🦦✨ OtterlyObsessedWithSemantics ✨🦦

This repo contains our submission for Semeval 2024 Task 4.
Feel free to take a deep dive 🦦🌊

Setup 🦦

Install Requirements 🚀

Using pip and Python3.10.0

pip install -r requirements.txt

Data 📊

1. Get data from the Task Site
2. Load data into data directory:

   data
   └── subtask1
       ├── dev.json
       ├── dev_unlabeled.json
       ├── train.json
       └── validation.json
   

3. Run preprocessing

    python3 -m src.classes.preprocess
   

Training 🏋️

Execute the notebook: src/tune_classification_model.ipynb
✨Parameters✨ can be changed in the run-config:

• dataset_style: Either cleaned or all_lower
• model_name: Huggingface identifier for the model (like bert-base-cased)
• use_custom_head: Whether to use the custom head we developed (Trueor False)
• use_hierarchy: Whether to use the hierarchy instead of only the leaves (True or False)
• extra_lazers: Whether to add additional linear layers in the custom head (True or False)
• weight_loss: Whether to weight the classes based on their inverse frequencies in the cross entropy loss calculation (Trueor False)
• epochs: Number of training epochs
• lr: Learning rate
• batch_size: For the GPU
• acc_steps: Accumulation steps
• seed: Set random seed
• limit: Only train on a subset of the data (int or None to use the full dataset)