This code is for paper "Explainable Natural Language Inference via Identifying Important Rationales".
-
To run this code, you should obtain the data from https://github.com/OanaMariaCamburu/e-SNLI, and put the datas in the 'datas/raw_datas/' directory. Then process these data by run the 'process_data.py';
-
go to the file of 'rationale_extra' and run the train.py and then the predict.py to predict the rationales;
-
go to the file of 'rationales_sel' and run the main_selector.py to selector the rationales;
-
go to the file of 'generator' and
4.1 finetune the GPT2 model: run the 'prepare_data_for_finetune.py' and 'GPT2_finetune_lm.py';
4.2 generate the NLEs by the finetuned model: run the 'prepare_data_for_generation.py' and 'GPT2_generate.py';
-
go to the file of 'classify', and run the 'prepare_data_task_pred.py' to prepare data for train. Then, run the 'main_classify.py' to obtain the prediction results.
Note: All parameters use default values, and you can also set them at runtime.
After running the above code correctly, you will get the following files in the 'datas/snli_data_dir/' directory
--train.json
--dev.json
--test.json
--train-rationales.json
--dev-rationales.json
--test-rationales.json
--train-with-hints.json
--dev-with-hints.json
--test-with-hints.json
--train-prompts.txt
--train-prompts.txt
--train-prompts.txt
--train-finetune.txt
--dev-finetune.txt
--train-gen.txt
--dev-gen.txt
--test-gen.txt
--train-r-with-e.txt
--dev-r-with-e.txt
--test-r-with-e.txt
This repo is built upon the following work:
LIREx: Augmenting Language Inference with Relevant Explanations.
https://github.com/zhaoxy92/LIREx
Many thanks to the authors and developers!