An intelligent system and method for assisting law enforcement based on large models and three-dimensional evaluation

Abstract

The application discloses an auxiliary law enforcement intelligent system and method based on a large model and three-dimensional evaluation, which comprises: an interactive intelligent agent module of a generative AI engine, which is used for converting unstructured police information into structured and compliant law enforcement suggestions; a three-dimensional evaluation system module, which is used for quantitatively evaluating law enforcement behaviors according to the law enforcement suggestions; a procedure compliance and discretion verification module, which is used for ensuring law enforcement procedure standardization and discretion result legality through a Drools rule engine; and a closed-loop feedback and self-evolution module, which is used for realizing system optimization according to backflow data. The application takes the generative AI engine as an intelligent core, forms a complete closed loop from law enforcement input, intelligent decision-making to effect evaluation and continuous optimization based on a three-dimensional evaluation system and a feedback and evolution mechanism.

Description

Technical Field

[0001] This invention relates to the field of computer information technology, and in particular to an intelligent system and method for assisting law enforcement based on large models and three-dimensional evaluation. Background Technology

[0002] The rapid pace of urbanization has led to a continuous increase in the urban population, posing a significant challenge to railway police enforcement: a relative shortage of manpower and complex enforcement situations. First, frontline enforcement relies heavily on the past experience of officers, resulting in significant differences in how different officers handle the same situation. The improper use of discretionary power leads to a lack of objectivity in enforcement, and procedural violations and inappropriate application of the law occur frequently. Meanwhile, the public's legal awareness is constantly improving, and their demands for standardized and transparent enforcement are increasing. Second, pain points in frontline enforcement, such as parties refusing to cooperate, inadequate response to emergencies, and unreasonable allocation of police resources, exacerbate enforcement conflicts and increase the probability of administrative review, administrative litigation, and complaints. Against this backdrop, leveraging technology to improve the scientific, standardized, and efficient nature of law enforcement has become an urgent need for the high-quality development of railway police enforcement. Summary of the Invention

[0003] This invention provides an intelligent system and method for assisting law enforcement based on large models and three-dimensional evaluation. The system takes a generative AI engine as its core and deeply integrates four core technologies: fine-grained legal entity recognition (BiLSTM-CRF), legal knowledge graph, LangChain multi-turn dialogue framework, and Drools rule engine. It covers a complete closed loop of law enforcement assistance, including "police information input - semantic parsing - knowledge retrieval - legal provision matching - discretionary suggestions - document generation - effect evaluation".

[0004] The technical solution of the present invention is as follows: This invention provides an auxiliary law enforcement intelligent system based on large models and three-dimensional evaluation, comprising: The interactive intelligent agent module of the generative AI engine is used to transform unstructured police information into structured, compliant law enforcement recommendations; The three-dimensional assessment system module is used to quantitatively evaluate law enforcement actions based on enforcement recommendations; The procedural compliance and discretion verification module is used to ensure the standardization of law enforcement procedures and the legality of discretionary results through the Drools rule engine; It also includes a closed-loop feedback and self-evolution module, used to optimize the system based on the feedback data.

[0005] Preferably, the interactive intelligent agent module of the generative AI engine includes a multimodal interaction and dialogue management module, a toolchain orchestration and scheduling module, a legal text retrieval chain module, a legal semantic hub module, and a legal knowledge hub module; The multimodal interaction and dialogue management module is used to receive the "seven Ws" elements of the alarm through the front-end interaction interface; retain the most recent K rounds of interaction according to the LangChain architecture; compress historical information that exceeds the window range into a structured summary; and store the intent classification, extracted entity slots and missing information fields of the current dialogue, which are used to guide subsequent follow-up questions or trigger the next step of the process. The toolchain orchestration and scheduling module is used by the LangChain architecture's Agent mechanism to automatically select and call tools in logical order based on the parsed user intent, and finally integrate multi-source results to generate law enforcement suggestions; the legal text retrieval chain module is used by LangChain to connect the large model with the legal knowledge graph and case vector library. The legal semantic central module is used to transform unstructured text into structured legal facts using the BiLSTM-CRF model; The legal knowledge hub module is used to organize legal knowledge in a graph structure to support intelligent analysis.

[0006] Preferably, the three-dimensional evaluation system module includes a compliance evaluation module, a public acceptance evaluation module, and an implementation effectiveness evaluation module.

[0007] Preferably, the legal semantic central module includes a fine-grained entity recognition module and a structured element output module. The fine-grained entity recognition module is used to capture contextual dependencies through a bidirectional long short-term memory network based on the received police report text sequence, and then optimize the global label sequence through a conditional random field layer to accurately extract key legal entities. The structured element output module is used to output a labeled sequence that directly transforms the police report description into structured data, providing input parameters for legal knowledge graph queries and legal provision matching.

[0008] Preferably, the legal knowledge central module includes a knowledge graph construction module, an automatic clause conflict detection module, and a semantic retrieval and hybrid query module; The knowledge graph construction module is used to organize scattered legal provisions, judicial interpretations, historical cases, and discretionary rules into a reasonable graph structure network. The automatic clause conflict detection module is used to detect whether newly entered legal provisions conflict with their superior laws by automatically comparing them, and to store the conflict relationship in a graph. When a query is performed, if a conflicting clause is involved, the system will actively issue a warning to ensure the accuracy of the application of the law. The semantic retrieval and hybrid query module is used to employ a hybrid strategy of "graph query + vector similarity retrieval" to improve the accuracy of legal provision matching.

[0009] This invention discloses an intelligent method for assisting law enforcement based on large models and three-dimensional assessment, comprising the following steps: Step 1: Transform unstructured police information into structured, compliant law enforcement recommendations; Step 2: Quantitatively evaluate the enforcement actions based on the enforcement recommendations; Step 3: Through the Drools rule engine, ensure the standardization of law enforcement procedures and the legality of discretionary results; Step 4: Optimize the system based on the reflux data.

[0010] Preferably, step 1 includes the following steps: Step 101), receive the seven elements of the alarm through the front-end interactive interface; Step 102): Based on the LangChain architecture, retain the most recent K rounds of interaction, compress historical information that exceeds the window range into a structured summary; store the intent classification, extracted entity slots and missing information fields of the current dialogue, which are used to guide subsequent follow-up questions or trigger the next step of the process; Step 103) The LangChain architecture's Agent mechanism automatically selects and calls tools in logical order based on the parsed user intent, and finally integrates multi-source results to generate law enforcement recommendations; Step 104) uses the BiLSTM-CRF model to transform unstructured text into structured legal facts; Step 105) Organizes legal knowledge in a graph structure to support intelligent analysis.

[0011] Preferably, it also includes step 106), which is used by LangChain to connect the large model with the legal knowledge graph and the case vector library.

