📝 Simultaneous Translation CZ → EN

Simul-Translation is a research-oriented toolkit for text-to-text Simultaneous Machine Translation (SimulMT) with support for dataset preparation, model training, fine-tuning, evaluation, and visualization. It includes modular components for alignment-based policies, local agreement strategies, and output analysis.

The main script main.py allows you to run model evaluation experiments with different configurations. It supports various command-line arguments to customize the behavior of the script.

The Data_preparation folder contains the create_dataset.py script for preparing our training and testing datasets.

The Finetuning folder contains the finetune_nmt_backbone.py script for fine-tuning the NMT backbone model on a dataset.

The AlignAtt folder contains the definitions of align att and local agreement policies and runs the model using them.

The Evaluation folder contains the scripts for the evaluation metrics, which can be used independently or by running the main script.

The Visualization folder contains scripts for parsing results, creating tables, and plotting results.

The AlignAttOutputs folder contains the output logs from the grid search over local agreement and align att parameters, parsed results, and final results.

🛠️ Technologies Used

---

🔧 Installation

Each module contains its own requirements.txt. To install dependencies for a module, run:

pip install -r <module>/requirements.txt

Example:

pip install -r Data_preparation/requirements.txt

Alternitavely, to install all dependencies for all modules, run:

pip install -r requirements.txt

---

⚙️ Command-line Arguments for main.py

📜 Usage Example

python main.py \
  --dataset_path ./Data_preparation/cleaned_eval_dataset.jsonl \
  --experiment_type alignatt \
  --layers 3 \
  --heads 0 1 2 3 4 5 \
  --model_id 0 \
  --device cuda \
  --output_file results.jsonl

---

📂 Project Structure

Simul-Translation/
├── .gitignore
├── main.py
├── README.md
├── report.pdf
├── requirements.txt
├── slides.pdf
├── AlignAtt/
│   ├── README.md
│   ├── alignatt.py
│   ├── analyze_dataset.py
│   ├── get_data.py
│   ├── local_agreement.py
│   ├── requirements.txt
│   └── translate.py
├── AlignAttOutputs/
│   ├── output_logs_grid_search_local_agreement/
│   ├── parsed/
│   └── results/
├── Data_preparation/
│   ├── README.md
│   ├── cleaned_eval_dataset.jsonl
│   ├── create_dataset.py
│   ├── iwslt2024_cs_devset.json
│   └── requirements.txt
├── Evaluation/
│   ├── README.md
│   ├── requirements.txt
│   └── simueval.py
├── Finetuning/
│   ├── README.md
│   ├── finetune_nmt_backbone.py
│   ├── finetuning.log
│   └── requirements.txt
└── Visualization/
    ├── RESULTS.md
    ├── RESULTS_OVERALL.md
    ├── best.json
    ├── create_tables.py
    ├── parse_results.py
    └── plot_results.py