# Multi-Class Cyberbullying Detection: Classical vs. Transformers
> Explore a comparative NLP study on detecting cyberbullying using Logistic Regression and DistilBERT, including semantic analysis and ethical trade-offs.

Tags: machine-learning, nlp, cyberbullying-detection, distilbert, data-science, text-classification, academic-project, artificial-intelligence
## Problem Formulation
- Online harassment requires multi-class detection for identity-targeted abuse (age, religion, ethnicity).
- Research focuses on comparing Classical ML and Transformer-based models (DistilBERT).

## Methodology & Dataset
- **Dataset:** 47,459 labeled English tweets across 6 balanced classes.
- **Dual Pipeline:** 
  - Tier 1: Basic cleaning for DistilBERT.
  - Tier 2: Advanced normalization (lemmatization) for TF-IDF/Classical models.

## Model Benchmarks
- **Naïve Bayes:** 77.17% Accuracy, 0.7644 Macro F1.
- **Logistic Regression:** 82.68% Accuracy, 0.8268 Macro F1.
- **DistilBERT:** 86.29% Accuracy, 0.8607 Macro F1.
- DistilBERT showed the most significant gains in ambiguous categories like 'not_cyberbullying' (+8.8 pp F1).

## Semantic Cohesion Analysis
- Identity-specific classes (age, ethnicity) show high intra-class cosine similarity (0.71-0.74).
- 'General' or 'not_cyberbullying' classes are more dispersed (0.44-0.47) and share overlapping embedding space.

## Technical Trade-offs & Ethics
- **Scalability:** DistilBERT is 55x slower in inference than Logistic Regression.
- **Ethics:** Discussion on representation bias, pre-training bias, and the risk of misclassifying reclaimed language.
---
This presentation was created with [Bobr AI](https://bobr.ai) — an AI presentation generator.