Method for assisting administrative law enforcement case document entry by using large language model

Abstract

The invention relates to the technical field of data processing. The method for assisting the entry of the administrative law enforcement case document by using the large language model comprises the following steps: carrying out legal constraint embedding processing on an original multi-modal feature based on a preset administrative litigation evidence rule knowledge graph to obtain a multi-modal feature with a legal constraint label; performing legal fact fragment generation processing on the multi-modal features with the legal constraint labels to obtain semantic description fragments with legal attribute labels; performing evidence three-property dynamic verification processing on the semantic description fragment with the legal attribute tag to obtain a verification result with a risk type tag; and performing legal document synthesis processing on the verification result with the risk type mark, and generating a standard case document with the legal risk mark and an independent evidence defect report, so as to improve the integrity of an evidence chain, reduce the omission ratio of program defects and avoid the evidence effectiveness risk of the document.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a method for assisting administrative law enforcement case document input by using a large language model. BACKGROUND

[0002] With the deep application of artificial intelligence technology in the field of judicial administration, the law enforcement case document automatic generation system based on a large language model is gradually replacing the traditional manual recording mode.

[0003] The traditional technology adopts a multi-modal separation scheme for independently processing video and text evidence, but this method does not establish a cross-modal correlation verification mechanism, and there is a risk of evidence chain rupture; when embedding legal rules, the traditional technology adopts a static identification means of regular expression matching keywords, but this method cannot dynamically quantify the procedural legality, resulting in difficulty in identifying implicit procedural defects; in addition, the traditional technology relies on a static rule base for evidence three-property verification, but this method is difficult to handle complex scenarios such as missing physical evidence chain and text record contradiction, multi-modal time sequence deviation, etc., resulting in insufficient effectiveness of document evidence. SUMMARY

[0004] Therefore, it is necessary to provide a method for assisting administrative law enforcement case document input by using a large language model to improve the completeness of the evidence chain, reduce the procedural defect detection rate, and avoid the risk of document evidence effectiveness.

[0005] In a first aspect, the present application provides a method for assisting administrative law enforcement case document input by using a large language model, which comprises:

[0006] performing visual time sequence analysis and processing on the input law enforcement record video stream to obtain visual action sequence features; performing text analysis and processing on the input associated text data to obtain structured text features;

[0007] performing multi-modal fusion processing on the visual action sequence features and the structured text features to obtain original multi-modal features; based on a pre-set administrative litigation evidence rule knowledge graph, performing legal constraint embedding processing on the original multi-modal features to obtain multi-modal features with legal constraint labels;

[0008] performing legal fact segment generation processing on the multi-modal features with legal constraint labels to obtain semantic description segments with legal attribute labels;

[0009] performing evidence three-property dynamic verification processing on the semantic description segments with legal attribute labels to obtain verification results with risk type labels;

[0010] performing legal document synthesis processing on the verification results with risk type labels to generate standard case documents with legal risk annotations and independent evidence defect reports.

[0011] In an embodiment, the semantic description segment with the legal attribute label is subjected to evidence tripartite dynamic checking processing to obtain a checking result with a risk type label, including:

[0012] The semantic description segment with the legal attribute label is subjected to program behavior matching processing based on a preset legality rule set using the following formula to generate a legality risk label:

[0013]

[0014] wherein, represents the legality risk label, represents the total number of legality rules, represents the legal weight coefficient of rule , and represents the characteristic matrix of the semantic description segment with the legal attribute label, represents the program behavior template vector of rule , and represents a nonlinear activation function, represents a legal semantic distance function, represents the legal behavior threshold vector of rule , and represents the rule violation judgment threshold of rule , and represents an indicator function, represents a transpose;

[0015] The semantic description segment with the legal attribute label is subjected to language neutrality analysis processing based on a preset objectivity rule set to generate an objectivity risk label;

[0016] The semantic description segment with the legal attribute label is subjected to evidence chain integrity verification processing based on a preset relevance rule set to generate a relevance flaw label;

[0017] The legality risk label, the objectivity risk label, and the relevance flaw label are subjected to dynamic weight fusion processing to generate the checking result with the risk type label.

[0018] In an embodiment, the semantic description segment with the legal attribute label is subjected to evidence chain integrity verification processing based on a preset relevance rule set to generate a relevance flaw label, including:

[0019] The event time sequence in the semantic description segment with the legal attribute label is subjected to logical gap detection processing based on a preset time relevance rule using the following formula to generate a time chain break label:

[0020] wherein, Indicates a time-broken chain marker. Indicates the start timestamp of the event sequence. Indicates the end timestamp of the event sequence. Indicates time Event description vector, Indicates the time interval between adjacent events. Indicates the threshold of time correlation. Indicates an indicator function;

[0021] Based on preset behavioral evidence association rules, the matching degree analysis of law enforcement actions in semantic description fragments and corresponding physical evidence descriptions is performed to generate physical evidence missing markers.

[0022] Keyframe back-tracing is performed on the visual action sequence features to obtain a video feature description;

[0023] Text element extraction processing is performed on semantic description fragments with legal attribute tags to obtain text record descriptions;

[0024] The following formula is used to perform consistency verification between video feature descriptions and text record descriptions, generating evidence conflict markers:

[0025]

[0026] in, Indicates a conflict of evidence marker. Indicates the total number of event nodes. Represents video feature description semantic encoding vector, Represents a text record description semantic encoding vector, Indicates distance from Manhattan. Represents the smoothing factor. This represents the node confidence score. Indicates the time decay coefficient. Indicates the video-to-text timestamp discrepancy. This represents the Sigmoid activation function. Indicates the event node index;

[0027] Based on a preset risk level matrix, the severity of defects in time chain break markers, missing physical evidence markers, and conflicting evidence markers is assessed, and relevant defect markers are generated.

[0028] In one embodiment, based on a pre-defined knowledge graph of administrative litigation evidence rules, the original multimodal features are subjected to legal constraint embedding processing to obtain multimodal features with legal constraint labels, including:

[0029] The legal provisions in the knowledge graph of evidence rules in administrative litigation are classified according to their urgency, and a dynamic set of weight coefficients is generated.

[0030] The original multimodal features are processed to separate visual and textual components, generating visual legal feature vectors and textual legal feature vectors;

[0031] Based on the clause priority in the dynamic weight coefficient set, the visual legal feature vector and the textual legal feature vector are differentially weighted to generate a weighted feature representation;

[0032] Based on the preset legal attribute mapping rules, the weighted feature representation is subjected to clause matching processing to generate initial legal constraint labels;

[0033] The following formula is used to perform cross-modal consistency verification on the initial legal constraint labels, eliminating conflicting tags between visual and textual features, and generating consistent legal constraint labels:

[0034]

[0035] in, Labels indicating consistency and legal constraints Indicates the initial legal binding label. Represents element-wise multiplication. Represents the conflict attenuation coefficient. This represents the gradient of the label with respect to visual features. This represents the gradient of the label with respect to the text features. Represents the squared Euclidean distance. Represents an exponential function;

[0036] Based on preset mandatory compliance clauses, key procedural behavior enhancement processing is performed on the consistency legal constraint label to generate multimodal features with legal constraint labels.

[0037] In one embodiment, the multimodal features labeled with legal constraints are processed to generate legal fact fragments, resulting in semantic description fragments labeled with legal attributes, including:

[0038] Based on a pre-set law enforcement process template, multimodal features with legal constraint labels are processed for legal procedure stage identification, and procedure node boundary markers are generated.

[0039] Based on the boundary markers of program nodes, spatiotemporal correlation analysis is performed on the visual and textual components of multimodal features with legal constraint labels within the same program node to generate an evidence chain completeness score.

[0040] Based on the evidence chain completeness score, the legal meaning of the multimodal features with legal constraint labels in each program node is analyzed to generate initial legal attribute labels;

[0041] By using a pre-defined legal domain language model, program nodes with initial legal attribute labels are processed by natural language conversion to generate semantic description fragments with legal attribute labels.

[0042] In one embodiment, based on the evidence chain completeness score, the multimodal features with legal constraint labels within each program node are subjected to legal meaning parsing to generate initial legal attribute labels, including:

[0043] The core legal elements of the multimodal features with legal constraint labels within the program nodes are separated to generate legal element feature groups.

[0044] Based on the evidence chain completeness score, the confidence level of the legal element feature group is adjusted to generate the element features with confidence level adjustment.

