By: Chenglei Si (River Valley High School)
XLNet has achieved impressive gains on RACE recently. You may refer to my other repo: https://github.com/NoviScl/XLNet_DREAM to see how to use XLNet for multiple-choice machine comprehension problems. Huggingface has updated their work pytorch_trainsformers, please refer to their repo for the documentation and more details of the new version.
This work is based on Pytorch implementation of BERT (https://github.com/huggingface/pytorch-pretrained-BERT). I adapted the original BERT model to work on multiple choice machine comprehension.
The code is tested with Python3.6 and Pytorch 1.0.0.
- Download the dataset and unzip it. The default dataset directory is ./RACE
- Run
./run.sh
I did some tuning and find the following hyperparameters to work reasonally well:
BERT_base: batch size: 32, learning rate: 5e-5, training epoch: 3
BERT_large: batch size: 8, learning rate: 1e-5 (DO NOT SET IT TOO LARGE), training epoch: 2
| Model | RACE | RACE-M | RACE-H |
|---|---|---|---|
| BERT_base | 65.0 | 71.7 | 62.3 |
| BERT_large | 67.9 | 75.6 | 64.7 |
You can compare them with other results on the leaderboard.
BERT large achieves the current (Jan 2019) best result. Looking forward to new models that can beat BERT!
I have written a short report in this repo describing the details.