Faithfulness Tests for Natural Language Explanations

• Existing explanations can be unfaithful, with saliency maps sometimes independent of training data or model parameters and other explanation methods also questioned.Source: §1 · p. 1
• NLEs are more expressive and human-readable than saliency maps, but their free text can include words absent from the input.Source: §1 · p. 1
• Because NLEs lack known ground-truth reasons and direct input-to-explanation mappings, existing faithfulness tests for saliency maps cannot be directly applied.Source: §1 · p. 1
• The first test inserts reasons that change model predictions and checks whether the generated NLE reflects those counterfactual interventions.Source: §1 · p. 1
• The second test reconstructs inputs from reasons stated in generated NLEs and checks whether the reconstructed inputs produce the corresponding prediction.Source: §1 · p. 1

2.1 The Counterfactual Test: Are NLE models faithful to reasons for counterfactual pre-

• The counterfactual test inserts a contiguous word set W that changes the model’s prediction, then checks whether the counterfactual NLE mentions at least one inserted word.The editor generates candidate insertions for a target counterfactual label, while evaluation measures the percentage of instances satisfying this mismatch condition.Source: §2.1 · p. 2
• The test assumes inserted edits should be reflected in NLEs, although confounding changes in the model’s attention may also explain prediction changes and are left for future work.The authors hypothesize such confounding is rare, especially for insertions rather than deletions.Source: §2.1 · p. 2
• The counterfactual metric is not a complete measure of overall NLE faithfulness because it covers only counterfactual reasons and depends on the editor’s ability to find them.Failure to find many counterfactual reasons can make NLEs appear more faithful.Source: §2.1 · pp. 2–3
• Table 2 reports counterfactual-test results for random, counterfactual-editor, and combined setups, measuring prediction-changing insert success and unfaithfulness among changed predictions and across the full test set.Source: Table 2 · p. 3
• Input reconstruction tests whether reasons extracted from an NLE form a new input that receives the same prediction as the original input.Because NLEs lack a direct token-to-reason mapping, reconstruction functions are task-dependent; they were built for e-SNLI and ComVE but not CoS-E.Source: §2.1 · p. 3

3 Experiments

• The study compares four setups formed by multi-task versus single-task training and reasoning versus rationalizing generation.The setups are MT-Re, MT-Ra, ST-Re, and ST-Ra; rationalizing models condition NLE generation on the predicted label.Source: §3 · pp. 3–4
• The experiments use e-SNLI, CoS-E, and ComVE, covering textual entailment, commonsense question answering, and commonsense reasoning.A T5-base model is used for both the NLE models and the counterfactual editor.Source: §3 · p. 4

3.1 Results

• The counterfactual editor generally finds more prediction-changing insertions than the random baseline, producing higher detected unfaithfulness.The random baseline’s words are conjectured to be rarer in the datasets than the editor’s learned insertions.Source: §3.1 · p. 4
• 37.04%–59.04% of instances were unfaithful to counterfactual interventions across datasets and model setups.The counterfactual editor found prediction-changing insertions for up to 56.70% of e-SNLI instances, while the combined interventions reached 59.04%.Source: §3.1 · pp. 4–5
• Input reconstruction found up to 14% unfaithful NLEs for e-SNLI and up to 40% for ComVE, despite reconstructing only some e-SNLI inputs.Inputs were reconstructed for up to 4487 of 10K e-SNLI instances and all ComVE test instances.Source: §3.1 · p. 5
• All four model types show similar faithfulness results without a consistent ranking across datasets and tests.Reasoning models tend to be less faithful than rationalizing models in most cases, but the pattern is not universal.Source: §3.1 · p. 5

4 Related Work

• Saliency-map diagnostics measure prediction changes after removing or manipulating important input regions, relying on token-to-score correspondence.Existing approaches include comprehensiveness, sufficiency, ROAR, and adversarial input manipulations.Source: §4 · p. 5
• These saliency-map tests cannot be directly applied to NLEs because NLEs may contain text absent from the input.NLEs lack the one-to-one correspondence between explanation elements and input regions assumed by saliency diagnostics.Source: §4 · p. 5
• Existing NLE studies often evaluate plausibility or whether an observer can simulate model outputs from generated explanations.Observer-based evaluation may use agents or humans, whereas plausibility alone does not test faithfulness.Source: §4 · p. 5
• The proposed tests complement prior sanity checks by providing quantitative metrics for NLE faithfulness.Wiegreffe et al. (2021) use pass/fail tests based on robustness and shared important tokens; the proposed tests assess different properties.Source: §4 · p. 5

5 Summary and Outlook

• The paper introduces two tests to evaluate the faithfulness of NLE models.Source: §5 · p. 5
• All four high-level NLE model setups are prone to generate unfaithful NLEs, reinforcing the need for evidence of faithfulness.Source: §5 · p. 5
• The tests can be used to evaluate emerging NLE models and inspire complementary faithfulness tests.Source: §5 · p. 5

• Perfect performance on these tests may still coexist with unfaithful NLEs because the tests are not comprehensive.Source: Limitations · p. 6
• The counterfactual test may miss faithful explanations that reflect other relevant prediction factors rather than the inserted counterfactual reasons.Source: Limitations · p. 6
• Random and editor-generated insertions can lack semantic coherence, motivating future methods that generate coherent candidates while exposing unfaithful NLEs.Source: Limitations · p. 6
• The reconstruction test relies on task-dependent hand-crafted heuristics, and no rule-based reconstruction was possible for CoS-E.Source: Limitations · p. 6
• Learned reconstruction functions trained with a small number of annotations are proposed as a route toward applying the second test across datasets.Source: Limitations · p. 6

Random insertion baseline

• Table 4 shows random-insertion examples across three datasets, with changed input parts and interventions marked separately to identify mismatched NLEs or predictions.Source: Table 4 · p. 9

Counterfactual editor

• Table 5 shows counterfactual-editor examples across three datasets, using marked edits and mismatched outcomes to identify unfaithful NLEs.Source: Table 5 · p. 9

Model performance

• Table 7 reports model performance using Acc for task prediction and BLEU for generated explanations against gold human explanations.Source: Table 7 · p. 10