A knowledge graph-based intelligent analysis system for criminal procedure

Abstract

This invention relates to the field of judicial data processing, specifically a knowledge graph-based intelligent analysis system for criminal proceedings. The system includes a multi-source data acquisition module, an entity and relation extraction module, a knowledge fusion and ontology construction module, a temporal and evidence chain graph construction module, a rule base and interpretable reasoning engine module, a risk warning and compliance detection module, and a visualization and interaction module. In this invention, a temporal evidence chain graph with temporal uncertainty and evidence support strength values ​​is constructed by fusing multi-source, multimodal data. Interpretable reasoning is performed using a hierarchical graph pattern that includes necessary, optional enhancement, and exclusionary subgraph patterns, automatically generating compliance verification results and risk warning reports. The system possesses cross-case correlation analysis, federated learning privacy protection, and user feedback-driven online learning capabilities, improving the accuracy, transparency, and security of intelligent criminal proceedings analysis throughout the entire process.

Description

Technical Field

[0001] This invention relates to the field of judicial data processing, and in particular to an intelligent analysis system for criminal proceedings based on knowledge graphs. Background Technology

[0002] With the development of judicial big data and artificial intelligence technologies, an increasing number of studies and applications are representing legal knowledge as knowledge graphs to assist in legal retrieval, question answering, and decision support. For example, existing technologies have constructed legal knowledge graphs and mapped user questions to semantic graphs, answering legal queries by matching subgraphs with the legal knowledge graph. Methods for constructing knowledge graphs based on criminal judicial documents have also been disclosed, achieving structured extraction of sentencing factors from massive amounts of judgment documents through natural language processing, entity relation extraction, and knowledge fusion. Furthermore, technologies utilize graph neural networks to perform relational reasoning on criminal case knowledge graphs to discover overlooked case details. Meanwhile, evidence chain analysis technology has also become a research hotspot in judicial assistance: for example, CN120163229A discloses an evidence chain generation method based on historical case knowledge graphs, verifying and supplementing the evidence chain of the current case by comparing standard temporal rules and evidence event timestamps.

[0003] However, the aforementioned technologies often have limitations: first, they are limited to textual data and lack the fusion of multimodal information (such as audio and video); second, they mainly focus on knowledge graph construction or single-mode reasoning, failing to fully utilize temporal evidence chains and interpretable rule-based reasoning; and third, they lack cross-case correlation analysis and privacy protection mechanisms. Therefore, an innovative system is needed to combine multi-source, multimodal data with ontology from the professional field of criminal cases, and to achieve intelligent analysis and visualization of criminal proceedings through comprehensive technologies such as knowledge graph construction, temporal modeling, rule-based reasoning, and risk warning. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing an intelligent analysis system for criminal proceedings based on knowledge graphs.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: an intelligent analysis system for criminal procedure based on knowledge graphs, comprising:

[0006] The multi-source data acquisition module is used to collect multimodal data from heterogeneous judicial data sources, extract timestamps and perform time-series indexing and encoding to generate structured case data streams;

[0007] The entity and relation extraction module is used to extract case entities and semantic relations using a domain-fine-tuned entity and relation joint extraction model, and generate a primary structured knowledge set.

[0008] The knowledge fusion and ontology construction module is used to perform entity alignment and disambiguation based on the criminal procedure ontology template, merge identical entities, and construct a knowledge graph.

[0009] The time sequence and evidence chain graph construction module is used to extract event and evidence entities, perform constraint propagation reasoning on fuzzy timestamps to determine the feasible time interval, generate multi-branch time sequence with uncertainty labeling, and calculate the evidence support strength of directed edges according to the evidence type weight and the number of corroborating sources to generate a time sequence evidence chain graph.

[0010] The rule base and interpretable reasoning engine module stores hierarchical graph patterns with necessary, optional and exclusive subgraph patterns. It generates reasoning paths by matching necessary patterns with the evidence chain graph, calculates path confidence based on the strength of edge evidence support, and outputs primary and backup reasoning paths and compliance verification results.

[0011] The risk warning and compliance detection module is used to identify risk points based on compliance verification results, calculate and sort risk indices, and generate risk warning reports.

[0012] The visualization and interaction module is used to transform knowledge graphs, evidence chain graphs, and early warning reports into graphical interactive interfaces and receive user feedback.

[0013] As a further aspect of the present invention, the multi-source data acquisition module includes:

[0014] The multimodal data acquisition submodule is used to acquire the original multimodal data in parallel from multiple data sources, including at least a court judgment database, a police information system, and on-site evidence collection equipment, through a preset authorized data interface.

[0015] The spatiotemporal index module is used to extract the creation timestamp and geographic coordinate information embedded in the original multimodal data, encode them into spatiotemporal metadata, and attach them to the corresponding data fragments.

