Explanation Extraction from Hierarchical Classification Frameworks for Long Legal Documents

Nishchal Prasad, Taoufiq Dkaki, Mohand BoughanemNAACL 2024 · DOI

TL;DR

This paper introduces Ob-HEx, a novel post-hoc explanation method for hierarchical classification models applied to long legal documents. Ob-HEx addresses the "black box" problem of these models by identifying key sentences that contribute to predictions, making legal AI systems more transparent and reliable.

What Problem Does This Paper Address?

Hierarchical classification frameworks are powerful tools for processing long legal documents, but their "black box" nature makes it difficult to understand why they make a particular prediction. This lack of transparency is a significant hurdle, especially in high-stakes fields like law, where trust and accountability are paramount. Existing explanation methods often struggle with the complexity of hierarchical structures and the sheer length of legal texts. The authors aim to bridge this gap by developing a method that can extract meaningful explanations from these complex models without requiring additional training data.

Methods

The core of the paper is the Ob-HEx (Occlusion-based Hierarchical Explanation-extractor) algorithm. Ob-HEx is a post-hoc explanation technique that works by systematically perturbing the input data and observing the impact on the model's predictions. Specifically, it uses an occlusion sensitivity impact function to measure how much a prediction changes when parts of the input (sentences or chunks) are masked.

Key aspects of the method include:

  • Hierarchical Scoring: Ob-HEx applies this occlusion sensitivity analysis across the different levels of the hierarchical model, from the top-level classification down to individual sentences.
  • Normalized Weighted Occlusion-Sensitivity Score: A novel scoring mechanism is introduced to rank the importance of input fragments. This score considers the impact of occluding a part of the input and also weighs it by the importance of that part in the higher levels of the hierarchy. This ensures that explanations are aligned across all levels.
  • Extractive Approach: Ob-HEx extracts actual sentences from the original document to serve as explanations, making them directly interpretable by legal professionals.

Datasets

The primary dataset used for evaluation is ILDCExpert from Malik et al. (2021). This dataset consists of unstructured English case transcripts from the Supreme Court of India, with decisions removed and gold explanations provided by legal experts. The gold explanations are ranked from 1 to 10, indicating their relevance to the judgment.

Key Insights and Ideas

  • Bridging the Black Box Gap: Ob-HEx successfully provides interpretable explanations for complex hierarchical models, a critical step towards building trust in AI for legal applications.
  • Hierarchical Sensitivity is Key: The method's strength lies in its ability to leverage the hierarchical structure of the models, scoring explanations not just based on their impact at the sentence level but also considering their influence on higher-level predictions.
  • Extractive Explanations are Practical: By extracting actual sentences, Ob-HEx provides explanations that are directly understandable and usable by legal experts, rather than abstract feature importance scores.
  • Model-Specific Focus: The analysis of the "S scores" reveals that even when trained on full documents, hierarchical models tend to focus on specific parts of the text (e.g., the end of the document), and Ob-HEx effectively visualizes these focus points.

Key Results

The paper demonstrates that Ob-HEx achieves competitive or improved performance compared to a baseline explanation method (CJPE) across various evaluation metrics (ROUGE-1, ROUGE-2, ROUGE-L, BERTScore, and Jaccard similarity).

  • Improved Explanation Quality: Ob-HEx consistently produced explanations that were more semantically similar to expert-provided explanations, even when the explanations were constrained to shorter lengths.
  • Shorter, Better-Ranked Explanations: For the same base hierarchical model, Ob-HEx provided shorter and better-ranked sentences as explanations than the CJPE baseline.
  • New Benchmarks: The authors report achieving new benchmarks on the ILDCExpert dataset with their Ob-HEx method. For instance, in "long explanations" (ranks 1-3), Ob-HEx showed an average metric points gain of up to 3.93 over CJPE for the XLNet+BiGRU model.

Limitations and Future Work

  • Model Dependency: Explanations generated by Ob-HEx are inherently dependent on the trained hierarchical model. If the model's weights are updated, the explanations may change.
  • No Ground Truth for Explanations: Since expert explanations can vary, there isn't a single "ground truth" for explanations, making direct comparison challenging. The authors addressed this by using explanations from multiple experts.
  • No Human Evaluation: The paper acknowledges the lack of human evaluation to definitively prove the meaningfulness of sentences extracted by Ob-HEx that were missed by CJPE. This is suggested as a direction for future research.
  • Future Work: The authors propose using Ob-HEx as a selective re-training strategy for hierarchical models to analyze its effects on predictions. They also aim to implement Ob-HEx in other domains that utilize hierarchical frameworks.

Summary updated

BibTeX
@inproceedings{prasad2024explanation,
  author = {Nishchal Prasad and Taoufiq Dkaki and Mohand Boughanem},
  title = {{Explanation Extraction from Hierarchical Classification Frameworks for Long Legal Documents}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2024},
  doi = {10.18653/v1/2024.findings-naacl.76},
}
Figure