[0012] Preferably, step 3 includes: Step 301), Data feedback step: The evaluation score after each law enforcement action is completed, together with the "seven Ws" characteristics of the handled police incident and the system's suggestions, forms a structured feedback record and is stored in the user feedback database; Step 302), Effectiveness Monitoring Step: Track the changing trends of average Accessibility scores and procedural violation rates for various types of incidents; Step 303), Incremental fine-tuning trigger step: When the public acceptance score of a certain type of police situation declines continuously, or when a certain type of procedural violation prompt is frequently ignored, the system will submit the relevant case data to the manual review team for review. After review, the data will be used to optimize the association weight of the LangChain Prompt template or knowledge graph node in that scenario, forming a continuous improvement closed loop of "handling - evaluation - optimization".

[0013] The beneficial effects of the present invention may include: This invention combines intelligent agent development technology with the standardized requirements of various law enforcement scenarios on railway sites, verifies it in actual law enforcement processes, and enables real-time feedback of evaluation data to a large model language library, forming a self-evolution mechanism to continuously optimize law enforcement effectiveness. This invention pioneers a three-dimensional evaluation system of "compliance, public acceptance, and execution effectiveness." This system breaks through the limitations of traditional AI law enforcement, which "emphasizes technology but neglects compliance," and is the first evaluation system in China capable of quantifying the social impact of law enforcement. Compliance is evaluated from three dimensions: legal basis, procedural norms, and data security. Public acceptance is calculated based on a quantitative model (ACCEPTANCE). Execution effectiveness comprehensively considers three indicators: handling efficiency, conflict resolution, and social impact. The weights of each indicator are determined using scientific methods such as the Delphi method and the Satie 1-9 scale. The final score reflects the effectiveness of the law enforcement action, and the scoring data is fed back into the big data language library for self-evolution. This three-dimensional evaluation system enables continuous improvement and enhancement of auxiliary law enforcement functions.

[0014] Currently, frontline law enforcement faces numerous challenges, including insufficient manpower, uneven enforcement capabilities, and a lack of objectivity in law enforcement. Meanwhile, the public's demands for standardized law enforcement are increasing, making traditional methods inadequate for the needs of law enforcement in the new era. AI technology's advantages in legal entity recognition, intelligent analysis, and efficient interaction make it an effective means to address these pain points and improve enforcement effectiveness. This invention verifies technical feasibility, data feasibility, and practical application effects. The performance indicators of each technical module meet the needs of frontline law enforcement. The application of intelligent agents significantly improves enforcement efficiency, reduces procedural violations, and increases public acceptance, providing solid practical support for AI-enabled law enforcement activities.

[0015] General-purpose AI products cannot meet the professional needs of public security law enforcement. This invention focuses on the specificities of law enforcement scenarios, using the "seven Ws" (what, why, why) of police incidents as core inputs and legal norms, on-site handling suggestions, and police equipment suggestions as core outputs, achieving deep integration with frontline law enforcement processes. This invention innovatively combines four major technical modules: BiLSTM-CRF, legal knowledge graph, LangChain, and the Drools rule engine, constructing a dedicated legal knowledge graph and a large model language library. This enables professional and scenario-based law enforcement assistance, effectively solving the problem of the disconnect between traditional AI products and law enforcement practice.

[0016] The synergy of "technology-law-society" is key to improving the effectiveness of AI-assisted law enforcement. AI-enabled law enforcement is not merely a technological application, but a deep integration of technology, law, and society. Neglecting any dimension will affect the actual application effect of the intelligent agent. This research pioneered a three-dimensional evaluation system of "compliance-public acceptance-implementation effect," which achieves quantitative evaluation of the legal, efficiency, and social effects of law enforcement, breaking through the limitations of traditional AI law enforcement that "emphasizes technology over compliance" and "emphasizes effect over acceptance." This system not only scientifically measures the effectiveness of law enforcement but also forms a self-evolution mechanism for the intelligent agent through data feedback, promoting a closed loop of "disposal-evaluation-optimization" in law enforcement work and achieving the synergistic development of technology, law, and society.

[0017] Overall, this invention effectively meets the practical needs of frontline law enforcement. Its technical and evaluation systems are highly innovative and practical, effectively improving the scientific, standardized, and efficient nature of law enforcement and providing a new path for the implementation of the strategy of strengthening policing through science and technology. Its application can not only address many pain points in current frontline law enforcement but also promote the transformation of public security law enforcement towards intelligence, professionalism, and modernization, enhancing the combat effectiveness and public trust in law enforcement teams, and achieving a unity of legal and social effects.

[0018] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings.

[0019] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a structural block diagram of an auxiliary law enforcement intelligent system based on a large model and three-dimensional evaluation according to the present invention. Figure 2 This is a structural block diagram of the interactive intelligent agent module of the generative AI engine of the present invention. Figure 3 This is a structural block diagram of the three-dimensional evaluation system module of the present invention; Figure 4 This is a structural block diagram of the legal knowledge central module of the present invention; Figure 5 This is a flowchart of an embodiment of an intelligent law enforcement assistance method based on a large model and three-dimensional evaluation according to the present invention; Figure 6 This is a flowchart of step 1 of the present invention; Figure 7 This is a flowchart of step 3 of the present invention; Figure 8 This is a flowchart illustrating an embodiment of an intelligent auxiliary law enforcement method based on a large model and three-dimensional evaluation according to the present invention. Detailed Implementation

[0021] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0022] In one embodiment, such as Figure 1 As shown, this embodiment provides an auxiliary law enforcement intelligent system based on a large model and three-dimensional evaluation, including: an interactive intelligent agent module with a generative AI engine, used to transform unstructured police information into structured and compliant law enforcement suggestions; a three-dimensional evaluation system module, used to quantitatively evaluate law enforcement behavior based on the law enforcement suggestions; a procedural compliance and discretion verification module, used to ensure the standardization of law enforcement procedures and the legality of discretion results through the Drools rule engine; and a closed-loop feedback and self-evolution module, used to optimize the system based on feedback data.

[0023] The technical architecture of this invention consists of three layers: Interaction Layer: This layer handles multi-turn dialogue between police officers and the system, supporting multimodal input such as text and voice, and is driven by the LangChain framework for dialogue state management and tool orchestration. Analysis Layer: This layer uses a BiLSTM-CRF model as its core to identify legal entities and the "seven W's" (What, What, Why, What, How) elements, and invokes a legal knowledge graph for associative reasoning and legal provision matching. Verification and Feedback Layer: This layer uses the Drools rule engine to provide real-time early warnings of procedural compliance, and the ACCEPTANCE three-dimensional evaluation system drives continuous system evolution.

[0024] LangChain is the core framework for implementing conversational interaction and multi-component collaboration in the system, and it undertakes the following three main functions: 1. Dialogue State Management In frontline law enforcement scenarios, police officers' descriptions are often fragmented and logically disjointed (e.g., reporting the location first, then adding the person's condition, and finally detailing the tools used). LangChain addresses this issue through the following components: ConversationBufferWindowMemory: Retains the most recent K rounds of interaction (e.g., K=5 rounds) to ensure the contextual coherence of the current conversation; ConversationSummaryMemory: Compresses historical information that exceeds the window range into a structured summary (including time, location, behavioral characteristics, etc.), retaining key information while controlling token consumption; Dialogue state representation: The system maintains a JSON object that stores the intent category, extracted entity slots, and missing information fields of the current dialogue, which is used to guide follow-up questions or trigger the next step of the process.

