Legal Experts Disagree With Rationale Extraction Techniques for Explaining ECtHR Case Outcome Classification

1. Problem Statement

The application of Large Language Models (LLMs) and deep learning in the legal domain faces a critical trust deficit due to a lack of interpretability. While models can predict legal outcomes (e.g., whether the European Court of Human Rights (ECtHR) will find a violation of rights), their "black-box" nature makes it difficult to verify if their reasoning aligns with legal principles.

Existing research attempts to explain these predictions using rationale extraction techniques (identifying text fragments that justify a model's output). However, two major gaps exist:

1. Lack of Expert Validation: Most evaluations rely on quantitative metrics (faithfulness) or automated checks, rarely involving legal experts to verify if the extracted rationales are actually plausible or sufficient for legal decision-making.
2. Data Quality Issues: Existing ECtHR datasets used for training and evaluation contain significant "gold-label" errors, such as misclassifying inadmissible cases or cases with implicit violations as "non-violations."

The paper asks: Do legal experts find model-agnostic rationale extraction techniques capable of explaining violation predictions? and Can LLMs substitute costly human expert evaluations?

2. Methodology

A. New ECtHR Dataset Construction

The authors identified critical flaws in previous datasets (Chalkidis et al., 2019a, 2021), where inadmissible cases and cases with implicit violations were incorrectly labeled as non-violations.

• Data Collection: Scraped ~390k cases from the HUDOC database (up to Dec 2025).
• Filtering: Applied strict legal filters to exclude inadmissible cases, cases struck out of the list, and those with ambiguous jurisdiction.
• Result: A new, high-quality dataset of 15,221 positive cases (violations) and 1,573 negative cases (non-violations). A balanced subset was created for experiments.
• Task: Binary classification (Violation vs. Non-violation) using LEGAL-BERT in a hierarchical configuration (HIER-BERT) to handle long documents.

B. Rationale Extraction Techniques

The study compares two model-agnostic interpretability methods:

1. Masked Rationale Creation (MaRC): An optimization-based approach that learns an input mask to preserve the model's prediction using only the most informative tokens, enforcing constraints for sufficiency, comprehensiveness, and compactness.
2. Flexible Instance-Specific Rationale Extraction (ISR): An ensemble-based approach combining various feature-attribution methods (Attention, Integrated Gradients, LIME, DeepLift, etc.) to select the best rationale configuration for each instance.

C. Evaluation Framework

The authors propose a three-pronged evaluation framework:

1. Quantitative Faithfulness Metrics:
   • Normalized Sufficiency (NormSuff): Measures if the rationale alone maintains the model's confidence.
   • Normalized Comprehensiveness (NormComp): Measures if removing the rationale degrades performance.
   • F1 Impact: Changes in F1 scores when using only rationales (Suff-F1) or masking them out (Comp-F1).
2. Qualitative Expert Analysis:
   • Setup: Two legal experts (A and B) evaluated 10 specific ECtHR cases (Articles 6 and 8).
   • Expert A: Extracted "gold" rationales based solely on case facts.
   • Expert B: Evaluated rationales from MaRC, ISR, and Expert A on Global Support (do they support the outcome?) and Global Sufficiency (are they enough to decide the outcome?).
3. LLM-as-a-Judge:
   • Three open-weight LLMs (Llama-3.1-70B, Mixtral-8x7B, Gemma-2-27B) were tested to see if they could replicate Expert B's judgments.
   • Evaluated in both zero-shot and few-shot settings using Cohen's Kappa to measure agreement with human experts.

3. Key Contributions

1. Curated ECtHR Dataset: The release of a new, legally vetted dataset that corrects historical labeling errors regarding inadmissibility and implicit violations, providing a more reliable benchmark for legal NLP.
2. Comparative Framework: A standardized framework for evaluating interpretability that combines quantitative faithfulness metrics with rigorous, global-context legal expert analysis.
3. Empirical Evidence of Divergence: The study provides concrete evidence that high faithfulness scores do not equate to legal plausibility. Models extract text fragments that mathematically support their prediction but are often semantically broken or legally irrelevant to human experts.
4. LLM-as-a-Judge Feasibility Study: A critical assessment of using LLMs to evaluate interpretability in law, finding that while they show some alignment with experts, their reliability is currently insufficient for high-stakes legal evaluation due to inconsistency and hallucination risks.

4. Key Results

Quantitative Results

• Faithfulness: MaRC achieved the highest NormSuff (0.61) and NormComp (0.78) scores, suggesting its rationales are mathematically sufficient to drive the model's prediction.
• Performance Paradox: There was a disconnect between faithfulness and utility. For instance, the Attention mechanism (α) had low faithfulness scores but achieved the best Suff-F1 (0.73), indicating that faithfulness metrics alone are insufficient to judge the quality of rationales for downstream tasks.

Qualitative Expert Results (The Core Finding)

• Expert Rejection: Despite high faithfulness scores, legal experts rejected the vast majority of model-generated rationales.
  • ISR: 9 out of 10 documents failed both support and sufficiency criteria.
  • MaRC: 7 out of 10 documents failed.
• Nature of Failure: Experts described model rationales as "broken snippets," consisting of random nouns without verbs, interrupted clauses, or irrelevant dates (e.g., "non-enforcement (section 74)" or "000.8. on 28").
• Context Loss: Unlike human experts who extract coherent paragraphs or whole arguments, models fragmented the text, often cutting off critical legal context (subjects, verbs, or procedural history), rendering the rationales legally meaningless.

LLM-as-a-Judge Results

• Agreement: LLMs showed "fair" agreement (Kappa $\in [0.2, 0.4]$) with human experts in few-shot settings, particularly Llama and Mixtral.
• Reliability: However, bootstrapped confidence intervals were wide, and LLMs were inconsistent (disagreeing with themselves on the same document). The study concludes that LLMs are not yet reliable enough to replace human experts for evaluating interpretability in legal domains.

5. Significance and Implications

• Trustworthiness in Legal AI: The paper challenges the assumption that high-performing interpretability metrics (like sufficiency/comprehensiveness) guarantee that a model is reasoning correctly. In high-stakes legal domains, a model can be "faithful" to its own internal logic while being "unfaithful" to legal reality.
• Limitations of Current XAI: Current model-agnostic extraction techniques are ill-suited for complex legal documents where reasoning relies on long-range dependencies and coherent narrative structures, not just isolated token importance.
• Future Directions: The findings suggest that future legal NLP systems must move beyond simple rationale extraction toward task-specific, inherently interpretable models or hybrid systems that incorporate legal reasoning structures. Furthermore, the "LLM-as-a-Judge" paradigm requires significant refinement before it can be trusted for evaluating legal reasoning.

In summary, the paper demonstrates a critical gap between computational interpretability (what the model thinks it used) and legal interpretability (what a human lawyer needs to justify a decision), urging caution in deploying current black-box legal AI systems without deeper structural changes.