BIDAF

Reimplementation Bi-Directional Attention Flow for Machine Comprehension

First of all, re-implement the BIDAF model with SQuAD dataset

and then, modifying model architecture to solve Multi-Sentence Reading Comprehension dataset

Execution

For training your own model,

python main.py --is_train True

for evaluating your trained model

python main.py --is_train False

Model Graph loss flow

Dataset

(1) SQuAD dataset v1.1

(2) Multi-Sentence Reading Comprehension dataset (Multi-RC)

1. ~6k high-quality multiple-choice RC questions
2. 7 different sources, allowing more diversity in content, style, and possible question types.
3. Questions aren't constrained in single correct answer.

Version Type

1. tensorflow version: 1.9.0
2. python version: 3.6.0

Reference

• Model - https://arxiv.org/pdf/1611.01603.pdf
• Dataset - http://www.aclweb.org/anthology/N18-1023