[0025] 2. Toolchain Orchestration and Scheduling A single question from a police officer may trigger coordinated calls from multiple backend modules. LangChain's Agent + Tools mechanism encapsulates these modules into callable tools: law_query_tool: Receives entity query parameters and calls the knowledge graph to retrieve relevant legal provisions; case_retrieve_tool: Retrieves historical case libraries based on semantic vector similarity; procedure_check_tool: Invokes the Drools rules engine to perform a preliminary compliance audit of the program; report_gen_tool: Generates draft legal documents from structured data; LangChain's ReAct Agent (reasoning-action loop) is responsible for automatically selecting and sequentially calling the aforementioned tools based on user intent, ultimately integrating multi-source results to generate law enforcement recommendation texts that can be directly referenced by police officers.

[0026] 3. Legal provisions search chain LangChain's RetrievalQA Chain connects large models with legal knowledge graphs and case vector libraries. The specific process is as follows: Police officers input natural language questions; The system vectorizes the problem using the Embedding model and retrieves the k most relevant legal entries or case texts from the Milvus vector database; The search results are used as contextual injection into the Prompt template to guide the large model in generating answers with accurate citations and traceable sources.

[0027] The estimated response time for a single query is less than 0.3 seconds, which is about 5 times more efficient than traditional keyword retrieval methods.

[0028] Figure 2 As shown, the interactive intelligent agent module of the generative AI engine includes a multimodal interaction and dialogue management module, a toolchain orchestration and scheduling module, a legal text retrieval chain module, a legal semantic hub module, and a legal knowledge hub module.

[0029] The multimodal interaction and dialogue management module receives the "seven Ws" elements (when, where, who, what, why, what, and what) of police reports input by officers in text or voice form through a front-end interface. It uses `ConversationBufferWindowMemory` to retain the most recent K rounds of dialogue to maintain contextual coherence; and utilizes `ConversationSummaryMemory` to compress historical information and control resource consumption. The system maintains a structured dialogue state representation to track user intent and extracted information slots, guiding subsequent follow-up questions.

[0030] The toolchain orchestration and scheduling module encapsulates underlying functionalities into callable tools, including legal knowledge graph query tools, historical case retrieval tools, procedural compliance pre-review tools, and document generation tools. LangChain's Agent mechanism is responsible for automatically selecting and logically invoking these tools based on the parsed user intent, ultimately integrating multi-source results to generate enforcement recommendations.

[0031] The legal semantic central module includes a fine-grained entity recognition module and a structured element output module. The fine-grained entity recognition module captures contextual dependencies from the received police report text sequence using a bidirectional long short-term memory network, and then optimizes the global label sequence using a conditional random field layer to accurately extract key legal entities. The structured element output module outputs a labeled sequence that directly transforms the police report description into structured data, providing input parameters for legal knowledge graph queries and legal provision matching. The legal semantic central module is fundamental to the system's understanding of police report text and employs a BiLSTM-CRF model to transform unstructured text into structured legal facts. Its functional flow is as follows: Fine-grained entity recognition: The model receives a sequence of police reports, captures contextual dependencies through a bidirectional long short-term memory network, and then optimizes the global label sequence through a conditional random field layer to accurately extract key legal entities such as “behavioral subject”, “illegal elements”, “involved items”, and “time and place”.

[0032] Structured element output: The labeled sequence output by the model directly transforms the police situation description into structured data, providing accurate input parameters for subsequent legal knowledge graph queries and legal provision matching, and solving the problem of recognizing colloquial and messy sentence structures in law enforcement texts.

[0033] BiLSTM-CRF is the core pre-module of the system, enabling the transformation of unstructured police reports into structured legal facts, and is used for fine-grained legal entity recognition (Fine-grained Legal NER). The BiLSTM-CRF model is the core algorithm of the legal semantic hub. Its core structure consists of a bidirectional long short-term memory network (BiLSTM) and a conditional random field (CRF): BiLSTM can bidirectionally capture the semantic context and dependencies of the text, excelling at handling long texts and complex sentences; CRF, as the top-level sequence layer, can learn the logical rules between labels, avoiding incorrect labeling and out-of-order recognition. This model can accurately complete entity extraction, keyword recognition, and element extraction, effectively solving the recognition challenges of messy, colloquial, and non-standard expressions in public security business texts. Compared with ordinary models, it has stronger error correction capabilities, can easily identify key information, reduce invalid and redundant data, and provides core support for knowledge graph construction, realizing the structured transformation of unstructured text. It is a core pre-module algorithm for knowledge graph construction, intelligent classification, and text filtering, replacing manual text element analysis and significantly reducing the manual cost of data processing.

[0034] In one embodiment, based on the needs of frontline law enforcement by public security officers, the BiLSTM-CRF model needs to identify the following key entity types: Example of the meaning of entity category The perpetrator who committed the illegal act was "Zhang Mou". The core elements of a violation are described as "intentionally damaging public or private property." The consequences of the illegal act were "a loss of RMB 2,000". The time element of the violation was "2:00 AM today". Location element: The illegal act occurred in front of Building 3, XX Residential Area. The illegal act involving the item or tool "baseball bat" The victim's or damaged rights: "Li's car" The training corpus comes from historical case files and police reports, and is sequence-annotated by legal professionals (BIO annotation system). The amount of data after annotation is expected to be in the tens of thousands, covering common types of public security and criminal cases.

[0035] BiLSTM-CRF, or Bidirectional Long Short-Term Memory Network-Conditional Random Field model, is a natural language processing model that combines BiLSTM and CRF models. It possesses both the text sequence feature extraction capabilities of BiLSTM and the global optimization capabilities of CRF for sequence labeling. This invention utilizes the BiLSTM-CRF model for fine-grained legal entity recognition. Taking the "seven W's" elements of a police report as sequence input, it accurately extracts key elements from legal provisions, such as the subject of the act, the elements of the violation, the severity of the penalty, and the proposed enforcement measures. This achieves a high accuracy of 87.3% in parsing legal provisions, providing core data support for the application of law and enforcement analysis in this invention. The BiLSTM-CRF model includes: BiLSTM layer: Employs a bidirectional long short-term memory network to capture contextual dependencies in law enforcement text from both the preceding and following directions. The input is a sequence of sentence-level vectors pre-trained with Word2Vec or BERT, and the output is the emission probability of each character on various labels.

[0036] CRF layer: A conditional random field is superimposed on the output of a BiLSTM to model the global transition probability of the label sequence. CRF can effectively constrain illegal label transitions (e.g., "I-time" should not be immediately followed by "B-location"), improving the overall rationality of the labeled sequence.

[0037] Output: The final output is a labeled sequence with entity category labels, which can be directly used for subsequent knowledge graph queries and legal provision matching.