[0045] Based on preset legal attribute determination rules, the credibility-adjusted element features are subjected to legal attribute matching processing to generate initial legal attribute labels.

[0046] In one embodiment, the verification results marked with risk type are processed into legal documents to generate standard case documents with attached legal risk labels and independent evidence defect reports, including:

[0047] Based on the pre-defined logical relationships of law enforcement events, the verification results marked with risk types are sorted chronologically to generate structured event descriptions;

[0048] Based on a pre-defined database of risk levels and legal provisions, the structured event descriptions are matched with legal risk types to generate a risk labeling index.

[0049] Based on a pre-set standard document template library, the structured event description is formatted and converted to generate basic document content segments;

[0050] Based on the risk-marking index, the basic document content segments are marked and inserted to generate standard case documents with attached legal risk markings.

[0051] The risk markers in the verification results with risk type markings are processed for key information filtering to generate an independent evidence flaw report.

[0052] Secondly, this application also provides a system for inputting administrative law enforcement case documents using a large language model, the system including:

[0053] The multimodal feature extraction module is used to perform visual temporal analysis on the input law enforcement video stream to obtain visual action sequence features; and to perform text parsing on the input associated text data to obtain structured text features.

[0054] The legal constraint embedding module is used to perform multimodal fusion processing on visual action sequence features and structured text features to obtain the original multimodal features; based on the preset administrative litigation evidence rule knowledge graph, the original multimodal features are subjected to legal constraint embedding processing to obtain multimodal features with legal constraint labels;

[0055] The legal fact generation module is used to process multimodal features with legal constraint labels to generate legal fact fragments, resulting in semantic description fragments with legal attribute labels.

[0056] The evidence three-dimensional verification module is used to perform dynamic evidence three-dimensional verification processing on semantic description fragments with legal attribute tags, and obtain verification results with risk type tags.

[0057] The document synthesis output module is used to process the verification results marked with risk types into legal documents, generating standard case documents with legal risk labels and independent evidence defect reports.

[0058] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the methods in the first aspect of this application.

[0059] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of any of the methods in the first aspect of this application.

[0060] The method for assisting in the entry of administrative law enforcement case documents provided in this application uses a large language model to obtain visual action sequence features by performing visual temporal analysis on law enforcement record video streams and text parsing on associated text data to obtain structured text features. The above two types of features are fused in a multimodal manner to obtain original multimodal features. Combined with a pre-set knowledge graph of administrative litigation evidence rules, the original multimodal features are embedded with legal constraints to generate multimodal features with legal constraint labels. Then, legal fact fragments are generated from these features to obtain semantic description fragments with legal attribute labels. Subsequently, the three-dimensional verification of evidence is dynamically verified to obtain verification results with risk type labels. Based on the verification results, legal documents are synthesized to generate standard case documents with legal risk labels and independent evidence defect reports.

[0061] The above series of processes enable the effective association of multimodal information related to law enforcement, thereby improving the integrity of the evidence chain. Through the embedding of legal constraints and the layer-by-layer dynamic verification of the three properties, the omission of procedural flaws is reduced, thereby lowering the rate of missed procedural flaw detection. At the same time, standardized document generation and clear risk labeling also help to avoid potential risks to the validity of document evidence. Attached Figure Description

[0062] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0063] Figure 1 This is a flowchart of a method for assisting in the entry of administrative law enforcement case documents using a large language model, as described in one embodiment of the present invention.

[0064] Figure 2 This is a flowchart illustrating the process of generating legal fact fragments from multimodal features labeled with legal constraints, as described in one embodiment of the present invention, to obtain semantic description fragments labeled with legal attributes.

[0065] Figure 3 This is a structural diagram of a system for assisting in the entry of administrative law enforcement case documents using a large language model, according to one embodiment of the present invention. Detailed Implementation

[0066] To make the above-mentioned objects, features, and advantages of this application more apparent and understandable, the specific embodiments of this application will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0067] First, the application scenarios of the embodiments of this application are described. In the embodiments of this application, a method for using a large language model to assist in the entry of administrative law enforcement case documents is provided, applicable to scenarios such as document entry in market supervision and law enforcement cases, document organization in environmental protection law enforcement cases, material archiving in transportation law enforcement cases, and record compilation in urban management law enforcement cases.

[0068] In illustrative purposes, the method for assisting in the entry of administrative law enforcement case documents using a large language model provided in this application embodiment can also be applied to application scenarios such as emergency management law enforcement case document processing, cultural market law enforcement case material compilation, and standardized entry of labor security supervision case documents. This is only an example and does not limit the specific application scenarios.

[0069] like Figure 1 As shown, this application provides a method for assisted entry of administrative law enforcement case documents using a large language model. The method includes:

[0070] S101: Perform visual temporal analysis on the input law enforcement video stream to obtain visual action sequence features; perform text parsing on the input associated text data to obtain structured text features.

[0071] For example, the input law enforcement video stream undergoes video signal denoising processing to filter out invalid information caused by environmental interference and equipment noise. Then, frame sequence synchronization calibration is performed to ensure the temporal continuity and integrity of the video frames. Based on the behavioral logic of the law enforcement scenario, visual information is extracted frame by frame from the calibrated video frame sequence. This identifies the actions of the law enforcement subject, the behavior of the involved parties, and key elements of the scene environment. Temporal correlation analysis is performed on the extracted visual information according to its chronological evolution, clarifying the logical connections between different actions and scenes. By abstracting and structurally integrating the correlated visual information, visual action sequence features are obtained.

[0072] The input related text data undergoes text format standardization processing, unifying text encoding, layout specifications, and expression forms, and removing redundant content and formatting interference information irrelevant to the law enforcement case. For the standardized related text data, a systematic text content analysis is performed, extracting basic case information, law enforcement procedure records, evidence descriptions, and legal basis elements segment by segment. Then, according to preset structured classification rules, the extracted text elements are logically categorized and hierarchically organized to ensure the logical coherence and relevance of the text information, resulting in structured text features.

[0073] S102: Perform multimodal fusion processing on visual action sequence features and structured text features to obtain original multimodal features; based on the pre-set administrative litigation evidence rule knowledge graph, perform legal constraint embedding processing on the original multimodal features to obtain multimodal features with legal constraint labels.

[0074] For example, cross-modal semantic alignment processing is performed on visual action sequence features and structured text features to ensure that the two types of features maintain consistency in the core semantics and time dimension of law enforcement event description. Through feature association analysis, the inherent logical relationship between law enforcement behavior information in visual action sequence features and factual record information in structured text features is mined. Based on semantic association, the two types of features are deeply integrated to fuse the scattered visual and text information into a unified feature expression form, thus obtaining the original multimodal features.

[0075] The core legal constraint elements of the pre-defined knowledge graph of administrative litigation evidence rules are analyzed, including procedural compliance requirements and evidence form norms. The original multimodal features are matched one by one with the above-mentioned legal constraint elements, corresponding constraint labels are added to the original multimodal features and adaptability is verified to obtain multimodal features with legal constraint labels.

[0076] S103: Perform legal fact fragment generation processing on multimodal features with legal constraint labels to obtain semantic description fragments with legal attribute labels.

[0077] For example, core information substantially related to law enforcement cases is filtered from multimodal features labeled with legal constraints, eliminating unnecessary interference. The corresponding legal attribute categories are matched based on the constraint labels to clarify the legal characterization of the core information. The core information is then organized according to the logical flow of the law enforcement event to ensure compliance with legal factual expression norms. This information is then transformed into natural language fragments that conform to the expression habits of legal documents, labeled with corresponding legal attribute categories, resulting in semantically descriptive fragments with legal attribute labels.

[0078] S104: Perform dynamic verification of the three aspects of evidence on semantic description fragments with legal attribute tags to obtain verification results with risk type tags.

[0079] For example, based on pre-defined legality norms, the procedural compliance and suitability for law enforcement authority of semantic descriptive fragments tagged with legal attributes are verified. Subjective assumptions and biased descriptions in semantic descriptive fragments tagged with legal attributes are screened according to objectivity standards. The relevance of semantic descriptive fragments tagged with legal attributes to the core facts of the case, as well as the logical consistency and coherence of the evidence chain, are examined according to relevance rules.

[0080] Define the risk types and scope of non-compliance, non-objectivity, and irrelevance issues found in the three-dimensional verification process, associate and bind the risk definition results with the corresponding descriptive fragments, form a system risk record, and obtain verification results marked with risk types.