[0016] The formatting processing submodule is used to perform unified encoding format conversion on the text data in the original multimodal data, and to perform key event segment detection and text description generation on the audio and video data through a pre-trained audio and video analysis model, and to aggregate them into the structured case data stream.

[0017] As a further aspect of the present invention, the formatting processing submodule further includes:

[0018] The speaker's voiceprint feature vector is extracted from the audio stream of the audio and video data, and the visual feature vector of the key frame is extracted from the video stream. The voiceprint feature vector and the visual feature vector are stored as attributes of the corresponding evidence entity in the criminal procedure domain knowledge graph, so that the rule base and the interpretable reasoning engine module can perform cross-modal evidence corroboration reasoning based on the attributes.

[0019] As a further aspect of the present invention, the knowledge fusion and ontology construction module includes:

[0020] The ontology template alignment submodule is used to load the preset criminal procedure domain ontology template, which defines the core concept hierarchy and its attribute constraints.

[0021] The semantic disambiguation submodule is used to calculate the first similarity between each entity in the primary structured knowledge set and the concept in the ontology template, as well as the second similarity between different entities. Based on a preset similarity threshold, multiple entity names pointing to the same objective object are associated with a globally unique identifier to form disambiguated entities.

[0022] The knowledge fusion submodule is used to deduplicatize and merge attribute information and relation triples related to the disambiguated entity from different data sources based on the globally unique identifier, and store the merged entities and relations into a graph database to construct the knowledge graph of the criminal procedure domain.

[0023] As a further aspect of the present invention, in the rule base and the interpretable reasoning engine module:

[0024] The necessary subgraph pattern defines the minimum node-edge combination that must be completely matched in the temporal evidence chain graph;

[0025] The optional enhanced subgraph pattern defines an additional node-edge structure that can improve the confidence of the corresponding inference path after a successful match;

[0026] The exclusion subgraph pattern defines a node-edge structure in which the corresponding reasoning path is automatically truncated if it is matched in the temporal evidence chain graph.

[0027] The reasoning path starts with the factual nodes of the case, passes through the evidence nodes and is supported by directed connecting edges with evidence support strength values, leading to the legal conclusion nodes, forming a complete interpretable causal reasoning chain.

[0028] As a further aspect of the present invention, the system also includes a cross-case fusion module, used for:

[0029] For each entity node in the aforementioned knowledge graph of criminal procedure, a semantic fingerprint vector is generated based on its core attribute set;

[0030] Identify entity node pairs in the temporal evidence chain graph from different cases whose semantic fingerprint similarity exceeds a preset threshold, and mark the entity node pairs as candidate anchor entities;

[0031] For the conflicting attribute values ​​of the candidate anchor entities in different cases, the conflict is automatically resolved based on the source credibility level and time sequence of the attribute values, and the attribute values ​​with higher credibility are retained.

[0032] Using the resolved anchor entities as connection points, the temporal evidence chain graphs of multiple cases are spliced ​​together to generate a unified temporal evidence chain graph across cases. Subgraph pattern matching across case boundaries is then performed on the unified temporal evidence chain graph to discover related reasoning paths that cannot be identified in a single case graph.

[0033] As a further aspect of the present invention, the risk warning and compliance detection module includes:

[0034] The pattern matching submodule is used to pre-store the program violation risk pattern library, compare the compliance verification results and the knowledge graph in the criminal procedure field with the patterns in the program violation risk pattern library, and mark the problem nodes.

[0035] The risk index calculation submodule is used to perform weighted calculations on each of the problem nodes based on the weight of its associated program type, the duration of program delay, and the strength of evidence support of the associated evidence chain nodes, to obtain a quantitative risk index.

[0036] The priority sorting and completion suggestion submodule is used to sort all the problem nodes in descending order according to the quantitative risk index, and at the same time identify the missing necessary procedure nodes in the current process based on the pre-stored compliance process standard template, and generate the risk warning report containing suggestions for automatic completion of missing procedures.

[0037] As a further aspect of the present invention, the system further includes:

[0038] The security and privacy protection module integrates a federated learning architecture. When multiple judicial data participants collaboratively build a shared knowledge graph, it controls each participating node to execute the construction of sub-knowledge graphs locally and performs cross-domain interaction through encrypted parameters or desensitized model gradient information to complete ontology alignment and knowledge fusion operations.

[0039] The training and online learning module is used to capture user feedback instructions received by the visualization interaction module, use the user feedback instructions as supervision signals to incrementally update the parameters of the entity relationship joint extraction model, and dynamically adjust the matching weights of necessary subgraph patterns, optional enhanced subgraph patterns, and excluded subgraph patterns in the inference rule template.

