This is source code, which is implementation for submission: 'Exploiting Question-Context Interaction for Machine Reading Comprehension'. We modify this code base on Hugging Face framework for Question Answering: https://huggingface.co/. Our study conduct the extensive experiments on 2 datasets as: Squad 2.0 vs News QA. To reproduce the results, follow exactly the following steps:
- Prepare Data
- For Squad 2.0: Train set vs Dev set can be easily found and downloaded from link: https://rajpurkar.github.io/SQuAD-explorer/
- For NewsQa: To convert NewsQa dataset to Squad 2.0 format, please follow strictly code in this link: https://github.com/amazon-research/qa-dataset-converter
- Train
- After preparing sucessfully data in Squad 2.0 format. For NewsQa, we could have three json files: newsqa_train.json , newsqa_dev.json , newsqa_test.json. Put all of them in same path with 'run_newqa.sh' file. Run command line: 'bash run_newsqa.sh' to start training process.
- Test
- To evaluate performance on test set, we only need change --do_train ---> --do_eval and set --dev_set=path_to_test_set. Run command line: 'bash test_newqa.sh'