[0081] S105: Perform legal document synthesis processing on the verification results marked with risk type to generate standard case documents with legal risk labels and independent evidence defect reports.

[0082] For example, the basic facts of the case, the description of evidence, and the risk markers are extracted from the verification results marked with risk types, distinguishing between core information and risk details. Following standard case document specifications, the basic facts of the case and the description of evidence are organized according to the legal format, and the risk markers are precisely linked to their corresponding positions in the document, forming a draft document with annotations.

[0083] Extract risk details separately and organize them into a clear and concise explanation of defects according to the three categories of evidence. Optimize the format and wording of the initial draft of the annotated document, improve the logic and completeness of the defect explanation, and generate a standard case document with attached legal risk annotations and an independent evidence defect report.

[0084] One embodiment of this application provides a method for assisting in the entry of administrative law enforcement case documents using a large language model. This method obtains visual action sequence features by performing visual temporal analysis on law enforcement record video streams and obtains structured text features by performing text parsing on associated text data. The two types of features are then fused in a multimodal manner to obtain original multimodal features. These original multimodal features are then combined with a pre-set knowledge graph of administrative litigation evidence rules to embed legal constraints into the original multimodal features, generating multimodal features with legal constraint labels. Subsequently, legal fact fragments are generated from these features to obtain semantic description fragments with legal attribute labels. The three characteristics of evidence are then dynamically verified to obtain verification results with risk type tags. Based on these verification results, legal documents are synthesized to generate standard case documents with legal risk labels and independent evidence defect reports.

[0085] The above series of processes enable the effective association of multimodal information related to law enforcement, thereby improving the integrity of the evidence chain. Through the embedding of legal constraints and the layer-by-layer dynamic verification of the three properties, the omission of procedural flaws is reduced, thereby lowering the rate of missed procedural flaw detection. At the same time, standardized document generation and clear risk labeling also help to avoid potential risks to the validity of document evidence.

[0086] In one embodiment, the semantic description fragment with legal attribute tags is subjected to dynamic verification of the three aspects of evidence to obtain a verification result with risk type marking, including:

[0087] (1) Using the following formula, based on a preset set of legality rules, perform procedural behavior matching on semantic description fragments with legal attribute tags to generate legality risk markers:

[0088]

[0089] in, Indicates a legality risk marker. Indicates the total number of legality rules. Representation rules The legal weighting coefficient, This represents the feature matrix of semantic description fragments with legal attribute labels. Representation rules The template vector of program behavior, Represents a non-linear activation function. Represents the legal semantic distance function. Representation rules The legal behavior threshold vector, Representation rules The threshold for determining violations Indicates an indicator function, This indicates transpose.

[0090] For example, a preset set of legality rules is invoked to extract the feature matrix of semantic description fragments with legal attribute labels. This feature matrix is ​​then matched with the program behavior template vector corresponding to the rule in the preset set of legality rules. The distance between the feature matrix of the semantic description fragment with legal attribute labels and the legal behavior threshold vector corresponding to the rule is calculated using a legal semantic distance function. It is then determined whether the distance exceeds the violation judgment threshold corresponding to the rule. Combined with the legal weight coefficient corresponding to the rule, the matching result is processed by a non-linear activation function to generate a legality risk label.

[0091] The preset legality rule set includes compliance requirements for law enforcement procedures and norms for the application of authority. The semantic description fragment feature matrix with legal attribute tags is a feature set formed after structuring the semantic description fragments with legal attribute tags. The procedural behavior template vector is the feature vector of the standard procedural behavior corresponding to each rule in the preset legality rule set.

[0092] (2) Based on the preset set of objective rules, the semantic description fragments with legal attribute tags are subjected to language neutrality analysis and processing to generate objective risk tags.

[0093] For example, a preset set of objective rules is invoked to analyze the content of semantic description fragments with legal attribute tags segment by segment, and to screen out content that does not meet the neutrality requirements, such as subjective and arbitrary statements and biased descriptions. It is then determined whether the expression of the fragment meets the language neutrality standard in the preset set of objective rules, and the content that does not meet the standard is marked as a risk point, generating an objective risk mark.

[0094] The preset set of objectivity rules includes standards for language neutrality in law enforcement records and objective norms for factual descriptions; language neutrality analysis and processing is a process of verifying the subjective bias and emotional tone of the expression.

[0095] (3) Based on the preset set of association rules, the semantic description fragments with legal attribute tags are subjected to evidence chain integrity verification processing to generate association defect markers.

[0096] For example, a preset set of correlation rules is invoked to check the degree of correlation between semantic description fragments with legal attribute tags and the core facts of law enforcement cases. At the same time, the logical connection between different semantic description fragments with legal attribute tags is checked to determine whether there are breaks or omissions in the chain of evidence. Content with insufficient correlation or flaws in the chain of evidence is marked as flaw points, and correlation flaw tags are generated.

[0097] The pre-defined set of relevance rules includes standards for the relevance of evidence to the facts of the case, and norms for the logical coherence of the chain of evidence; the verification process for the integrity of the chain of evidence is a process of checking the relevance between evidence and the completeness of the overall chain.

[0098] (4) Perform dynamic weight fusion processing on the legality risk mark, the objective risk mark and the related defect mark to generate the verification result with risk type mark.

[0099] For example, legality risk markers, objectivity risk markers, and relevance defect markers are integrated. Based on the degree of impact of the risk corresponding to each marker on the case documents, corresponding dynamic weights are configured. Different types of markers are integrated according to the configured dynamic weights. The integrated risk content is then categorized to clarify the category corresponding to each risk and generate verification results with risk type markers.

[0100] Among them, dynamic weight fusion processing is an integration method that adjusts the weight ratio based on the actual impact of different types of risks; the verification result with risk type marking is a set of verification information containing various types of risks and their corresponding types.

[0101] In one embodiment, based on a preset set of relevance rules, the semantic description fragments with legal attribute tags are subjected to evidence chain integrity verification processing to generate relevance defect markers, including:

[0102] (1) Using the following formula, based on the preset time correlation rules, perform logical gap detection processing on the event sequence of semantic description fragments with legal attribute tags to generate time chain break markers:

[0103] in, Indicates a time-broken chain marker. Indicates the start timestamp of the event sequence. Indicates the end timestamp of the event sequence. Indicates time Event description vector, Indicates the time interval between adjacent events. Indicates the threshold of time correlation. This indicates an indicator function.

[0104] For example, a preset time-related rule is invoked to identify all event-related statements from semantic description fragments with legal attribute tags, and these statements are arranged into an ordered event sequence according to their chronological order of appearance. The time information corresponding to the first event in the event sequence is determined as the start timestamp of the event sequence, and the time information corresponding to the last event in the event sequence is determined as the end timestamp of the event sequence. The event description content corresponding to each time point in the event sequence is extracted one by one, and the above content is converted into structured event description information. Then, the difference between the time points corresponding to two adjacent events in the event sequence is calculated to obtain the time interval between adjacent events.

[0105] The system retrieves a time correlation threshold adapted to the current law enforcement scenario from preset time correlation rules. Through logical gap detection, each time interval between adjacent events is compared with the time correlation threshold to determine if the interval exceeds it. Intervals between adjacent events exceeding the threshold are marked as breakpoints, and the corresponding event location information is recorded. All breakpoint annotation information is integrated to form a structured marking result, generating a time breakpoint marker.

[0106] The preset time correlation rules include reasonable interval standards for event timing under different law enforcement scenarios and criteria for judging the continuity of event sequences; semantic description fragments with legal attribute tags are law enforcement-related natural language descriptions with legal attributes already labeled; event sequences are sets of law enforcement events arranged by time dimension; logical gap detection is the process of checking whether the event timing intervals comply with the standards; and time chain break markers are the marking results of time gap defects in event sequences.

[0107] (2) Based on the preset behavioral evidence association rules, the matching degree analysis of law enforcement actions in semantic description fragments and corresponding evidence descriptions is performed to generate evidence missing markers.

[0108] For example, a pre-defined behavioral evidence association rule is invoked to extract the corresponding descriptions of all law enforcement actions from semantic description fragments with legal attribute tags, generating a list of law enforcement action descriptions. From this fragment, all evidence-related descriptions are extracted to generate a list of evidence descriptions. The correspondence between law enforcement actions and evidence in the pre-defined behavioral evidence association rule is compared with the content in the evidence description list, one by one. It is determined whether each law enforcement action has a corresponding evidence description. Law enforcement actions that do not match a corresponding evidence description are marked as defects, and the specific description of the law enforcement action is recorded. The marking information of all defects is integrated to generate a missing evidence marker.