[0040] As a further aspect of the present invention, when the computer program is executed by the processor, it implements the functions of each module in the knowledge graph-based intelligent analysis system for criminal proceedings as described in any one of claims 1 to 8.

[0041] As a further aspect of the present invention, an electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, when the processor executes the computer program, it implements the functions of each module in the knowledge graph-based intelligent analysis system for criminal proceedings as described in any one of claims 1 to 8.

[0042] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0043] This invention integrates voiceprint and visual feature vectors extracted from audio and video streams into a knowledge graph as entity attributes, achieving deep semantic alignment and automatic corroboration reasoning for cross-modal evidence, effectively expanding the information dimension of fact-finding. The system employs a hierarchical reasoning engine with necessary, optional enhancement, and exclusionary subgraph patterns, automatically truncating conflicting paths while constructing an interpretable causal reasoning chain from case facts through evidence chain nodes to legal conclusions, significantly improving the transparency and credibility of analytical conclusions. Simultaneously, the integration of a federated learning architecture and a user-feedback-driven online learning mechanism ensures privacy, security, and compliance under multi-institutional data collaboration, while continuously transforming expert feedback into supervisory signals, driving dynamic iterative optimization of the underlying model and reasoning rule weights. This gives the system continuous self-evolution capabilities, improving the accuracy and practicality of intelligent criminal litigation analysis throughout the entire process. Attached Figure Description

[0044] Figure 1 This is a system flowchart of the present invention. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0046] Please see Figure 1 A knowledge graph-based intelligent analysis system for criminal procedure includes:

[0047] Implementation of the multi-source data acquisition module: The system accesses various data sources in parallel (such as the national criminal judgment document database, local public security police data, and on-site law enforcement video equipment) through pre-defined authorized interfaces to acquire multimodal raw data of the target case. For the acquired text, audio, and video data, the spatiotemporal indexing module extracts and normalizes the embedded timestamps and geographic coordinates, appending them to the corresponding data segments as spatiotemporal metadata. The formatting processing module performs unified encoding conversion on the text data, converting heterogeneous text (WORD, PDF, TXT, etc.) to UTF-8 encoding; for video data, it uses a pre-trained video event detection model (such as CNN+LSTM-based security behavior detection) to extract keyframes and generate text descriptions; for audio streams, it uses a speaker recognition model to extract voiceprint features. Through these steps, the multimodal raw data is converted into a structured data stream with spatiotemporal labels. For example, from a crime scene video, the system identifies key action segments, extracts the corresponding time and location, and generates a text description of "the suspect showing a dagger to the victim." From police recordings of witness interviews, the system extracts speaker characteristics from the audio and transcribes key information. Ultimately, the structured case data stream aggregates various evidentiary information in a unified data format and is indexed and encoded chronologically, providing a rich information foundation for downstream analysis modules.

[0048] Implementation of the Entity and Relation Extraction Module: This module takes structured case data streams as input and uses a joint entity and relation extraction model fine-tuned based on criminal procedure corpora for knowledge extraction. For example, a pre-trained language model (BERT, RoBERTa, or a model specific to the judicial domain) is fine-tuned using corpora such as criminal case documents and investigation records to identify entity types (e.g., "suspect," "victim," "evidence," "time point") and semantic relationships between entities (e.g., "involved," "possess," "occurred in," etc.). During implementation, the model first segments and encodes the text data, then applies sequence labeling and relation classification techniques to extract "entity-relation-entity" triples. For example, for the sentence "On December 1, 2021, suspect Zhang committed theft at victim Li's home," the model extracts the following triples: <case event, occurrence time, December 1, 2021>, <suspect Zhang, committed, theft>, <victim Li, victim, case>. These initial triples constitute a preliminary structured knowledge set. For audio and video descriptions, they can be converted into text through speech recognition and video description before extraction. Simultaneously, the speaker's voiceprint and visual feature vectors extracted from the audio and video streams can be stored as additional attributes of the corresponding entities, supporting subsequent cross-modal evidence corroboration reasoning. This module outputs preliminary knowledge data containing all extracted entities and their relationships, providing input for the knowledge fusion module.

