Law element hierarchical extraction method and system based on large model and knowledge graph

Abstract

The invention relates to the technical field of artificial intelligence and judicial big data, and particularly provides a legal element hierarchical extraction method and system based on a big model and a knowledge graph. The method comprises the steps of obtaining basic information and a fact text of a case; automatically matching and confusing criminal name groups Gi according to the basic information and the fact text; reading four elements and element metadata of the confused crime name group Gi through the knowledge graph KG; carrying out multi-crime-name element merging and unified field modeling, and generating a confused crime name group template with annotations; constructing cue words and calling a large language model LLM for reasoning so as to extract case elements of the confused crime name group template; according to the case elements, LLM output is analyzed, verification is carried out, and a confused criminal name group element table is generated and stored, and the method realizes automatic combination, automatic annotation and template extraction of multiple criminal name key groups.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and judicial big data technology, and in particular to a method and system for hierarchical extraction of legal elements based on large models and knowledge graphs. Background Technology

[0002] With the advancement of online court judgments and judicial transparency, a vast number of criminal judgments have been generated in judicial practice. These documents are mostly written in natural language, with loose structures, long lengths, and scattered information. Structured processing of these judgments has become a fundamental task in the application of judicial big data.

[0003] In criminal law theory and judicial practice, the determination of a crime typically revolves around the four elements of criminal law: the object of the crime, namely the legally protected interests and the victims; the objective aspect, including the nature of the act, the harmful act, the harmful result, the constitutive pattern, the characteristics of the act, and the stage of the act; the subject of the crime, including the type of subject, the age of criminal responsibility, and the capacity for criminal responsibility; and the subjective aspect, including the form of culpability, the purpose of the crime, the element of knowledge, the motive, and the mental state. In specific cases, many types of crimes overlap and intersect in these elements, especially some typical and easily confused crimes, such as fraud and contract fraud; embezzlement and misappropriation of funds. These crimes share commonalities in terms of harmful act, constitutive pattern, the existence of a contractual relationship, the object of the crime, and the purpose of the crime, but differ in key constitutive elements. An inappropriate characterization will directly affect the severity of the sentence and adversely affect the credibility of the judiciary.

[0004] In the existing technology, (1) the traditional extraction method based on rules and keywords: through manually set rules, regular expressions, keyword dictionaries, etc., fields such as defendant information, crime name, amount involved, and time are extracted from documents; for complex elements such as behavior description and subjective malice, it usually relies on keyword triggering and human experience for preliminary judgment. For example, in "Research and Implementation of Relationship Extraction Technology for Knowledge Elements in Legal Texts", the approach of using rules and dictionaries as the basis for entity recognition and combining semi-supervised template matching to iteratively extract relationships is adopted. However, this method has the following disadvantages: ① The cost of rule writing and maintenance is high, and it relies heavily on the experience of domain experts. Once the legal provisions or trial standards change, large-scale adjustments are required; ② It has limited ability to identify complex sentence structures and implicit elements (such as "knowing" and "illegal possession purpose"), and it is difficult to handle long-distance dependencies and implicit expressions; ③ It is difficult to naturally express the four elements of criminal law and the multi-level and multi-dimensional structured information such as the confused crime groups; ④ It generally only extracts around a certain established crime and cannot perform element filling and comparative analysis on multiple confused crime groups in the same case in parallel.

[0005] (2) Sequence labeling and classification methods based on deep learning: CRF, BiLSTM, BERT and other models are used to perform entity recognition, element classification and sentence-level label prediction on legal documents. For example, the research on the construction technology of criminal law knowledge graph proposes the JLB-BiLSTM-CRF model, which uses BERT to enhance representation and perform entity recognition; the research on the generative named entity recognition framework for Chinese legal domain explores the use of a sequence-to-sequence generative framework for legal entity recognition. This approach has the following problems: ① Existing models mostly adopt a flat label system, which is difficult to directly carry the hierarchical structure of four elements-sub-elements-specific crimes; ② The model output is usually a label ID or a brief category, which requires one or more rounds of rule and normalization processing to map to a unified and comparable set of elements; ③ When facing multiple easily confused crimes, it is often necessary to design labels or models separately for each crime or a small range of crimes, which is not conducive to comparison and transfer under a unified framework; ④ The corresponding labeling cost is high. Once a new crime or a new element system is introduced, it is often necessary to re-label and retrain, which limits the scalability.

[0006] (3) Extraction method based on large language model + simple JSON template: In recent years, attempts have gradually emerged in practice to extract elements from legal texts using general large language models (such as GPT-type models). The usual practice is to write explanatory prompts for the model and attach a simple JSON template. This method has the following shortcomings: ① The field values ​​lack strong constraints, and the model is prone to generating a large number of synonymous, ambiguous or non-standardized expressions, which are difficult to directly map to a fixed set of legal elements; ② In order to map these free texts to a finite and stable set of elements, a large number of rules, mapping tables and manual post-processing are required; ③ Existing JSON templates are mostly based on the perspective of a single case or a single crime, and have not built a unified comparison template structure for the confused crime group.

[0007] (4) Preliminary attempts at knowledge graph + template: Some studies have constructed knowledge graphs in the field of criminal law that are of the type of crime-constituent elements-legal provisions-judicial interpretation. For example, "From Graph to Word Bag: Introducing DomainKnowledge to Confusing Charge Prediction" uses a graph structure to show the common elements and distinguishing features between different crimes, and then automatically filters them to form a bag of words, and uses the bag of words to supervise the model's attention to improve the accuracy of predicting confusing crimes. This type of approach has achieved certain results in the task of predicting confusing crimes, but it still has the following limitations: ① Insufficient structured output capability: The method remains at the bag of words level and fails to generate a structured template that conforms to the thinking habits of legal professionals. First, the constituent elements of law are a system with time sequence and logical connection, rather than a simple set of keywords. Key elements in judgment documents (such as "knowingly") are often reflected through continuous contexts, and may not have corresponding words. Keywords are disconnected from facts. Second, the method itself contains a pre-judgment, relying on specific keywords to distinguish crimes, which is actually making a qualitative judgment in advance. For example, it's impossible to explain whether key elements of contract fraud versus fraud, such as "during the signing / performance of the contract," are met. ② Fundamental Misalignment: Technology-Driven vs. Problem-Driven Approaches applying readily available NLP technologies (knowledge graphs, attention mechanisms) to legal scenarios instead of addressing pain points in judicial practice. Legal judgment is not keyword matching; simplifying medical diagnosis to symptom keyword retrieval is technically feasible but lacks professional interpretability. Summary of the Invention

[0008] In view of this, the present invention provides a hierarchical extraction method and system for legal elements based on large models and knowledge graphs, which can realize automatic merging, automatic annotation and template extraction of multiple crime elements groups.

