SinhSafe: Multi-Model Detection of Cyberbullying and Hate Speech

<img src="./images/thumbnail.png" width="200">

Team

• e20397, Thilakasiri P.D., email

Supervisors

• Dr. Eng. Sampath Deegalla, email

Table of content

1. Abstract
2. Related works
3. Methodology
4. Experiment Setup and Implementation
5. Results and Analysis
6. Conclusion
7. Publications
8. Links

Abstract

SinhSafe is an advanced abusive language detection framework specifically designed for the linguistic complexities of the Sinhala language and its code-mixed variant, Singlish. Addressing the critical gap in low-resource language safety tools, this project leverages state-of-the-art multilingual transformer models, specifically XLM-RoBERTa (XLM-R), SinBERT, and SinhLlama, to classify social media content into three distinct categories: Harassment, Offensive, and Normal.

A key innovation of SinhSafe is its 3-Model Ensemble Pseudo-Labeling pipeline, which is designed to generate a massive, high-quality corpus of ~65,000 documents from unlabelled data. By establishing a rigorous ground truth through Inter-Annotator Agreement and achieving a peak accuracy of 80.49% in cross-validation, SinhSafe aims to provide nuanced content moderation solutions for Sri Lankan social media platforms, promoting safer online communities through AI-driven intervention.

Related works

Research into Sinhala cyberbullying detection has been constrained by a lack of diverse datasets and resources, a challenge highlighted by Viswanath & Kumar (2023).

Datasets & Classification Taxonomy Most existing work, such as the SOLD and semi-SOLD datasets by Ranasinghe et al. (2022), focuses on binary classification (Offensive vs. Not Offensive). We identified that binary labels are often insufficient for nuanced moderation. Consequently, our project draws significant inspiration from Mathew et al. (2021) and their HateXplain dataset. Following their approach, we moved beyond binary detection to a 3-class taxonomy (Harassment, Offensive, Normal), allowing for more fine-grained analysis of harmful content.

Model Architectures In the domain of model selection, Dhananjaya et al. (2022) demonstrated that monolingual models like SinBERT often outperform multilingual alternatives on small, specific Sinhala datasets. Conversely, the recent introduction of SinLlama by Aravinda et al. (2025) provides a foundation for testing Large Language Model (LLM) capabilities in Sinhala. SinhSafe benchmarks these distinct architectures (Multilingual XLM-R vs. Monolingual SinBERT vs. SinLlama) to determine the optimal approach.

Methodology

The SinhSafe framework consists of three main stages:

1. Data Collection & Curation: Aggregating a consolidated dataset from multiple sources and establishing a rigorous Ground Truth using Manual Annotation and Inter-Annotator Agreement (Peer Review).
2. Taxonomy Definition (The “Umbrella” Approach):
   • Harassment: An umbrella term covering both Cyberbullying and Hate Speech. It targets intent to harm, threats, and self-harm encouragement.
   • Offensive: General vulgarity, crude jokes, or profanity without malicious targeted intent.
   • Normal: Non-abusive content.
3. Hybrid Preprocessing Pipeline:
   • Google Transliteration API: High-accuracy online transliteration for normalizing Singlish content into Sinhala script.
   • Ensemble Strategy: Using the consensus of XLM-R, SinLlama, and SinBERT to pseudo-label unlabelled data.

Experiment Setup and Implementation

The models were trained using the Hugging Face transformers library on an NVIDIA RTX 3090 GPU. We utilized a custom classification head on the XLM-RoBERTa Large architecture to prevent overfitting:

• Dense Layer: A fully connected linear layer with Tanh activation to extract non-linear semantic features from the 1024-dimensional input vector.
• Dropout Layer: Implemented with a probability of 0.1 to randomly deactivate neurons during training. This “Anti-Misfit” mechanism forces the model to learn robust linguistic patterns rather than memorizing specific training examples.
• Optimization: AdamW optimizer with a learning rate of $2 \times 10^{-5}$ and a 500-step linear warmup.

Results and Analysis

We validated our model using 5-Fold Stratified Cross-Validation to ensure stability across different data splits. This rigorous testing confirmed that the model is robust to the linguistic variance found in social media comments.

Phase 1 Results: XLM-RoBERTa (Large)

Phase 2: Comparative Benchmarking (In Progress) Experiments are currently underway for SinBERT and SinhLlama. Once completed, their performance will be tabulated here to determine the most effective architecture for the final ensemble.

Conclusion

SinhSafe successfully establishes a robust pipeline for Sinhala cyberbullying detection, moving beyond simple binary classification to a nuanced Harassment vs. Offensive taxonomy. Preliminary results with XLM-RoBERTa Large indicate that fine-grained classification in low-resource settings is highly achievable, with a peak accuracy of 80.49%. The ongoing comparison with SinBERT and SinhLlama will provide a definitive benchmark for the research community, identifying the optimal model architecture for protecting the Sri Lankan digital ecosystem.

Publications

Links

• Project Repository
• Project Page
• Department of Computer Engineering
• University of Peradeniya