A Systematic Comparison between Extractive Self-Explanations and Human Rationales in Text Classification

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing, Data Science & Analytics · Depth: Expert, medium

Summary

A systematic comparison by Stephanie Brandl and Oliver Eberle, presented at TrustNLP 2026 in San Diego, California (pages 563–583), evaluates the quality of extractive self-explanations generated by instruction-tuned Large Language Models (LLMs) in text classification. The study assesses these input rationales for human plausibility across three tasks: sentiment classification (including Danish and Italian translations), forced labor detection, and claim verification, utilizing a newly collected human rationale dataset for Climate-Fever. Researchers analyzed four open-weight LLMs, comparing their self-explanations against human annotations and post-hoc attribution methods. Key findings indicate that alignment between self-explanations and human rationales varies significantly with text length and task complexity. While self-explanations provide faithful token-level rationales, post-hoc attribution methods often highlight structural and formatting tokens, revealing distinct explanation strategies.

Key takeaway

For NLP Engineers developing explainable AI, consider instruction-tuned LLM self-explanations as a viable method for generating faithful token-level rationales. Be aware that their alignment with human rationales is sensitive to text length and task complexity, requiring careful evaluation for your specific application. Prioritize self-explanations when you need content-focused rationales, as post-hoc attribution methods tend to emphasize structural elements instead.

Key insights

LLM self-explanations offer faithful rationales, differing from post-hoc methods, but their human alignment varies with task complexity.

Principles

• Self-explanation alignment with humans varies by text length and task complexity.
• LLM self-explanations provide faithful token-level rationales.
• Post-hoc attribution methods emphasize structural tokens.

Method

Compared LLM self-explanations (input rationales) against human annotations and post-hoc attribution for plausibility and faithfulness across three text classification tasks, including a new human rationale dataset.

In practice

• Evaluate LLM self-explanations for human plausibility.
• Consider task complexity when assessing explanation alignment.
• Distinguish self-explanations from post-hoc attribution.

Topics

• Extractive Self-Explanations
• Human Rationales
• Text Classification
• LLM Interpretability
• Post-hoc Attribution
• Explainable AI

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Faithfulness Serum: Mitigating the Faithfulness Gap in Textual Explanations of LLM Decisions via Attribution Guidance — LLM-generated explanations often lack epistemic faithfulness, but a new method improves it via attention-level guidance.
• Reasoning as Pattern Matching: Shared Mechanisms in Human and LLM Everyday Reasoning — Human and LLM everyday reasoning aligns with content-sensitive pattern-matching, challenging abstract world model assumptions.
• Self-Evaluation Is Already There: Eliciting Latent Judge Calibration in Base LLMs with Minimal Data — Base LLMs inherently predict external judge scores, a latent ability efficiently elicited by Self-Evaluation Elicitation (SEE) with minimal data.
• Introspective Coupling: Self-Explanation Training Tracks Behavioral Change Despite Fixed Supervision — Language models can develop faithful self-explanations using fixed, potentially outdated, counterfactual training data.
• Applied Explainability for Large Language Models: A Comparative Study — Gradient-based explainability provides stable LLM insights, while attention is efficient but less aligned with predictions.
• The LLM Fallacy: Misattribution in AI-Assisted Cognitive Workflows — The LLM fallacy describes users misattributing AI-assisted output to their own competence, distorting self-perception.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.