[0109] The pre-defined rules for associating physical evidence with law enforcement actions include the required types of physical evidence for different law enforcement actions and the matching standards between law enforcement actions and physical evidence. The law enforcement action description list is a collection of records of law enforcement actions within a semantically labeled descriptive fragment. The physical evidence description list is a collection of records of physical evidence information within that fragment; the physical evidence missing marker is a record of any missing physical evidence corresponding to a law enforcement action.

[0110] (3) Perform keyframe back-tracing processing on the visual action sequence features to obtain video feature descriptions.

[0111] For example, visual action sequence features are retrieved, and core keywords corresponding to the event are extracted from semantic description fragments with legal attribute tags, including the type of law enforcement behavior and the identification of the involved parties. Based on these core keywords, the frame interval corresponding to the event is located in the visual action sequence features. Frame content that clearly reflects the key information of the event is selected from this frame interval as keyframes. Visual features are extracted from each keyframe, covering action outlines, scene elements, object states, etc. The extracted keyframe features are concatenated and integrated in chronological order. The integrated feature content is structured and organized to form an information set containing event time, action details, and scene information, resulting in a video feature description.

[0112] Among them, visual action sequence features are the feature set obtained after the law enforcement recording video stream has undergone visual temporal analysis and processing; keyframes are the frame content containing the core information of the event in the visual action sequence features; and video feature description is the structured information formed after feature extraction and sorting of keyframe content.

[0113] (4) Extract text elements from semantic description fragments with legal attribute tags to obtain text record descriptions.

[0114] For example, semantic description fragments with legal attribute tags are broken down sentence by sentence to obtain several independent sentence units. Content recognition is performed on each sentence unit to extract relevant event elements, including the event subject, actions, time of occurrence, and setting. These extracted event elements are then categorized according to preset recording standards into categories such as subject, action, time, and setting. The categorized event elements are logically analyzed to clarify the corresponding relationships between different elements. Finally, the analyzed elements are integrated into a well-structured information set with clear hierarchical levels to obtain the text record description.

[0115] Among them, semantic description fragments with legal attribute tags are law enforcement-related natural language descriptions that have been labeled with legal attributes; sentence units are the basic language units after the fragment is broken down; and text record descriptions are structured information formed after the extracted event elements are organized.

[0116] (5) Use the following formula to perform consistency verification between video feature descriptions and text record descriptions, and generate evidence conflict markers:

[0117]

[0118] in, Indicates a conflict of evidence marker. Indicates the total number of event nodes. Represents video feature description semantic encoding vector, Represents a text record description semantic encoding vector, Indicates distance from Manhattan. Represents the smoothing factor. This represents the node confidence score. Indicates the time decay coefficient. Indicates the video-to-text timestamp discrepancy. This represents the Sigmoid activation function. Indicates the event node index.

[0119] For example, semantic encoding is performed on video feature descriptions to convert them into corresponding semantically encoded content. Semantic encoding is also performed on text record descriptions to convert them into corresponding semantically encoded content. The semantic distance between these two types of semantically encoded content is calculated to measure the degree of semantic difference between them. The corresponding node confidence score and video-text timestamp deviation are retrieved from event node information. A smoothing factor is then used to adjust the calculated semantic distance to mitigate the interference of outlier values.

[0120] The adjusted result is fused with the correction value corresponding to the time decay coefficient to adapt to the impact of time deviation. The fused result is then input into an activation function for feature transformation to enhance the discriminative power of semantic differences. It is determined whether the transformed result exceeds a preset consistency threshold; if so, the corresponding content is marked as a conflict point, and the event node corresponding to the conflict point is recorded. The annotation information of all conflict points is integrated to generate evidence conflict markers.

[0121] Among them, semantic encoding content is a vector representation formed by encoding the features of video feature description and text record description; semantic distance is an indicator to measure the semantic difference between the two types of encoded content; activation function is a tool for feature transformation of fusion results; evidence conflict label is a labeling result that records inconsistencies between video feature description and text record description.

[0122] (6) Based on the preset risk level matrix, the severity of defects of time chain break markers, physical evidence missing markers and evidence conflict markers is assessed and generated as relevant defect markers.

[0123] For example, a preset risk level matrix is ​​invoked to read the specific content of time chain break markers, missing evidence markers, and evidence conflict markers. The severity level of each marker is determined by comparing it with the severity grading standards corresponding to different defect types in the preset risk level matrix. The impact weight of each defect level on the integrity of the evidence chain is evaluated, and the weight ratio is adjusted according to the core nature of the defect. The defects corresponding to different markers are integrated and sorted according to their impact weights to clarify the primary and secondary relationships of the defects. Based on the integrated results, each defect is labeled with both type and level according to the labeling specifications in the preset risk level matrix. All labeled defect information is integrated into a structured set to generate related defect markers.

[0124] The preset risk level matrix includes severity level standards for different defects, weight specifications for the evidence chain of defect impact, etc.; the defect severity level is the classification result of the degree of defect impact; the related defect mark is a set of marks after integrating all related defects.

[0125] In one embodiment, based on a pre-defined knowledge graph of administrative litigation evidence rules, the original multimodal features are subjected to legal constraint embedding processing to obtain multimodal features with legal constraint labels, including:

[0126] (1) The legal provisions in the knowledge graph of evidence rules in administrative litigation are classified according to their urgency and a dynamic weight coefficient set is generated.

[0127] For example, the knowledge graph of evidence rules in administrative litigation is retrieved, all legal clauses contained therein are extracted, and the legal clauses are hierarchically divided according to the urgency of the enforcement procedures and the degree of compliance impact of the clauses. The clauses at each level are assigned corresponding weight values, and the weight values ​​of all clauses are integrated to form a structured set, generating a dynamic weight coefficient set.

[0128] The knowledge graph of evidence rules in administrative litigation includes statutory provisions and compliance norms related to evidence in administrative litigation; the urgency level classification is a hierarchical division based on the enforcement priority and compliance impact of the provisions; and the dynamic weight coefficient set is a set of weight values ​​corresponding to provisions with different urgency levels.

[0129] (2) Perform visual and text component separation processing on the original multimodal features to generate visual legal feature vectors and text legal feature vectors.

[0130] For example, the original multimodal features are retrieved, and visually relevant feature components derived from law enforcement video streams and text-related feature components derived from associated text data are identified. Feature extraction is performed on these two types of components respectively. The extracted visually relevant feature components are encoded into vector form to generate a visual legal feature vector. The extracted text-related feature components are encoded into vector form to generate a text legal feature vector.

[0131] Among them, the original multimodal features are a unified feature expression after the fusion of visual action sequence features and structured text features; the separation of visual and text components is the process of distinguishing components from different sources in the original multimodal features; the visual legal feature vector is the vector representation of visually related feature components; and the textual legal feature vector is the vector representation of textually related feature components.

[0132] (3) Based on the priority of the clauses in the dynamic weight coefficient set, the visual legal feature vector and the textual legal feature vector are differentially weighted to generate a weighted feature representation.

[0133] For example, a dynamic weight coefficient set is retrieved, the priority weight of each legal clause is determined, and the visual legal feature vector is associated with and matched with the corresponding priority weight. The visual legal feature vector is then adjusted according to the weight value. Simultaneously, the textual legal feature vector is associated with and matched with the corresponding priority weight. The textual legal feature vector is then adjusted according to the weight value. The adjusted visual legal feature vector and textual legal feature vector are then integrated to generate a weighted feature representation.

[0134] Among them, the priority of the clauses in the dynamic weight coefficient set is the weight ranking of clauses with different urgency levels; the differentiated weighting process is the process of adjusting different feature vectors according to the weights; the weighted feature representation is the feature form after weight adjustment and integration.

[0135] (4) Based on the preset legal attribute mapping rules, perform clause matching processing on the weighted feature representation to generate initial legal constraint labels.

[0136] For example, the preset legal attribute mapping rules are retrieved, and the feature content in the weighted feature representation is compared one by one with the corresponding legal clauses in the preset legal attribute mapping rules to determine the legal clauses that the weighted feature representation is compatible with. The compatible legal clauses are converted into tag form, and these tags are integrated to form a structured set to generate initial legal constraint tags.

