A chainer implementation of Bi-directional Attention Flow for Reading Comprehension (arxiv:1611.01603)
Requirements:
- Python 3.5+ (verified)
- Chainer 3.1.0+ (verified)
- The preprocessed SQuAD data from the original BiDAF (github)
To run,
python train.py
For debugging with a small data:
python train.py --debug_mode
| dataset | EM (%) | F1 (%) | loss |
|---|---|---|---|
| train (this repo, 10k) | 65.3 | 79.3 | 1.83 |
| dev (this repo, 10k) | 55.1 | 70.1 | 2.81 |
| dev (original) | (67.7) | (77.3) | (-) |