Hypothesis Prediction Using XLM-Roberta

Overview

This project implements a multilingual Natural Language Inference (NLI) pipeline using the XLM-Roberta model. It predicts the relationship between a premise and a hypothesis (contradiction, neutral, entailment) across multiple languages. The workflow leverages PyTorch and HuggingFace Transformers for model training, evaluation, and prediction.

Results & Performance

• Achieved >91% accuracy on the test submission data.
• Kaggle Data: Contradictory, My Dear Watson
• My Code on Kaggle: Hypothesis Prediction Using XLM-Roberta

Features

• Multilingual NLI using XLM-Roberta (XNLI)
• Stratified train/eval split by language and label
• Weighted loss for handling class imbalance
• Custom PyTorch Dataset and Sampler for balanced multilingual batches
• Training loop with accuracy tracking and model checkpointing
• Evaluation with classification report and confusion matrix
• Test set prediction and CSV submission generation
• Data visualization for language distribution

Project Structure

• hypothesis-prediction-using-xlm-roberta.ipynb: Main Jupyter notebook containing all code and workflow
• LICENSE: Project license
• README.md: Project documentation

Requirements

• Python 3.7+
• PyTorch
• Transformers (HuggingFace)
• pandas, numpy, seaborn, matplotlib, scikit-learn

Install dependencies (in notebook or terminal):

!pip install transformers torch pandas numpy seaborn matplotlib scikit-learn

Data

• Train/Eval Data: train.csv from Kaggle's "Contradictory, My Dear Watson" competition
• Test Data: test.csv from the same source

Usage

1. Open the notebook: hypothesis-prediction-using-xlm-roberta.ipynb
2. Run all cells: The notebook will install dependencies, load data, preprocess, train, evaluate, and generate predictions.
3. Model Training: The notebook trains XLM-Roberta on the NLI task, saving the best model based on evaluation accuracy.
4. Evaluation: Prints classification report and confusion matrix for validation set.
5. Test Prediction: Loads test data, makes predictions, and saves results to submission.csv.

Key Code Sections

• Imports & Setup: Installs and imports required libraries
• Data Loading & Visualization: Loads CSVs, analyzes and visualizes language distribution
• Model & Tokenizer: Loads XLM-Roberta model and tokenizer
• Custom Dataset & Sampler: Defines classes for balanced multilingual batching
• Training Loop: Trains model, tracks metrics, saves best checkpoint
• Evaluation & Reporting: Evaluates model, prints metrics
• Test Prediction: Generates predictions for test set

Customization

• Adjust MODEL_NAME to use different transformer models
• Change max_length, batch_size, or num_epochs for experimentation
• Modify data paths for local or cloud environments

Results

• The notebook provides detailed metrics and visualizations for model performance across languages.
• Final predictions are saved in submission.csv for Kaggle submission.

License

This project is licensed under the terms of the LICENSE file.

References

• HuggingFace Transformers
• Kaggle Contradictory, My Dear Watson
• XLM-Roberta Model