[0009] In a first aspect, the present invention provides a hierarchical extraction method for legal elements based on large models and knowledge graphs, the method comprising:

[0010] Step 1: Obtain basic information and factual text of the case; Step 2: Based on the basic information and factual text, automatically match the confusing charge group Gi; Step 3: Read the four elements and metadata of the confusing crime group Gi through the knowledge graph KG; Step 4: Merge multiple crime elements and model unified fields to generate annotated templates for confusing crime groups; Step 5: Construct cue words and call the Large Language Model (LLM) for reasoning to extract case elements from the confused charge group template; Step 6: Based on the case elements, parse the LLM output, perform verification, and generate and store the element table of the confused crime group.

[0011] Optionally, the basic information in step 1 includes the case number (case_no), the charge (charge), the facts of the crime (FD), and the court's opinion (Hold_that).

[0012] Optionally, step 2 includes: Based on the case charge or the previous coarse-grained classification results, the corresponding confusing charge group Gi is automatically determined by the mapping relationship of [charge → confusing charge group]. If no confusing charge group is matched, the process ends directly or enters other processing paths.

[0013] Optionally, step 3 includes: Using the knowledge graph KG, we read the four elements and element metadata corresponding to each crime in the confused crime group Gi. The read content includes the candidate value set content and annotation note information; the element metadata includes the nature of the behavior, the harmful behavior, the characteristics of the behavior, the knowledge element, and the standard of completion. The knowledge graph KG used is organized as follows: The root node represents the criminal law knowledge graph; the child nodes represent specific crimes. Under the crime node, it is further broken down into: The object of a crime, its objective aspect; the subject of a crime, its subjective aspect; the standard for completion of a crime. Record at the leaf nodes of each element: The content is the set of candidate values ​​for the current element under the current charge; the note is the annotation information for the content of the current element, which is used to supplement the legal connotation and applicable boundaries briefly expressed in the content, and is used to generate the description field in the confusing charge group template.

[0014] Optionally, step 4 includes: The crime element structure is recursively merged. Through multi-crime element merging and field alignment algorithms, the nature of the behavior, harmful behavior, behavior characteristics, known elements, and completed standards are horizontally aligned and deduplicated. For each field, the candidate value set of each crime is retained to form a unified, multi-level element metadata structure. A JSON template is generated based on the element metadata, and annotation information is automatically added to each field in the template to obtain an annotated confusion charge group template for large language model inference.

[0015] Optionally, step 5 includes: The descriptive text, the annotated template for the confusing charge group, and the case fact FD are taken as input and submitted to the large language model. The request is to populate the elements for each charge in the confusing charge group according to the template structure, and to limit the field values ​​to the set of candidate values ​​given in the template annotation.

[0016] Optionally, step 6 includes: Using regular expressions, extract JSON paragraphs or strip markers from the LLM output text, and call the JSON parsing function to convert them into a dictionary structure; perform field integrity checks and candidate value validity checks on the parsing results; organize the results that pass the checks into a confusing crime group element table and save it for subsequent analysis and modeling.

[0017] Optionally, step 6 may be followed by downstream applications; The elements table of confused charges will be provided to the given charge / characterization auxiliary analysis module, the similar case retrieval module, and the conviction prediction or sentencing prediction model, respectively, and will be used to determine the charge requirements that a case meets, to search by element similarity, and as feature input.

[0018] Secondly, this invention provides a hierarchical legal element extraction system based on a large model and knowledge graph, the system comprising: The data preprocessing module is used to obtain basic information and factual text of the case; The automatic matching module for confusing charge groups is used to automatically match the confusing charge group Gi based on basic information and factual text; The knowledge graph management and field metadata extraction module is used to store the criminal law knowledge graph KG and read the four elements and element metadata of the confused crime group Gi through the knowledge graph KG; The module for merging multiple crime elements and confusing crime groups is used for merging multiple crime elements and unifying field modeling, generating annotated confusing crime group templates; The Large Language Model (LLM) prompt construction and reasoning module is used to construct prompt words and call the Large Language Model (LLM) for reasoning to extract case elements from the confused crime group template. The parsing output and legality verification module is used to parse the LLM output based on case elements, perform verification, and generate and store the element table of the confusing crime group.

[0019] The technical solution provided by this invention includes the following steps: acquiring basic case information and factual text; automatically matching a confused charge group Gi based on the basic information and factual text; reading the four elements and element metadata of the confused charge group Gi through a knowledge graph (KG); merging multiple charge elements and modeling unified fields to generate an annotated confused charge group template; constructing prompt words and calling a large language model (LLM) for reasoning to extract case elements from the confused charge group template; parsing the LLM output based on the case elements, performing verification, generating and storing a confused charge group element table. This method achieves automatic merging, automatic annotation, and templated extraction of multiple charge element groups. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 A flowchart illustrating the hierarchical extraction method of legal elements based on large models and knowledge graphs provided in this embodiment of the invention; Figure 2 A schematic diagram of the confusing crime group and its element structure in a knowledge graph, provided for embodiments of the present invention; Figure 3 This is a schematic diagram of a hierarchical legal element extraction system based on a large model and knowledge graph, provided in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” used in the embodiments of this invention are also intended to include the plural forms unless the context clearly indicates otherwise.

[0024] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0025] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0026] The present invention provides a hierarchical extraction method for legal elements based on large models and knowledge graphs, as follows Figure 1 and Figure 2 shown. This method includes: Step 1: Obtain the basic information and factual text of the case.

[0027] In an embodiment of the present invention, the basic information in Step 1 includes the case number case_no, the case charge charge, the factual description of the case FD, and the court's opinion Hold_that.