[0038] After training and validation with labeled data, the BiLSTM-CRF model achieved an entity recognition accuracy (F1) of 87.3%, meeting the requirements for high-precision extraction of key elements in law enforcement analysis scenarios. It will continue to be fine-tuned and improved as new data from real cases are added.

[0039] The legal text retrieval chain module is used by LangChain to connect the large model with the legal knowledge graph and case vector library. LangChain's RetrievalQA Chain connects the large model with the legal knowledge graph and case vector library. The specific process is as follows: Police officers input natural language questions; The system vectorizes the problem using an embedding model and retrieves the k most relevant legal entries or case texts from the Milvus vector database. The search results are used as contextual injection into the Prompt template to guide the large model in generating answers with accurate citations and traceable sources.

[0040] The estimated response time for a single query is less than 0.3 seconds, which is about 5 times more efficient than traditional keyword search methods.

[0041] like Figure 4 As shown, the legal knowledge hub module includes a knowledge graph construction module, an automatic clause conflict detection module, and a semantic retrieval and hybrid query module.

[0042] The knowledge graph construction module organizes scattered legal provisions, judicial interpretations, historical cases, and discretionary rules into a reasonable graph structure network. Based on the existing legal system and case resources, the graph construction scale is approximately 12,000 entity relationships, and can be continuously expanded based on the integration of local regulations. Its core node types include: Legal provisions: specific to the article, clause, and item, such as Article 49 of the Law on Penalties for Administration of Public Security; Judicial interpretation milestones: Judicial interpretation documents issued by the Supreme People's Court and the Supreme People's Procuratorate; Enforcement case nodes: historical real cases that have been anonymized; Constituent elements: The core elements constituting an illegal act, such as "subjective intent" and "serious circumstances"; Penalty types and details: such as "warning," "fine," and "administrative detention"; Its core relationship types include: Related terms (RELATED_TO): Terms related within the same law or across laws.

[0043] Basis of higher-level laws (BASED_ON): The basis for the formulation of lower-level laws or local regulations.

[0044] Applied Case (APPLIED_IN): A certain law applies in this case.

[0045] Interpretations of the constituent elements: Judicial interpretations provide specific explanations of the constituent elements of a particular provision.

[0046] Penalty discretionary provisions: The types and ranges of penalties stipulated in the provisions.

[0047] LangChain is an application development framework for large language models. Its core function is to connect and collaborate with external data and tools, supporting multi-turn conversational interactions and intelligent execution of complex tasks. This study uses the LangChain framework to construct a legal text retrieval chain, connecting the large model's language database with a legal knowledge graph and a law enforcement case database. This enables the agent to accurately retrieve relevant legal texts and cases based on natural language information input by law enforcement personnel, achieving context-sensitive multi-turn dialogue, improving interaction efficiency and judgment accuracy. The estimated response time is <0.3s, approximately 5 times faster than traditional retrieval methods.

[0048] LangChain consists of a large model foundation and prompt word templates, RAG (Resource Aggregator), link scheduling and memory modules, tool invocation, and data parsing. It addresses the knowledge lag issue of large models by connecting to local private data, public security business documents, and a self-built knowledge graph to achieve precise question answering, retrieval, and reasoning within a specific domain. Simultaneously, it can connect fragmented data, understand user needs based on context, support multi-turn dialogues, conditional filtering, and related content push, and lightweightly schedule various algorithms and database interfaces, reducing the development difficulty of multi-technology integration. In public security scenarios, this architecture can adapt to specific business application needs, building enhanced retrieval generation capabilities. Its modular design also facilitates later iterative optimization, providing flexible and efficient intelligent interaction support for public security operations.

[0049] The automatic clause conflict detection module is used to detect whether newly entered legal provisions conflict with their superior laws by automatically comparing them, and to store the conflict relationship in a graph. When a query is performed, if a conflicting clause is involved, the system will actively issue a warning to ensure the accuracy of the application of the law. The semantic retrieval and hybrid query module is used to improve the accuracy of legal provision matching by adopting a hybrid strategy of "graph query + vector similarity retrieval".

[0050] During the knowledge graph construction process, the system incorporates automatic detection logic based on the hierarchical rule base of the Legislation Law. By automatically comparing newly entered legal provisions with their superior laws, the system can identify the following conflict types: conflicts where the penalty exceeds the scope authorized by the superior law; conflicts where the characterization of the behavior (criminal / non-criminal) is inconsistent with the superior law; and conflicts regarding the priority of application between special and general provisions. The detection accuracy rate reaches 93.7%, and the detection results are marked as CONFLICT_WITH and stored in the knowledge graph, with proactive warnings provided when police officers query the information.

[0051] The semantic retrieval and hybrid query module employs a hybrid strategy of "graph query + vector similarity retrieval" to improve the accuracy of legal provision matching. Queries: Perform precise structured queries in Neo4j using the Cypher language, such as "find articles related to 'intentional damage to property' and whose penalties are based on the provisions of the Public Security Administration Punishment Law".

[0052] Vector retrieval: Each legal text and case summary generates a 1024-dimensional embedding vector using the bge-large-zh-v1.5 model and stores it in the Milvus vector database. When the police officer's natural language input cannot be accurately mapped to a graph node, semantic-level fuzzy matching is achieved through cosine similarity calculation.

[0053] The Legal Knowledge Hub module organizes legal knowledge in a graph structure to support intelligent analysis. This module forms the knowledge foundation of the entire system, organizing legal knowledge in a graph structure to support intelligent analysis. Its functional flow is as follows:

[0054] Knowledge Fusion and Retrieval: The system performs precise graph query and semantic vector retrieval in parallel. On the one hand, it uses graph query language to accurately search within a knowledge graph composed of nodes and their relationships, such as legal provisions, judicial interpretations, historical cases, and constituent elements. On the other hand, it vectorizes the input question and matches the most relevant legal provisions or cases in the vector database based on semantic similarity.

[0055] Automatic Conflict Warning: The system incorporates a hierarchy of legal force rules based on the "Legislation Law of the People's Republic of China." When a new regulation or local ordinance is entered, the system automatically detects whether it conflicts with higher-level laws and stores the conflict relationships in a graph. When police officers conduct inquiries, if conflicting clauses are involved, the system will proactively issue a warning to ensure the accuracy of legal application.

[0056] like Figure 3 As shown, the three-dimensional evaluation system module is the core innovation of this system, breaking through the limitations of traditional AI applications that emphasize technology while neglecting evaluation. It quantitatively evaluates law enforcement actions from three dimensions: legal, social, and efficiency, forming a closed-loop feedback loop that drives the model's self-evolution. The three-dimensional evaluation system module includes a compliance evaluation module, a public acceptance evaluation module, and an implementation effectiveness evaluation module.

[0057] The compliance assessment module aims to evaluate the substantive and procedural legality of law enforcement actions. It is scored by the head of the legal department or the case handler based on body camera footage and system logs. The assessment functions and standards are as follows: Legal compliance: Assess whether the legal provisions cited in the enforcement process are accurate and appropriate, and whether the application of discretionary power is within the statutory range.

