Intelligent semantic understanding-based case factor analysis and processing method

Abstract

The invention discloses a case factor intelligent analysis and processing method based on intelligent semantic understanding, and belongs to the field of artificial intelligence, natural language processing and public security judicial informationization, and the method comprises the following steps: 1, preprocessing public security judicial document information, and supporting text analysis, format standardization and structured data conversion; 2, carrying out refined extraction on case elements based on a large language model, and identifying case constituent elements, hazard degrees and processing results; 3, fusing logical reasoning of a public security judicial knowledge graph, and enhancing cross-cause semantic understanding and reasoning accuracy; and 4, performing intelligent auxiliary processing, generating class case recommendation and penalty interval prediction, and performing interactive regulation question and answer. According to the method, efficient, accurate and automatic case factor analysis is successfully realized, and the public security judicial document processing efficiency and decision accuracy are remarkably improved.

Description

Technical Field

[0001] The patent of this invention relates to the fields of artificial intelligence and public security and judicial informatization, and specifically is a method for analyzing and processing case elements based on intelligent semantic understanding. Background Art

[0002] Public security and judicial documents, such as interrogation records, investigation reports, and confidential materials, contain a wealth of critical case information and involve complex elements, including the subject, subjective intent, means of committing crimes, location of the incident, and the outcome of the disposition. Currently, the analysis and extraction of case elements primarily rely on manual annotation and retrieval, which is inefficient, costly, and prone to errors. While rule-based element extraction methods can reduce the manual burden, they are limited by predefined templates and keyword matching mechanisms, making them difficult to adapt to the heterogeneous nature of document formats and the diversity of case types. The continued surge in massive amounts of public security and judicial data further exacerbates the urgent need for efficient and accurate analysis.

[0003] Large-scale language models represented by ChatGPT and DeepSeek have demonstrated significant advantages in intelligent semantic understanding and information extraction. However, the adaptability of existing technologies in the vertical field of public security and justice still has obvious limitations. General large-scale language models lack professional training in the field of public security and justice, and it is difficult to accurately understand public security and judicial terminology, legal logic and case background. The semantic parsing and logical reasoning capabilities in cross-case scenarios are insufficient, making it difficult to cope with the differences in elements and complex relationships between different case types. In addition, existing methods have significant shortcomings in document processing, data security protection and human-computer interaction design, which limits their in-depth application in public security and judicial practice.

[0004] In summary, the invention of a case factor analysis and processing method based on intelligent semantic understanding provides an innovative solution for public security and judicial document processing and has high practical value. Summary of the Invention

[0005] To achieve the above objectives, the present invention provides a technical solution: a case factor analysis and processing method based on intelligent semantic understanding, which specifically includes the following steps:

[0006] Step S1: Preprocessing of multimodal information of public security and judicial documents: using the character recognition module to parse the image or PDF data of documents such as interrogation records, investigation reports or confidential materials, and analyze the text content and layout features. Regular expression matching and syntax analyzer are used to standardize the document structure and generate structured documents containing titles, party information, fact findings, and inquiry content. Construct a knowledge base in the field of public security and justice, integrating a professional dictionary of public security and justice, an entity type library, and a legal interpretation association rule library. The entity type library contains multiple types of legal element entity labels (such as case subjects, subjective aspects, case objects, objective behaviors, criminal tools, evidence types, legal provisions, processing results, etc.), and the legal interpretation association rule library covers multi-level legal provision citation relationships;

[0007] Step S2: Jointly extract multi-granular case elements. Based on a pre-trained general large-scale language model, domain-adaptive fine-tuning strategies are used to learn public security and judicial semantic features. The fine-tuning data consists of a cross-case training set consisting of tens of thousands of annotated interrogation transcripts, investigation reports, or confidential documents. A hierarchical prompt template is designed to extract "subject-behavior-result" triples in the factual description section and extract the association between the treatment result and the legal basis in the treatment reason section. A multi-head attention mechanism is used to fuse text features with knowledge base entity vectors, outputting a structured element set containing the case subject, subjective intent, illegal and criminal means, degree of harm, and treatment result.

