This is source code for EMNLP 2020 'Towards Interpretable Reasoning over Paragraph Effects in Situation'.
Code is based on hugging face transformer v2.3.0, to install all dependencies, including external libraries, or a link to such resources, following commands will help you to install easily.
pip install .
pip install allennlp==0.9.0
pip install -r requirements.txt
If you have problems about installing above package, especially the first line. The best way is go to Huggingface, find the version 2.3.0 then download it.
After that , replace all the files in the examples and src/transformers/data/ with this repositories. Then continue installation.
Be careful with Pytorch version, make sure it is compatible with both allennlp and transformers. (suggest v1.5.1)
We provide official data, auxiliary labeled data and 5-fold cross-validation data in here.
Finetuned Model will be uploaded in few days
The following script is used for evaluation models. If you want to train a model by yourself, just need to modify this script a bit.
# Evaluation example for Interpretable Reasoning
CUDA_VISIBLE_DEVICES=0 python examples/run_ropes_interpretable.py --model_type roberta --model_name_or_path /path/to/model --do_eval --do_lower_case --data_dir /path/to/data --predict_file /path/to/file --max_seq_length 512 --doc_stride 64 --output_dir /path/to/output --overwrite_output_dir --gradient_accumulation_steps 1 --grounding_type s_first --overwrite_cache
# Evaluation example for Answer Prediction
CUDA_VISIBLE_DEVICES=0 python examples/run_ropes.py --model_type roberta --model_name_or_path /path/to/model --do_eval --do_lower_case --data_dir /path/to/data --predict_file /path/to/file --max_seq_length 384 --doc_stride 128 --max_answer_length 9 --output_dir /path/to/output --grounding_type synthetic_2nd --multi_answer --overwrite_cache --overwrite_output_dir