Intruduction
This project is my code for AiChallenger2018 Opinion Questions Machine Reading Comprehension. This project mainly has 2 models which are implemented in Tensorflow.
- Model-1 is based on QANet, but I rewrite it in some details.
- Model-2 is based on capsuleNet which mainly from freefuiiismyname's project
Dependencies
- Python 3.6
- Tensorflow 1.9.0
- tqdm
- gensim
Data Sample
{ “query_id”:1, “query”:“维生c可以长期吃吗”, “url”: “xxx”, “passage”: “每天吃的维生素的量没有超过推荐量的话是没有太大问题的。”, “alternatives”:”可以|不可以|无法确定”, “answer”:“可以” }
Performance
I train each model on one GTX1080ti for 30 epochs, and report the best Performance on dev set. We finally run Model-1 on testa, the accuracy is 73.2.
| Model | Accuracy |
|---|---|
| Model-1(ensembled) | 73.66 |
| Model-2(ensembled) | 73.85 |
| 1&2 ensembled | 76.62 |
Project Structure
- capsuleNet： Model-2's codes
- data: Model-2's data
- QANet: Model-1's codes and data
- start.sh: example for usage
- vote_ser_new_word.py: vote ensemble file
Details
you can see more details in /QANet/README.md and /capsuleNet/README.md
Reference
some codes are borrowed from :