Awesome

NILE

Reference code for ACL20 paper - NILE : Natural Language Inference with Faithful Natural Language Explanations.

Dependencies

The code was written with, or depends on:

• Python 3.6
• Pytorch 1.4.0

Running the code

1. Create a virtualenv and install dependecies

   virtualenv -p python3.6 env
   source env/bin/activate
   pip install -r requirements.txt
   

2. Fetch data and pre-process. This will create add files in dataset_snli and dataset_mnli folders.

   bash fetch_data.sh  
   python prepare_train_test.py --dataset snli --create_data  --filter_repetitions
   python prepare_train_test.py --dataset mnli --create_data  --filter_repetitions
   

3. Fine-tuning langauge models using e-SNLI, for entailment, contradiction and neutral explanations. 'all' is trained to produce a comparable single-explanation ETPA baseline, and can be skippd in this and subsequent steps if only reproducing NILE.

   bash run_finetune_gpt2m.sh 0 entailment 2
   bash run_finetune_gpt2m.sh 0 contradiction 2
   bash run_finetune_gpt2m.sh 0 neutral 2
   bash run_finetune_gpt2m.sh 0 all 2
   

4. Generate explanations using the fine-tuned langauge models, where <dataset> can be snli or mnli, and <split> is train/dev/test for SNLI and dev/dev_mm for MNLI.

   bash run_generate_gpt2m.sh 0 entailment <dataset>  all <split>
   bash run_generate_gpt2m.sh 0 contradiction <dataset> all <split>
   bash run_generate_gpt2m.sh 0 neutral <dataset> all <split>
   bash run_generate_gpt2m.sh 0 all <dataset> all <split>
   

5. Merge generated explanation

   python prepare_train_test.py --dataset <dataset> --merge_data --split <split> --input_prefix gpt2_m_
   

   To merge for the single-explanatoin baseline, run

   python prepare_train_test.py --dataset <dataset> --merge_data --split <split> --input_prefix gpt2_m_  --merge_single
   

6. Train classifiers on the generated explantions, models are saved at saved_clf.

   NILE-PH

   bash run_clf.sh 0 <seed> snli independent independent gpt2_m_ train dev _ _ _
   bash run_clf.sh 0 <seed> snli aggregate aggregate gpt2_m_ train dev _ _ _
   bash run_clf.sh 0 <seed> snli append append gpt2_m_ train dev _ _ _
   

   NILE-NS

   bash run_clf.sh 0 <seed> snli instance_independent instance_independent gpt2_m_ train dev _ _ _
   bash run_clf.sh 0 <seed> snli instance_aggregate instance_aggregate gpt2_m_ train dev _ _ _
   bash run_clf.sh 0 <seed> snli instance_append instance_append gpt2_m_ train dev _ _ _
   

   NILE

   bash run_clf.sh 0 <seed> snli instance_independent instance_independent gpt2_m_ train dev sample _ _
   bash run_clf.sh 0 <seed> snli instance_aggregate instance_aggregate gpt2_m_ train dev sample _ _
   

   NILE post-hoc

   bash run_clf.sh 0 <seed> snli instance instance gpt2_m_ train dev sample _ _
   

   Single-explanation baseline

   bash run_clf.sh 0 <seed> snli Explanation_1 Explanation_1 gpt2_m_ train dev sample _ _
   

7. Evaluate a trained classifier for label accuracy, <model_path> is the path of a model saved in the previous step, and <model> can be independent, aggregate, append, instance_independent, instance_aggregate, and instance for NILE variants, and all_explanation for the single-explanation baseline

   bash run_clf.sh 0 <seed> snli <model> <model> gpt2_m_ _ test _ _ <model_path>
   

Example explanations

Generated explanations on the e-SNLI dev and test sets are present at ./generated_explanations/*.gz. Please unzip before using (gunzip <filename>). The generated explanations are present in the entailment_explanation,contradiction_explanation,neutral_explanation columns in a csv format.

Pre-trained models

We are sharing pre-trained label-specific generators. We are also sharing pre-trained classifiers for NILE-PH Append and NILE Independent architectures.

Citation

If you use this code, please consider citing:

[1] Sawan Kumar and Partha Talukdar. 2020. NILE : Natural language inference with faithful natural language explanations. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8730–8742, Online. Association for Computational Linguistics. [bibtex]

Contact

For any clarification, comments, or suggestions please create an issue or contact sawankumar@iisc.ac.in