zotown/JUREX-4E

JUREX

Source code and data for JUREX-4E: Juridical Expert-Annotated Four-Element Knowledge Base for Legal Reasoning

Code: https://github.com/THUlawtech/JUREX

Overview

• Dataset Structure
• Annotation
• Experiment
  • Similar Charge Distinction
  • Legal Case Retrieval
• Requirements
• License

Dataset Structure

JUREX-4E is the first part of our curated expert knowledge base(mind map structure), focusing on the four elements of criminal charges.

data
- law  # legal texts
  - crime2article.json  # mapping from crime to article
  - criminal_law.json  # criminal articles
- jurex4e.json  # The JSON-formatted four elements converted from a mind map structure. The "title" represents the content of each hierarchical level, and "topic" represents the content of the next level.
- flattened_jurex4e.json  # The JSON-formatted four elements obtained by simply concatenating all levels.

An example:

the original mind map: jurex4e.json:

    {
        "投放危险物质罪": {
            "犯罪客体": [
                {
                    "title": "公共安全，即不特定多数人的生命、健康和重大公私财产安全"
                }
            ],
            "客观方面": [
                {
                    "title": "行为人实施了投放毒害性、放射性、传染病病原体等物质，危害公共安全的行为。",
                    "topics": [
                        {
                            "title": "假的定投放虚假物质罪（属于扰乱公共秩序罪）。"
                        },
                        {
                            "title": "危险物质分类【T12/45",
                            "topics": [
                                {
                                    "title": "(1)“投放毒害性、放射性、传染病病原体等物质”，是指向公共饮用的水源以及出售的食品、饮料或者牲畜、禽类的饮水池、饲料等物品中投放毒害性、放射性、传染病病原体等物质的行为。"
                                },
                                {
                                    "title": "(2)“毒害性物质”，是指能够对人体或者动物产生毒害作用的有毒物质，包括化学性、生物性、微生物类有毒物质。"
                                },
                                {
                                    "title": "(3)“放射性物质”，是指含铀、镭、钻等放射性元素，可能对人体、动物或者环境产生严重辐射危害的物质，包括能够产生裂变反应或者聚合反应的核材料。"
                                },
                                {
                                    "title": "(4)“传染病病原体”，是指能够通过在人体或者动物体内适当的环境中繁殖，从而使人体或者动物感染传染病，甚至造成传染病扩散的细菌、霉菌、毒种、病毒等。"
                                }
                            ]
                        }
                    ]
                }
            ],
            "犯罪主体": [
                {
                    "title": "一般主体，包括已满14周岁不满16周岁的未成年人"
                }
            ],
            "主观方面": [
                {
                    "title": "故意，行为人出于何种动机不影响本罪成立"
                }
            ]
        }
    },

flattened_jurex4e.json:

    "投放危险物质罪": {
        "犯罪客体": "# 公共安全，即不特定多数人的生命、健康和重大公私财产安全",
        "客观方面": "# 行为人实施了投放毒害性、放射性、传染病病原体等物质，危害公共安全的行为。\n## 假的定投放虚假物质罪（属于扰乱公共秩序罪）。\n## 危险物质分类【T12/45\n### (1)“投放毒害性、放射性、传染病病原体等物质”，是指向公共饮用的水源以及出售的食品、饮料或者牲畜、禽类的饮水池、饲料等物品中投放毒害性、放射性、传染病病原体等物质的行为。\n### (2)“毒害性物质”，是指能够对人体或者动物产生毒害作用的有毒物质，包括化学性、生物性、微生物类有毒物质。\n### (3)“放射性物质”，是指含铀、镭、钻等放射性元素，可能对人体、动物或者环境产生严重辐射危害的物质，包括能够产生裂变反应或者聚合反应的核材料。\n### (4)“传染病病原体”，是指能够通过在人体或者动物体内适当的环境中繁殖，从而使人体或者动物感染传染病，甚至造成传染病扩散的细菌、霉菌、毒种、病毒等。",
        "犯罪主体": "# 一般主体，包括已满14周岁不满16周岁的未成年人",
        "主观方面": "# 故意，行为人出于何种动机不影响本罪成立"
    },

Annotation

JUREX-4E is an expert-annotated knowledge base covering 155 criminal charges. It is structured through a progressive hierarchical annotation framework that prioritizes legal source validity and employs diverse legal interpretation methods to ensure comprehensiveness and authority.

*Thick arrows indicate the primary level where a particular interpretive method is applied, while dashed arrows represent its supplementary use at that level.