[0058] Procedural compliance: Check whether the statutory procedures of the enforcement action are complete and whether the chain of evidence is complete and without loopholes.

[0059] Data security compliance: Verify whether the system audit logs are complete and traceable to ensure data security during law enforcement.

[0060] The public acceptance assessment module aims to quantify the social impact of law enforcement actions. Based on the ACCEPTANCE model, it allows individuals involved in a police incident to rate the actions afterward via a terminal. The assessment functions and dimensions are as follows: Law enforcement transparency: Assess whether the legal basis for law enforcement is disclosed to the parties involved and whether the algorithmic decision-making is explainable.

[0061] Fairness of law enforcement: Assess whether there are significant differences in the rate of misjudgment among different groups in law enforcement actions.

[0062] Enforcement participation and redress: Assess whether public feedback channels are accessible and whether the parties involved have clear and accessible guidance on redress avenues such as administrative review and litigation.

[0063] The execution effectiveness evaluation module aims to measure the efficiency and effectiveness of law enforcement actions, and is conducted by law enforcement officers themselves after handling a police incident. The evaluation functions and indicators are as follows: Response efficiency: Assess the time and resources consumed from receiving the alarm to the completion of the incident response.

[0064] Conflict resolution: Assess whether the on-site conflict has been effectively resolved and whether the parties involved have reached a preliminary agreement.

[0065] Social impact: Assess the potential subsequent social impact of this enforcement action and its outcome, and whether it represents the best course of action.

[0066] A closed-loop feedback and self-evolution module is used to optimize the system based on feedback data. This includes: Data feedback and monitoring: After each enforcement action, the evaluation score, incident characteristics, and system suggestions are compiled into a structured feedback record and stored in the user feedback database. The system control panel monitors the average scores and violation rates of various incidents in real time.

[0067] Incremental model optimization: When the system detects a continuous decline in scores for a specific scenario or frequent neglect of certain risk warnings, it triggers an optimization process. Relevant cases, after manual review, are used to fine-tune the prompt word templates of the interactive agent, optimize the node weights of the knowledge graph, or update the rule base, forming a continuous improvement loop from "handling" to "evaluation" to "optimization," enabling the system to adapt to the dynamic development of law enforcement practices.

[0068] In one embodiment, the Drools rule engine is the core verification module that ensures the standardization of law enforcement procedures and the legality of discretionary results. It operates independently of the generative AI engine and conducts compliance reviews on all output suggestions. The Drools rule engine adopts a production-oriented rule architecture, with its core consisting of a rule base, working memory, inference engine, conflict scheduling, and executor. In public security business scenarios, it can transform vague business standards into machine-recognizable hard rules, enabling automatic screening, intelligent analysis, and risk warning. It supports the modification, addition, and decommissioning of rules at any time, adapting to policy and business changes without code modifications. Simultaneously, it can perform multi-condition combination filtering, compliance verification, and anomaly identification on data, outputting standardized judgment results. This meets the requirements of standardized law enforcement, closed-loop management, and traceable results, playing a crucial role in business standardization, lightweight operation and maintenance, and closed-loop processes. It effectively solves the problems of human judgment bias and experience-based office work, reduces reliance on technology, and aligns with the compliance needs of public security operations.

[0069] Based on regulations such as the "Procedures for Handling Administrative Cases by Public Security Organs", the "Procedures for Handling Criminal Cases by Public Security Organs", and the "Administrative Enforcement Law of the People's Republic of China", this invention extracts 18 common procedural violation scenarios and converts them into Drools rules (.drl files).

[0070] Example rules: Rule 1: Identification must be made before inspection. ``` rule "Check_Show_Identity" when $case : Case(isInspection == true, showIdentity == false) then $case.addViolation("Work identification must be presented before conducting on-site inspections"); end ``` Rule 2: Penalties shall not exceed the scope of discretion. ``` rule "Penalty_Within_Range" when $advice:Advice(penaltyDays>legalMaxDays||penaltyAmount>egalMaxAmount ) then $advice.setWarning("This discretionary recommendation exceeds the scope of legal authority; please review it again"); end ``` The Drools rule engine operates in two modes: 1) Real-time mode (streaming review): During the use of the intelligent agent by police officers, the Drools rule engine embeds an interactive link to pre-verify each suggestion output by the intelligent agent. If a violation rule is triggered, the system will provide a clear risk warning on the front end, but will not prevent the police officer from making the final decision. 2) Batch mode (post-event sampling inspection): Historical law enforcement records and system operation logs are periodically scanned to generate compliance analysis reports for review and training reference by the legal department.

[0071] This invention uses the Drools rule engine to ensure the standardization of law enforcement procedures and the legality of discretionary results; its procedural violation rate is expected to decrease from the current 18.7% to about 2.3%, and the number of administrative reconsideration and administrative litigation is expected to decrease by about 12%.

[0072] The three-dimensional evaluation system of this invention is the core innovation that distinguishes the "Smart Law Engine" from other general-purpose legal AIs. It consists of three dimensions: legality, public acceptance, and enforcement effectiveness. Legality assessment (weight 0.4): includes five secondary indicators: matching degree of law enforcement basis, reasonableness of discretionary power, completeness rate of statutory procedures, completeness of evidence chain and completeness of system audit logs. The full score is 100 points, and the score is given by the head of the legal department or the person in charge of the case based on the law enforcement recorder video and system logs.

[0073] Public acceptance assessment (weight 0.3): calculated based on the ACCEPTANCE model, i.e., ACCEPTANCE = 0.35 × law enforcement transparency + 0.25 × law enforcement fairness + 0.2 × law enforcement participation + 0.2 × law enforcement remedies, with a maximum score of 100 points. The score is given by the parties involved in the incident through a terminal afterward.

[0074] Law enforcement effectiveness evaluation (weight 0.3): includes three indicators: handling efficiency (0.3), degree of conflict resolution (0.35), and social impact (0.35), with a full score of 100 points. It is a self-evaluation conducted by law enforcement officers after the incident has been handled.

[0075] The final comprehensive score W = 0.4 × legality + 0.3 × public acceptance + 0.3 × enforcement effectiveness, and the corresponding score range is given as the rating level (Excellent 90-100, Good 80-89, Satisfactory 70-79, Needs Improvement 60-69, Unsatisfactory below 60).

[0076] This invention achieves closed-loop feedback and system optimization based on reflux data, while also enabling effect monitoring. Data Reflux: The evaluation score after each enforcement action, along with the "seven Ws" characteristics of the handled incident and system suggestions, forms a structured feedback record, stored in the user feedback database. Effect Monitoring: The system has a built-in monitoring panel that tracks the average acceptance score and procedural violation rate trends for various types of incidents. Incremental Fine-tuning Trigger: When the public acceptance score for a certain type of incident (such as "noise disturbance disputes") continuously declines, or when a certain type of procedural violation prompt is frequently ignored, the system submits relevant case data to a manual review team for verification. This data is then used to optimize the association weights of the LangChain Prompt template or knowledge graph nodes in that scenario, forming a continuous improvement closed loop of "handling—evaluation—optimization".