[0008] Step S3: Public security judicial logical reasoning and contradiction prediction, constructing a dynamic public security judicial interpretation knowledge graph, the nodes contain three types of entities: legal provisions, case elements, and evidence types, and the edge relationships define the applicable conditions of the provisions, the element proof path, and the closure rules of the evidence chain. Use the graph neural network reasoning mechanism to verify the legality of the extracted elements: verify the integrity of the case constituent elements through subgraph matching, detect missing elements and generate completion suggestions. Apply the rule reasoning module to judge the adaptability of the processing results and the legal provisions, output the legal application conflict warning, and combine the probability reasoning model to calculate the credibility of the evidence chain. When the support of key evidence is lower than the threshold, a review prompt is triggered;

[0009] Step S4: Intelligent Assisted Processing and Interpretable Output: Structured elements are input into the penalty prediction model, which includes a cascaded MLP classifier. This model outputs penalty intervals and probability distributions, and visualizes the weights of influencing factors. Combined with a similar case recommendation engine, this model uses cosine similarity of element vectors to match historical case libraries, returning a comparative analysis table of several similar cases and their key handling points. An interactive regulatory Q&A interface is generated, supporting automatic parsing of natural language queries. When returning answers, the underlying regulatory text and relevant case snippets are displayed simultaneously.

[0010] Preferably, the character recognition module in step S1 includes the following steps:

[0011] Step 1.1 Text region detection

[0012] A modified network based on YOLOv5, by optimizing the feature extraction layer and adding a rotated bounding box prediction head, is used to detect and locate tilted text regions in public security and judicial documents such as interrogation transcripts, investigation reports, and confidential materials. This improved network incorporates a multi-scale feature fusion module to enhance detection capabilities for text of varying sizes and angles, and incorporates a tilt angle loss function to improve the accuracy of rotated bounding box positioning. This step addresses common issues with tilted and low-resolution scans of public security and judicial documents, accurately identifying the bounding coordinates of text regions and providing high-quality input data for subsequent text recognition.

[0013] Step 1.2 Text Recognition

[0014] Combining a generative adversarial network (GAN) and an attention mechanism, adversarial generative repair is performed on blurred or stamp-covered public security and judicial documents, generating clear text. The GAN uses a generator to reconstruct blurred text areas, while a discriminator optimizes the authenticity of the repaired text. The attention mechanism focuses on key text areas, prioritizing text segments containing case elements such as the parties' names, time, and location. This step addresses common issues such as stamp obstruction and blurred ink in public security and judicial documents, significantly improving text recognition accuracy and generating structured text sequences to support subsequent semantic analysis.

[0015] Step 1.3 Layout Analysis

[0016] A graph convolutional network (GCN) is used to extract coordinate metadata for multimodal elements such as forms, signatures, and seals in public security and judicial documents, and analyze their semantic relationships for subsequent case analysis. GCN constructs a topological graph of the document layout, identifies form boundaries, signature locations, and seal types, and aggregates contextual information about multimodal elements by combining node features. The generated coordinate metadata includes element type, location, and associations, supporting structured extraction and semantic reasoning of case elements.

[0017] Preferably, the multi-granularity case factor joint extraction method in step S2 includes the following steps:

[0018] Step 2.1 Model fine-tuning

[0019] Using a pre-trained general-purpose large-scale language model, through a domain-adaptive fine-tuning strategy, and using a training set containing tens of thousands of cross-case annotated documents , learning the semantic features unique to the public security and judicial fields. The fine-tuning optimization goal is to minimize the loss function ,in is the cross entropy loss, is the domain adaptation item, is the regularization term, and are weight coefficients to improve the model's understanding of public security and judicial terminology and case logic;

[0020] Step 2.2 Tip template design

[0021] Designing a layered prompt template ,in The extraction of "subject-behavior-result" triples for fact description paragraphs is defined as , Indicates the subject of the case, Indicates behavioral relationships. Indicates harmful consequences; The processing result used to process the reason paragraph is associated with the legal basis and is defined as , Indicates the processing result. Represents the legal basis, which is used to guide the model to accurately locate case elements of different granularities;

[0022] Step 2.3 Feature extraction

[0023] Based on the fine-tuned language model, combined with the hierarchical prompt template, triple extraction is performed in the fact description paragraph, and the output set , perform association extraction in the processing reason paragraph and output the set , generating a preliminary element set including the subject of the case, subjective intent, illegal and criminal means, degree of harm, and handling results;

[0024] Step 2.4 Feature fusion and output

[0025] Adopt multi-head attention mechanism to fuse text feature vectors and the entity vector of the public security and judicial knowledge base , calculate the fusion features ,in Represents a multi-head attention function to enhance the semantic relevance between elements. Based on fusion features , through the decoding function , combined with the prompt template , mapping the fusion features into a structured feature set ,in Provides a semantic representation basis, It provides explicit output for subsequent semantic reasoning and auxiliary processing.