[0028] preprocessed_data = { "case_no": "2020 Liao 0422 Xing Chu X Hao", "charge": "Fraud", "FD": "Upon examination, in July 2019, Peng XX1, the father of the victim Peng XX, found the defendant Zhang XX and asked Zhang XX to help introduce a girlfriend for Peng XX. The defendant Zhang XX then falsely claimed that he had a niece named 'Li XX' and communicated with Peng XX in the name of 'Li XX' through WeChat. From July 2019 to January 2020, during the communication between the defendant Zhang XX and Peng XX, the defendant Zhang XX repeatedly fabricated reasons such as seeing a doctor and buying medicine, and demanded Peng XX to transfer money to him. During this period, Peng XX transferred a total of 41,921.00 yuan to Zhang XX through WeChat transfers.", "Hold_that": "The court believes that the defendant Zhang XX, for the purpose of illegal possession, by fabricating facts, defrauded others of property, and the amount involved is relatively large. The facts of the crime are clear, and the evidence is solid and sufficient. His act has constituted the crime of fraud, and the accusation of the public prosecution agency is established, and the court supports it. The defendant Zhang XX was summoned to the case and could truthfully confess the facts of the crime after surrender, which is a confession. Therefore, the public prosecution agency's opinion that he has the circumstance of confession is adopted by the court, and he can be given a lighter punishment according to law. The defendant Zhang XX voluntarily and truthfully confessed his crime, admitted the facts of the crime charged, and was willing to accept punishment. The sentencing recommendation of the public prosecution agency is appropriate. In accordance with Article 266, Article 47, Article 52, Article 61, Article 64, Article 67, Paragraph 3 of the Criminal Law of the People's Republic of China, Article 15 of the Criminal Procedure Law of the People's Republic of China, and Article 1 of the Interpretation of the Supreme People's Court and the Supreme People's Procuratorate on Several Issues Concerning the Specific Application of Law in Handling Criminal Cases of Fraud, the judgment is as follows:", }.

[0029] Step 2: Automatically match the group of confused charges Gi based on the basic information and factual text.

[0030] matched_groups = ["G1"], G1: {Fraud, Contract Fraud}.

[0031] In this embodiment of the invention, step 2 includes: Based on the case charge (e.g., labeled as "fraud") or the previous coarse-grained classification results, the corresponding confused charge group Gi is automatically determined according to the mapping relationship of "charge → confused charge group". For example, G1={fraud, contract fraud}, G2={traffic accident, dangerous driving} or G3={fraud, embezzlement}. If no confused charge group is matched, the process ends directly or enters other processing paths.

[0032] Generate a template for elements with confusing charges and automatic annotations. Based on the element metadata: organize a set of candidate values ​​for each field; summarize the note information into a brief description; generate an element template with "unified field + charge column" and add "optional value + description" annotations after the field to constrain subsequent extraction.

[0033] Step 3: Read the four elements and metadata of the confusion charge group Gi through the knowledge graph KG.

[0034] In this embodiment of the invention, step 3 includes: Using the knowledge graph KG, we read the four elements and element metadata corresponding to each crime in the confused crime group Gi. The read content includes the candidate value set content and annotation note information; the element metadata includes the nature of the behavior, the harmful behavior, the characteristics of the behavior, the knowledge element, and the standard of completion. The knowledge graph KG used is organized as follows: The root node is a criminal law knowledge graph; the child nodes are specific crimes, such as traffic accident crime, fraud, contract fraud, etc. Under the crime node, it is further broken down into: The object of a crime includes the objective aspect (including the nature of the act, the harmful act, the harmful result, the constitutive pattern, the characteristics of the act, the stage of the act, etc.), the subject of the crime (subject type, age of criminal responsibility, capacity for criminal responsibility, etc.), and the subjective aspect (form of culpability, criminal purpose, element of knowledge, etc.); the standard for completion of the crime. Record at the leaf nodes of each element: The content is the set of candidate values ​​for the current element under the current crime; the note is the annotation information for the content of the current element, which is used to supplement the legal connotation and applicable boundaries of the brief expression in the content. It is used to generate the explanatory field in the template of confusing crime groups, which helps to improve the understandability and consistency of the extraction results. For example, "the victim disposes of the property he has the right to dispose of to the other party" refers to judging whether there is a "disposal act". The crime of fraud requires the existence of a "disposal act".

[0035] { "Fraud": { "Object of the crime": [{"content": ["ownership of public and private property"],"note": ["......"]}], "Objective aspects": { "Behavioral Essence": {"content": ["Deceived to obtain"],"note": ["......"]}, "Harmful behavior": ["Fabricating facts...concealing the truth"], "Harmful Consequences": ["Risk of loss to public and private property......"], "Constitutional Pattern": [[".The perpetrator commits deceptive acts......the victim suffers property loss"]], "Behavioral characteristics": {"content": ["Victim's disposal of property......"], "note":["......"]}, "Stage of Behavior": ["The victim controls property...within their domain"] }, "Subject of the crime": [{"content": "natural person","note": ["......"]}], "Subjective aspect": { "Form of culpability": ["Direct intent"], "Criminal intent": ["illegally appropriating another person's property"] "Knowingly" element: ["Knowingly engaging in deception...disposing of property"] }, "Standard for Completion of the Crime": {"content": ["The actor controlled the property...loss"],"note":["......"]}, "Sentencing Factors": {"content": ["relatively large amount", "huge amount", "especially huge amount", "principal offender", "accomplice", "coerced accomplice", "instigator", "criminal preparation", "criminal abandonment", "attempted crime", "legitimate defense", "excessive defense"......], "note": ["......"]} }, "Contract fraud": { "Object of the Crime": [{"content": ["Public and Private Property Ownership......Contract Management Order"],"note":["......"]}], "Objective aspects": { "Behavioral Essence": {"content": ["Victim's disposal of property......", "Defrauding by using a contract"],"note": ["......"]}, "Harmful Acts": ["Fabricating facts...receiving stolen goods and fleeing"], "Harmful Consequences": ["Property loss risk...disruption of economic order"], "Constitutional Pattern": [[".The perpetrator deceives...the victim suffers financial loss"]], "Behavioral Characteristics": {"content": ["Contractual Relevance"],"note": ["......"]}, "Action Stage": {"content": ["During the signing / performance of the contract"], "note": ["......"]} }, "Subject of the crime": [{"content": ["natural person......unit"],"note": ["......"]}], "Subjective aspect": { "Form of culpability": ["Direct intent"], "Criminal Purpose": ["Illegally appropriating the property of the other party to the contract"] "Knowingly" element: {"content": ["Knowingly lacking the ability to fulfill obligations...absconding"],"note":["......"]} }, "Standard for Completion of the Case": {"content": ["The victim's disposal of property based on the contract...losses"], "note":["......"]}, "Sentencing Factors": {"content": ["relatively large amount", "huge amount", "especially huge amount", "principal offender", "accomplice", "coerced accomplice", "instigator", "criminal preparation", "criminal abandonment", "attempted crime", "legitimate defense", "excessive defense"......], "note": ["......"]} }}}.