[0077] This invention begins with data input and builds a core capability foundation based on a large-scale model language library. This library synchronously drives three core modules to work collaboratively: First, a compliance validator (Drools rule engine) performs rule matching and compliance review of the data, ensuring that business processes comply with law enforcement standards. Second, an interactive intelligent agent (LangChain architecture) provides contextual understanding, multi-turn interaction, and business reasoning capabilities to support intelligent business processing. Third, a legal semantic hub (BiLSTM-CRF model) completes text element extraction and semantic parsing, achieving structured governance of unstructured data. The processing results from these three modules converge to form standardized data output, which is then compared and verified with historical data from the police information system. Cross-validation with historical data ensures the accuracy and consistency of the results, ultimately generating reliable business data and forming a complete business closed loop of "input - large-scale model driving - multi-engine collaboration - data output - historical verification - final output".

[0078] In one embodiment, the present invention discloses an intelligent method for assisting law enforcement based on large models and three-dimensional assessment, comprising the following steps: Step 1: Transform unstructured police information into structured, compliant law enforcement recommendations; Step 2: Quantitatively evaluate the enforcement actions based on the enforcement recommendations; specifically, this includes outputting enforcement decisions based on the enforcement recommendations, and conducting a three-dimensional quantitative evaluation of the output enforcement decisions, including compliance assessment, public acceptance assessment, and implementation effectiveness assessment. Step 3: Through the Drools rule engine, ensure the standardization of law enforcement procedures and the legality of discretionary results; Step 4: Optimize the system based on the reflux data.

[0079] like Figure 6 As shown, step 1 includes the following steps: Step 101), receive the seven elements of the alarm through the front-end interactive interface; Step 102): Based on the LangChain architecture, retain the most recent K rounds of interaction, compress historical information that exceeds the window range into a structured summary; store the intent classification, extracted entity slots and missing information fields of the current dialogue, which are used to guide subsequent follow-up questions or trigger the next step of the process; Step 103) The LangChain architecture's Agent mechanism automatically selects and calls tools in logical order based on the parsed user intent, and finally integrates multi-source results to generate law enforcement recommendations; Step 104) uses the BiLSTM-CRF model to transform unstructured text into structured legal facts; Step 105) Organize legal knowledge in a graph structure to support intelligent analysis; Step 106), LangChain connects the large model with the legal knowledge graph and case vector library.

[0080] like Figure 7 As shown, step 3 includes: Step 301), Data feedback step: The evaluation score after each law enforcement action is completed, together with the "seven Ws" characteristics of the handled police incident and the system's suggestions, forms a structured feedback record and is stored in the user feedback database; Step 302), Effectiveness Monitoring Step: Track the changing trends of average Accessibility scores and procedural violation rates for various types of incidents; Step 303), Incremental fine-tuning trigger step: When the public acceptance score of a certain type of police situation declines continuously, or when a certain type of procedural violation prompt is frequently ignored, the system will submit the relevant case data to the manual review team for review. After review, the data will be used to optimize the association weight of the LangChain Prompt template or knowledge graph node in that scenario, forming a continuous improvement closed loop of "handling - evaluation - optimization".

[0081] like Figure 8 As shown below, this is a complete use case demonstrating how the various technical modules work together: Scenario: A police report is received stating that "someone is causing trouble at a restaurant while intoxicated, smashing tables and chairs."

[0082] 1. Dialogue Input and Intent Understanding (LangChain): Police officers input police reports via voice or text. LangChain manages the dialogue context, receives and forwards questions. 2. Entity Recognition (BiLSTM-CRF): The model identifies the "seven W's": subject of the action (unclear), location (restaurant), action (disorderly behavior, vandalism), and items involved (tables and chairs), etc. 3. Legal provision retrieval and graph reasoning (knowledge graph): The system simultaneously performs precise Cypher query and vector similarity retrieval, matching Article 26 (picking quarrels and provoking trouble) and Article 49 (intentionally damaging property) of the Public Security Administration Punishment Law, and extracts related discretionary standards and similar cases from the graph; 4. Procedure Compliance Verification (Drools): The rule engine checks whether there are corresponding procedure requirements, such as the procedure specifications for restrictive sobering measures; 5. Suggestion Generation and Adjudication (LangChain Integration): All search results and verification results are aggregated into structured law enforcement suggestions, which are presented on the police officer's terminal, including applicable legal provisions, reference for discretionary range, suggestions on the number of officers to be dispatched, safety tips, etc. 6. Evaluation and Feedback (Three-Dimensional System): After the incident is handled, relevant parties score it, and the scores, along with the case data, are entered into the feedback database for the next training and optimization.

[0083] This invention is a software system built on generative AI technology, possessing autonomous interaction, intelligent judgment, and precise output capabilities. Guided by law enforcement needs, this invention can engage in multi-round dialogues with law enforcement personnel through natural language, receiving input information such as the "seven Ws" elements of a police report. Relying on legal knowledge graphs and large-scale model language libraries, it performs intelligent judgment and quickly outputs professional services such as legal norm citations, law enforcement guidance, police equipment recommendations, and legal document generation. Furthermore, it can achieve data feedback and self-evolution through a three-dimensional evaluation system. Possessing the core characteristics of scenario-based, professional, and closed-loop systems, it serves as a core carrier connecting AI technology and law enforcement practice.

[0084] The data sources of this invention mainly include three categories: legal and regulatory provisions, historical cases of frontline law enforcement, and police incident handling data. Legal and regulatory provisions can be legally obtained through web crawling technology from the websites of the Ministry of Public Security and the Supreme People's Court, and then standardized. Historical cases of frontline law enforcement and police incident handling data will be obtained through cooperation with local public security organs by signing confidentiality agreements, while employing federated learning technology to ensure that the original data does not leave the local servers of the public security organs, meeting data security and confidentiality requirements. For sensitive information, differential privacy and PETS privacy enhancement technologies are used for anonymization, strictly adhering to the "minimum necessary principle" in data collection to ensure the legality and feasibility of data collection. Furthermore, by opening up the large database corpus input structure, users are invited to supplement and update data based on local law enforcement realities, further addressing the issues of data maintenance and updates.

[0085] The ACCEPTANCE model, constructed in this study, is a computational model for quantifying public acceptance of law enforcement. Its core is to comprehensively assess public acceptance of law enforcement work from four dimensions: transparency, fairness, participation, and access to redress. The model's calculation formula is as follows: ACCEPTANCE = 0.35 × Law Enforcement Transparency Score + 0.25 × Law Enforcement Fairness Score + 0.2 × Law Enforcement Participation Score + 0.2 × Law Enforcement Remedy Score. The weights of each dimension were determined using the Satie-9 scale combined with questionnaire survey results, with a score range of 0-100. This model transforms the abstract concept of public acceptance into quantifiable and concrete indicators, objectively reflecting the social impact of law enforcement work and providing a scientific basis for evaluating law enforcement effectiveness.