[0026] Preferably, the construction and reasoning method of the dynamic public security judicial interpretation knowledge graph in step S3 includes the following steps:

[0027] Step 3.1 Knowledge Graph Construction

[0028] Knowledge graph construction: building a dynamic public security and judicial interpretation knowledge graph , where the node set Contains regulatory provisions , Case Elements and type of evidence Three types of entities, edge sets Defined as the conditions for application of the provisions , element proof path and the rule of evidence chain closure , extracting "if-then" conditional rules through dependency analysis of legal provisions Establish a mapping relationship between clauses and constituent elements, and use the TransR knowledge representation learning model to embed case elements into the vector space , and set up a dynamic update interface to trigger subgraph reconstruction and association rule regeneration when new public security judicial interpretations are released;

[0029] Step 3.2 Graph Neural Network Verification

[0030] Adopting the graph neural network reasoning mechanism, the extracted case elements are verified for legitimacy, the integrity of the case elements is verified through the subgraph matching algorithm, and the node feature updates are calculated. , detect missing elements and generate completion suggestions ;

[0031] Step 3.3 Rule Reasoning and Conflict Warning

[0032] Apply rule reasoning module to calculate processing results and regulatory basis The fitness function ,when Output regulatory conflict warning , combined with the edge relationship of the public security and judicial knowledge graph and Analyze the causes of conflict;

[0033] Step 3.4 Probabilistic Reasoning and Review Tips

[0034] Combined with the probability reasoning model, calculate the credibility of the evidence chain ,in Indicates the Item of evidence in the evidence subset The conditional probability under The weight of evidence is dynamically adjusted based on the relevance of evidence and the input of new evidence. Set the credibility threshold ,when Trigger key evidence review prompt , prompts include the type of missing evidence and suggestions for completion , through the Bayesian update rule Incorporating new evidence Reassess credibility and support public security and judicial personnel in optimizing the chain of evidence and improving processing accuracy.

[0035] Preferably, the intelligent assisted processing and interpretable output method in step S4 includes the following steps:

[0036] Step 4.1 Penalty Prediction

[0037] The structured elements are input into the penalty prediction model, which contains a cascaded multi-layer perceptron (MLP) classifier. First, the first-level network extracts continuous features such as penalty amount and degree of injury, and calculates the baseline penalty prediction value. Then, the second-level network processes discrete circumstances such as surrender and recidivism to calculate the penalty adjustment range. Finally, the attention mechanism is introduced through the fusion module to dynamically allocate feature weights , output the penalty interval and probability distribution including upper and lower limits , used to assist public security and judicial processing;

[0038] Step 4.2 Recommendation of similar cases

[0039] Combined with the similar case recommendation engine, by calculating the current element vector Vector with historical case library Cosine similarity of ,match several similar cases from the public security and judicial historical case database, return a comparative analysis table of key processing points, and integrate new processing data into the case database through a dynamic update interface to maintain the timeliness of recommendations;

[0040] Step 4.3 Visualization and Interaction

[0041] Generate weight visualization module to present the weight of influencing factors in the form of charts The contribution degree of public security and judicial personnel is supported to intuitively evaluate the basis for handling; at the same time, an interactive legal question-and-answer interface is built to support natural language queries. , automatically parse the query intent through the semantic parsing model and generate answers , and simultaneously display the legal basis provisions and related case fragments to improve the efficiency of public security and judicial processing.

[0042] Compared with the prior art, the present invention has the following advantages and effects:

[0043] 1. The present invention proposes a case factor analysis and processing method based on intelligent semantic understanding. It adopts advanced intelligent semantic understanding technology, significantly improving the adaptability and accuracy of extracting elements from public security and judicial documents, and surpassing the limitations of traditional manual labeling and rule matching methods.

[0044] 2. The present invention adopts efficient logical reasoning and interactive processing support to enhance the depth and transparency of public security and judicial analysis, and is superior to the shortcomings of existing technologies in complex reasoning and human-computer collaboration. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a framework diagram of the system of the present invention.

[0046] Figure 2 This is the flow chart of the multimodal information preprocessing module for public security and judicial documents.

[0047] Figure 3 Schematic diagram of the multi-granularity case factor joint extraction module.

[0048] Figure 4 This is a schematic diagram of the public security and judicial logic reasoning and contradiction prediction module.

