Inoculation by Fine-Tuning

Code for Inoculation by Fine-Tuning: A Method for Analyzing Challenge Datasets, to appear at NAACL 2019.

Table of Contents

• Installation
• Training the models used for inoculation
• Re-running the NLI Challenge Dataset Experiments
  • Decomposable Attention
  • ESIM
  • ESIM + Character-Level Component
• Re-running the Adversarial SQuAD Experiments
  • BiDAF
  • QANet
• References

Installation

This project was developed in Python 3.6, with the AllenNLP framework.

Conda will set up a virtual environment with the exact version of Python used for development along with all the dependencies needed to run the code.

1. Download and install Conda.

2. Change your directory to your clone of this repo.

   cd inoculation-by-finetuning

3. Create a Conda environment with Python 3.6 .

   conda create -n inoculation python=3.6

4. Now activate the Conda environment. You will need to activate the Conda environment in each terminal in which you want to run code from this repo.

   source activate inoculation

5. Install the required dependencies.

   pip install -r requirements.txt

Training the models used for inoculation

You can download the pretrained models we used in the paper with the links below:

• Decomposable Attention (Matched, Mismatched)
• ESIM (Matched, Mismatched)
• ESIM + character-level component (Matched, Mismatched)
• BiDAF
• QANet

AllenNLP configuration files for training the NLI models used for inoculation can be found in inoculation-by-finetuning/training_configs/nli_stress_test/ .

In particular, the Decomposable Attention model was trained with decomposable_attention_original_matched.jsonnet and decomposable_attention_original_mismatched.jsonnet, the ESIM model was trained with esim_character_level_matched.json and esim_character_level_mismatched.json, and the ESIM + character-level component model was trained with esim_character_level_matched.json and esim_character_level_mismatched.json.

AllenNLP configuration files for training the reading comprehension models used for inoculation can be found in inoculation-by-finetuning/training_configs/adversarial_squad/ .

The BiDAF model was trained with bidaf_original.json, and the QANet model was trained with qanet_original.jsonnet.

To train each of these models again, use the allennlp train command like so:

allennlp train <path to config file> -s <path to directory to save results>

Re-running the NLI Challenge Dataset Experiments

To reproduce the inoculation results for each of the NLI models (Decomposable Attention, ESIM, ESIM + character-level component), use the scripts in scripts/nli_stress_test/ . These scripts look for configuration files in training_configs/nli_stress_test/.

Example running commands for each model are given below.

Decomposable Attention

To rerun inoculation of the decomposable attention model on, say, the matched negation dataset, we'd simply run:

./scripts/nli_stress_test/tune_decomposable_attention_fine_tune_lr.sh matched_negation_adversary

where matched_negation_adversary is a folder containing configs at training_configs/nli_stress_test/decomposable_attention_original/.

As another example, if we wanted to run the inoculation of the decomposable attention model on the mismatched spelling advesary, we can use:

./scripts/nli_stress_test/tune_decomposable_attention_fine_tune_lr.sh mismatched_spelling_adversary

where mismatched_spelling_adversary is another directory name in training_configs/nli_stress_test/decomposable_attention_original/.

These scripts fine-tune the decomposable attention model on each of the fine-tuning dataset sizes for a range of learning rates, saving the results to disk.

ESIM

To rerun inoculation of the ESIM model on, say, the matched negation dataset, we'd simply run:

./scripts/nli_stress_test/tune_esim_fine_tune_lr.sh matched_negation_adversary

where matched_negation_adversary is a folder containing configs at training_configs/nli_stress_test/esim_original/.

As another example, if we wanted to run the inoculation of the ESIM model on the mismatched spelling advesary, we can use:

./scripts/nli_stress_test/tune_esim_fine_tune_lr.sh mismatched_spelling_adversary

where mismatched_spelling_adversary is another directory name in training_configs/nli_stress_test/esim_original/.

These scripts fine-tune the ESIM model on each of the fine-tuning dataset sizes for a range of learning rates, saving the results to disk.

ESIM + Character-Level Component

To rerun inoculation of the ESIM model augmented with a character-level component on the matched and mismatched spelling adversary datasets, we can run the following commands:

./scripts/nli_stress_test/tune_esim_character_level_fine_tune_lr.sh matched_spelling_adversary

./scripts/nli_stress_test/tune_esim_character_level_fine_tune_lr.sh mismatched_spelling_adversary

One again, note that matched_spelling_adversary and mismatched_spelling_adversary are the names of folder containing configs at training_configs/nli_stress_test/esim_character_level/.

Re-running the Adversarial SQuAD Experiments

To reproduce the inoculation results for each of the reading comprehension models (BiDAF and QANet), use the scripts in scripts/adversarial_squad/ . These scripts look for configuration files in training_configs/adversarial_squad/.

BiDAF

To rerun BiDAF inoculation experiments, run:

./scripts/adversarial_squad/tune_bidaf_fine_tune_lr.sh

QANet

To rerun QANet inoculation experiments, run:

./scripts/adversarial_squad/tune_qanet_fine_tune_lr.sh

References

@InProceedings{liu-schwartz-smith:2019:NAACL,
  author    = {Liu, Nelson F.  and  Schwartz, Roy  and  Smith, Noah A.},
  title     = {Inoculation by Fine-Tuning: A Method for Analyzing Challenge Datasets},
  booktitle = {Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies},
  year      = {2019}
}