.
N-ATLaS: A very large, powerful model based on Llama 3.

The Surprise Result:
You might expect the biggest, most expensive model (N-ATLaS) to win. And while it did perform very well, the M2M100 model (which is much smaller and lighter) performed just as well, if not better, in many cases. It was like a small, agile sports car beating a heavy, fuel-guzzling truck in a race. This is great news because smaller models are cheaper and faster to run.

Does Cleaning the Text Actually Help?

The authors didn't just stop at cleaning the text; they wanted to see if a clean room actually helps the workers (the AI tools) do their jobs better. They tested three scenarios:

Sorting Text (Genre Detection): When asked to sort stories by type (e.g., news vs. literature), the AI struggled with the messy text. Once the text was cleaned, the AI's performance bounced back to its original, high level.
Translation (Hausa to English): This was the most dramatic. When translating messy Hausa, the translation quality dropped significantly. After cleaning the Hausa text first, the English translations became much more accurate. It's like trying to translate a recipe written in crayon scribbles versus one written in neat handwriting; the neat version yields a much better dish.
Answering Questions (LLMs): When asking AI chatbots questions in Hausa, the messy text caused the bots to get confused or give wrong answers. Cleaning the text helped the bots understand the questions better, though some complex tasks (like math) remained tricky even after cleaning.

The Limitations: The Janitor Isn't Perfect

The authors are honest about the flaws. Their "digital janitor" isn't perfect.

Context Confusion: Sometimes the AI fixes a spelling error but changes the meaning. For example, it might fix a name but use a version of the name common in a different language, rather than the specific Hausa version.
The "Unfixable" Errors: In some cases, the noise was so severe that the AI couldn't guess the original meaning, and the sentence became gibberish.

The Takeaway

This paper provides a blueprint for cleaning up the digital Hausa language. By creating a massive, synthetic dataset and training a smart, efficient AI model, the authors have shown that we can automatically fix writing errors. This doesn't just make text look nicer; it acts as a foundation, allowing other tools (like translators and chatbots) to finally work effectively for millions of Hausa speakers.

They have made their "messy-to-clean" dataset and their code public, so other researchers can use these tools to help not just Hausa, but other languages that face similar digital growing pains.

Technical Summary: Automatic Correction of Writing Anomalies in Hausa Texts

Problem Statement
Hausa, a major regional language in West Africa with approximately 80 million speakers, faces significant challenges in Natural Language Processing (NLP) due to the prevalence of writing anomalies in digital texts. These anomalies, particularly on social media, include incorrect character substitutions (e.g., replacing Hausa-specific hooked letters like ɓ, ɗ, ƙ, ƴ with standard Latin equivalents) and spacing errors (removal or addition of spaces). While humans can easily disambiguate these errors, they introduce noise that degrades the performance of downstream NLP tasks such as machine translation, text classification, and Large Language Model (LLM) prompting. The scarcity of high-quality, annotated datasets for Hausa exacerbates this issue, creating a "chicken-and-egg" problem where the lack of clean data hinders the development of correction tools, and the lack of tools prevents the generation of clean data.

Methodology
The authors propose a pipeline to automatically correct these anomalies using transformer-based models fine-tuned on a synthetically generated dataset.

Data Collection and Synthesis:
- Clean Corpus: A base of over 400,000 clean Hausa sentences was aggregated from public sources, including Hausa Wikipedia, Wikimedia, and translated texts from the MultiplEYE project (covering genres like argumentative, literary, and institutional texts).
- Noise Modeling: To create a parallel noisy-clean dataset, the authors analyzed approximately 10,000 Hausa tweets from the NaijaSenti dataset. They identified error patterns (character substitutions, spacing issues, deletions, duplications) using a lexicon-based approach and normalized Levenshtein distance clustering.
- Synthetic Generation: These error distributions were used to inject synthetic noise into the clean corpus, generating over 400,000 noisy-clean sentence pairs. The synthetic noise was calibrated to match the normalized Levenshtein distance distribution of the real Twitter data (Jensen–Shannon distance of 0.14), ensuring the dataset mimics realistic writing errors.
Model Architecture and Training:
- The task was framed as a sequence-to-sequence translation problem (noisy input $\to$ clean output).
- Several multilingual and African-specific transformer models were fine-tuned on the synthetic dataset:
  - M2M100 (418M parameters)
  - AfriTeVA (Base, Small, Large versions)
  - UBC-NLP/cheetah-base (1.2B parameters)
  - AfriMBART and mBART-Large-50
  - mT5-Base
  - NCAIR1/N-ATLaS (8B parameters, based on Llama-3)
- Baselines included a copy baseline, an LSTM sequence-to-sequence model, and GPT-5.1 (zero-shot).

Key Contributions

Dataset Creation: The release of a large-scale, publicly available parallel dataset of over 400,000 noisy-clean Hausa sentence pairs, constructed via synthetic noise generation to overcome data scarcity.
Tool Development: The creation of the first set of transformer-based tools specifically designed for correcting writing anomalies in Hausa.
Empirical Analysis: A comprehensive evaluation of various model architectures and their pretraining strategies on the Hausa correction task.
Downstream Impact Assessment: A quantitative analysis demonstrating how text correction improves performance in downstream tasks, including genre detection, machine translation, and LLM instruction following.

Results

Correction Performance: The M2M100 (418M) model achieved state-of-the-art results, outperforming larger models like AfriTeVA and UBC-NLP/cheetah-base, and performing on par with the massive 8B NCAIR1/N-ATLaS model. This suggests that a model pre-trained on many-to-many translation tasks (like M2M100) is highly effective for monolingual text normalization.
Human Evaluation: On a sample of 500 predictions, 71.6% were deemed correct by a native speaker. While the model successfully preserved coherence, it occasionally introduced semantic shifts (e.g., changing tense or person) or failed to correct named entities to Hausa-specific standards (e.g., using "Pridnistria" instead of "Pridnestrovie").
Downstream Task Impact:
- Genre Detection: Cleaning the data restored classifier accuracy to levels comparable to the original clean data (81% vs. 80% original, compared to 78% on noisy data).
- Machine Translation: Corrected inputs consistently yielded better translation scores (BLEU and MetricX) across Google Translate, NLLB-200, and GPT-5.1 compared to noisy inputs, though the recovery was not perfect.
- Instruction Following (LLMs): Proprietary LLMs showed significant sensitivity to noise. Cleaning inputs improved performance in Named Entity Recognition (F1 score recovered from 0.19 to 0.37) and various reasoning tasks (AfriMMLU, AfriMGSM), though gains varied by task type.

Significance and Claims
The paper argues that automatic text correction is a critical preprocessing step for low-resource languages like Hausa. It demonstrates that smaller, efficiently fine-tuned models (like M2M100) can rival much larger instruction-tuned models in correcting orthographic errors. The authors claim that improving text quality through these methods has a tangible, positive impact on the reliability of downstream NLP applications. By providing a methodology and a public dataset, the work aims to advance NLP capabilities for Hausa and offers a transferable framework for addressing similar orthographic challenges in other low-resource languages. The authors remain modest, acknowledging limitations such as the reliance on social media data for noise modeling and the fact that automatic cleaning cannot always perfectly restore original meaning or handle complex named entity variations.