mlsanityData to 7 Oct 2026

Investigating the Benefits of Free-Form Rationales

arXiv 2206.11083 ↗First posted 25 May 2022cs.CL cs.AIcited by 13

Not on the map. Why?

Abstract (arXiv)

Free-form rationales aim to aid model interpretability by supplying the background knowledge that can help understand model decisions. Crowdsourced rationales are provided for commonsense QA instances in popular datasets such as CoS-E and ECQA, but their utility remains under-investigated. We present human studies which show that ECQA rationales indeed provide additional background information to understand a decision, while over 88% of CoS-E rationales do not. Inspired by this finding, we ask: can the additional context provided by free-form rationales benefit models, similar to human users? We investigate the utility of rationales as an additional source of supervision, by varying the quantity and quality of rationales during training. After controlling for instances where rationales leak the correct answer while not providing additional background knowledge, we find that incorporating only 5% of rationales during training can boost model performance by 47.22% for CoS-E and 57.14% for ECQA during inference. Moreover, we also show that rationale quality matters: compared to crowdsourced rationales, T5-generated rationales provide not only weaker supervision to models, but are also not helpful for humans in aiding model interpretability.

Key facts. Cited by 13 papers in the map (data through 7 Oct 2026). Builds on 39 of its 52 references that are in the map. First posted on arXiv 25 May 2022.

Builds on

39 of 52 references.

References · 39

  1. 2019Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformercited by 14,491
  2. 2020Shortcut Learning in Deep Neural Networkscited by 1,928
  3. 2018CommonsenseQA: A Question Answering Challenge Targeting Commonsense Knowledgecited by 1,793
  4. 2020Towards Faithfully Interpretable NLP Systems: How should we define and evaluate faithfulness?cited by 590
  5. 2019ERASER: A Benchmark to Evaluate Rationalized NLP Modelscited by 523
  6. 2018Towards Robust Interpretability with Self-Explaining Neural Networkscited by 505
  7. 2018e-SNLI: Natural Language Inference with Natural Language Explanationscited by 468
  8. 2019Social Bias Frames: Reasoning about Social and Power Implications of Languagecited by 377
  9. 2019Explain Yourself! Leveraging Language Models for Commonsense Reasoningcited by 367
  10. 2016Rationalizing Neural Predictionscited by 281
  11. 2018Manipulating and Measuring Model Interpretabilitycited by 256
  12. 2018Textual Explanations for Self-Driving Vehiclescited by 226
27 more
  1. 2020Evaluating Explainable AI: Which Algorithmic Explanations Help Users Predict Model Behavior?cited by 189
  2. 2020Measuring Association Between Labels and Free-Text Rationalescited by 159
  3. 2016Generating Visual Explanationscited by 156
  4. 2020WT5?! Training Text-to-Text Models to Explain their Predictionscited by 148
  5. 2021Teach Me to Explain: A Review of Datasets for Explainable Natural Language Processingcited by 126
  6. 2021Reframing Human-AI Collaboration for Generating Free-Text Explanationscited by 125
  7. 2020NILE : Natural Language Inference with Faithful Natural Language Explanationscited by 96
  8. 2020Leakage-Adjusted Simulatability: Can Models Generate Non-Trivial Explanations of Their Behavior in Natural Language?cited by 95
  9. 2019QuaRTz: An Open-Domain Dataset of Qualitative Relationship Questionscited by 89
  10. 2021e-ViL: A Dataset and Benchmark for Natural Language Explanations in Vision-Language Taskscited by 84
  11. 2021Few-Shot Self-Rationalization with Natural Language Promptscited by 83
  12. 2020Aligning Faithful Interpretations with their Social Attributioncited by 79
  13. 2021When Can Models Learn From Explanations? A Formal Framework for Understanding the Roles of Explanation Datacited by 75
  14. 2021Contrastive Explanations for Model Interpretabilitycited by 75
  15. 2019Summit: Scaling Deep Learning Interpretability by Visualizing Activation and Attribution Summarizationscited by 72
  16. 2021Prompting Contrastive Explanations for Commonsense Reasoning Taskscited by 45
  17. 2021SelfExplain: A Self-Explaining Architecture for Neural Text Classifierscited by 44
  18. 2021A Study of Automatic Metrics for the Evaluation of Natural Language Explanationscited by 34
  19. 2019Do Human Rationales Improve Machine Explanations?cited by 33
  20. 2021On the Diversity and Limits of Human Explanationscited by 30
  21. 2019Align, Mask and Select: A Simple Method for Incorporating Commonsense Knowledge into Language Representation Modelscited by 28
  22. 2020Learning to Rationalize for Nonmonotonic Reasoning with Distant Supervisioncited by 27
  23. 2021What to Learn, and How: Toward Effective Learning from Rationalescited by 25
  24. 2020LIREx: Augmenting Language Inference with Relevant Explanationcited by 24
  25. 2020FiD-Ex: Improving Sequence-to-Sequence Models for Extractive Rationale Generationcited by 21
  26. 2020Why do you think that? Exploring Faithful Sentence-Level Rationales Without Supervisioncited by 15
  27. 2021Exploring Distantly-Labeled Rationales in Neural Network Modelscited by 8

Built on by

Cited by year

Citing papers per year: 2022: 3; 2023: 6; 2024: 3; 2025: 1.

13 papers.

Sort:

By area:

  1. 2023SCOTT: Self-Consistent Chain-of-Thought Distillationcited by 103
  2. 2022PINTO: Faithful Language Reasoning Using Prompt-Generated Rationalescited by 56
  3. 2023HARE: Explainable Hate Speech Detection with Step-by-Step Reasoningcited by 30
  4. 2024On Behalf of the Stakeholders: Trends in NLP Model Interpretability in the Era of LLMscited by 24
  5. 2022REV: Information-Theoretic Evaluation of Free-Text Rationalescited by 22
  6. 2024Evaluating Human Alignment and Model Faithfulness of LLM Rationalecited by 12
  7. 2023Beyond Labels: Empowering Human Annotators with Natural Language Explanations through a Novel Active-Learning Architecturecited by 10
  8. 2023Modeling subjectivity (by Mimicking Annotator Annotation) in toxic comment identification across diverse communitiescited by 6
  9. 2025QE4PE: Word-level Quality Estimation for Human Post-Editingcited by 6
  10. 2023Explanation-aware Soft Ensemble Empowers Large Language Model In-context Learningcited by 4
  11. 2023Identifying Self-Disclosures of Use, Misuse and Addiction in Community-based Social Media Postscited by 2
  12. 2024Okay, Let's Do This! Modeling Event Coreference with Generated Rationales and Knowledge Distillationcited by 2
1 more
  1. 2022KNIFE: Distilling Reasoning Knowledge From Free-Text Rationalescited by 1