Can Rationalization Improve Robustness?

This repository contains the code for our NAACL'22 paper Can Rationalization Improve Robustness?. You can also check out our slides here or the NAACL 2022 talk here.

Quick Links

• Overview
• Main Results
• Run Code
  • Requirements
  • Training
  • Attacks
• Bugs or Questions?
• Citation
• Acknowledgment

Overview

We explore the intersection between interpretability and robustness in the popular predict-then-rationalize framework. Extractive rationalization is a way to expose its inner reasoning process by selecting a subset of input text (upper part of the figure). Ideally, the rationalizer would select a set of tokens that "explains" the prediction. A good rationalizer should select non-spurious texts (e.g., stopwords or dataset specific biases) and hence is expected to fence off added attack/spurious text (bottom part of the figure). We examine a wide range of models and datasets and find that the rationalization may be promising in providing robustness benefit to regular full-context models. We also caustion the findings with analysis on cases where the the models are sensitive to the choice of the attack.

Below are some examples of our robustness tests. The blue sentences indicate the query/question, the red sentences indicate the attack (transformed from the query/question), and the yellow highlights indicating the gound truth rationales.

Main Results

Below is the main result table of the original performance (Ori) and the attacked performance (Att). The models are:

• FC = full context model
• VIB = the Variational Information Buttleneck model
• SPECTRA = the Sparse Structured Text Rationalization model
• *-sup = models using the human provided rationales as supervision

	FEVER (Ori / Att)	MultiRC (Ori / Att)	SQuAD (Ori / Att)	Beer (Ori / Att)	Hotel (Ori / Att)
FC	90.7 / 77.9	70.7 / 63.0	87.2 / 59.1	93.8 / 59.5	99.5 / 79.3
VIB	87.8 / 82.6	65.4 / 63.6	77.1 / 56.5	93.8 / 88.0	94.0 / 59.3
SPECTRA	84.0 / 76.5	63.8 / 63.3	65.5 / 45.5	95.4 / 89.7	94.5 / 51.3
FC-sup	91.9 / 77.1	71.5 / 64.0	87.0 / 57.3	-	-
VIB-sup	90.2 / 81.4	68.7 / 63.7	86.5 / 56.5	-	-

Run Code

Requirements

To install the dependencies, please run

pip install -r requirements.txt

Configure rrtl/config.py to point the base path to your repo.

Training

To train, simply run run_{dataset}.sh {model_option}. {model_option} takes one of these values [fc | vib | spectra].

For example, to run the SPECTRA model on FEVER (make sure to set your repo path base_path in the run script):

./scripts/run_fever.sh spectra

Below are the arguments specific to the VIB models:

• --tau: temperature
• --pi: rationale sparsity
• --beta: coefficient of the KL term (strength of the KL regularization)
• --gamma: coefficient for the human rationale supervision loss
• --use-gold-rationale: if set to True, the VIB-sup model will be trained
• --use-neg-rationale: if set to True, the ART model will be trained

Below are the arguments specific to the SPECTRA models:

• --budget: (int) absolute number of tokens or sentences to select as rationales
• --budget_ratio: (float) the ratio of rationales to predict over the number of sentences in the context
• --temperature: temperature term
• --solver_iter: number of iterations to run the LP-SparseMAP solver

Attacks

Run the checkpoint on the attack dataset (FEVER/MultiRC):

First run python -m rr.attacks.aug_and_cache_data --dataset-name fever --aug-method addsent --attack-dir addsent --insert_pos 0 to cache the attacks (see rr/attacks/aug_and_cache_data.py for other options).

python -m rr.eval.run_eval --model-name {MODEL_CHECKPOINT_DIR} \
                           --bottleneck-type {vib | vib_semi | full | full_multitask} \
                           --exp-dir {fever | multirc} \
                           --attack-dir addsent_pos0 \
                           --output-to-tmp                           

Here vib_semi corresponds to VIB-sup, full to FC, and full_multitask to FC-sup in the paper.

Run the checkpoint on the attack dataset (SQuAD):

python -m rrtl.run_squad_eval --dataset-name squad-addonesent --load-path /path/to/checkpoint.pt

Run the checkpoint on the attack dataset (Beer/Hotel)

To cache the attack, first run python -m rrtl.attacks.sentiment_attack to cache the attacks. Second, run:

python -m rrtl.analysis.eraser.run_token_level_analysis --dataset-name {beer | hotel} \
                                                        --eval-mode stats \
                                                        --load-path /path/to/checkpoint.pt

Hyperparameters

Dataset	Epoch	Learning Rate	Batch Size
FEVER	10	5e-5	32
MultiRC	10	5e-5	32
SQuAD	3	1e-5	32
Beer	20	5e-5	64
Hotel	20	5e-5	64

Model variants such as *-sup models or ART training may have slightly different hyperparameters. Please refer to the training script for the detailed setting.

Bugs or Questions

If you have any questions about our paper, please reach out to Howard Chen (howardchen@cs.princeton.edu). Also, feel free to open an issue for bug reporting.

Citation

@inproceedings{chen2022RationaleRobustness,
   title={Can Rationalization Improve Robustness?},
   author={Chen, Howard and He, Jacqueline and Narasimhan, Karthik and Chen, Danqi},
   booktitle={North American Chapter of the Association for Computational Linguistics (NAACL)},
   year={2022}
}

Acknowledgment

We thank the authors of An Information Bottleneck Approach for Controlling Conciseness in Rationale Extraction for their implementations that we adapt in our VIB models.