[0036] Step 4: Merge multiple crime elements and model unified fields to generate annotated templates for confusing crime groups.

[0037] In this embodiment of the invention, step 4 includes: The crime element structure is recursively merged. Through multi-crime element merging and field alignment algorithms, the nature of the behavior, harmful behavior, behavior characteristics, known elements, and completed standards are horizontally aligned and deduplicated. For each field, the candidate value set of each crime is retained to form a unified, multi-level element metadata structure. A JSON template is generated based on the element metadata, and annotation information (such as optional values ​​and descriptions) is automatically appended to each field in the template to obtain an annotated confusion charge group template for large language model inference.

[0038] This invention designs the element template for the confusing crime group as a comparison table with unified element fields and columns by crime, and introduces an automatic annotation mechanism on this basis to form the element template for the confusing crime group and the automatic annotation structure.

[0039] The overall structure can be summarized as follows: Rows: representing unified element fields, such as the nature of the behavior, harmful behavior, behavioral characteristics, known elements, and completion standards; Columns: representing the various crimes in the confused crime group, such as fraud and contract fraud; Cell content: the set of candidate values ​​for a specific crime under a specific element field, along with a brief description. After extracting specific cases, the cell will mark the values ​​actually selected for that case.

[0040] In JSON format, it can be represented as the following example structure: { "Confusing Charges" group: ["Fraud", "Contract Fraud"] Case Element Table: { "Fraud": { "Object of Crime": [ { "content": "Ownership of public and private property" "note": [ "

Public and Private Property Ownership

Deceived Payment

Behavioral Characteristics

Completed Case Standard

A relatively large amount

A huge amount

An especially huge amount

Juvenile delinquency

Elderly delinquency

[0041] In this embodiment of the invention, as shown in Table 1, the table is presented in tabular form: the header is the crime within the confused crime group, the left side is the unified element field, and each cell displays the optional values, descriptions, and other information; in practical applications, the specific element values ​​selected by the model for the case can be indicated by underlines, bolding, or other marking methods.

[0042] Table 1. Templates and Automatic Annotations for Elements of the Confusing Charges Group

[0043] The above templates are automatically generated by the multi-crime element merging and field metadata extraction module and the automatic annotation template generation module, eliminating the need for manual writing of each item. They can also be automatically updated as the knowledge graph and the configuration of the confused crime groups change, ensuring the maintainability and scalability of the system.

[0044] Metadata after merging the crimes of fraud and contract fraud; {"Merged metadata":{ "Object of the Crime":{ "Candidate value":[{ "value": ["ownership of public and private property"], "source_charges":["fraud","contract fraud"], "note": ["

Public and Private Property Ownership

Public and Private Property Ownership

Economic Order (Contract Management Order)

Economic Order (Contract Management Order)

Economic Order (Contract Management Order)

Deceiving to Obtain

Contract Fraud

Contract Fraud

The victim disposes of property they have the right to dispose of to the other party

The victim is aware of the actual existence of the disposed property

Contractual Relevance

Contractual Relevance

Knowledge Element

Knowledge Element

Completed Case Standard

Completed Standard

A relatively large amount

A huge amount

An especially huge amount

Juvenile delinquency

Elderly delinquency

[0045] Generate annotated templates for obfuscated charge groups; { "Object of Crime": { "value": "", "FD": "", "Note": { Optional values: ["Public and private property ownership", "Economic order (contract management order)"], "Explanation": "

Public and Private Property Ownership

Public and Private Property Ownership

Economic Order (Contract Management Order)

Economic Order (Contract Management Order)

Deceptive Acquisition

Deceptive Acquisition Through Contracts

Deceptive Acquisition Through Contracts

Natural Person

Entity

Illegally appropriating another's property

Illegally appropriating the property of the other party to a contract

Completed Standard

Completed Standard

Relatively large amount

Huge amount

Particularly huge amount

Juvenile delinquency

Elderly delinquency

[0046] Step 5: Construct cue words and call the Large Language Model (LLM) for reasoning to extract case elements from the confused charge template.

[0047] In this embodiment of the invention, step 5 includes: The descriptive text, the annotated template for the confusing charge group, and the case fact FD are taken as input and submitted to the large language model. The request is to populate the elements for each charge in the confusing charge group according to the template structure, and to limit the field values ​​to the set of candidate values ​​given in the template annotation.

[0048] Template-based case element extraction and task instructions; The FD and Hold_that fields must respectively reference the original text of the facts of the crime (FD) and the court's opinion (Hold_that): ① Precise citation of corresponding candidate values: Enclose the FD and Hold_that segments in double quotes (e.g., "He repeatedly fabricated reasons such as seeing a doctor and buying medicine, and asked Peng to transfer money to him," and "The defendant Zhang was summoned to the case and was able to truthfully confess the facts of the crime after being brought to justice, which constitutes a confession. Therefore, the public prosecutor's opinion that he has a confession is accepted by this court"); ② No direct evidence: Write "No direct evidence," val The explanations in the ue reference notes require understanding of optional values ​​and that they must be strictly selected from "Notes → Optional Values": For example, the behavioral characteristic value must be one of the following: ① "The victim disposes of property that he has the right to dispose of to the other party", "The victim is aware of the actual existence of the disposed property"; ② "Contractual relevance" elements of the crime (object of the crime, objective aspect, subject of the crime, subjective aspect, standard of completion) are based solely on FD analysis and do not involve Hold_that; only the "Sentencing Factors" part requires a combination of FD and Hold_that.

