Abstract
Distinguishing between principals and accessories in multi-defendant criminal cases is a critical task in legal judgment prediction. Traditional NLP methods often rely on statistical features like name frequency, which fail to capture the underlying narrative power dynamics. This paper proposes a novel approach using a learning-to-rank framework that combines behavioral semantic analysis with statistical linguistic features. By treating legal fact descriptions as narratives, the model identifies primary actors through Match-LSTM architectures and syntactic complexity measures. Evaluated on a dataset of over 15,000 Chinese criminal cases, the proposed method significantly outperforms frequency-based baselines, achieving an F1-score of 79.77%.
Introduction
In the field of Legal AI, most existing research focuses on simple cases involving a single defendant. However, real-world crime is often collaborative, requiring legal systems to distinguish between the principal (the mastermind or major player) and the accessory (someone playing a secondary or auxiliary role). Identifying these roles automatically is challenging because it requires narrative comprehension rather than simple keyword spotting. A mastermind might be mentioned only once as "ordering" a crime, while a henchman might appear in every sentence describing the physical execution. Traditional methods relying on frequency or dispersion often fail to capture these power dynamics.
Methodology
The authors treat fact descriptions as narratives and defendants as roles within those stories, employing a learning-to-rank framework (specifically RankNet) to score defendants based on their "importance" or impact on the case. The model utilizes two primary feature families:
- Behavioral Semantic Features ($f_{semantic}$): The model extracts activity fragments (sentences containing a defendant's name and at least one verb). It uses a
match-lstmarchitecture to measure how well an individual's actions align with the core narrative of the crime, treating the full description as a "hypothesis" and the fragments as "premises." - Statistical Characteristics ($f_{statistical}$): This includes syntactic complexity (sentence length, verb phrase depth), cooperation mode (identifying hierarchy through dependency parsing, e.g., "A instructed B"), and order/frequency of appearance.
Dataset
The study utilized a real-world dataset of 15,312 criminal cases involving more than two defendants, sourced from Chinese legal disclosure websites. From these, 41,342 paired samples were generated to train the ranking model. Approximately 67% of the cases involved a single principal.
Results and Discussion
The proposed model significantly outperformed traditional baselines across all metrics. As shown in Table 1, the full model achieved a Macro F1-score of 79.77%, compared to 65.10% for a simple frequency-based approach.
Ablation studies confirmed the superiority of semantic features. As detailed in Table 2, the semantic feature family ($f_{semantic}$) achieved a higher F1-score (75.93) on its own than the entire suite of statistical features combined (73.97).
Key insights from the results suggest that what a person does is more indicative of their legal role than how often they are mentioned. The "Cooperation Mode" feature proved particularly effective at capturing social hierarchies by identifying linguistic cues like "instructed," which denote master-slave relationships even when the principal has a low appearance frequency.
Conclusion
By framing legal fact descriptions as narratives, this research bridges the gap between general NLP story-understanding and specialized legal reasoning. Future work will focus on reducing manual feature engineering in favor of a completely semantic-based approach and exploring more nuanced hierarchies in organized crime syndicates where multiple tiers of leadership may exist.
Summary updated