[0049] The implementation of the knowledge fusion and ontology construction module: This module performs semantic alignment and disambiguation on primary structured knowledge based on a pre-set ontology template for the criminal procedure domain (containing core concept levels and attribute constraints such as "case", "litigation participants", "evidence materials", and "statutory procedures"). Specifically, the ontology template alignment submodule first loads the subject ontology structure of relevant laws and regulations and cases in the judicial field, such as the Criminal Law and typical cases, and matches entities with ontology concepts; the semantic disambiguation submodule calculates the similarity between each entity and ontology concept in the primary knowledge set (using entity name, attributes, edit distance of context, or word vector similarity) and the similarity between entities, and sets a threshold to determine if they are the same entity. For example, if "Zhang San" in document 1 and "Zhang Sanmou" in document 2 are highly similar in attributes and context, they are considered to refer to the same objective person; at the same time, if the same fingerprint evidence appears with slight differences in marking in different documents, it is also unified into a "disambiguated entity" through matching. The knowledge fusion submodule uses globally unique identifiers (UUIDs) to deduplicate and merge attributes and triples related to entities from different sources that have already been disambiguated. For example, it merges records of the same victim's information appearing in police systems and court documents; and it integrates multiple "suspect-time-place" relationships into a single spatiotemporal connection. The merged entities and relationships are stored in a graph database (such as Neo4j), ultimately constructing a complete knowledge graph for the criminal procedure domain. This graph comprehensively describes the factual elements of a case and their semantic connections through entity nodes and relation edges, providing a structured foundation for subsequent temporal analysis and reasoning.

[0050] Implementation of the Time Sequence and Evidence Chain Graph Construction Module: This module receives the constructed criminal procedure knowledge graph and performs time sequence modeling on event entities and evidence entities that carry time information. For each event entity (e.g., "case filing," "service of indictment," "trial commencement") and evidence entity (e.g., "physical evidence")... Witness testimony First, its timestamp attribute is extracted. If some events only have a rough time (e.g., "Winter 2021"), then constraint propagation reasoning is performed based on other related entities (e.g., the specific time of evidence collection): the possible time interval of the event is inferred through the event relationship network. In a preferred embodiment, the constraint propagation reasoning adopts a method based on... This document describes the implementation of a constraint propagation algorithm based on Allen's time interval algebra. Allen's time interval algebra defines 13 basic relationships between two time intervals, including "before," "after," "contains," "during," "overlaps," "overlapped by," "starts," "started by," "finishes," "finished by," "equals," "meets," and "metby." The system first establishes a time constraint network based on known relationships between events (such as the program-defined order and causal relationships), representing the time interval of each event as a tuple. ,in The earliest possible time, This is the latest possible time. The specific reasoning process is as follows: Step 1, for each event entity carrying a precise timestamp (such as "Case filing date: November 15, 2021"), initialize its time interval to a precise value. For event entities with only a rough time description (such as "time of the crime: winter of 2021"), initialize them to a wide interval according to a preset time granularity template (such as...). Step two: Traverse all relation edges between event entities and determine the Allen relation based on the relation type. For example, there is a "before" relation between "filing a case" and "investigation" (filing a case must begin before the investigation starts), and there is an "after" or "meets" relation between "scene investigation" and "time of the crime". Step three: For each pair of event entities with an Allen relation constraint, according to... The composition table of relations performs interval shrinking. Specifically, if events... The time interval is ,event The time interval is ,and "before" Then, after constraint propagation, the update is: ;like "during" Then, after constraint propagation, the update is: and Step four: Iterate through steps two and three until the time intervals of all event entities no longer change (i.e., the constraint fixed point is reached), or the preset maximum number of iterations is reached. After constraint propagation reasoning, each event entity with insufficient timestamp precision obtains a feasible time interval. The system further calculates the timing uncertainty index. This is defined as the ratio of the length of the feasible time interval to the total time span of the case, i.e. ,in The time span of a case from the earliest event to the latest event. The smaller the value, the more precise the time of the event. A higher value indicates higher time uncertainty. When multiple constraint paths impose time constraints on the same event, the intersection of the feasible intervals derived from each path is taken as the final feasible time interval. If the intersection is empty (i.e., constraint conflict), the system marks the event as a "time constraint conflict" state and retains the intervals derived from each path as alternative feasible time intervals, while reducing the confidence of the relevant reasoning paths for the event. For example, if "evidence extraction" is connected to the precise "time of the crime," this time can be passed and the time interval of "evidence extraction" can be restricted. After this processing, an event sequence with time uncertainty labeling is generated. Subsequently, according to the procedural order defined by the ontology (e.g., "case filing → investigation → prosecution → trial") and the causal relationship between entities, directed connection edges are created in the graph to form a directed time sequence graph. For each directed edge, the evidence support strength value is calculated as an edge attribute based on the type weight of the evidence entity connected by the edge and the number of independent corroborating sources. In a preferred embodiment, the evidence support strength value... The formula for calculating (EvidenceSupportValue) is as follows:

[0051] .

[0052] in: The weight for the evidence type is obtained from a preset evidence type weight table based on the evidence type to which the evidence entity belongs. The number of independently corroborating sources refers to the number of evidentiary entities that point to the same factual conclusion and whose sources are independent of each other. This is the reliability coefficient of evidence, assessed based on the source of the evidence (e.g., evidence obtained by state authorities is more reliable than evidence provided by individuals) and the integrity of the evidence's preservation. Its value ranges from [value missing]. The aforementioned evidence type weighting table is pre-set according to the rules of evidence in criminal procedure. For example, the weight of physical evidence is 0.95, documentary evidence is 0.90, expert opinions are 0.85, audiovisual materials are 0.80, witness testimony is 0.60, the suspect's confession and defense are 0.50, and the victim's statement is 0.55. These weight values ​​can be configured and adjusted according to specific application scenarios.

