Multilingual-Sentiment-Classifier

This repository contains an end-to-end Natural Language Processing (NLP) pipeline for multilingual sentiment analysis using the pre-trained BERT model nlptown/bert-base-multilingual-uncased-sentiment. The pipeline includes data preprocessing, model fine-tuning with Hugging Face's Trainer, and evaluation metrics. The fine-tuned model classifies text sentiment as negative, neutral, or positive, achieving a test accuracy of 77% and a test F1 score of 76%. Due to limited computational resources, the training process used only 84,000 examples (out of 1.26 million available) and took approximately 1 hour on a T4 GPU.

Features

• Data Preprocessing: Filters datasets to target languages (English, Spanish, French, German), converts star ratings to sentiment polarity, and performs stratified sampling to maintain class balance.
• Tokenization: Prepares text data for the BERT model by applying padding, truncation, and a maximum length of 128 tokens.
• Model Fine-Tuning: Utilizes nlptown/bert-base-multilingual-uncased-sentiment for sequence classification with custom training configurations.
• Evaluation Metrics: Reports accuracy, precision, recall, and F1-score on training, development, and test sets.
• Custom Callbacks: Includes early stopping for optimal performance.
• Final Predictions: Outputs predictions for test data in a structured CSV format.

Why nlptown/bert-base-multilingual-uncased-sentiment?

This model is specifically designed for multilingual sentiment analysis and offers the following advantages:

• Multilingual Support: Supports multiple languages, including English, Spanish, French, and German, making it well-suited for diverse datasets like the Amazon Reviews Corpus.
• Pre-Trained for Sentiment Analysis: Fine-tuned for sentiment classification tasks, reducing the need for extensive retraining and ensuring reliable performance out of the box.
• Robust Contextual Understanding: Leverages the BERT architecture for better handling of complex sentence structures and nuanced sentiments across different languages.

Instructions

1. Ensure you have the required libraries installed (transformers, datasets, pandas, numpy, scikit-learn).
2. Place the training, validation, and test datasets (train.csv, validation.csv, test.csv) in the appropriate directory.
3. Run the script sequentially to preprocess data, fine-tune the model, and evaluate its performance.
4. The final predictions will be saved in a CSV file in the specified output directory.

Dependencies

• transformers: For tokenization and fine-tuning the BERT model.
• datasets: For handling datasets compatible with Hugging Face's Trainer.
• pandas: For data manipulation and processing.
• numpy: For numerical computations.
• scikit-learn: For evaluation metrics such as accuracy, precision, recall, and F1-score.

Notes

• Ensure the file paths in the script match the locations of your datasets.
• Training arguments can be adjusted to explore different configurations (e.g., learning rate, batch size, epochs).
• The random_state = 42 is used throughout to ensure reproducibility.

Author

Christina Joslin

Acknowledgements

• Data provided by the Multilingual Amazon Reviews Corpus posted on Kaggle.
• Citation: Phillip Keung, Yichao Lu, György Szarvas, and Noah A. Smith. “The Multilingual Amazon Reviews Corpus.” In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, 2020.https://www.amazon.science/publications/the-multilingual-amazon-reviews-corpus
• Special thanks to the open-source community for contributing the libraries and tools used in this project.