Repair-QA: No That's Not What I Meant: Handling Third Position Repair in Conversational Question Answering

This repo contains the dataset and implementation for the paper: No That's Not What I Meant: Handling Third Position Repair in Conversational Question Answering. Vevake Balaraman, Arash Eshghi, Ioannis Konstas and Ioannis Papaioannou. SIGdial 2023. PDF

Dataset

The dataset can be found data folder

Packages

To install required packages for this repository,

pip install -r requirements.txt

Train the model

1. Download Repair-QA dataset and place it under folder data
2. Modify utils/config.py to point to the right folders.
3. Start training using the following command

   python engine.py
   

Predict for a given input file

1. To make predictions for a input file.

   python predict.py <input_file> <prediction_savename>
   

Note: The input file should consists of a list of dictionaries with each item consisting of {"Context": list, "Question": str}

Compute evaluation metrics on the prediction file

1. Compute metrics on the evaluation file

   python src/compute_metrics.py <prediction_file>