[0053] The number of independent corroborating sources The calculation rule is as follows: Count the number of evidence entities pointing to the same target factual node that have different source institutions or different collection times. For example, if the fact that "the suspect was at the scene" is supported by three independent sources: "on-site fingerprints" (source: forensic identification center), "surveillance video" (source: security system), and "eyewitness testimony" (source: police station record), then... For example: Suppose there is a directed edge connecting "on-site fingerprint evidence" to the fact node "the suspect was at the scene". The evidence type of this physical evidence is "physical evidence". Looking up the table... This fact is supported by three independent sources. The fingerprint was extracted and preserved by the National Forensic Science Center according to standard procedures, and its reliability coefficient is [not specified]. .but The system normalizes this value to... The interval is used as the final edge attribute value, and the normalization method is as follows: ,Right now The resulting temporal evidence chain diagram is a directed graph with time attributes and confidence level labels. It not only presents the possible chronological order of events but also reflects the strength of support between different pieces of evidence, providing rich spatiotemporal and confidence information for subsequent reasoning.

[0054] Implementation of the rule base and interpretable reasoning engine module: This module predefines multiple graph pattern reasoning rules, designed based on criminal procedures and evidence logic. Each rule is divided into three sub-patterns: necessary subgraph pattern (core node-edge structure required for case fact reasoning), optional enhanced subgraph pattern (additional structure that increases path confidence upon successful matching), and exclusionary subgraph pattern (conflicting structure that proves the path is unreasonable if matched). For example, a necessary causal relationship can be defined between "the defendant's arrival at the scene" and "placement of physical evidence," while "break in the chain of evidence" can be used as an exclusionary pattern. Once the chronological evidence chain graph is generated, the system uses a graph matching algorithm to perform subgraph matching between the necessary subgraph pattern of each rule and the graph to identify factual relationships. In a preferred embodiment, the subgraph matching adopts a method based on... An improved subgraph isomorphic matching method for the algorithm. The algorithm is a depth-first search subgraph isomorphism determination algorithm that completes most of the matching task in polynomial time by maintaining a state space and feasibility pruning rules. Considering the special characteristics of the temporal evidence chain graph in this scheme, the standard... The algorithm is improved in the following ways: First, node type constraint matching. In the algorithm's feasibility assessment step, a node type consistency check is added: a mapping relationship is only allowed if a candidate node in the graph and its corresponding node in the rule subgraph pattern belong to the same or compatible entity type (e.g., a "suspect" type node can only match a "perpetrator" type node in the rule). Secondly, edge attribute threshold filtering is implemented. During the matching process, a threshold filter is applied to the evidence support strength value of directed connected edges in the graph: only when the edge's... An edge only participates in subgraph matching when its value exceeds a preset minimum matching threshold (default is 0.3). This filtering mechanism effectively eliminates noisy edges with weak supporting evidence, improving matching efficiency and result quality. Third, temporal consistency constraints. For subgraph patterns involving temporal relationships, an additional check is performed during matching to ensure that the feasible time intervals of candidate event nodes in the graph meet the constraints. Relationship constraints. For example, if the rule subgraph pattern requires events... "before" incident Then, during matching, verification is performed. Is the upper bound of the feasible time interval of B no greater than the lower bound of the feasible time interval of B? Fourth, the matching strategy of hierarchical subgraph patterns. For each inference rule template, the system first performs matching of necessary subgraph patterns; after successful matching, it then performs matching of optional enhanced subgraph patterns—if the optional pattern also matches successfully, the comprehensive confidence of the corresponding inference path is multiplied by an enhancement coefficient (default is 1.2, upper limit is 1.0); finally, it performs matching of exclusionary subgraph patterns—if the exclusionary pattern matches successfully, the inference path is marked as a "conflicting path" and automatically truncated, and does not participate in the subsequent confidence ranking. When a match is successful, the matched actual entities and relationships are mapped to graph patterns to generate an inference path (e.g., from "time node of the crime" through "existence of physical evidence" to "guilty conclusion"). If there are multiple inference paths, the system uses a weighted algorithm (such as product or weighted sum) to calculate the comprehensive confidence of each path based on the evidence support strength value of the edges on each path, and distinguishes between the "main inference path" (highest confidence) and the "alternative inference path". Simultaneously, the system determines process compliance based on whether the endpoint of the path meets legal conclusions, automatically generating compliance verification results and a visualized reasoning chain. This module's reasoning process not only provides chain-like logical analysis but also displays the credibility of evidence at each step of the path, enhancing interpretability. For example, the main path in the diagram might show "Zhang was present → fingerprints found at the scene → fingerprints matched the suspect," with relevant evidence strengths marked. If a path is excluded from pattern matching (e.g., contradictory statements are found), that path will automatically be interrupted and marked as invalid.