[0049] Compliant output of large models: { "Object of Crime": { "value": ["ownership of public and private property"], "FD": "The defendant, Zhang, falsely claimed to have a niece named 'Li,' and used this identity to communicate with the victim, Peng, via WeChat. He fabricated reasons such as seeking medical treatment and buying medicine, and defrauded Peng of 41,921.00 yuan transferred via WeChat." "Note": {}} "Objective aspects": { "The essence of behavior": { "value": ["obtained through deception"], "FD": "Zhang fabricated the identity of 'Li' and details of seeing a doctor and buying medicine, causing the victim, Peng, to have a mistaken understanding, and based on this mistake, Peng disposed of property through WeChat transfers." "Note": {}} "Behavioral characteristics": { "value": ["The victim disposes of property that they possessed and had the right to dispose of to the other party"] "FD": "The victim, Peng, voluntarily transferred funds from his bank account to the defendant Zhang's account via WeChat, thus completing the disposal and transfer of property." "Note": {}}} "Subject of the crime": { "value": ["natural person"], "FD": "“Defendant Zhang Moumou”, "Note": { Optional values: ["natural person", "organization"], "Explanation": "

Natural Person

Entity

Illegally appropriating another's property

Illegally appropriating the property of the other party to a contract

Relatively large amount

Huge amount

Particularly huge amount

Juvenile delinquency

Elderly delinquency

[0050] Step 6: Based on the case elements, parse the LLM output, perform verification, and generate and store the element table of the confused crime group.

[0051] In this embodiment of the invention, step 6 includes: Using regular expressions, extract JSON paragraphs or strip markers from the LLM output text, and call the JSON parsing function to convert them into a dictionary structure; perform field integrity checks and candidate value validity checks on the parsing results; organize the results that pass the checks into a confusing crime group element table and save it for subsequent analysis and modeling.

[0052] Results parsing and validity verification: Extract the JSON structure from the model output; check whether the fields are complete, whether the field values ​​belong to the preset candidate set, and whether the data format meets the requirements; discard or re-extract outputs that do not meet the requirements.

[0053] Generate and store the table of elements for the group of confused charges. Organize the verified results into a unified "Table of Elements for the Group of Confused Charges", which separately lists the element filling situations from the perspectives of fraud and contract fraud; store the element table together with information such as the case number for subsequent query and analysis.

[0054] Generate and store the table of elements for Group G1 of confused charges; { "case_no": "Criminal Initial Judgment No. X, 2020, Liaoning 0422", "charge": "Fraud", "confusion_group": ["Fraud", "Contract Fraud"], "elements": { "Fraud": { "Object of the crime": { "value": ["The ownership of public and private property"], "FD": "“Peng paid Zhang a total of 41,921.00 yuan through WeChat transfers.”", "Basis for matching": "FD directly indicates that the property of the victim has been damaged, fully meeting the definition of the object of fraud that infringes on the ownership of public and private property."}, "Subject of the crime": { "value": ["Natural person"], "FD": "“The defendant Zhang某某”", "Basis for matching": "FD clearly refers to the individual “the defendant Zhang某某” throughout the text, meeting the requirements of the natural person subject."}, "Objective aspect": { "Nature of the act": { "value": ["Obtaining by deception"], "FD": "“Claiming to have a niece named ‘李某某’”, “Repeatedly fabricating reasons such as seeing a doctor and buying medicine, and asking Peng to transfer money to him”", "Basis for matching": "FD fully demonstrates the typical fraud act chain of ‘fabricating facts → causing the victim to have a wrong perception → disposing of property based on the wrong perception’."}}, "Subjective aspect": { "Form of guilt": { "value": ["Direct intent"], "FD": "“Claiming to have a niece named ‘李某某’”, “Repeatedly fabricating reasons such as seeing a doctor and buying medicine”", "Matching Criteria": "The proactive and continuous 'falsehood' and 'fabrication' in FD (Financial Deception) indicate that the perpetrator held a hopeful or indifferent attitude towards the fraudulent act and its consequences, which meets the characteristics of direct intent." "Criminal Purpose": { "value": ["illegally appropriating other people's property"], "FD": No direct evidence. "Matching Basis": "FD describes a pattern of demanding money under fabricated reasons, strongly suggesting an 'intent to illegally possess,' though not explicitly stated in writing. This intent has been judicially determined by Hold_that." "Known element": { "value": ["knowingly engaging in deceptive behavior"], "FD": "“false claim” or "fictional" "Matching Criteria": "The descriptions of behavior such as 'falsely claiming' and 'fabricating' in the FD itself indicate that the perpetrator was aware of the deceptive nature of their actions." "Completion Standard": { "value": ["Property actually controlled by the perpetrator", "Property loss suffered by the victim"] "FD": "Peng paid Zhang a total of 41,921.00 yuan via WeChat transfer." "Matching Basis": "FD clearly stated that the funds had been 'paid' to the defendant, indicating that he had actual control of the property, the victim suffered property loss, and the crime was completed." Sentencing factors: { "value": ["substantial amount", "confess", "plead guilty and accept punishment"], "FD&Hold_that": "“Peng paid Zhang a total of 41,921.00 yuan via WeChat transfer.” (FD) and “Defrauding others of a substantial amount of money.” (Hold_that); “The defendant Zhang was summoned to the station.” (FD) and “After being apprehended, he truthfully confessed to the crime, which constitutes a confession.” (Hold_that); “The defendant Zhang voluntarily and truthfully confessed to his crimes, admitted the charged facts, and was willing to accept punishment.” (Hold_that) "Matching Basis": "All circumstances are based on a combination of the factual basis of FD and the judicial determination of Hold_that. The amount deemed 'relatively large' (41,921 yuan) was determined in FD and characterized by Hold_that." "Contract fraud": { "Object of Crime": { "value": ["Public and private property ownership", "Economic order (contract management order)"], "FD": "Peng paid Zhang a total of 41,921.00 yuan via WeChat transfer." "Matching Basis": "FD only reflects an infringement of property rights, without any factual description involving 'contract' or 'economic order,' therefore it does not meet the complex object requirement of the crime of contract fraud." "Subject of the crime": { "value": ["natural person"], "FD": "“Defendant Zhang Moumou”, "Matching Criteria": "FD indicates that the perpetrator is a natural person, which meets the requirement of a natural person as the subject of the crime of contract fraud." "Objective aspects": { "The essence of behavior": { "value": ["defrauded by using a contract"], "FD": No direct evidence. "Matching Criteria": "FD describes deception and money transfers between individuals based on social relationships, without mentioning the signing, performance, or exploitation of any form of contract, and is unrelated to the core behavior of 'exploiting contracts.'" "Subjective aspect": { "Forms of culpability": { "value": ["directly intentional"], "FD": "“Falsely claimed to have a niece named 'Li Moumou'”, and “repeatedly fabricated reasons for seeking medical treatment and purchasing medicine”. "Matching Criteria": "The conduct can be presumed to be intentional, but this intent is not specifically aimed at 'using the contract' to commit fraud." "Criminal Purpose": { "value": ["illegally possessing the property of the other party to the contract"], "FD": No direct evidence. "Matching Criteria": "Neither FD nor Hold_that indicates the existence of a specific 'contractual counterparty,' suggesting that the criminal intent is directed at an unspecified individual, rather than a contractual party." "Known element": { "value": ["Knowing that there is no ability or intention to perform the contract"] "FD": No direct evidence. "Matching Basis": "FD did not involve any premises or commitments related to 'performance,' therefore it is impossible to determine the actor's level of awareness regarding this." "Completion Standard": { "value": ["Victim disposes of property based on contract", "Perpetrator actually controls property", "Victim suffers property loss"] "FD": "Peng paid Zhang a total of 41,921.00 yuan via WeChat transfer." "Matching Basis": "FD only satisfies the two points of 'control of property' and 'property loss,' while the core requirement of 'disposal of property based on contract' lacks factual basis." Sentencing factors: { "value": ["substantial amount", "confess", "plead guilty and accept punishment"], "FD&Hold_that": "“Peng paid Zhang a total of 41,921.00 yuan via WeChat transfer.” (FD) and “fraudulently obtained property from others, amounting to a substantial sum.” (Hold_that) "Matching Criteria": "The amount (41,921 yuan) meets the threshold for 'relatively large amount' (≥20,000 yuan for individuals) in the crime of contract fraud, but this circumstance is shared by both crimes and is not unique to contract fraud, and there are no other circumstances specific to this crime (such as corporate crime). 'The defendant Zhang was summoned to the case.' (FD) and 'After being apprehended, he was able to truthfully confess to the crime, which constitutes a confession.' (Hold_that); 'The defendant Zhang voluntarily and truthfully confessed to his crimes, admitted the charged crimes, and was willing to accept punishment.' (Hold_that)."

