[Paper Review] Explaining Legal Concepts with Augmented Large Language Models (GPT-4)

Jaromír Šavelka, Kevin D. Ashley+3 more|arXiv (Cornell University)|Jun 15, 2023

Artificial Intelligence in LawSocial Sciences19 citations

TL;DR

This paper compares direct GPT-4 explanations of statutory terms with augmented GPT-4 explanations that incorporate retrieved case-law sentences, finding augmentation improves factuality and overall quality while reducing hallucinations.

ABSTRACT

Interpreting the meaning of legal open-textured terms is a key task of legal professionals. An important source for this interpretation is how the term was applied in previous court cases. In this paper, we evaluate the performance of GPT-4 in generating factually accurate, clear and relevant explanations of terms in legislation. We compare the performance of a baseline setup, where GPT-4 is directly asked to explain a legal term, to an augmented approach, where a legal information retrieval module is used to provide relevant context to the model, in the form of sentences from case law. We found that the direct application of GPT-4 yields explanations that appear to be of very high quality on their surface. However, detailed analysis uncovered limitations in terms of the factual accuracy of the explanations. Further, we found that the augmentation leads to improved quality, and appears to eliminate the issue of hallucination, where models invent incorrect statements. These findings open the door to the building of systems that can autonomously retrieve relevant sentences from case law and condense them into a useful explanation for legal scholars, educators or practicing lawyers alike.

Motivation & Objective

• Assess GPT-4's ability to explain open-textured legal terms from statutory provisions to legal professionals.
• Evaluate limitations of direct GPT-4 explanations in terms of factual accuracy and reliance on training data.
• Test whether augmenting GPT-4 with legal information retrieval (case-law sentences) reduces hallucinations and improves explanation quality.
• Demonstrate a pipeline that retrieves explanatory sentences from case law and condenses them into explanations.
• Provide benchmarking on whether augmented GPT-4 outperforms baseline GPT-4 in professional legal settings.

Proposed method

• Baseline: prompt GPT-4 directly to explain a term from a source provision without external context.
• Augmented: retrieve high-value explanatory sentences from case law referencing the term and inject them into GPT-4's prompt.
• Use a statutory interpretation data set with 42 terms and 1,853 high-value sentences for augmentation.
• Generate two explanations per term: a short (1 sentence) and a long (10 sentences) version.
• Have two legal scholars annotate pairwise explanations across five quality dimensions.
• Compare factuality, clarity, relevance, information richness, and on-pointedness between baseline and augmented outputs.

Figure 2: System Architectures Diagrams. The top part shows the baseline directly applying the LLM. The bottom part describes the components of the augmented architecture that relying on the information retrieval component.

Experimental results

Research questions

• RQ1 What are the limitations of generating explanations directly with GPT-4 for statutory interpretation?
• RQ2 Does augmenting GPT-4 with relevant case-law sentences improve explanation quality across factuality, clarity, relevance, information richness, and on-pointedness?

Key findings

• Augmented GPT-4 explanations are generally preferred over baseline for short and long explanations across annotators.
• Augmented explanations eliminate non-existent-citation and misrepresentation issues observed in the baseline’s factuality assessments.
• Augmented explanations improve clarity, relevance, information richness, and on-pointedness compared to baseline.
• Baseline explanations exhibit hallucinations and citation inaccuracies, though many citations are real; content often misrepresents cases.
• Augmentation does not fully remove issues when the information retrieval component provides irrelevant or misleading case-law content; high-quality IR is crucial.
• Overall, augmented LLMs show promise for auto-generating accurate summaries of statutory term interpretations for legal education and practice.

Figure 3: Short Explanation Preferences. Red corresponds to the preferences for the explanations generated by the baseline system while green indicates preferences for the explanations coming from the augmented LLM. The yellow/orange informs about the number of instances where no preference was indi