[0055] Implementation of the Risk Warning and Compliance Detection Module: This module includes a pattern matching submodule, a risk index calculation submodule, and a priority ranking and completion suggestion submodule. The pattern matching submodule pre-maintains a database of procedural violation risk patterns (such as common issues like "exceeding bail period" and "missing evidence collection steps"), compares compliance verification results and knowledge graph content with the risk patterns, and marks potential risk nodes (e.g., failure to proceed according to procedural steps at a certain point in time). The risk index calculation submodule quantifies and weights each problem node according to its corresponding procedural type weight (important procedures have higher weight), procedural delay duration, and the strength of supporting evidence, obtaining a risk index value. In a preferred embodiment, the risk index... The calculation formula is as follows:

[0056] .

[0057] in: This is the problem node; The procedural type weight reflects the importance of this procedural stage in the criminal proceedings. It is obtained from a pre-set procedural type weight table. For example, the procedural weight of the investigation stage is 0.8, the prosecution review stage is 0.9, and the trial stage is 1.0. The program dwell time is defined as the difference between the current time and the time that the program node should complete (if it is negative, it is 0). The maximum allowed duration of detention is preset, and the value varies depending on the type of procedure (e.g., the detention period for investigation is the corresponding statutory time limit). This represents the average strength of evidence support across all evidence chain edges associated with the problematic node, reflecting the sufficiency of the current evidence support. Its value ranges from [value missing]. ; , , The preset weighting coefficients satisfy... In one implementation, take =0.4, =0.35, =0.25. (Calculated value) The value range is A higher value indicates a higher risk. The system calculates based on... The value categorizes risk into three levels: High risk (red alert) It is classified as medium risk (yellow alert). Low risk (blue alert). For example: Suppose the "review and prosecution" stage in a case has exceeded the time limit by 15 days ( =15 days), the statutory period for this stage is 45 days ( =45 days), program type weight =0.9, average support strength of related evidence =0.6. Take =0.4, =0.35, =0.25, then The issue is classified as medium risk. The priority sorting submodule sorts all problem nodes in descending order of risk index and automatically identifies missing necessary procedural nodes by combining pre-stored compliance process standard templates. For example, if it is found that the subject of investigation has not been transferred for review and prosecution for a long time, the template suggests that steps such as "notification of prosecution" may be missing. Finally, a risk warning report is output, which lists each risk node, risk index, relevant evidence, and supplementary suggestions (such as "suggestion to supplement missing investigation report documents"). Through this mechanism, investigators can quickly locate procedural violations and receive targeted corrective suggestions.

[0058] Implementation of the Visual Interaction Module: This module transforms knowledge graphs, evidence chain graphs, and risk reports into a graphical interface for users. The system can create interactive graphs based on graph databases and web technologies (such as the front-end D3.js visualization library), allowing users to view detailed information such as case entity nodes, reasoning paths, and evidence content. For risk reports, the interface displays high-risk aspects and improvement suggestions in list or flowchart format. The module also receives user feedback instructions on the visualization results (such as confirming the correctness or exclusion of a reasoning path) and transmits this feedback back to the backend, providing monitoring signals for the online learning module.

[0059] Implementation of the cross-case fusion module: To discover connections that are difficult to identify in a single case, this system supports cross-case analysis. Specifically, for each entity node in the knowledge graph, a semantic fingerprint vector is generated based on its core attribute set. In a preferred embodiment, the core attribute set is defined differently according to the entity type. For entities of the "litigation participant" type, its core attribute set includes: name (or former name), gender, date of birth, ID card number, and biometric encoding (such as fingerprint encoding, DNA encoding); for entities of the "evidence material" type, its core attribute set includes: evidence type, evidence number, extraction time, extraction location, and evidence content summary; for entities of the "case event" type, its core attribute set includes: case type, time of incident, location of incident, charge involved, and case keywords. The semantic fingerprint vector is generated by normalizing the attribute values ​​in the entity's core attribute set and concatenating them into an attribute vector, which is then mapped to a fixed-dimensional semantic embedding space through a pre-trained attribute encoder. Specifically, for textual attributes (such as name and case keywords), pre-trained text embedding models (such as Transformer-based sentence encoders) are used to generate text embedding vectors; for structured attributes (such as date of birth and time of incident), they are encoded into normalized numerical vectors; for biometric features (such as fingerprints), their original feature vectors are used directly. The embedding vectors of the above attributes are then weighted and concatenated to obtain the semantic fingerprint vector of the entity. ,in For fingerprint vector dimensions (e.g.) =256). During weighted concatenation, the weights of different attributes are pre-set according to their distinguishing ability. For example, the weight of the ID card number is higher than that of the name (because names may be duplicated), and the weight of the biometric code is higher than that of the birth date. The semantic fingerprint similarity is calculated using cosine similarity:

