Unmasking Cyberbullying in Bengali: A Deep Dive into Digital Harassment

The Challenge

Cyberbullying is a growing problem online, and it's not limited to English-speaking communities. In the Bengali-speaking world, it's a significant issue, yet there's little research on the topic. A recent study aimed to change that by analyzing over 70,000 social media comments in Bengali to better understand and detect cyberbullying.

The Approach

Data Cleaning and Categorization

• The researchers first cleaned the data, sorting comments into positive and negative categories.
• They then applied Latent Dirichlet Allocation (LDA) to identify patterns in the negative comments, focusing on topics like age, gender, ethnicity, and religion.

Model Testing

To determine the best method for detecting cyberbullying, they tested several models:

• Support Vector Machine (SVM)
• XGBoost
• CNN+BiLSTM+GRU
• Multilingual BERT (mBERT)

Results:

• mBERT achieved the highest accuracy at 92%.
• The hybrid model CNN+BiLSTM+GRU was close behind at 91%.

Model Optimization

The researchers further improved accuracy by:

• Adding BERT embeddings to CNN and ANN, boosting accuracy to 93%.
• Using Local Interpretable Model-agnostic Explanations (LIME) to ensure the models were fair and transparent.

The Impact

This study marks a significant advancement in detecting cyberbullying in Bengali. However, the fight against cyberbullying is far from over. As a complex issue, continuous improvement in detection methods is crucial.