An academic research project investigating a novel approach to remove sentiment during machine translation using a reinforcement learning framework.
Traditional methods for text style transfer often rely on back-translation to create a style-neutral latent representation of a sentence. This project proposed a more direct and innovative approach: fine-tuning an English-to-Japanese translation model to purge sentiment (a form of style) in a single translation step.
The goal was to create a style-agnostic latent representation in the target language (Japanese) by optimizing the model with a custom reward function that encourages semantic preservation while maximizing sentiment ambiguity.
In this research-focused group project, my role centered on data management and providing essential linguistic expertise:
- Data Curation & Processing: I was responsible for sourcing, curating, and processing the key datasets for this project. This included the English Yelp dataset (as a source of sentiment-rich text) and the Japanese CHABSA dataset (intended for the style re-application phase).
- Japanese Linguistic Expertise: I provided critical Japanese language support to the team. This involved evaluating the quality, coherence, and nuances of machine-generated translations, which was essential for tackling the significant linguistic challenges between English and Japanese.
- Project Collaboration: I participated in project planning, literature review, and the analysis of the baseline back-translation model's performance.
Our proposed method involved fine-tuning a pre-trained Helsinki-NLP English-to-Japanese model. The fine-tuning was designed as a Reinforcement Learning task, where the reward function was based on the output of a BERT-based Japanese sentiment classifier.
- Languages: Python
- Frameworks & Models: Hugging Face Transformers, BERT, Helsinki-NLP Models
- Conceptual Framework: Reinforcement Learning (RL)
While the RL training phase presented significant challenges (as detailed in the report), this project provided deep insights into the complexities of cross-lingual style transfer and fine-grained control of text generation.
For a complete description of our proposed methodology, experimental setup, and detailed analysis, please see the full final report.
This project was a collaborative research effort with my teammates: [Annanya, Jayden Serenari, Noah Bright].