[0055] In this embodiment of the invention, downstream applications are also included after step 6; The elements table of confused charges will be provided to the given charge / characterization auxiliary analysis module, the similar case retrieval module, and the conviction prediction or sentencing prediction model, respectively, and will be used to determine the charge requirements that a case meets, to search by element similarity, and as feature input.

[0056] This invention provides a hierarchical extraction system for legal elements based on large models and knowledge graphs, such as... Figure 3 As shown, the system includes: The data preprocessing module is used to obtain basic information and factual text of the case; read case records from the original judgment documents or preprocessed datasets; extract information such as case number (case_no), charge, and criminal facts (FD); and filter out samples that are obviously too short or marked as "investigate further" in the case details section, making them unsuitable for extraction.

[0057] The automatic matching module for confusing charge groups is used to automatically match confusing charge groups Gi based on basic information and factual text; maintain the mapping relationship between [charge → confusing charge group], for example: fraud / contract fraud → confusing charge group G1; traffic accident / dangerous driving → confusing charge group G2, etc.; automatically determine the set of confusing charge groups to which the current case belongs or needs to be compared and analyzed based on the charge information of the case or based on the pre-trained coarse-grained charge prediction results; and support the situation where one case corresponds to multiple confusing charge groups.

[0058] The knowledge graph management and field metadata extraction module is used to store the criminal law knowledge graph KG, and to read the four elements and element metadata of the confused crime group Gi through the knowledge graph KG; it stores the criminal law knowledge graph KG, records the four elements structure and detailed elements of each crime (such as the nature of the behavior, harmful behavior, behavior characteristics, knowledge element, completion standard, etc.), as well as content / note information for extracting constraints and annotations; it maintains a predefined set of confused crime groups, such as {fraud, contract fraud}, {traffic accident, dangerous driving}, etc.; and it provides an interface to read the definition of the corresponding crime element by confused crime group.

[0059] The module for merging and confusing crime groups with multiple crime elements is used to merge multiple crime elements and unify field modeling, generating annotated templates for confusing crime groups. For a specific confusing crime group, it reads the four elements and detailed element structures of each crime in the current group from the knowledge graph. Through a merging algorithm, it aligns and unifies fields such as "behavioral nature, harmful behavior, and behavioral characteristics," and merges and removes duplicates of candidate values ​​by crime under each field. It adopts a recursive merging strategy for data structures such as list / dict / content / note to form a unified, multi-level element metadata structure.

[0060] In this embodiment of the invention, an automatic annotation template generation module is also included, which is used to: summarize the candidate value set of each field based on the merged element metadata; summarize the criminal facts or the court's opinion corresponding to the candidate values ​​of each field; summarize note information and form a brief explanatory text; automatically generate a multi-level JSON template with the core structure of "unified field + crime column", and write the following in the form of notes after the fields: "[Optional Values]", listing the candidate value set; "[Explanation]", giving a brief explanation formed by summarizing multiple notes.

[0061] The Large Language Model (LLM) prompt construction and reasoning module is used to construct prompt words and call the LLM for reasoning to extract case elements from the confused crime group template; combine explanatory text, confused crime group element template (including automatic annotations) and case crime facts FD to form prompt information for reasoning; call the LLM to generate JSON structured output that conforms to the template constraints; and require the model to select only the candidate value set given by the template when taking values ​​for each field.

[0062] The parsing output and legality verification module is used to parse the LLM output based on case elements, perform verification, generate and store the confusing crime group element table, extract JSON strings from model output (such as text containing code blocks) and parse them; check whether the fields are complete, whether the field values ​​belong to the preset candidate set, and whether the data format meets the requirements; and convert the results that pass the verification into a unified confusing crime group element table structure.