[0137] Among them, the preset legal attribute mapping rules include the correspondence between feature content and legal clauses; clause matching processing is the process of adapting features to legal clauses; and the initial legal constraint label is a set of legal clause labels generated after matching.

[0138] (5) Use the following formula to perform cross-modal consistency verification on the initial legal constraint labels, eliminate conflicting tags between visual features and text features, and generate consistent legal constraint labels:

[0139]

[0140] in, Labels indicating consistency and legal constraints Indicates the initial legal binding label. Represents element-wise multiplication. Represents the conflict attenuation coefficient. This represents the gradient of the label with respect to visual features. This represents the gradient of the label with respect to the text features. Represents the squared Euclidean distance. This represents an exponential function.

[0141] For example, the initial legal constraint label is retrieved, and the gradients of the visual features and text features corresponding to the initial legal constraint label are calculated respectively. The degree of difference between the two gradients is measured, and the degree of difference is adjusted by the conflict decay coefficient. The adjusted results are used to perform element-level correction on the initial legal constraint label, eliminating conflicting label content corresponding to visual features and text features. The corrected label content is integrated to generate a consistent legal constraint label.

[0142] Among them, cross-modal consistency verification is the process of checking whether the label content corresponding to visual and textual features is consistent; conflict markers are inconsistent content in the labels corresponding to visual and textual features; and consistency legal constraint labels are the set of labels after conflict elimination.

[0143] (6) Based on the preset mandatory compliance clauses, perform key procedural behavior enhancement processing on the consistency legal constraint label to generate multimodal features with legal constraint labels.

[0144] For example, the system retrieves preset mandatory compliance clauses, identifies key procedural behavior-related tags corresponding to the mandatory compliance clauses from the consistent legal constraint tags, enhances the feature weight of the above tags to increase their priority in the tag set, and associates and binds the enhanced consistent legal constraint tags with the original multimodal features to integrate them into a unified feature form, thereby generating multimodal features with legal constraint tags.

[0145] Among them, the preset mandatory compliance clauses are the legal procedural clauses that must be followed during the enforcement process; the key procedural behavior enhancement processing is the process of increasing the weight of mandatory compliance-related tags; and the multimodal features with legal constraint tags are the original multimodal features bound to legal constraint tags.

[0146] like Figure 2 As shown, legal fact fragment generation processing is performed on multimodal features labeled with legal constraints to obtain semantic description fragments with legal attribute labels, including:

[0147] S201: Based on a preset law enforcement process template, perform legal procedure stage identification processing on multimodal features with legal constraint labels, and generate procedure node boundary markers.

[0148] For example, a preset law enforcement process template is retrieved, and content related to legal procedures is extracted from the multimodal features labeled with legal constraints. The content range corresponding to different law enforcement procedure stages in the multimodal features labeled with legal constraints is identified by comparing the law enforcement stage categories in the preset law enforcement process template with the content range corresponding to different law enforcement procedure stages in the multimodal features labeled with legal constraints. The start and end ranges corresponding to each law enforcement procedure stage are determined, and the above range information is labeled and organized to generate procedure node boundary markers.

[0149] The preset law enforcement process template includes the standardized division of each legal procedural stage in the law enforcement process; the legal procedure stage identification and processing is the process of distinguishing the multimodal features with legal constraint labels corresponding to different law enforcement stages; and the procedure node boundary marker is the set of content scope labels corresponding to each law enforcement procedure stage.

[0150] S202: Based on the boundary markers of program nodes, perform spatiotemporal correlation analysis on the visual and textual components of multimodal features with legal constraint labels within the same program node to generate an evidence chain completeness score.

[0151] For example, the boundary markers of program nodes are retrieved to determine the range of multimodal features with legal constraint labels corresponding to the same program node. The visual and textual components of the multimodal features with legal constraint labels within this range are extracted. The degree of matching between the time information corresponding to the visual components and the time information corresponding to the textual components is analyzed. At the same time, the consistency of the factual content corresponding to the visual components and the textual components is correlated. Based on the comprehensive result of the time matching degree and the content consistency, the completeness of the evidence association within the program node is evaluated. The evaluation result is converted into the corresponding score form to generate an evidence chain completeness score.

[0152] Among them, the procedure node boundary marker is the content scope labeling of each law enforcement procedure stage; the spatiotemporal correlation analysis is the process of matching the temporal and content correlation between visual components and text components; and the evidence chain integrity score is the score result of the completeness of evidence correlation within the procedure node.

[0153] S203: Based on the evidence chain completeness score, perform legal meaning analysis on the multimodal features with legal constraint labels in each program node to generate initial legal attribute labels.

[0154] For example, the evidence chain completeness score is retrieved, and multimodal features with legal constraint labels within the procedural nodes whose scores meet the preset standards are selected. The legal constraint labels and associated factual information contained in these features are extracted. Combined with the legal clauses in the administrative litigation evidence rules knowledge graph, the legal attribute categories corresponding to the above information are parsed. The parsed legal attribute categories are converted into labels, and corresponding category labels are added to the multimodal features with legal constraint labels within each procedural node to generate initial legal attribute labels.

[0155] Among them, the evidence chain completeness score is the score of the completeness of the evidence association within the program node; the legal meaning analysis process is the process of analyzing the legal attribute categories corresponding to the feature information; and the initial legal attribute label is the set of legal attribute category identifiers corresponding to the program node features.

[0156] S204: Using a pre-defined legal domain language model, perform natural language conversion on program nodes with initial legal attribute labels to generate semantic description fragments with legal attribute labels.

[0157] For example, a pre-defined legal domain language model is retrieved, and the factual information corresponding to the program node with initial legal attribute labels and the initial legal attribute labels are input into the pre-defined legal domain language model. According to the normative expression requirements of legal documents, the characteristic information is converted into natural language text that conforms to legal expression habits. At the same time, the initial legal attribute labels are marked at the corresponding content positions of the natural language text. The converted text and the label content are integrated to form structured descriptive content and generate semantic descriptive fragments with legal attribute labels.

[0158] Among them, the pre-defined legal domain large language model is a natural language conversion model adapted to legal scenarios; natural language conversion processing is the process of converting feature information into legal normative text; and semantic description fragments with legal attribute tags are natural language descriptions of legal normative texts labeled with legal attribute categories.

[0159] In one embodiment, based on the evidence chain completeness score, the multimodal features with legal constraint labels within each program node are subjected to legal meaning parsing to generate initial legal attribute labels, including:

[0160] (1) Perform core legal element separation processing on the multimodal features with legal constraint labels in the program node to generate legal element feature groups.

[0161] For example, the scope corresponding to the program node is determined, and multimodal features with legal constraint labels are extracted within that scope. Core legal elements covering law enforcement entities, involved behaviors, legal basis, and factual details are identified and separated from the multimodal features with legal constraint labels. Core legal elements of the same category are classified and integrated to form several sets of features, and legal element feature groups are generated.

[0162] Among them, the multimodal features with legal constraint labels within the program nodes are multimodal features that correspond to a certain stage of law enforcement procedures and are bound with legal constraint labels; the core legal element separation processing is the process of extracting key legal-related content from these features; and the legal element feature group is a feature set after integrating similar core legal elements.

[0163] (2) Based on the evidence chain completeness score, the confidence level of the legal element feature group is adjusted to generate the element features with adjusted confidence level.

[0164] For example, the evidence chain completeness score is retrieved, and the evidence chain completeness score is associated with each element in the legal element feature group. Based on the evidence chain completeness score, the confidence weight of each element in the legal element feature group is adjusted, strengthening the confidence ratio of elements with higher scores and weakening the confidence ratio of elements with lower scores. The adjusted elements are then integrated to generate element features with adjusted confidence.

[0165] Among them, the evidence chain completeness score is the score result of the completeness of the evidence association within the procedural node; the confidence correction process is the process of adjusting the confidence weight of the elements based on the score; the confidence-adjusted element features are the set of legal element features optimized by confidence.

[0166] (3) Based on the preset legal attribute determination rules, the feature characteristics of the credibility adjusted are subjected to legal attribute matching processing to generate initial legal attribute labels.

[0167] For example, the preset legal attribute determination rules are retrieved, and the credibility-adjusted element features are compared one by one with the legal attribute categories specified in the preset legal attribute determination rules to determine the legal attribute category corresponding to each element. The matched legal attribute categories are integrated, converted into tags, and associated with the corresponding element content to generate initial legal attribute tags.