[0086] The three-dimensional evaluation system, a pioneering quantitative law enforcement evaluation system in this embodiment, is one of the core innovations of the research: a three-dimensional evaluation system of "compliance-public acceptance-implementation effect." This system breaks through the limitation of traditional AI law enforcement, which "emphasizes technology but neglects compliance." Compliance is evaluated from three dimensions: legal basis, procedural norms, and data security. Public acceptance is quantitatively calculated based on the ACCEPTANCE model. Implementation effect comprehensively considers three indicators: handling efficiency, conflict resolution, and social impact. The weights of each indicator are determined using scientific methods such as the Delphi method and the Satie 1-9 scale. Evaluation data is fed back to the large model language library in real time, forming the self-evolution mechanism of this invention and achieving continuous optimization of law enforcement effectiveness. The calculation is shown in Formula 1.

[0087] W total score = 0.4 × legality score + 0.3 × public acceptance score + 0.3 × enforcement effectiveness score.

[0088] The legality scoring involves inviting heads of legal departments or public security case-handling departments to review and score videos recorded by body cameras. Scores are based on the legal compliance of the actual police situation and the prompts generated by the AI ​​agent, the compliance of the law enforcement procedures, and the compliance of data security. Based on these three points, this embodiment proposes the indicators shown in Table 1.

[0089] Table 1 Three-dimensional evaluation index table

[0090] The above indicators were generated through brainstorming and survey experiments. This embodiment invited leaders from frontline police investigation departments and law professors from police academies to discuss and generate the above indicators, focusing on the emphasis and content elements of legality. This embodiment first determined the indicator content: legal compliance, procedural compliance, and data security compliance. Because law enforcement involves substantive and procedural justice, based on practical operational needs, this embodiment involved 10 people brainstorming based on the weights of the three. The 10 people assigned scores (0-100, unit: %) to the three indicators and then calculated the average, finally obtaining W: Legal Compliance = 52, Procedural Compliance = 42, and Data Security Compliance = 6. Further discussion resulted in Legal Compliance = 50, Procedural Compliance = 40, and Data Security Compliance = 10. This embodiment comprehensively sets a scoring standard (range 0-100) for each secondary indicator, ultimately forming the legality score W.

[0091] Public acceptance scoring involves inviting parties involved in police incidents to rate their actions. The calculation model is a weighted average of the ACCEPTANCE model proposed in this embodiment, i.e., ACCEPTANCE = 0.35 × law enforcement transparency score + 0.25 × law enforcement fairness score + 0.2 × law enforcement participation score + 0.2 × law enforcement redress score. Table 2 shows the scoring criteria based on the ACCEPTANCE calculation model.

[0092] Table 2. ACCEPTANCE Scoring Criteria

[0093] The above indicators were generated using brainstorming, survey experiments, and the Saadi 1-9 scale. This embodiment invited non-police workers to participate in discussions, conducted research from the perspective of those subject to law enforcement, and collected data to generate the above indicators. Based on questionnaire results and the "Regulations on Procedures for Handling Administrative / Criminal Cases by Public Security Organs" and the "Standards for Receiving and Handling Police Calls by People's Police of Public Security Organs," the indicator content was first determined: law enforcement transparency, law enforcement fairness, law enforcement participation, and law enforcement remedies. Based on this, this embodiment established an ACCEPTANCE scoring model using the Saadi 1-9 scale, deriving the following order: Law Enforcement Transparency > Law Enforcement Fairness > Law Enforcement Fairness = Law Enforcement Participation. By assigning scores (0-100, unit: %) to the four indicators and then calculating the average, the final results were: W Law Enforcement Transparency = 35, W Law Enforcement Fairness = 25, W Law Enforcement Participation = 20, and W Law Enforcement Remedies = 20. Finally, a scoring standard was formed by distributing questionnaires to 50 respondents, and a public acceptance score was generated through calculation.

[0094] The scoring of law enforcement effectiveness mainly involves inviting law enforcement officers to give scores, which comprehensively consider three aspects: law enforcement efficiency, law enforcement quality, and social impact. Based on this, the indicators shown in Table 3 are proposed.

[0095] The above indicators were generated through brainstorming and survey experiments. In this embodiment, six frontline police officers were invited to participate in discussions, and their handling of police cases over the past three months was tracked and reconstructed. The above indicators were generated through brainstorming. In this embodiment, the brainstorming process first determined the indicator content: handling efficiency, degree of conflict resolution, and the degree of social impact caused by subsequent events in the later stages of the case. Based on this, this invention assigns scores (0-100, unit: %) to the three indicators and then calculates the average, finally obtaining handling efficiency = 30, degree of conflict resolution = 35, and social impact = 35. Finally, the law enforcement effect can be calculated using the formula: Law Enforcement Effect = 0.3 × Handling Efficiency Score + 0.35 × Conflict Resolution Score + 0.35 × Social Impact Score.

[0096] The final score W can be generated using the above three-dimensional indicators. Based on the score of W, the evaluation criteria are defined as shown in Table 4.

[0097] The final scoring system reflects the effectiveness of the enforcement action, and the scoring data is fed back into a big data lexicon for self-evolution. Therefore, this three-tiered evaluation system is crucial for the continuous improvement and enhancement of auxiliary enforcement functions.

[0098] Empirical tests were conducted on the core technical modules of this invention. The results show that the performance of each technical module meets the expected requirements: the BiLSTM-CRF model achieves a legal entity recognition accuracy of 87.3%, accurately extracting key elements from legal provisions to meet the needs of law enforcement analysis; the legal knowledge graph achieves a clause conflict detection rate of 93.7%, effectively identifying contradictions between local regulations and higher-level laws to ensure the accuracy of legal application; the LangChain framework has a natural language query response time of <0.3s, improving interaction efficiency by 5 times compared to traditional retrieval methods, enabling efficient interaction with law enforcement personnel; the Drools rule engine can provide real-time warnings for 18 types of procedural violations, significantly enhancing the compliance supervision of law enforcement procedures. Meanwhile, the overall technical effect of the coordinated operation of various technical modules is good. The intelligent agent can complete the analysis of the police situation and output law enforcement suggestions in a short time. From a technical perspective, it effectively solves the problems of mechanized application of law and lagging procedural supervision. The application effect of the three-dimensional evaluation system is verified through simulated law enforcement experiments. Ten typical police situations were selected, and law enforcement officers handled them using traditional methods and the intelligent agent system of this invention. The three-dimensional evaluation system was used to quantitatively score the effectiveness of the two methods. The results show that the law enforcement method using this invention has an average comprehensive score of 86.5 points, rated as good, while the traditional law enforcement method has an average comprehensive score of 62.3 points, rated as needing improvement. In terms of compliance, the procedural violation rate of the intelligent agent method is only 2.1%, far lower than the 19.2% of the traditional method; in terms of public acceptance, the average ACEPTANCE model score of the intelligent agent method is 88.2 points, significantly higher than the 60.5 points of the traditional method; in terms of execution effect, the intelligent agent method reduces the police situation handling time by an average of 7.3%, and the on-site conflict resolution rate reaches 95%, far higher than the 78% of the traditional method. Empirical results show that the three-dimensional evaluation system can scientifically quantify the effectiveness of law enforcement, and relying on the closed-loop feedback mechanism of the system, it can effectively promote the continuous optimization of law enforcement effectiveness.