[0049] Figure 5 Schematic diagram of the intelligent auxiliary processing and explainable output module. DETAILED DESCRIPTION

[0050] The present invention will be described in further detail below with reference to specific embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0051] like Figure 1 As shown, this embodiment discloses a case factor analysis and processing method based on intelligent semantic understanding, including the following steps:

[0052] Step 1: Preprocessing of multimodal information of public security and judicial documents. The multimodal information preprocessing module of public security and judicial documents is as follows: Figure 2 As shown in the figure, the text region detection is performed based on the improved YOLOv5 network, and a rotation bounding box prediction head is added to adapt to tilted text. The improved YOLOv5 enhances the detection capability of text regions of different sizes by introducing a multi-scale feature pyramid FPN. The rotation bounding box prediction head predicts the angle parameters of the text region. , the bounding box loss function is defined as:

[0053]

[0054] in is the rotation regression loss, is the confidence loss, and The detection process first inputs the document image or PDF data into the network, and the output includes the position coordinates and rotation angle , and generate a preliminary set of text regions.

[0055] Combining Generative Adversarial Networks (GANs) and attention mechanisms to repair blurred or stamp-covered text. The GAN generator uses a convolutional neural network to reconstruct the blurred area, and the discriminator uses the PatchGAN structure to determine the authenticity of the generated results. The adversarial loss is defined as:

[0056]

[0057] in is the generator output, is the discriminator result, is the real text. The attention mechanism calculates the weight matrix of the text region based on self-attention, which is defined as:

[0058]

[0059] in 、 、 are the query, key and value matrices after linear transformation, is the dimension coefficient, which is used to focus on key text information such as the party’s name and date.

[0060] After repair, a clear text sequence is generated, the obscured subject information is accurately restored, and the blurred time-related fields are effectively restored. The repaired text area is serialized and the text sequence is processed using a Transformer-based model. Input the repaired text sequence To the Transformer encoder, long-range context dependencies are captured through multi-head self-attention calculations. The encoder output sequence representation , through residual connection and layer normalization optimization, context-coherent serialized text is generated.

[0061] Graph Convolutional Network (GCN) is used to analyze the layout and extract text information elements. GCN first constructs a layout topology graph, with text areas, tables, signatures, and seals as nodes, and edges representing spatial adjacency. The node feature aggregation formula is defined as:

[0062]

[0063] in For the Layer node features, For nodes Neighborhood, is the weight matrix, is the activation function. GCN outputs the table boundary coordinates , signature location coordinates and seal area information.

[0064] Through text character recognition and layout analysis, the extraction of structured text and metadata is completed to support the intelligent element extraction, analysis and processing of subsequent documents.

[0065] Step 2: Joint extraction of multi-granularity case elements. The multi-granularity case element joint extraction module is as follows: Figure 3 As shown, domain adaptive fine-tuning is performed based on the pre-trained large language model, using the training set Optimize model parameters. The fine-tuning process uses a joint loss function defined as:

[0066]

[0067] in is the cross entropy loss, is the domain adaptation loss, is the regularization term, and are weight hyperparameters respectively. Model parameter update is performed by gradient descent method, which is defined as:

[0068]

[0069] in is the current parameter, is the learning rate, is the gradient of the loss function.

[0070] Designing a layered prompt template , corresponding to the fact layer, reason layer and regulation layer respectively. The template embedding process is generated through linear transformation and is defined as:

[0071]

[0072] in and are weight matrix and bias vector respectively. Based on the prompt template, the case elements are extracted through attention weighted calculation, which is defined as:

[0073]

[0074] in is the input hidden state, is the attention weight matrix, output attention distribution Used to guide feature extraction.

[0075] A multi-head attention mechanism is used to fuse text features and knowledge base features. The calculation process is:

[0076]

[0077] in , 、 、 are the query, key, and value projection matrices, respectively, is the output projection matrix, is the number of attention heads. The fused features are further optimized through residual connections and layer normalization, defined as:

[0078]

[0079] The decoding function is used to generate a set of structured features. The decoding process is based on conditional probability processing and is defined as:

[0080]

[0081] in For the elements, For the front elements, is the sequence length. The decoding function calculates the feature probability through the Softmax layer and is defined as:

[0082]

[0083] in and are weights and biases respectively. The final output is a set of structured features .

[0084] The feature set is optimized through the post-processing module. An orthogonal constraint is applied to ensure independence between features, defined as:

[0085]

[0086] in is the element matrix, is the unit matrix. The redundant elements are further eliminated by cosine similarity, which is defined as:

[0087]

[0088] in and are two element vectors respectively. The optimized set is projected generate is the redundant direction vector.