[0168] Among them, the preset legal attribute determination rules are the matching norms between legal elements and corresponding attribute categories; the legal attribute matching process is the process of adapting element features to legal attribute categories; and the initial legal attribute labels are the set of legal attribute category identifiers corresponding to element features.

[0169] In one embodiment, the verification results marked with risk type are processed into legal documents to generate standard case documents with attached legal risk labels and independent evidence defect reports, including:

[0170] (1) Based on the preset logical relationship of law enforcement events, the verification results marked with risk type are sorted by timeline to generate a structured event description.

[0171] For example, the preset logical relationship of law enforcement events is retrieved, and all content related to law enforcement events is extracted from the verification results marked with risk type. This includes the time information, behavior information and associated risk type marks of the events. The extracted law enforcement event content is arranged according to the time sequence and logical relationship specified in the preset logical relationship of law enforcement events. At the same time, the risk type mark information corresponding to the events is integrated to form a set of event information with clear hierarchy and generate a structured event description.

[0172] Among them, the preset logical relationship of law enforcement events includes the reasonable evolution order of various events in the law enforcement process and the logical association norms between events; the verification result with risk type marking is a set of verification information containing risk type marking; the timeline sorting process is the process of arranging the event content according to time and logic; the structured event description is a set of event information sorted according to the norms.

[0173] (2) Based on the preset risk level and legal provision association database, the structured event description is matched with the legal risk type to generate a risk labeling index.

[0174] For example, a preset risk level and legal provision association database is retrieved, and the associated risk type markers and corresponding event content are extracted from the structured event description. The correspondence between risk type and legal provision and risk level in the preset risk level and legal provision association database is compared to determine the risk level and associated legal provision corresponding to each risk type marker. The risk type marker, the location of the corresponding event content, the risk level, and the associated legal provision are linked and bound to form a clearly identified index set, and a risk labeling index is generated.

[0175] The preset risk level and legal provision association database includes matching standards for risk levels and associated legal provisions corresponding to different risk types; the structured event description is a collection of event information sorted according to standards; the legal risk type matching process is the process of adapting risk markers with corresponding levels and legal provisions; and the risk labeling index is a collection of association identifiers related to risk.

[0176] (3) Based on the preset standard document template library, the structured event description is format converted to generate basic document content segments.

[0177] For example, a preset standard document template library is retrieved, and a suitable standard document template is selected from the preset standard document template library according to the type of law enforcement case. The content in the structured event description is split and formatted according to the format requirements specified by the standard document template, while retaining the key information corresponding to the event, forming text paragraphs that conform to the document specifications, and generating basic document content paragraphs.

[0178] The preset standard document template library includes document format templates corresponding to different types of law enforcement cases; the structured event description is a collection of event information sorted according to standards; the format conversion process is the process of formatting the event information according to the document template format; and the basic document content paragraphs are document text paragraphs that conform to the template format.

[0179] (4) Based on the risk labeling index, the basic document content segments are labeled and inserted to generate standard case documents with attached legal risk labels.

[0180] For example, the risk labeling index is retrieved, the event content corresponding to the risk labeling index is located in the basic document content section, and the risk type, risk level, related legal provisions, and other information contained in the risk labeling index are inserted into the corresponding content position in a standardized labeling format. At the same time, the layout format of the basic document content section is adjusted to ensure the consistency between the labeling information and the document content, thereby generating a standard case document with attached legal risk labels.

[0181] Among them, the risk labeling index is a set of associated identifiers related to risk; the basic document content segment is a document text paragraph that conforms to the template format; the labeling insertion process is the process of adding risk information to the corresponding position in the document; and the standard case document with attached legal risk labeling is a standardized law enforcement case document that includes risk labeling.

[0182] (5) Perform key information screening on the risk markers in the verification results with risk type markers to generate an independent evidence defect report.

[0183] For example, extract all content corresponding to risk tags from the verification results marked with risk type, filter out key information such as risk type, evidence location corresponding to risk, and impact of risk on the chain of evidence, classify and organize the above key information according to the category of evidence defects, clarify the specific content and degree of impact of each defect, and format the organized information according to the standard report format to generate an independent evidence defect report.

[0184] Among them, the verification result with risk type marking is a set of verification information containing risk type marking; the key information screening and processing is the process of extracting the core content from the risk marking; and the independent evidence defect report is a standardized report that specifically records evidence defect information.

[0185] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0186] In one embodiment, such as Figure 3 As shown, this application also provides a system 300 for inputting administrative law enforcement case documents using a large language model. The system 300 includes:

[0187] The multimodal feature extraction module 301 is used to perform visual temporal analysis processing on the input law enforcement record video stream to obtain visual action sequence features; and to perform text parsing processing on the input associated text data to obtain structured text features.

[0188] The legal constraint embedding module 302 is used to perform multimodal fusion processing on visual action sequence features and structured text features to obtain original multimodal features; based on a preset administrative litigation evidence rule knowledge graph, the original multimodal features are subjected to legal constraint embedding processing to obtain multimodal features with legal constraint labels;

[0189] The legal fact generation module 303 is used to process multimodal features with legal constraint labels to generate legal fact fragments, and obtain semantic description fragments with legal attribute labels.

[0190] The evidence three-dimensional verification module 304 is used to perform dynamic evidence three-dimensional verification processing on semantic description fragments with legal attribute tags to obtain verification results with risk type tags.

[0191] The document synthesis output module 305 is used to process the verification results marked with risk type into legal documents, and generate standard case documents with legal risk labels and independent evidence defect reports.

[0192] Specifically, the multimodal feature extraction module 301 preprocesses the input law enforcement video stream, filtering out invalid information caused by environmental interference and equipment noise, and performs frame sequence synchronization calibration to ensure the continuity and integrity of the video frame timeline. Based on the behavioral logic of the law enforcement scenario, it extracts visual information frame by frame from the calibrated video frame sequence, identifying the actions of the law enforcement subject, the behavior of the involved parties, and key elements of the scene environment. It then performs temporal correlation analysis on the extracted visual information according to the chronological evolution order, clarifying the logical connection between different actions and scenes. Finally, it abstracts and structurally integrates the correlated visual information to obtain visual action sequence features.

[0193] Among them, law enforcement record video stream is video material recorded during administrative law enforcement; visual temporal analysis and processing is the process of analyzing visual information in the video according to the time dimension; visual action sequence features are the feature set of integrated law enforcement actions and temporal information.

[0194] The input related text data is standardized in terms of text format, unifying text encoding, typesetting specifications and expression forms, removing redundant content and formatting interference information that is irrelevant to law enforcement cases, and systematically parsing the text content of the standardized related text data, extracting basic case information, law enforcement procedure records, evidence descriptions and legal basis elements segment by segment, and logically classifying and hierarchically organizing the extracted text elements according to preset structured classification rules to ensure the logicality and relevance of the text information, thus obtaining structured text features.

[0195] Among them, associated text data refers to written materials generated during administrative law enforcement; text parsing and processing is the process of extracting and organizing text content; and structured text features are a set of text information features after classification and integration.

[0196] The legal constraint embedding module 302 performs cross-modal semantic alignment processing on visual action sequence features and structured text features to ensure that the two are consistent in the core semantics and time dimension of law enforcement event description. It explores the inherent logical relationship between law enforcement behavior information in visual action sequence features and factual record information in structured text features. Based on semantic correlation, it deeply integrates the two types of features, merging the scattered visual and text information into a unified feature expression form to obtain the original multimodal features.

[0197] Among them, multimodal fusion processing is the process of integrating features from different sources; the original multimodal features are a unified feature expression after the fusion of visual and textual features.

[0198] The system retrieves a pre-defined knowledge graph of administrative litigation evidence rules, analyzes the core legal constraints it covers, such as procedural compliance requirements, evidence form norms, and legal element limitations, and matches the original multimodal features with the extracted legal constraints one by one. It clarifies the legal clause requirements corresponding to each part of the information in the original multimodal features, adds corresponding constraint labels to the original multimodal features according to the category and applicable scenario of the legal constraints, and performs cross-dimensional consistency verification on the features with constraint labels to ensure the adaptability of the constraint labels and feature content, thus obtaining multimodal features with legal constraint labels.

[0199] Among them, the pre-set knowledge graph of administrative litigation evidence rules is a knowledge set containing legal clauses related to administrative litigation evidence; the legal constraint embedding process is the process of adding legal clause labels to features; and the multimodal features with legal constraint labels are the original multimodal features bound with legal constraint labels.

