Event Grounded Criminal Court View Generation with Cooperative (Large) Language Models

TL;DR

The paper tackles criminal court view generation by grounding summaries in fine-grained legal events extracted from case facts. It proposes Event Grounded Generation (EGG), a two-phase pipeline combining a LLM-based event extractor (trained via instruction-tuning on CJRC with LoRA) and a PLM-based court view generator that merges events with facts; a practical variant, EGG_free, uses contrastive learning to enable event-free inference. Empirical results on CJRC/CJO show that EGG improves generation quality over strong baselines and that EGG_free achieves comparable effectiveness with substantially faster inference, making it workable for resource-constrained settings. The work advances legal AI by demonstrating how cooperative LLMs and contrastive learning can yield more accurate, efficient court views while highlighting ethical considerations and future work on richer event-relationship modeling.

Abstract

With the development of legal intelligence, Criminal Court View Generation has attracted much attention as a crucial task of legal intelligence, which aims to generate concise and coherent texts that summarize case facts and provide explanations for verdicts. Existing researches explore the key information in case facts to yield the court views. Most of them employ a coarse-grained approach that partitions the facts into broad segments (e.g., verdict-related sentences) to make predictions. However, this approach fails to capture the complex details present in the case facts, such as various criminal elements and legal events. To this end, in this paper, we propose an Event Grounded Generation (EGG) method for criminal court view generation with cooperative (Large) Language Models, which introduces the fine-grained event information into the generation. Specifically, we first design a LLMs-based extraction method that can extract events in case facts without massive annotated events. Then, we incorporate the extracted events into court view generation by merging case facts and events. Besides, considering the computational burden posed by the use of LLMs in the extraction phase of EGG, we propose a LLMs-free EGG method that can eliminate the requirement for event extraction using LLMs in the inference phase. Extensive experimental results on a real-world dataset clearly validate the effectiveness of our proposed method.

Figures (6)

• Figure 1: Schematic of C3VG and our method EGG presented in this paper.
• Figure 2: (a). An example from the CJRC dataset, including two questions. (b). An example from the processed instruction dataset based on the Question1 in (a). (c). An example from the processed instruction dataset based on the Question2 in (a).
• Figure 3: The process of event extraction. Among them, the input $x$ is a legal document in the "crime of Affray" case type, and the questions $q^{\prime}$ is annotated for the "crime of Affray" case type (not only for the input $x$).
• Figure 4: Architecture of $\textmd{\textbf{EGG}}_{\textbf{free}}$.
• Figure 5: Model performance by the length of court views. Among them, C3VG is implemented with BART.