Myanmar Hate Speech Analysis with XLM-RoBERTa

The project explores how different NLP approaches handle Burmese hate speech analysis across sparse features, sequence learning, and transformer fine-tuning.

• Uses the myHateSpeech corpus to study sentence-level harmful language categories.
• Moves beyond a simple hate versus non-hate setup to compare nine-label prediction.
• Benchmarks sparse text features, sequence learning, and transformer fine-tuning in one workflow.
• Evaluates overall performance and class-balanced behavior instead of relying on one metric.

The myHateSpeech dataset contains Burmese/Myanmar sentences with sentence-level labels. The main task predicts one of nine categories: abusive, religious, racist, body-shaming, political, sexist, lethal, educational, and non-hate speech.

• The raw dataset is kept outside the GitHub repository because of dataset licensing and redistribution rules.
• The main experiment uses a nine-label hate speech setup.
• A binary hate versus non-hate label is also created as an early pipeline check.

The portfolio shows Burmese/Myanmar examples in a safe and non-offensive format so the input format is clear without repeating harmful content.

• Original harmful text is not displayed on the public page.
• The models still train on the labeled dataset used in the experiments.
• Each sentence maps to one sentence-level label before training.

The raw corpus includes word-level annotation tags such as /ab, /bo, and /ed|ab. These markers are removed from the model input while the sentence-level label is kept for training and evaluation.

• Keeps Burmese/Myanmar text and emojis instead of over-cleaning the corpus.
• Creates a processed CSV used by all training notebooks.
• Uses one consistent preprocessing pipeline across all model families.

The baseline setup uses character-level TF-IDF features with Logistic Regression and Linear SVM. Character n-grams help when Burmese spacing and tokenization are inconsistent.

• The baseline is fast, practical, and effective for repeated character patterns.
• It provides a realistic benchmark before moving to deeper architectures.
• It produced the strongest macro F1 in the final comparison.

The BiLSTM baseline treats each sentence as a token sequence and learns from word order before predicting one of the nine labels.

• The BiLSTM provides a deep learning sequence baseline.
• It reached moderate accuracy but weak macro F1.
• The results suggest difficulty with minority labels.

The transformer experiment fine-tunes xlm-roberta-base, a pretrained multilingual transformer model, for Burmese hate speech analysis.

• Uses pretrained multilingual knowledge instead of training from scratch.
• Self-attention helps the model focus on informative parts of each sentence.
• This model achieved the best overall accuracy and weighted F1.

The final comparison shows a clear tradeoff: XLM-RoBERTa leads on overall performance with 78.2% accuracy and 77.2% weighted F1, while TF-IDF achieved the strongest macro F1.

• XLM-RoBERTa delivered the strongest overall result.
• TF-IDF stayed competitive and led on macro F1.
• BiLSTM lagged on class-balanced performance.
• The comparison shows why metric choice matters for imbalanced label sets.

The confusion matrix is used to inspect class-level prediction behavior and identify where the transformer confuses similar or underrepresented categories.

• It shows which labels are predicted more reliably than others.
• It helps locate overlap between similar or sparse categories.
• This adds detail that summary metrics alone can miss.

The stack combines Python data processing, traditional NLP baselines, deep learning sequence modeling, and Hugging Face transformer fine-tuning.

This project is a portfolio case study, not a production moderation system.

• Hate speech detection can be sensitive to sarcasm, slang, coded language, and social context.
• Some labels have fewer examples, which affects macro F1 and minority-label reliability.
• The model may confuse categories with overlapping wording or similar harmful expressions.
• The dataset comes from a specific source and time period, so generalization should be tested further.
• The model should not be used for automatic moderation without additional validation and human review.

This project shows how traditional NLP baselines, sequence models, and transformers behave differently on imbalanced Burmese text tasks.

• Reinforced why strong baselines are important before using deeper models.
• Improved evaluation around macro F1, weighted F1, and class imbalance.
• Fine-tuned a pretrained multilingual transformer using Hugging Face Trainer.
• Connected model results with the practical limits of automated content analysis.