How resham-patel/hinglish-sentiment-analysis works
This is a Hinglish three-class sentiment-classification project. Its main workflow prepares labeled CONLL data, removes conflicting training examples, fine-tunes MuRIL with dev-set model selection, then evaluates once on the held-out test set. Separate branches provide interactive text classification, optional YouTube out-of-domain evaluation, and test-error analysis; the README also documents classical TF-IDF baselines. Source excerpts show the preparation, training, evaluation, and demo implementations; the README supplies the baseline and saved-model details.
Main entry points: train_transformer.py MuRIL fine-tuning · evaluate_final.py Final evaluation · prepare_data.py CONLL preparation · demo_app.py Gradio sentiment demo
Components 15
Data preparation
- SentiMix raw data
data/raw - CONLL preparation
src/prepare_data.py - Clean datasets
- Conflict removal
src/prepare_data.py - Conflict-free training data
Model training
- TF-IDF baselines
notebook/final.ipynb - MuRIL fine-tuning
src/train_transformer.py - Fine-tuned model
Inference and evaluation
- Gradio sentiment demo
src/demo_app.py - Final evaluation
src/evaluate_final.py - Test predictions
- YouTube comments
Error analysis
- Error analysis
src/error_analysis.py - Test error report
Other components
- Text user
Connections 16
- SentiMix raw data → CONLL preparation: provides input
src/prepare_data.py - CONLL preparation → Clean datasets: writes datasets
src/prepare_data.py - Clean datasets → Conflict removal: supplies training CSV
src/prepare_data.py - Conflict removal → Conflict-free training data: writes filtered data
src/prepare_data.py - Conflict-free training data → MuRIL fine-tuning: trains on
src/train_transformer.py - Clean datasets → MuRIL fine-tuning: selects on dev data
src/train_transformer.py - MuRIL fine-tuning → Fine-tuned model: saves model
README.md - Text user → Gradio sentiment demo: enters text
src/demo_app.py - Gradio sentiment demo → Fine-tuned model: loads and predicts
src/demo_app.py - Clean datasets → Final evaluation: provides test data
src/evaluate_final.py - Fine-tuned model → Final evaluation: loads model
src/evaluate_final.py - YouTube comments → Final evaluation: optional input
src/evaluate_final.py - Final evaluation → Test predictions: writes predictions
src/evaluate_final.py - Test predictions → Error analysis: analyzes mistakes
src/error_analysis.py - Error analysis → Test error report: writes errors
src/error_analysis.py - TF-IDF baselines → MuRIL fine-tuning: compares models
README.md