💬 NLP

Using Machine Learning to Enhance the Detection of Obfuscated Abusive Words in Swahili: A Focus on Child Safety

This study investigates the application of machine learning models, optimized with techniques like SMOTE, to detect obfuscated abusive Swahili language for child safety, while acknowledging that current findings are limited by dataset size and imbalance and calling for future work involving larger datasets and advanced techniques.

Original authors: Phyllis Nabangi, Abdul-Jalil Zakaria, Jema David Ndibwile

Published 2026-02-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Phyllis Nabangi, Abdul-Jalil Zakaria, Jema David Ndibwile

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 internet as a giant, bustling playground. For children, it's a place to learn, play, and make friends. But just like a real playground, it has shady corners where bullies hide. These bullies aren't just shouting mean words; they are wearing disguises. They swap letters for numbers, add random symbols, or spell things weirdly to trick the "playground monitors" (the safety software) into thinking they are saying something nice.

This paper is about building a smarter, more observant playground monitor that speaks Swahili, a language spoken by millions in East Africa but often ignored by big tech because it's considered a "low-resource" language (meaning there aren't many digital dictionaries or tools for it yet).

Here is the story of their research, broken down into simple parts:

1. The Problem: The "Code-Switching" Bullies

Children are spending more time online than ever. Unfortunately, bad actors are using the internet to bully, harass, and even exploit them. To avoid getting caught, these bullies use obfuscation.

The Analogy: Imagine a bully trying to sneak a note into class. Instead of writing "I hate you," they write "I h@te y0u" or "I h4te y0u." A human can read it easily, but a basic computer program might think, "Oh, that looks like a typo, not an insult!"
The Challenge: In Swahili, this is even harder because there are fewer computer tools to understand the language's nuances. The researchers wanted to teach a computer to see through these disguises.

2. The Toolkit: Teaching the Computer to "Read Between the Lines"

The team didn't just guess; they built a training program using Machine Learning. Think of this like training a new security guard.

The Data: They gathered 100 examples of Swahili abusive texts. Some were written normally, and some were "obfuscated" (disguised).
The Imbalance Problem: They had very few examples of the "disguised" bullies compared to the normal ones. It's like trying to teach a guard to spot a specific type of thief when you only have one photo of that thief and 100 photos of innocent people.
The Solution (SMOTE): To fix this, they used a technique called SMOTE. Imagine you have one photo of the thief. SMOTE is like a photocopier that creates slightly different versions of that photo (changing the lighting, the angle, the background) so the guard has more examples to study. This helps the computer learn better without needing millions of real-world examples.

3. The Contenders: Who Was the Best Guard?

They tested four different types of "guards" (algorithms) to see who could spot the disguised bullies best:

Logistic Regression: A steady, reliable guard who follows the rules strictly.
Support Vector Machine (SVM): A guard good at drawing clear lines between "good" and "bad" in complex situations.
Random Forest: A team of guards working together, voting on who is bad.
Decision Tree: A guard who asks a series of "Yes/No" questions to make a decision (e.g., "Does it have a symbol? Yes. Does it have a number? Yes. Is it bad?").

4. The Results: The Surprise Winner

When they ran the tests, the results were a bit surprising:

The Decision Tree won the race with a 99% accuracy. It was incredibly sharp at spotting the disguised words.
However, there's a catch. The researchers suspect the Decision Tree might be overfitting.
- The Analogy: Imagine a student who memorized the answers to a specific practice test perfectly but fails if you change the order of the questions. The Decision Tree might have just memorized the specific 100 examples they gave it, rather than learning the concept of what a bully looks like. It's too perfect, which is suspicious.
Logistic Regression and SVM scored around 87-88%. They weren't as flashy, but they were more balanced and likely better at handling new, unseen bullies in the future.
Random Forest actually did the worst (82%), likely because the group of guards got confused by the small amount of data they had to work with.

5. Why This Matters

This research is a crucial step toward Safety by Design.

Cultural Context: What is considered an insult in one culture might be a joke in another. By focusing on Swahili, the researchers are ensuring that safety tools aren't just built for English speakers but are culturally relevant for millions of African children.
The Goal: The ultimate goal isn't just to build a perfect model today, but to create a foundation. They want to gather more data (more examples of bullies) and use advanced techniques (like "Transfer Learning," where a model trained on English helps it learn Swahili) to make these systems robust.

The Bottom Line

The researchers successfully showed that computers can learn to spot Swahili bullies wearing masks. While their current "super-guard" (the Decision Tree) might be a bit too good at memorizing the test, the fact that they can do this at all is a huge win.

The takeaway: To keep kids safe online, we need safety tools that speak their language, understand their culture, and are smart enough to see through the bullies' disguises. This paper is a building block toward that safer future.

1. Problem Statement

The paper addresses the critical issue of cyberbullying and online sexual exploitation of children, specifically within the context of Swahili, a low-resource language (LRL).

