CycleTrans: Unpaired Language Translation with Adversarial and Cycle Consistency

CycleTrans is an advanced model for unpaired language translation, using adversarial training and cycle consistency to translate between languages without requiring parallel datasets. This model allows seamless translation between languages such as English and Italian using two separate, unaligned datasets.

🚀 Overview

CycleTrans is built to work with unpaired datasets—meaning you don't need exact translations for each sentence. It uses two generators and two discriminators inspired by Generative Adversarial Networks (GANs) to ensure high-quality translations. The model also incorporates cycle consistency to maintain translation accuracy between languages.

Key Features:

• Unpaired Data: Train the model with two separate datasets (e.g., English and Italian).
• Adversarial Training: The generators produce translations while the discriminators help improve their quality.
• Cycle Consistency: Ensures that translating between languages in both directions yields the same original sentence.
• Contrastive Loss: Aligns sentences in the shared embedding space for better translation quality.

🧠 How It Works

CycleTrans architecture consists of:

1. Shared Embedding Layer: A shared embedding that maps both English and Italian sentences into the same representation space.
2. Two Generators:
   • G_E2I (English to Italian): Translates from English to Italian.
   • G_I2E (Italian to English): Translates from Italian to English.
3. Two Discriminators:
   • D_E (English Discriminator): Ensures realistic English translations.
   • D_I (Italian Discriminator): Ensures realistic Italian translations.
4. Cycle Consistency: Ensures that translated sentences, when converted back to the original language, remain close to the initial sentence.
5. Adversarial and Contrastive Losses: Improve the quality of translations by leveraging adversarial training and sentence alignment.

💻 Getting Started

Running the Notebook:

To try out the model, you can run the CycleTrans notebook in Google Colab or locally:

Run the notebook for training, evaluation, and testing on your own datasets.

📚 Citation

If you use CycleTrans in your work, please consider citing this project:

@misc{cycletrans2024, author = Davide Brunori, title = {CycleTrans: Unpaired Language Translation with Adversarial and Cycle Consistency}, year = {2024}, url = {https://github.com/dixy52-beep/translation-CycleTRANS-and-Adversal-Network} }

💡 Contribute

CycleTrans is an open-source project! Feel free to fork, submit issues, and create pull requests. Contributions are always welcome!

• Improve the model
• Add new language support
• Update documentation
• Report bugs or suggest features

Publications

[Translating Unknown Languages is now possible with AI - medium article (2024)]