The statistics of JUREX-4E shows as follow:

	Mean	Median
Avg. Length	472.53	-
Subject	51.64	17
Object	36.01	25
Subjective Aspect	42.38	21
Objective Aspect	342.5	230

Experiment

We apply JUREX-4E to Similar Charge Distinction task and the Legal Case Retrieval task.

Similar Charge Distinction

We chose three 2-label classification groups in GCI dataset:

Charge Sets	Charges	Cases
F&E	Fraud & Extortion	3536 / 2149
E&MPF	Embezzlement & Misappropriation of Public Funds	2391 / 1998
AP&DD	Abuse of Power & Dereliction of Duty	1950 / 1938

We use an unified approach to introduce four-element descriptions. For each group of similar charges, the model receives charges' four-elements from JUREX-4E or generated by LLM to aid classification. Specifically, GPT-4o+FET_Expert relies on expert-annotated four-elements, while GPT-4o+FET_LLM relies on LLM-generated four-elements. The instruction format is consistent across methods, with only the [Four Elements of candidate charges] varying based on the source.

Prompt:

You are a lawyer specializing in criminal law. Based on Chinese criminal law, please determine which of the following candidate charges the given facts align with. The candidate charges and their corresponding four elements are as follows: [Four Elements of Candidate Charges]. The four elements represent the core factors for determining the constitution of a criminal charge. [The basic concepts of the Four-Element Theory] Please Compare the case facts to determine which charge's four elements they align with, thereby identifying the charge.

All experiments are conducted in a zero-shot setting, with the max_tokens set to 3,000 (or 10,000 for COT and MALR reasoning) and temperature set to 0 or 0.0001(In repeated experiments).

Model	F&E		E&MPF		AP&DD		Average
	Acc	F1	Acc	F1	Acc	F1	Acc	F1
GPT-4o	94.36	95.81	86.49	89.76	85.54	87.12	88.72	90.07
GPT-4o+Article	95.34	96.30	92.64	93.03	88.30	89.33	92.09	92.89
Legal-COT	94.99	96.27	90.50	90.99	87.81	88.14	89.95	90.85
MALR	94.62	95.82	86.99	86.98	87.86	88.68	89.82	90.49
GPT-4o+FETLLM	95.73	96.56	91.87	92.01	89.61	89.69	92.40	92.75
GPT-4o+FETExpert	96.06	96.69	92.57	93.05	90.53	90.62	93.05	93.45

Legal Case Retrieval

We propose the FET_{Expert_guided} method to enhance legal case retrieval by leveraging JUREX-4E. Our approach consists of three key steps:

1. Predicting Charges: A small LLM analyzes case facts to predict potential charges.
2. Matching Elements: The model retrieves corresponding four-element details from a curated legal knowledge base.
3. Analyzing Case Facts: Guided by the matched charges' four-element, another LLM generates case-specific four elements for each candidate. The final ranking combines the similarity of case four elements and case facts between the query and candidates.

Model	NDCG@10	NDCG@20	NDCG@30	R@1	R@5	R@10	R@20	R@30	MRR
BERT	0.1511	0.1794	0.1978	0.0199	0.0753	0.1299	0.2157	0.2579	0.1136
Legal-BERT	0.1300	0.1487	0.1649	0.0186	0.0542	0.1309	0.1822	0.2172	0.0573
Lawformer	0.2684	0.3049	0.3560	0.0432	0.1479	0.2330	0.3349	0.4683	0.1096
ChatLaw	0.2049	0.2328	0.2745	0.0353	0.1306	0.1913	0.2684	0.3751	0.1285
SAILER	0.3142	0.4133	0.4745	0.0539	0.1780	0.3442	0.5688	0.7092	0.1427
GEAR	*	*	*	0.0630	0.1706	0.3142	0.4625	*	0.2162
BGE	0.4737	0.5539	0.5937	0.0793	0.2945	0.4298	0.6500	0.7394	0.1926
FETLLM	0.5139	0.5862	0.6291	0.0980	0.2967	0.4769	0.6802	0.7828	0.2140
- base	0.3583	0.4293	0.4798	0.0506	0.2240	0.3644	0.5383	0.6652	0.1453
FETExpert_guided	0.5211	0.5920	0.6379	0.1024	0.3049	0.4883	0.6885	0.7967	0.2155
- base	0.3766	0.4584	0.5111	0.0715	0.1894	0.3709	0.5891	0.7203	0.1624

*SCR results. Bold fonts indicate leading results in each setting. * denotes that the indicator is not applicable to the current model.

License

MIT