[0063] In this embodiment of the invention, a result storage and downstream application module is also included to write the element table of the confused charge group corresponding to each case into a JSONL file or a relevant database row by row; and to provide a data interface for downstream systems such as conviction prediction, similar case retrieval, and sentencing recommendations to call and display.

[0064] This invention introduces an analytical perspective of confusing crime groups, treating a group of easily confused crimes (such as fraud and contract fraud) as a whole analytical unit. For a single case, it can simultaneously extract the constituent elements of each crime within the group, thereby supporting comparative analysis from the perspective of multiple crimes in a single case.

[0065] By utilizing knowledge graphs to automatically generate unified element templates for multiple crimes, the four elements and their detailed elements of each crime are modeled in the criminal law knowledge graph. Through multi-crime element merging and field alignment algorithms, a multi-level template framework with "unified field + crime column" as the core structure is automatically generated to achieve consistent modeling across crimes.

[0066] By leveraging the strong constraints of the large language model's output, post-processing is significantly simplified. Candidate values ​​and explanatory information from the knowledge graph are automatically written into template annotations. This not only improves semantic reasoning performance but also effectively suppresses the generation of illusions in the large model, guiding it to select from a predefined candidate set. By constraining the output space from the source, post-processing can be simplified to JSON parsing and validity verification, rather than relying on a large number of normalization rules.

[0067] An automatic matching and annotation mechanism for confusing charge groups is introduced. Based on the charge or preliminary prediction results of a case, it automatically matches possible confusing charge groups. At the same time, the content / note in the knowledge graph is automatically converted into field annotations of "optional values ​​+ legal explanations", forming an automatic annotation template for large language models, realizing the linkage between knowledge graph, template and model reasoning.

[0068] A group element table of confusing charges is formed that can directly serve the distinction of conviction and downstream modeling. Through the above method, a group element table of confusing charges with a unified structure and horizontal comparability is generated for each case, providing a high-quality structured feature foundation for the correction of conviction and characterization errors, case retrieval, judicial policy research and data-driven conviction / sentencing prediction models.

[0069] The key technical aspects to be protected by this invention include, but are not limited to, the following: 1. Definition and organization method of confusing crime groups: It proposes the idea of ​​extracting and comparing elements of several easily confused crimes as an analytical unit, and managing them uniformly in the system in the form of confusing crime groups.

[0070] 2. A multi-crime “four-element-factor” modeling method based on knowledge graph: In the knowledge graph, the four elements and their detailed elements, including sentencing factors, of each crime are uniformly modeled, and candidate values ​​and explanatory information are stored in the form of content / note for subsequent template generation and extraction constraints. This framework adopts an open design. The elements listed in this invention are only examples, and the system can be flexibly expanded and refined according to actual needs.

[0071] 3. Automatic matching mechanism for confusing charge groups: Starting from the charge information or charge prediction results of the case, the system automatically locates the confusing charge groups that should be included in the analysis, without the need for manual specification of the comparison charge set for each case.

[0072] 4. Multi-crime element merging and alignment algorithm: For knowledge graph fragments corresponding to confused crime groups, recursively merge and deduplicate at the dict, list, content / note levels to generate a unified element structure and realize the modeling method of "unified field + multi-crime candidate value set".

[0073] 5. The "Unified Field + Crime Name Column" Confused Crime Name Group Element Template Structure: Through field alignment and merging, an element comparison template is formed with unified fields as rows and crime names as columns, so that the same case can complete the element filling in from multiple crime name perspectives at the same time.

[0074] 6. Automatic annotation template generation mechanism: Automatically converts the content / not information in the knowledge graph into annotations (including a list of optional values ​​and brief descriptions) after the fields in the template, forming a constrained template for the extraction model, so that the model takes into account both standardized values ​​and legal meaning when it is generated.

[0075] 7. Automatic Template Generation and Constraint-Based Prompt Method: Based on the configuration of the confused crime group and the knowledge graph content, the method automatically generates element templates and embeds the templates (including candidate values, the corresponding criminal facts or the court's opinion and explanation) into the reasoning input. It clearly constrains the output structure and value range. The four elements of criminal law (object of crime, objective aspect, subject of crime, and subjective aspect) and the standard of completed crime are all based on semantic reasoning of criminal facts. Sentencing factors are extracted based on criminal facts and the court's opinion. This method performs excellently in the extraction of common conviction and sentencing factors, with an accuracy rate of over 99% in the extraction of sentencing factors. It also has good scalability and can effectively adapt to further expansion and refinement of the element system.

[0076] 8. Lightweight post-processing mechanism based on constraint output: Structured results can be generated by parsing JSON and validating the validity of candidate values, without relying on complex text normalization and multi-layer rule systems.

[0077] 9. Application interface design of "Confused Crime Group Element Table" in conviction differentiation and downstream modeling: The extracted results are provided to conviction prediction, case retrieval, sentencing suggestions and other systems in the form of "Confused Crime Group Element Table" as feature input and visualization basis.

[0078] 10. The end-to-end processing flow driven by the above template: the complete process design and implementation from case data input, automatic matching of confused crime groups, element merging and annotation template generation, to model extraction, result verification and storage.

[0079] Compared with the prior art, the present invention has the following advantages: 1. Introducing the perspective of confusing crime groups to achieve parallel extraction of elements of multiple crimes in one case: This invention no longer focuses on extracting elements around a single crime, but rather on confusing crime groups in judicial practice, generating element filling results for each crime group in the same case at the same time; it is conducive to directly comparing the differences in the constituent elements of easily confused crimes such as "fraud - contract fraud" and "traffic accident - dangerous driving" at the structured level.

[0080] 2. Knowledge Graph-Driven Unified Element Framework for Multiple Crimes: The framework unifies the four elements and their detailed components through a criminal law knowledge graph, and constructs an element template of "unified field + crime column" through a merging algorithm. This enables comparative analysis of the nature of the behavior, harmful behavior, subjective aspects, sentencing circumstances, etc., of each crime under the same field dimension, which conforms to the legal system of conviction and sentencing, and facilitates interpretability analysis and tracing of the root causes by judges, prosecutors, lawyers and other judicial practitioners.

