Usage

1. Clone the repository
2. Install the following dependencies:
   • NumPy
   • torch
   • Sklearn
   • transformers
   • evaluate
   • nnsight
   • PySpark
   • Pandas
   • matplotlib
   • seaborn

Data:

Twitter Emotion Dataset: https://www.kaggle.com/datasets/adhamelkomy/twitter-emotion-dataset/data

To get the model:

The models are available here - https://drive.google.com/drive/folders/1D7eiUCLUJFTEy0FTlyFbQirYR2dbRhTw?usp=drive_link .

There are two folders -

1. model/ - used for all predictions, run tests, etc.
2. model_with_tokenizer/ - used primarily in token_analysis

Ensure the two models are inside code/

To run the head masking experiment:

1. Run head_masking_samples.py to obtain the samples (the results from this are stored in data/samples if you do not want to install PySpark and run the file.)
2. Run head_masking.py to process the sentences obtained from step1 on each variant. The results are stored as npy files in the execution directory under the names head_masking_class_prob_diff.npy and multi_head_masking_class_prob_diff.py

To run the word flipping experiment:

1. Connect to a python/ ipynb kernel that has the libraries mentioned in the requirements.
2. Run the first five cells of code/token_analysis.ipynb
3. The remainder of the notebook compares the heatmaps of two sentences. Feel free to experiment by giving custom sentences.

To run the word replacement experiment:

1. Run common_words.py to obtain the 10 most common words per class and replace them with synonyms (the results from this are stored in data/replaced_words_text.csv if you do not want to install PySpark and run the file.)
2. Run the cells of code/synonym_analysis.ipynb to get the probability of the original and synonym-replaced sentences.

Nonstandard dependencies:

• PySpark for sample collection