[0200] The legal fact generation module 303 identifies legal procedure stages based on a preset law enforcement process template for multimodal features with legal constraint labels. It determines the content range corresponding to each procedure stage and generates procedure node boundary markers. Based on these markers, it performs spatiotemporal correlation analysis on the visual and textual components of the multimodal features with legal constraint labels within the same procedure node, assesses the completeness of evidence association, and generates an evidence chain integrity score. Based on this score, it parses the legal attribute categories corresponding to the multimodal features with legal constraint labels and generates initial legal attribute labels. Through a preset legal domain large language model, it converts the content with initial legal attribute labels into natural language fragments that conform to the expression habits of legal documents, while labeling the corresponding legal attribute categories, thus obtaining semantic description fragments with legal attribute labels.

[0201] Among them, the legal fact fragment generation process is the process of converting features into natural language descriptions with legal attributes; the semantic description fragments with legal attribute tags are law enforcement-related natural language content labeled with legal attributes.

[0202] The evidence verification module 304 verifies the three aspects of evidence. For semantic description fragments with legal attribute tags, it checks the compliance of the procedure and the suitability of law enforcement authority according to the preset legality norms. It screens out subjective assumptions and biased descriptions according to the objective standard, and examines the degree of relevance to the core facts of the case and the logical consistency between fragments according to the relevance rule. It defines the risk type and scope of non-compliance, non-objectivity, and non-relevance problems found in the three aspects of evidence verification. It links and binds the risk definition results with the corresponding content of the semantic description fragments with legal attribute tags to form a system risk record and obtains the verification results marked with risk type.

[0203] Among them, the dynamic verification process of the three aspects of evidence is the process of verifying the legality, objectivity and relevance of evidence; the verification result with risk type marking is a set of verification information containing risk type identifiers.

[0204] The document synthesis and output module 305, based on a preset logical relationship of law enforcement events, sorts the verification results marked with risk types in a timeline, extracts the content of law enforcement events, and generates a structured event description by chronological and logical arrangement. It matches the risk level and associated legal provisions corresponding to each risk mark with a preset risk level and legal provision association library, generating a risk label index. Based on a preset standard document template library, it selects a suitable template, converts the structured event description according to the template format to generate basic document content segments, inserts the risk label index into the corresponding positions of the basic document content segments, and generates a standard case document with attached legal risk labels. Simultaneously, it filters key risk information from the verification results marked with risk types, categorizes and typesets it according to the type of evidentiary flaw, and generates an independent evidentiary flaw report.

[0205] Among them, legal document synthesis processing is the process of converting verification results into standardized documents and reports; standard case documents with attached legal risk labels are law enforcement case documents that include risk labels; and independent evidence defect reports are special reports that record information on evidence defects.

[0206] The evidence verification module 304 is also used for:

[0207] Using the following formula, based on a preset set of legality rules, procedural behavior matching is performed on semantic description fragments with legal attribute tags to generate legality risk markers:

[0208]

[0209] in, Indicates a legality risk marker. Indicates the total number of legality rules. Representation rules The legal weighting coefficient, This represents the feature matrix of semantic description fragments with legal attribute labels. Representation rules The template vector of program behavior, Represents a non-linear activation function. Represents the legal semantic distance function. Representation rules The legal behavior threshold vector, Representation rules The threshold for determining violations Indicates an indicator function, Indicates transpose;

[0210] Based on a pre-defined set of objective rules, semantic description fragments with legal attribute tags are subjected to language neutrality analysis to generate objective risk markers.

[0211] Based on a pre-defined set of relevance rules, the semantic description fragments with legal attribute tags are subjected to evidence chain integrity verification processing to generate relevance defect markers.

[0212] The legality risk marker, objective risk marker, and related defect marker are dynamically weighted and fused to generate a verification result with risk type markers.

[0213] The evidence verification module 304 is also used for:

[0214] Using the following formula, based on preset time correlation rules, logical gap detection is performed on the event sequence of semantic description fragments with legal attribute tags to generate time chain break markers:

[0215] in, Indicates a time-broken chain marker. Indicates the start timestamp of the event sequence. Indicates the end timestamp of the event sequence. Indicates time Event description vector, Indicates the time interval between adjacent events. Indicates the threshold of time correlation. Indicates an indicator function;

[0216] Based on preset behavioral evidence association rules, the matching degree analysis of law enforcement actions in semantic description fragments and corresponding physical evidence descriptions is performed to generate physical evidence missing markers.

[0217] Keyframe back-tracing is performed on the visual action sequence features to obtain a video feature description;

[0218] Text element extraction processing is performed on semantic description fragments with legal attribute tags to obtain text record descriptions;

[0219] The following formula is used to perform consistency verification between video feature descriptions and text record descriptions, generating evidence conflict markers:

[0220]

[0221] in, Indicates a conflict of evidence marker. Indicates the total number of event nodes. Represents video feature description semantic encoding vector, Represents a text record description semantic encoding vector, Indicates distance from Manhattan. Represents the smoothing factor. This represents the node confidence score. Indicates the time decay coefficient. Indicates the video-to-text timestamp discrepancy. This represents the Sigmoid activation function. Indicates the event node index;

[0222] Based on a preset risk level matrix, the severity of defects in time chain break markers, missing physical evidence markers, and conflicting evidence markers is assessed, and relevant defect markers are generated.

[0223] The legal constraint embedding module 302 is also used for:

[0224] The legal provisions in the knowledge graph of evidence rules in administrative litigation are classified according to their urgency, and a dynamic set of weight coefficients is generated.

[0225] The original multimodal features are processed to separate visual and textual components, generating visual legal feature vectors and textual legal feature vectors;

[0226] Based on the clause priority in the dynamic weight coefficient set, the visual legal feature vector and the textual legal feature vector are differentially weighted to generate a weighted feature representation;

[0227] Based on the preset legal attribute mapping rules, the weighted feature representation is subjected to clause matching processing to generate initial legal constraint labels;

[0228] The following formula is used to perform cross-modal consistency verification on the initial legal constraint labels, eliminating conflicting tags between visual and textual features, and generating consistent legal constraint labels:

[0229]

[0230] in, Labels indicating consistency and legal constraints Indicates the initial legal binding label. Represents element-wise multiplication. Represents the conflict attenuation coefficient. This represents the gradient of the label with respect to visual features. This represents the gradient of the label with respect to the text features. Represents the squared Euclidean distance. Represents an exponential function;

[0231] Based on preset mandatory compliance clauses, key procedural behavior enhancement processing is performed on the consistency legal constraint label to generate multimodal features with legal constraint labels.

[0232] Legal fact generation module 303 is also used for:

[0233] Based on a pre-set law enforcement process template, multimodal features with legal constraint labels are processed for legal procedure stage identification, and procedure node boundary markers are generated.

[0234] Based on the boundary markers of program nodes, spatiotemporal correlation analysis is performed on the visual and textual components of multimodal features with legal constraint labels within the same program node to generate an evidence chain completeness score.

[0235] Based on the evidence chain completeness score, the legal meaning of the multimodal features with legal constraint labels in each program node is analyzed to generate initial legal attribute labels;

[0236] By using a pre-defined legal domain language model, program nodes with initial legal attribute labels are processed by natural language conversion to generate semantic description fragments with legal attribute labels.

[0237] Legal fact generation module 303 is also used for:

[0238] The core legal elements of the multimodal features with legal constraint labels within the program nodes are separated to generate legal element feature groups.

[0239] Based on the evidence chain completeness score, the confidence level of the legal element feature group is adjusted to generate the element features with confidence level adjustment.

[0240] Based on preset legal attribute determination rules, the credibility-adjusted element features are subjected to legal attribute matching processing to generate initial legal attribute labels.

[0241] The document synthesis output module 305 is also used for:

[0242] Based on the pre-defined logical relationships of law enforcement events, the verification results marked with risk types are sorted chronologically to generate structured event descriptions;

[0243] Based on a pre-defined database of risk levels and legal provisions, the structured event descriptions are matched with legal risk types to generate a risk labeling index.

[0244] Based on a pre-set standard document template library, the structured event description is formatted and converted to generate basic document content segments;

[0245] Based on the risk-marking index, the basic document content segments are marked and inserted to generate standard case documents with attached legal risk markings.

[0246] The risk markers in the verification results with risk type markings are processed for key information filtering to generate an independent evidence flaw report.

[0247] In one embodiment, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0248] In one embodiment, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above-described method embodiments.

