SQuAD_Sequential_Model

This project attempts to conduct a word-level BiLSTM to work in tandem with the bi-direction attention flow mechanism for machine comprehension problem. Experiments conducted on the Stanford Question Answering Dataset(SQuAD) attest the effectiveness of our models and others.

Dataset Description

Standford Question Answering Dataset(SQuAD) is the main dataset that we are going to use for this project.It is a reading comprehension dataset, consisting of questions posed by crowdworkers on a set of Wikipedia articles, where the answer to every question is a segment of text, or span, from the corresponding reading passage, or the question might be unanswerable.

Preprocessing

Create a new folder called data

$ mkdir data

Download the qa data into data file structure

data 
  |-- train-v1.1.json
  |-- dev-v1.1.json

$ python3 code/preprocess.py

Generate Preprocess files

data 
  |-- train.context
  |-- train.question
  |-- train.answer
  |
  |-- dev.context
  |-- dev.question
  |--dev.answer

Download Pre-trained model under data

data 
  |-- glove.6B
        |-- glove.6B.100d.txt
        |-- glove.6B.200d.txt
  		...

Run the RNN Model

Install required packages

pip install tensorflow
pip install numpy

Run the program

$ python code/main.py --experiment_name=bidaf_best --dropout=0.15 --batch_size=60 --hidden_size_encoder=150 --embedding_size=100 --do_char_embed=False --add_highway_layer=True --rnet_attention=False --bidaf_attention=True --answer_pointer_RNET=False --smart_span=True --hidden_size_modeling=150 --mode=train

Experiment Results

Model	F1	EM
LSTM + Bidaf	68.42%	53.69%
LSTM + Attn	47.57%	34.66%
DRQA	78.9%	69.4%
BERT	91.8%	85.1%