EEG-to-Text-Project

Natural Language Processing (CS5624) Course Project in Virginia Tech

Preprocessing the data

Add the mat files to the dataset/raw folder. Eg. dataset/raw/Task1_SR/{subject_mat_files} The dataset/raw folder should have three tasks' data: Task1_SR, Task2_NR, Task3_TSR.

Run

bash scripts/preprocess_datasets.sh

This will save the pickle files to dataset/processed directory

Training the decoding model

After processing the datasets run

bash scripts/train.sh

This will train the decoder model for converting EEG to text embeddings. The models are saved in save_data/checkpoints/ folder.

Evaluate the decoding model

To evaluate the trained model run

bash scripts/eval.sh

This will save an output file with the target and predicted sentences in saave_data/eval_results folder.

#Zero-shot sentiment classification model In contrast to the code you need to run using shell script, for the zero-shot sentiment classification model, you can just run jupyter notebook.

train_eval_sentiment_textbased.ipynb: this includes code for training and evaluating zero-shot classification model. train_eval_sentiment_baseline.ipynb: this includes codes for training and evaluating baseline classification model.