[0249] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0250] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for assisting administrative law enforcement case document entry using a large language model, characterized in that, The method comprises: performing visual timing analysis processing on the input law enforcement record video stream to obtain visual action sequence features; performing text analysis processing on the input associated text data to obtain structured text features; performing multi-modal fusion processing on the visual action sequence features and the structured text features to obtain original multi-modal features; performing legal constraint embedding processing on the original multi-modal features based on a preset administrative litigation evidence rule knowledge graph to obtain multi-modal features with legal constraint labels; performing legal fact segment generation processing on the multi-modal features with legal constraint labels to obtain semantic description segments with legal attribute labels; performing evidence three-property dynamic checking processing on the semantic description segments with legal attribute labels to obtain a checking result with a risk type mark; performing legal document synthesis processing on the checking result with the risk type mark to generate standard case documents and independent evidence defect reports with legal risk annotations.

2. The method for assisting administrative law enforcement case document entry with a large language model according to claim 1, characterized in that, The evidence three-property dynamic checking processing on the semantic description segments with legal attribute labels to obtain a checking result with a risk type mark comprises: performing program behavior matching processing on the semantic description segments with legal attribute labels based on a preset legality rule set using the following formula to generate legality risk marks: wherein, denotes a legality risk marker, denotes a total number of legality rules, denotes a rule with a legal weight coefficient, denotes a semanticized description segment feature matrix with a legal attribute tag, denotes a program behavior template vector of a rule , denotes a nonlinear activation function, denotes a legal semantic distance function, denotes a legal behavior threshold vector of a rule , denotes a rule with a violation judgment threshold, denotes an indicator function, denotes a transpose; performing language neutrality analysis processing on the semantic description segments with legal attribute labels based on a preset objectivity rule set to generate objectivity risk marks; performing evidence chain integrity verification processing on the semantic description segments with legal attribute labels based on a preset relevance rule set to generate relevance defect marks; performing dynamic weight fusion processing on the legality risk marks, the objectivity risk marks, and the relevance defect marks to generate the checking result with the risk type mark.

3. The method for assisting administrative law enforcement case document entry with a large language model according to claim 2, characterized in that, The evidence chain integrity verification processing on the semantic description segments with legal attribute labels based on a preset relevance rule set to generate relevance defect marks comprises: Using the following formula, based on the preset time relevance rules, the logical gap detection processing is performed on the event timing in the semantic description segment with the legal attribute label, and a time break mark is generated: wherein, denotes a time break marker, denotes an event sequence start timestamp, denotes an event sequence end timestamp, denotes a time event description vector, denotes a neighboring event time interval, denotes a time correlation threshold, denotes an indicator function; performing matching degree analysis processing on law enforcement actions in the semantic description segments and corresponding physical evidence descriptions based on a preset behavior and physical evidence relevance rule to generate a physical evidence missing mark; performing key frame backtracking processing on the visual action sequence features to obtain video feature descriptions; performing text element extraction processing on the semantic description segments with legal attribute labels to obtain text record descriptions; performing consistency checking processing on the video feature descriptions and the text record descriptions using the following formula to generate an evidence conflict mark: in, Indicates a conflict of evidence marker. Indicates the total number of event nodes. Represents video feature description semantic encoding vector, Represents a text record description semantic encoding vector, Indicates distance from Manhattan. Represents the smoothing factor. This represents the node confidence score. Indicates the time decay coefficient. Indicates the video-to-text timestamp discrepancy. This represents the Sigmoid activation function. Indicates the event node index; performing defect severity evaluation processing on the time chain break mark, the physical evidence missing mark, and the evidence conflict mark based on a preset risk level matrix to generate the relevance defect mark.

4. The method for assisting administrative law enforcement case document entry with a large language model according to claim 1, characterized in that, The legal constraint embedding processing on the original multi-modal features based on a preset administrative litigation evidence rule knowledge graph to obtain multi-modal features with legal constraint labels comprises: performing emergency level classification processing on legal provisions in the administrative litigation evidence rule knowledge graph to generate a dynamic weight coefficient set; The original multi-modal feature is subjected to visual and text component separation processing to generate a visual legal feature vector and a text legal feature vector; According to the clause priority in the dynamic weight coefficient set, the visual legal feature vector and the text legal feature vector are subjected to differential weighting processing to generate a weighted feature representation; Based on a preset legal attribute mapping rule, the weighted feature representation is subjected to clause matching processing to generate an initial legal constraint label; Based on a preset legal attribute mapping rule, the weighted feature representation is subjected to clause matching processing to generate an initial legal constraint label; wherein, represents a consistency legal constraint tag, represents an initial legal constraint tag, represents an element-level multiplication, represents a conflict decay coefficient, represents a gradient of a label on a visual feature, represents a gradient of a label on a text feature, represents a squared Euclidean distance, represents an exponential function; Based on a preset legal attribute mapping rule, the weighted feature representation is subjected to clause matching processing to generate an initial legal constraint label; 5. The method for assisting administrative law enforcement case document entry with a large language model according to claim 1, characterized in that, The multi-modal feature with the legal constraint label is subjected to legal fact segment generation processing to obtain a semantic description segment with a legal attribute label, including: Based on a preset law enforcement process template, the multi-modal feature with the legal constraint label is subjected to legal procedure stage identification processing to generate a procedure node boundary marker; Based on the procedure node boundary marker, the visual component and the text component in the multi-modal feature with the legal constraint label within the same procedure node are subjected to spatio-temporal correlation analysis processing to generate an evidence chain completeness score; Based on the evidence chain completeness score, the multi-modal feature with the legal constraint label within each procedure node is subjected to legal significance analysis processing to generate an initial legal attribute label; The procedure node with the initial legal attribute label is subjected to natural language conversion processing by a preset legal domain large language model to generate the semantic description segment with the legal attribute label.

6. The method for assisting administrative law enforcement case document entry with a large language model according to claim 5, characterized in that, The multi-modal feature with the legal constraint label within each procedure node is subjected to legal significance analysis processing based on the evidence chain completeness score to generate an initial legal attribute label, including: The multi-modal feature with the legal constraint label within the procedure node is subjected to core legal element separation processing to generate a legal element feature group; According to the evidence chain completeness score, the legal element feature group is subjected to confidence correction processing to generate a credibility-adjusted element feature; Based on a preset legal attribute determination rule, the credibility-adjusted element feature is subjected to legal attribute matching processing to generate the initial legal attribute label.

7. The method for assisting administrative law enforcement case document entry with a large language model according to claim 1, characterized in that, The verification result with the risk type label is subjected to legal document synthesis processing to generate a standard case document with a legal risk label and an independent evidence defect report, including: Based on a preset law enforcement event logical relationship, the verification result with the risk type label is subjected to timeline sorting processing to generate a structured event description; Based on a preset risk level and article association library, the structured event description is subjected to legal risk type matching processing to generate a risk label index; Based on a preset standard document template library, the structured event description is subjected to format conversion processing to generate a basic document content segment; Based on the risk annotation index, the basic document content section is annotated and inserted to generate a standard case document with legal risk annotations; The risk markers in the risk type marked verification result are subjected to key information screening processing to generate an independent evidence flaw report.

8. A system for assisting administrative law enforcement case document entry with a large language model, characterized in that, The system comprises: A multi-modal feature extraction module is configured to perform visual time sequence analysis on an input law enforcement record video stream to obtain visual action sequence features, and perform text analysis on associated text data to obtain structured text features; A legal constraint embedding module is configured to perform multi-modal fusion processing on the visual action sequence features and the structured text features to obtain original multi-modal features, and perform legal constraint embedding processing on the original multi-modal features based on a pre-set administrative litigation evidence rule knowledge graph to obtain multi-modal features with legal constraint labels; A legal fact generation module is configured to perform legal fact segment generation processing on the multi-modal features with legal constraint labels to obtain semantic description segments with legal attribute labels; An evidence three-property verification module is configured to perform evidence three-property dynamic verification processing on the semantic description segments with legal attribute labels to obtain a verification result with risk type markers; A document synthesis output module is configured to perform legal document synthesis processing on the verification result with risk type markers to generate a standard case document with legal risk annotations and an independent evidence flaw report. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8. The processor executes the computer program to realize the steps of the method for assisting administrative law enforcement case document entry using a large language model according to any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the steps of the method for assisting administrative law enforcement case document entry using a large language model according to any one of claims 1 to 7.

Images