[0099] This invention can supplement the legal framework in cases where frontline law enforcement officers may lack legal recourse, thus providing a solid basis for enforcement and gaining the trust of those involved. It also analyzes similar incidents based on historical information to arrive at more realistic conclusions, avoiding rigid enforcement due to over-application of the law. Secondly, for emergencies, it uses historical information as a foundation, employing the "seven Ws" (Who, What, Why, Why) to make predictions about the scene, providing advance guidance for law enforcement officers. It effectively addresses both insufficient police force and a high volume of incidents, offering reasonable solutions for police deployment through recommendations, thereby better defining responsibilities, reducing decision-making pressure, improving incident response efficiency, reducing response time, and increasing rest time.

[0100] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0101] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An intelligent law enforcement assistance system based on large models and three-dimensional evaluation, characterized in that: include: The interactive intelligent agent module of the generative AI engine is used to transform unstructured police information into structured, compliant law enforcement recommendations; The three-dimensional assessment system module is used to quantitatively evaluate law enforcement actions based on enforcement recommendations; The procedural compliance and discretion verification module is used to ensure the standardization of law enforcement procedures and the legality of discretionary results through the Drools rule engine; It also includes a closed-loop feedback and self-evolution module, used to optimize the system based on the feedback data.

2. The intelligent law enforcement assistance system based on large model and three-dimensional evaluation as described in claim 1, characterized in that: The interactive intelligent agent module of the generative AI engine includes a multimodal interaction and dialogue management module, a toolchain orchestration and scheduling module, a legal text retrieval chain module, a legal semantic hub module, and a legal knowledge hub module; The multimodal interaction and dialogue management module is used to receive the "seven Ws" elements of the alarm through the front-end interaction interface; retain the most recent K rounds of interaction according to the LangChain architecture; compress historical information that exceeds the window range into a structured summary; and store the intent classification, extracted entity slots and missing information fields of the current dialogue, which are used to guide subsequent follow-up questions or trigger the next step of the process. The toolchain orchestration and scheduling module is used by the LangChain architecture's Agent mechanism to automatically select and call tools in logical order based on the parsed user intent, and finally integrate multi-source results to generate law enforcement recommendations. The legal text retrieval chain module is used by LangChain to connect the large model with the legal knowledge graph and case vector library; The legal semantic central module is used to transform unstructured text into structured legal facts using the BiLSTM-CRF model; The legal knowledge hub module is used to organize legal knowledge in a graph structure to support intelligent analysis.

3. The intelligent law enforcement assistance system based on large model and three-dimensional evaluation as described in claim 1, characterized in that: The three-dimensional evaluation system module includes a compliance evaluation module, a public acceptance evaluation module, and an implementation effectiveness evaluation module.

4. The intelligent law enforcement assistance system based on large model and three-dimensional evaluation as described in claim 1, characterized in that: The legal semantic central module includes a fine-grained entity recognition module and a structured element output module. The fine-grained entity recognition module is used to capture contextual dependencies through a bidirectional long short-term memory network based on the received police report text sequence, and then optimize the global label sequence through a conditional random field layer to accurately extract key legal entities. The structured element output module outputs a labeled sequence that directly transforms the police report description into structured data, providing input parameters for legal knowledge graph queries and legal provision matching.

5. The intelligent law enforcement assistance system based on large model and three-dimensional evaluation as described in claim 1, characterized in that: The legal knowledge central module includes a knowledge graph construction module, an automatic clause conflict detection module, and a semantic retrieval and hybrid query module; The knowledge graph construction module is used to organize scattered legal provisions, judicial interpretations, historical cases, and discretionary rules into a reasonable graph structure network. The automatic clause conflict detection module is used to detect whether newly entered legal provisions conflict with their superior laws by automatically comparing them, and to store the conflict relationship in a graph. When a query is performed, if a conflicting clause is involved, the system will actively issue a warning to ensure the accuracy of the application of the law. The semantic retrieval and hybrid query module is used to employ a hybrid strategy of "graph query + vector similarity retrieval" to improve the accuracy of legal provision matching.

6. The intelligent auxiliary law enforcement method based on large model and three-dimensional evaluation as described in claim 1, characterized in that: Includes the following steps: Step 1: Transform unstructured police information into structured, compliant law enforcement recommendations; Step 2: Quantitatively evaluate the enforcement actions based on the enforcement recommendations; Step 3: Through the Drools rule engine, ensure the standardization of law enforcement procedures and the legality of discretionary results; Step 4: Optimize the system based on the reflux data.

7. The intelligent auxiliary law enforcement method based on large model and three-dimensional evaluation as described in claim 6, characterized in that: It also includes the following steps. Step 1 includes the following steps: Step 101), receive the "seven questions" elements of the alarm through the front-end interaction interface; Step 102): Based on the LangChain architecture, retain the most recent K rounds of interaction, compress historical information that exceeds the window range into a structured summary; store the intent classification, extracted entity slots and missing information fields of the current dialogue, which are used to guide subsequent follow-up questions or trigger the next step of the process; Step 103) The LangChain architecture's Agent mechanism automatically selects and calls tools in logical order based on the parsed user intent, and finally integrates multi-source results to generate law enforcement recommendations; Step 104) uses the BiLSTM-CRF model to transform unstructured text into structured legal facts; Step 105) Organizes legal knowledge in a graph structure to support intelligent analysis.

8. The intelligent auxiliary law enforcement method based on large model and three-dimensional evaluation as described in claim 6, characterized in that: It also includes step 106, which is used by LangChain to connect the large model with the legal knowledge graph and the case vector library.

9. The intelligent auxiliary law enforcement method based on large model and three-dimensional evaluation as described in claim 6, characterized in that: Step 3 includes: Step 301), Data feedback step: The evaluation score after each law enforcement action is completed, together with the "seven Ws" characteristics of the handled police incident and the system's suggestions, forms a structured feedback record and is stored in the user feedback database; Step 302), Effectiveness Monitoring Step: Track the changing trends of average Accessibility scores and procedural violation rates for various types of incidents; Step 303), Incremental Fine-tuning Trigger Step: When the public acceptance score of a certain type of police incident continues to decline, or when a certain type of procedural violation prompt is frequently ignored, the system will submit the relevant case data to the manual review team for review. After review, the data will be used to optimize the association weight of the LangChain Prompt template or knowledge graph node in that scenario, forming a continuous improvement closed loop of "handling - evaluation - optimization".

Images