[0081] 3. Automatic matching of confusing charge groups, improving the system's automation level: The system can automatically match the corresponding confusing charge groups based on the case's charge or coarse-grained prediction results, without the need for manual comparison of each case; suitable for large-scale batch processing and case handling assistance scenarios.

[0082] 4. Automatic merging and annotation of multiple crime elements, automatic template generation, facilitating expansion and maintenance: The element template is automatically generated by the knowledge graph and merging algorithm, and the fields and candidate values ​​can be automatically adjusted as the knowledge graph is updated; the automatic annotation template converts content / not information into a structure of "optional value + legal explanation", which makes it easier for the model to accurately understand the meaning of each element during extraction; it is easy to extend to other confusing crime groups or other areas of legal application.

[0083] 5. The generated output is constrained by the template, and the post-processing workload is small: The template clearly lists the candidate value set and brief description of each field, guiding the model to select within the predetermined value range; the post-processing mainly involves JSON parsing and candidate value validity verification, without the need for large-scale natural language normalization and complex rule matching, which helps to reduce the overall system complexity.

[0084] 6. The output results can directly serve the determination of guilt and downstream modeling: The output "Confused Crime Group Element Table" is a structured and comparable data representation that can be directly used as feature input for models such as: conviction prediction and conviction anomaly warning; case retrieval and case cluster analysis; research on crime boundaries and scope of application; and sentencing recommendations.

[0085] The technical solution provided by this invention includes the following steps: acquiring basic case information and factual text; automatically matching a confused charge group Gi based on the basic information and factual text; reading the four elements and element metadata of the confused charge group Gi through a knowledge graph (KG); merging multiple charge elements and modeling unified fields to generate an annotated confused charge group template; constructing prompt words and calling a large language model (LLM) for reasoning to extract case elements from the confused charge group template; parsing the LLM output based on the case elements, performing verification, generating and storing a confused charge group element table. This method achieves automatic merging, automatic annotation, and templated extraction of multiple charge element groups.

[0086] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A hierarchical extraction method for legal elements based on large models and knowledge graphs, characterized in that, The method includes: Step 1: Obtain basic information and factual text of the case; Step 2: Based on the basic information and factual text, automatically match the confusing charge group Gi; Step 3: Read the four elements and metadata of the confusing crime group Gi through the knowledge graph KG; Step 4: Merge multiple crime elements and model unified fields to generate annotated templates for confusing crime groups; Step 5: Construct cue words and call the Large Language Model (LLM) for reasoning to extract case elements from the confused charge group template; Step 6: Based on the case elements, parse the LLM output, perform verification, and generate and store the element table of the confused crime group.

2. The method according to claim 1, characterized in that, The basic information in step 1 includes the case number (case_no), the charge (charge), the facts of the crime (FD), and the court's opinion (Hold_that).

3. The method according to claim 2, characterized in that, Step 2 includes: Based on the case charge or the previous coarse-grained classification results, the corresponding confusing charge group Gi is automatically determined by the mapping relationship of [charge → confusing charge group]. If no confusing charge group is matched, the process ends directly or enters other processing paths.

4. The method according to claim 3, characterized in that, Step 3 includes: Using the knowledge graph KG, we read the four elements and element metadata corresponding to each crime in the confused crime group Gi. The read content includes the candidate value set content and annotation note information; the element metadata includes the nature of the behavior, the harmful behavior, the characteristics of the behavior, the knowledge element, and the standard of completion. The knowledge graph KG used is organized as follows: The root node represents the criminal law knowledge graph; the child nodes represent specific crimes. Under the crime node, it is further broken down into: The object of a crime, its objective aspect; the subject of a crime, its subjective aspect; the standard for completion of a crime. Record at the leaf nodes of each element: The content is the set of candidate values ​​for the current element under the current charge; the note is the annotation information for the content of the current element, which is used to supplement the legal connotation and applicable boundaries briefly expressed in the content, and is used to generate the description field in the confusing charge group template.

5. The method according to claim 4, characterized in that, Step 4 includes: The crime element structure is recursively merged. Through multi-crime element merging and field alignment algorithms, the nature of the behavior, harmful behavior, behavior characteristics, known elements, and completed standards are horizontally aligned and deduplicated. For each field, the candidate value set of each crime is retained to form a unified, multi-level element metadata structure. A JSON template is generated based on the element metadata, and annotation information is automatically added to each field in the template to obtain an annotated confusion charge group template for large language model inference.

6. The method according to claim 5, characterized in that, Step 5 includes: The descriptive text, the annotated template for the confusing charge group, and the case fact FD are taken as input and submitted to the large language model. The request is to populate the elements for each charge in the confusing charge group according to the template structure, and to limit the field values ​​to the set of candidate values ​​given in the template annotation.

7. The method according to claim 6, characterized in that, Step 6 includes: Using regular expressions, extract JSON paragraphs or strip markers from the LLM output text, and call the JSON parsing function to convert them into a dictionary structure; perform field integrity checks and candidate value validity checks on the parsing results; organize the results that pass the checks into a confusing crime group element table and save it for subsequent analysis and modeling.

8. The method according to claim 7, characterized in that, Step 6 is followed by downstream applications; The elements table of confused charges will be provided to the given charge / characterization auxiliary analysis module, the similar case retrieval module, and the conviction prediction or sentencing prediction model, respectively, and will be used to determine the charge requirements that a case meets, to search by element similarity, and as feature input.

9. A hierarchical extraction system for legal elements based on large models and knowledge graphs, characterized in that, The system includes: The data preprocessing module is used to obtain basic information and factual text of the case; The automatic matching module for confusing charge groups is used to automatically match the confusing charge group Gi based on basic information and factual text; The knowledge graph management and field metadata extraction module is used to store the criminal law knowledge graph KG and read the four elements and element metadata of the confused crime group Gi through the knowledge graph KG; The module for merging multiple crime elements and confusing crime groups is used for merging multiple crime elements and unifying field modeling, generating annotated confusing crime group templates; The Large Language Model (LLM) prompt construction and reasoning module is used to construct prompt words and call the Large Language Model (LLM) for reasoning to extract case elements from the confused crime group template. The parsing output and legality verification module is used to parse the LLM output based on case elements, perform verification, and generate and store the element table of the confusing crime group.

Images