This project demonstrates fine-tuning a BERT-based multilingual model for sentiment classification across multiple languages using the Amazon Reviews Multilingual dataset. The notebook explores different transfer learning approaches, including zero-shot, few-shot, training transfer, and test transfer methods, comparing their performance on diverse language datasets.
- Goal: Fine-tune a multilingual model to classify product reviews as positive or negative across multiple languages.
- Dataset: Amazon Reviews Multilingual.
- Model: BERT-multilingual.
- Reference:2020.emnlp-main.369.pdf
- Clone this repository:
git clone https://github.com/Hazel1763/Projects.git- Install the required packages:
pip install -r requirements.txt- Open the notebook and run cells to reproduce the results.
⚠️ Note: A GPU environment is recommended to reduce runtime.
- Explore additional transfer learning techniques (e.g. data augmentation, back-translation).
- Increase test sample size in generative models to enhance the robustness of performance comparisons.
- Evaluate performance on more languages, especially low-resource languages.
- The amazon_reviews_multi dataset used in this project is sourced from the Hugging Face datasets library, based on the publicly available Amazon product reviews dataset.
- The bert-base-multilingual-cased model used in this project is provided by Hugging Face Model Hub, originally developed by Google AI.
If you have any questions or suggestions, feel free to reach out.