AI6127 DEEP NEURAL NETWORK FOR NATURAL LANGUAGE PROCESSING

SQuAD baseline BiDAF model

Set up environment

To setup and install dependencies, clone or download this repository into your local machine. Create and activate conda environment for squad baseline with the following command.

conda env create -f environment.yml
conda activate squad_baseline

Next, download the file data.zip containing the processed dataset, embedding vectors and GloVE pre-trained word vectors glove.840d.300d.txt from the following link:

https://drive.google.com/open?id=1cPpDSnXKm-Grh7yW8nU3EQq1skfVBBnm

Unzip the file and place the entire /data folder in the same folder as the script files. If you setup correctly, the scripts and data shoud be in the following structure:

args.py
train.py
models.py
layers.py
util.py
/data/char_emb.json
/data/char2idx.json
/data/word_emb.json
/data/word2idx.json
/data/dev.npz
/data/train.npz
/data/test.npz
/data/glove.840B.300d/glove.840B.300d.txt

If all the files are present, you're good to go.

Train and evaluate

1. To train baseline model WITH character-level embedding, use this command:

python train.py --name <label>

Example:

python train.py --name baseline_char_embed

2. To train baseline model WITHOUT character-level embedding, use this command:

python train.py --name <label> --use_char_emb 0

Example:

python train.py --name baseline_without_char_embed --use_char_emb 0

About our code

This code package is adapted from SQuAD baseline by Chris Chute

https://github.com/chrischute/squad.git

Scripts available here:

args.py : available arguments
train.py : main training script
models.py: BiDAF model
layers.py: implementation of different layers in the model
util.py : utilities script