Robustness Perturbation Eval

robustness-perturbation-eval

Evaluating the Robustness of Neural Language Models to Input Perturbations — Moradi et al. (2021) (arXiv:2108.12237, 2021)

What this evaluates

Evaluates the robustness of neural language models to non-adversarial character- and word-level input perturbations (e.g., typos, deletions, synonyms) while preserving semantic meaning.

Datasets

• TC, SA, NER, SS, QA (unspecified downstream datasets) — total ?; splits: test (-1); repo https://github.com/mmoradi-iut/NLP-perturbation

Metrics

• accuracy (primary) — range: percent
  • Percentage of correctly predicted labels on perturbed test samples compared to ground truth labels.
• absolute performance decrease — range: percent
  • Difference between baseline accuracy on unperturbed test sets and accuracy on perturbed samples.

Input / output format

Input: Text samples from downstream NLP tasks with applied character-level or word-level perturbations.

Output: Predicted labels or scores for each perturbed input instance.

Scoring recipe

accuracy = sum(1 for pred, gold in zip(predictions, gold_labels) if pred == gold) / len(gold_labels) * 100
drop = baseline_accuracy - accuracy
return accuracy, drop

Common pitfalls

• Three word-level perturbations (Deletion, Negation, Replacement with Synonyms) require manual curation to filter meaningless samples or adjust labels, as they may change text meaning.
• Performance is highly sensitive to the Perturbation Per Sample (PPS) hyperparameter, with results reported for PPS=1 in main tables but tested up to PPS=4.
• Models are evaluated in a black-box manner without fine-tuning on perturbed data; results reflect zero-shot robustness on test sets.

Evidence (verbatim from paper)

Table 4: Performance of the language models on the test sets and word-level perturbed samples of the downstream tasks. For every task and every perturbation method, the highest score is shown in bold face.

Citation

@misc{moradi2021evaluating,
  title={Evaluating the Robustness of Neural Language Models to Input Perturbations},
  author={Moradi et al. (2021)},
  year={2021},
  note={arXiv:2108.12237}
}

• arXiv: 2108.12237