[0060] ;

[0061] when When the value exceeds a preset threshold (default 0.85), the corresponding entity node pair is marked as a candidate anchor entity. The system compares the similarity of entity fingerprints in different cases, and entity pairs exceeding the threshold are marked as candidate anchor entities. For candidate pairs, if there are conflicts in different cases (such as inconsistencies in the age or identity information of the same person in two trials), they are automatically resolved according to the credibility level and time sequence of each source: generally, attribute values ​​from higher-level agencies or more recent materials are retained. Then, using the resolved entities as bridges, the temporal evidence chain graphs of multiple cases are spliced ​​at the anchor points to generate a unified temporal evidence chain graph across cases. Graph pattern matching reasoning is re-executed on this unified graph to discover multi-case related reasoning paths that cannot be captured in a single case. For example, if the same physical evidence or suspect shows similar paths in different cases, the system can alert investigators to possible case-related connections.

[0062] Implementation of the Security and Privacy Protection Module: Considering scenarios involving multiple judicial data participants, this system adopts a federated learning architecture to protect sensitive information. Each participating node (such as different provincial procuratorates or public security bureaus) constructs a sub-knowledge graph locally, synchronizing ontology alignment and knowledge fusion information with the central server through parameter encryption or desensitization gradient methods, avoiding direct exchange of raw data. While ensuring local privacy, this achieves shared ontology consistency and graph merging. The training and online learning module is responsible for continuously optimizing the model using user feedback from visual interactions. Specifically, the system uses user operations and annotations (such as confirmation and negation of inference paths) as supervisory signals, incrementally updating the parameters of the entity relation joint extraction model to better suit actual application scenarios; simultaneously, it dynamically adjusts the matching weights of necessary, optional, and excluded patterns in the inference rule template based on feedback, achieving continuous adaptation of the model and rules.

[0063] Each of the above modules can be implemented by a computer program stored in a computer-readable storage medium and run on a general-purpose hardware platform. The various functions of the system are decoupled through modular design. In specific implementation, the above modules can be divided into multiple sub-devices or deployed as software components on servers or terminals.

[0064] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A knowledge graph-based intelligent analysis system for criminal procedure, characterized in that, include: The multi-source data acquisition module is used to collect multimodal data from heterogeneous judicial data sources, extract timestamps and perform time-series indexing and encoding to generate structured case data streams; The entity and relation extraction module is used to extract case entities and semantic relations using a domain-fine-tuned entity and relation joint extraction model, and generate a primary structured knowledge set. The knowledge fusion and ontology construction module is used to perform entity alignment and disambiguation based on the criminal procedure ontology template, merge identical entities, and construct a knowledge graph. The time sequence and evidence chain graph construction module is used to extract event and evidence entities, perform constraint propagation reasoning on fuzzy timestamps to determine the feasible time interval, generate multi-branch time sequence with uncertainty labeling, and calculate the evidence support strength of directed edges according to the evidence type weight and the number of corroborating sources to generate a time sequence evidence chain graph. The rule base and interpretable reasoning engine module stores hierarchical graph patterns with necessary, optional and exclusive subgraph patterns. It generates reasoning paths by matching necessary patterns with the evidence chain graph, calculates path confidence based on the strength of edge evidence support, and outputs primary and backup reasoning paths and compliance verification results. The risk warning and compliance detection module is used to identify risk points based on compliance verification results, calculate and sort risk indices, and generate risk warning reports. The visualization and interaction module is used to transform knowledge graphs, evidence chain graphs, and early warning reports into graphical interactive interfaces and receive user feedback.

2. The knowledge graph-based intelligent analysis system for criminal procedure as described in claim 1, characterized in that, The multi-source data acquisition module includes: The multimodal data acquisition submodule is used to acquire the original multimodal data in parallel from multiple data sources, including at least a court judgment database, a police information system, and on-site evidence collection equipment, through a preset authorized data interface. The spatiotemporal index module is used to extract the creation timestamp and geographic coordinate information embedded in the original multimodal data, encode them into spatiotemporal metadata, and attach them to the corresponding data fragments. The formatting processing submodule is used to perform unified encoding format conversion on the text data in the original multimodal data, and to perform key event segment detection and text description generation on the audio and video data through a pre-trained audio and video analysis model, and to aggregate them into the structured case data stream.

