BiDAF (Bidirectional Attention Flow)

This repository contains the implementation of the Bidirectional Attention Flow (BiDAF) model for machine reading comprehension and question-answering tasks. The BiDAF model is designed to process context and query to provide precise answers from a given passage.

Repository Structure

• Bidaf.ipynb: Jupyter Notebook containing the implementation and training workflow for the BiDAF model.
• Bidaf.py: Python script version of the BiDAF implementation for users who prefer non-notebook environments.
• Data.zip: Compressed dataset file containing the training and validation data for BiDAF. Unzip this file before running the code.
• bidafglove_tv.npy: Pre-trained GloVe embeddings used for training the BiDAF model.
• bidaftrain.pkl: Serialized training dataset in Python pickle format.
• bidafvalid.pkl: Serialized validation dataset in Python pickle format.

Features

• Implementation of BiDAF model for question-answering tasks.
• Utilizes pre-trained GloVe embeddings for word representation.
• Support for training and validation workflows.
• Jupyter Notebook for experimentation and Python script for streamlined execution.

Setup and Usage

Prerequisites

• Python 3.7 or higher
• Required Python libraries:
  • numpy
  • pandas
  • torch
  • tqdm
  • pickle

Install the dependencies using:

pip install -r requirements.txt

Dataset

• Unzip the Data.zip file to extract the training and validation datasets.
• Ensure the .pkl and .npy files are in the same directory as the scripts or notebooks.

Running the Code

1. For interactive experimentation, open Bidaf.ipynb in Jupyter Notebook:

   jupyter notebook Bidaf.ipynb

2. For command-line execution, use the Bidaf.py script:

   python Bidaf.py

Results and Evaluation

The model's performance can be evaluated using standard metrics like Exact Match (EM) and F1 score. These scores are calculated on the validation dataset.