Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Legal experts disagree with rationale extraction techniques for explaining ECtHR case outcome classification

Mahammad Namazov, Tomáš Koref, Ivan Habernal

TL;DR

This work questions the reliability of model-derived rationales in legal outcome prediction and introduces a standardized comparative framework combining quantitative faithfulness metrics with legal expert and LLM-based qualitative evaluations. By constructing a new ECtHR dataset and applying two model-agnostic rationale-extraction methods (MaRC and ISR), the study demonstrates that model-generated reasons often diverge from legal experts’ grounds and can be unreadable or insufficient. Quantitative results show MaRC achieves higher sufficiency while ISR often boosts comprehensiveness, yet experts find most rationales lacking in real-world plausibility, and LLM-based judgments exhibit limited reliability. The paper highlights the need for closer legal collaboration in designing interpretability techniques and robust, scalable evaluation methods, providing guidelines and data to propel future progress in legal NLP interpretability.

Abstract

Interpretability is critical for applications of large language models in the legal domain which requires trust and transparency. While some studies develop task-specific approaches, other use the classification model's parameters to explain the decisions. However, which technique explains the legal outcome prediction best remains an open question. To address this challenge, we propose a comparative analysis framework for model-agnostic interpretability techniques. Among these, we employ two rationale extraction methods, which justify outcomes with human-interpretable and concise text fragments (i.e., rationales) from the given input text. We conduct comparison by evaluating faithfulness-via normalized sufficiency and comprehensiveness metrics along with plausibility-by asking legal experts to evaluate extracted rationales. We further assess the feasibility of LLM-as-a-Judge using legal expert evaluation results. We show that the model's "reasons" for predicting a violation differ substantially from those of legal experts, despite highly promising quantitative analysis results and reasonable downstream classification performance. The source code of our experiments is publicly available at https://github.com/trusthlt/IntEval.

Legal experts disagree with rationale extraction techniques for explaining ECtHR case outcome classification

TL;DR

This work questions the reliability of model-derived rationales in legal outcome prediction and introduces a standardized comparative framework combining quantitative faithfulness metrics with legal expert and LLM-based qualitative evaluations. By constructing a new ECtHR dataset and applying two model-agnostic rationale-extraction methods (MaRC and ISR), the study demonstrates that model-generated reasons often diverge from legal experts’ grounds and can be unreadable or insufficient. Quantitative results show MaRC achieves higher sufficiency while ISR often boosts comprehensiveness, yet experts find most rationales lacking in real-world plausibility, and LLM-based judgments exhibit limited reliability. The paper highlights the need for closer legal collaboration in designing interpretability techniques and robust, scalable evaluation methods, providing guidelines and data to propel future progress in legal NLP interpretability.

Abstract

Interpretability is critical for applications of large language models in the legal domain which requires trust and transparency. While some studies develop task-specific approaches, other use the classification model's parameters to explain the decisions. However, which technique explains the legal outcome prediction best remains an open question. To address this challenge, we propose a comparative analysis framework for model-agnostic interpretability techniques. Among these, we employ two rationale extraction methods, which justify outcomes with human-interpretable and concise text fragments (i.e., rationales) from the given input text. We conduct comparison by evaluating faithfulness-via normalized sufficiency and comprehensiveness metrics along with plausibility-by asking legal experts to evaluate extracted rationales. We further assess the feasibility of LLM-as-a-Judge using legal expert evaluation results. We show that the model's "reasons" for predicting a violation differ substantially from those of legal experts, despite highly promising quantitative analysis results and reasonable downstream classification performance. The source code of our experiments is publicly available at https://github.com/trusthlt/IntEval.
Paper Structure (66 sections, 7 equations, 11 figures, 11 tables)

This paper contains 66 sections, 7 equations, 11 figures, 11 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Interpretability methods
  4. MaRC
  5. Sufficiency
  6. Comprehensiveness
  7. Compactness
  8. Flexible Instance-Specific Rationale Extraction (ISR)
  9. Downstream legal task
  10. ECtHR dataset
  11. Violation detection in the ECtHR cases
  12. Comparative evaluation
  13. Quantitative analysis
  14. Normalized sufficiency
  15. Normalized comprehensiveness
  16. ...and 51 more sections

Figures (11)

  • Figure 1: The scheme of the qualitative analysis, where expert A extracts rationales along with the techniques, and expert B and LLMs (in LLM-as-a-Judge setting) evaluate all of the extracted rationales.
  • Figure 2: Rationales extracted by ISR from the case of https://hudoc.echr.coe.int/eng?i=001-175482.
  • Figure 3: Extracted rationales by MaRC (a), ISR (b) and expert A (c), from the same span of the case of https://hudoc.echr.coe.int/eng?i=001-228385. As mentioned in the legal analysis part, MaRC extracts longer rationales than ISR, usually. However, it does not always mean that it carries supportive information for the outcome.
  • Figure 4: Example from concatenated rationales by MaRC (a), ISR (b) and expert A (c), from the case of https://hudoc.echr.coe.int/eng?i=001-228385. Expert B assesses global support and sufficiency criteria using such documents.
  • Figure 5: Extracted rationales by MaRC (a), ISR (b) and expert A (c), from the same span of the case of https://hudoc.echr.coe.int/eng?i=001-92582. As mentioned in the legal analysis part, MaRC extracts longer rationales than ISR, usually. However, it does not always mean that it carries supportive information for the outcome.
  • ...and 6 more figures