3. The knowledge graph-based intelligent analysis system for criminal proceedings according to claim 2, characterized in that, The formatting processing submodule further includes: The speaker's voiceprint feature vector is extracted from the audio stream of the audio and video data, and the visual feature vector of the key frame is extracted from the video stream. The voiceprint feature vector and the visual feature vector are stored as attributes of the corresponding evidence entity in the criminal procedure domain knowledge graph, so that the rule base and the interpretable reasoning engine module can perform cross-modal evidence corroboration reasoning based on the attributes.

4. The knowledge graph-based intelligent analysis system for criminal procedure as described in claim 1, characterized in that, The knowledge fusion and ontology construction module includes: The ontology template alignment submodule is used to load the preset criminal procedure domain ontology template, which defines the core concept hierarchy and its attribute constraints. The semantic disambiguation submodule is used to calculate the first similarity between each entity in the primary structured knowledge set and the concept in the ontology template, as well as the second similarity between different entities. Based on a preset similarity threshold, multiple entity names pointing to the same objective object are associated with a globally unique identifier to form disambiguated entities. The knowledge fusion submodule is used to deduplicatize and merge attribute information and relation triples related to the disambiguated entity from different data sources based on the globally unique identifier, and store the merged entities and relations into a graph database to construct the knowledge graph of the criminal procedure domain.

5. The knowledge graph-based intelligent analysis system for criminal proceedings according to claim 1, characterized in that, In the rule base and interpretable reasoning engine module: The necessary subgraph pattern defines the minimum node-edge combination that must be completely matched in the temporal evidence chain graph; The optional enhanced subgraph pattern defines an additional node-edge structure that can improve the confidence of the corresponding inference path after a successful match; The exclusion subgraph pattern defines a node-edge structure in which the corresponding reasoning path is automatically truncated if it is matched in the temporal evidence chain graph. The reasoning path starts with the factual nodes of the case, passes through the evidence nodes and is supported by directed connecting edges with evidence support strength values, leading to the legal conclusion nodes, forming a complete interpretable causal reasoning chain.

6. The knowledge graph-based intelligent analysis system for criminal proceedings according to claim 1, characterized in that: The system also includes a cross-case fusion module, used for: For each entity node in the aforementioned knowledge graph of criminal procedure, a semantic fingerprint vector is generated based on its core attribute set; Identify entity node pairs in the temporal evidence chain graph from different cases whose semantic fingerprint similarity exceeds a preset threshold, and mark the entity node pairs as candidate anchor entities; For the conflicting attribute values ​​of the candidate anchor entities in different cases, the conflict is automatically resolved based on the source credibility level and time sequence of the attribute values, and the attribute values ​​with higher credibility are retained. Using the resolved anchor entities as connection points, the temporal evidence chain graphs of multiple cases are spliced ​​together to generate a unified temporal evidence chain graph across cases. Subgraph pattern matching across case boundaries is then performed on the unified temporal evidence chain graph to discover related reasoning paths that cannot be identified in a single case graph.

7. The knowledge graph-based intelligent analysis system for criminal procedure as described in claim 1, characterized in that: The risk warning and compliance detection module includes: The pattern matching submodule is used to pre-store the program violation risk pattern library, compare the compliance verification results and the knowledge graph in the criminal procedure field with the patterns in the program violation risk pattern library, and mark the problem nodes. The risk index calculation submodule is used to perform weighted calculations on each of the problem nodes based on the weight of its associated program type, the duration of program delay, and the strength of evidence support of the associated evidence chain nodes, to obtain a quantitative risk index. The priority sorting and completion suggestion submodule is used to sort all the problem nodes in descending order according to the quantitative risk index, and at the same time identify the missing necessary procedure nodes in the current process based on the pre-stored compliance process standard template, and generate the risk warning report containing suggestions for automatic completion of missing procedures.

8. The knowledge graph-based intelligent analysis system for criminal procedure as described in claim 1, characterized in that, The system also includes: The security and privacy protection module integrates a federated learning architecture. When multiple judicial data participants collaboratively build a shared knowledge graph, it controls each participating node to execute the construction of sub-knowledge graphs locally and performs cross-domain interaction through encrypted parameters or desensitized model gradient information to complete ontology alignment and knowledge fusion operations. The training and online learning module is used to capture user feedback instructions received by the visualization interaction module, use the user feedback instructions as supervision signals to incrementally update the parameters of the entity relationship joint extraction model, and dynamically adjust the matching weights of necessary subgraph patterns, optional enhanced subgraph patterns, and excluded subgraph patterns in the inference rule template.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the functions of each module in the knowledge graph-based intelligent analysis system for criminal proceedings as described in any one of claims 1 to 8.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the functions of each module in the knowledge graph-based intelligent analysis system for criminal proceedings as described in any one of claims 1 to 8.

Images