The Challenge: Children are increasingly exposed to online abuse, including obfuscated (manipulated) language designed to evade detection filters.
The Gap: Most existing detection systems focus on high-resource languages (HRLs) like English. LRLs like Swahili suffer from a lack of annotated datasets, linguistic tools, and culturally specific models.
Specific Focus: The study targets obfuscated abusive words in Swahili. Obfuscation involves altering text (e.g., misspellings, character substitution) to bypass standard keyword filters, making detection significantly harder than identifying standard abusive text.

2. Methodology

Data Collection and Preprocessing

Dataset: The authors utilized a dataset comprising 100 Swahili abusive texts.
- Labels: Binary classification where 1 represents obfuscated abusive text and 0 represents non-obfuscated abusive text.
- Note: The dataset is noted as small and imbalanced, posing a challenge for training robust models.
Preprocessing:
- TF-IDF (Term Frequency-Inverse Document Frequency): Text was converted into numerical vectors using TF-IDF to capture word importance and relevance within the documents.
- SMOTE (Synthetic Minority Over-sampling Technique): To address data imbalance, SMOTE was applied to the minority class (obfuscated texts) to artificially generate synthetic samples, ensuring a balanced training set.

Machine Learning Models

Four traditional machine learning algorithms were employed and optimized via rigorous parameter tuning:

Support Vector Machines (SVM)
Logistic Regression
Decision Trees
Random Forest

Evaluation Framework

Validation: 5-fold cross-validation was used to ensure result robustness and prevent overfitting.
Metrics: Performance was evaluated using Accuracy, Precision, Recall, and F1-Score.
Mathematical Formulation: The problem was framed as a binary classification function $y = f(x)$ , where $x$ is the input text and $y$ is the binary output (obfuscated vs. non-obfuscated).

3. Key Results

The models were evaluated based on mean cross-validation scores. The results revealed a significant disparity between ensemble methods and single-tree models on this specific small dataset.

Model	Precision	Recall	F1-Score	Mean Accuracy
Decision Tree	0.98	1.00	0.99	99%
Logistic Regression	0.88	0.88	0.88	88%
SVM	0.86	0.88	0.87	87%
Random Forest	0.79	0.86	0.83	82%

Key Observations:

Decision Tree Dominance: The Decision Tree achieved the highest performance (99% accuracy, perfect recall). However, the authors note this is likely due to overfitting, where the model memorized the small training set rather than learning generalizable patterns.
Random Forest Underperformance: Contrary to expectations, the Random Forest (an ensemble method designed to reduce overfitting) performed the worst (82%). The authors attribute this to the small dataset size, which prevents ensemble methods from effectively leveraging their averaging capabilities, and the inherent randomness in feature selection.
Balanced Models: Logistic Regression and SVM showed stable, balanced performance (approx. 87-88% accuracy), making them potentially more reliable for real-world deployment where generalizability is crucial.

4. Key Contributions

LRL Focus: This is one of the first studies to apply machine learning specifically to obfuscated abusive language in Swahili, addressing a significant gap in NLP research for African languages.
Obfuscation Detection: Unlike previous works that focused on standard abusive text, this study specifically targets the nuanced challenge of obfuscated language used to evade filters.
Methodological Benchmarking: The paper provides a comparative analysis of traditional ML models (SVM, LR, DT, RF) on a small, imbalanced Swahili dataset, highlighting the risks of overfitting in low-resource scenarios.
Cultural Context: It emphasizes the necessity of culturally aware detection systems, noting that abusive language varies significantly across cultures and cannot be detected using models trained solely on Western data.

5. Significance and Future Work

Significance:
The study demonstrates that while machine learning can effectively detect obfuscated Swahili abuse, the scarcity of data is a primary bottleneck. The high performance of the Decision Tree serves as a proof-of-concept but highlights the urgent need for larger, diverse datasets to prevent overfitting and ensure the system works for unseen data.

Future Directions:

Dataset Expansion: Creating larger, more diverse, and culturally representative datasets for Swahili.
Advanced Architectures: Moving beyond traditional ML to explore Deep Learning models (e.g., LSTM, BERT) and Transfer Learning (using pre-trained multilingual models like XLM-R) to leverage knowledge from high-resource languages.
Overfitting Mitigation: Applying techniques like pruning to Decision Trees and exploring hybrid models.
Multimodal Integration: Investigating the use of behavioral biometrics (e.g., touch gestures) and multimodal data to complement text-based detection.
Real-world Deployment: Validating these models in practical settings to create "Safety by Design" systems that protect children in the Global South.

In conclusion, the paper establishes a foundational step toward safer online environments for Swahili-speaking children by proving that ML can detect obfuscated abuse, while simultaneously calling for a community-driven effort to build the necessary linguistic resources to scale these solutions.