[0089] By jointly extracting multi-granularity elements, the structured representation of document elements is completed to support subsequent logical reasoning and processing analysis.

[0090] Step 3: Public security and judicial logic reasoning and contradiction prediction. The public security and judicial logic reasoning and contradiction prediction module is as follows: Figure 4 As shown, building a dynamic knowledge graph , node set Contains regulatory provisions , Case Elements and type of evidence , edge set Including the applicable conditions of the provisions , element proof path and the rule of evidence chain closure Using the TransR model for embedding representation, we calculate the vector representation of entities and relations, which is defined as:

[0091] and

[0092] in and are the original vectors of the head entity and the tail entity respectively, is the relationship vector, is a relation-specific projection matrix. Graph neural network GNN is used for legitimacy verification and node feature matrix is ​​initialized. Based on the embedding representation, the node feature update is performed through the message passing formula, which is defined as:

[0093]

[0094] in For nodes Neighborhood, is the normalization constant, is the adjacency weight, is the self-loop weight, is the activation function. Subgraph matching is done by cosine similarity Measuring the structural consistency between nodes and generating a completion suggestion vector .

[0095] Use the rule reasoning module to calculate the degree of fit between the processing results and the regulatory basis. Define the fitness function:

[0096]

[0097] in is the result vector, is the regulation vector, is the mapping matrix. When , conflict detection is triggered and conflict score is calculated . Further optimization through logical constraints is defined as:

[0098]

[0099] in is the fitness matrix, is the conflict direction vector.

[0100] The credibility of the evidence chain is evaluated using a probabilistic reasoning model. The initial probability distribution of evidence is defined as:

[0101]

[0102] in Embedded for evidence, is the weight matrix. The joint probability calculation is defined as:

[0103]

[0104] in is the weight of evidence, For the characteristics of evidence, is the weight parameter. When the probability of triggering a review .

[0105] Finally, the knowledge graph is adjusted through graph embedding optimization. The global consistency loss of the graph is calculated and defined as:

[0106]

[0107] in is the attention coefficient, and the updated embedding is .

[0108] Through public security and judicial logical reasoning and contradiction prediction, the legitimacy verification and risk warning of case elements are completed to support intelligent analysis of public security and judicial processing.

[0109] Step 4: Intelligent auxiliary processing and interpretable output. The intelligent auxiliary processing and interpretable output module is as follows: Figure 5 As shown in Figure 2, the structured elements are input into the penalty prediction model and a cascaded multi-layer perceptron (MLP) is used for prediction. The first-level MLP calculates the baseline penalty result, which is defined as:

[0110]

[0111] in is the input feature vector, and are weight matrices and biases respectively. The second-level MLP calculates the adjustment amplitude, which is defined as:

[0112]

[0113] in is the first-level output, and are weights and biases respectively, is the hyperbolic tangent function. The fusion module assigns weights through the attention mechanism and finally outputs the penalty probability distribution .

[0114] Build a similar case recommendation engine and match based on cosine similarity. Define the similarity calculation process:

[0115]

[0116] in is the feature vector of the current case, is the feature vector of public security and judicial historical cases. The recommendation process is determined by weighted sorting and is defined as:

[0117]

[0118] in is the weight coefficient, is the number of candidate cases. The redundancy removal process is further optimized by orthogonal projection, which is defined as:

[0119]

[0120] in is the score matrix, Redundant direction.

[0121] Construct a visual representation of the weights and calculate the contribution as a percentage, defined as:

[0122]

[0123] The weight distribution graph is drawn by normalization, which is defined as:

[0124]

[0125] The visualization loss function is defined as:

[0126]

[0127] in For the target distribution, optimization is done by gradient update conduct.

[0128] Finally, an interactive question-answering interface is built based on natural language queries. Generate answers. Query embeddings are calculated as follows:

[0129]

[0130] in and are weights and biases respectively. Answer generation is obtained by calculating the conditional probability, which is defined as:

[0131]

[0132] in is the historical context vector, and are weights and biases respectively. The confidence evaluation is calculated as follows and is defined as:

[0133]

[0134] in is the confidence weight.

[0135] Finally, the optimization process outputs consistency. The global processing score is calculated and defined as:

[0136]

[0137] in To process weights, To process vectors, is the number of processes. The consistency constraint is minimized as follows:

[0138]

[0139] Updated by conduct.

[0140] Through intelligent assisted processing and explainable output, the automated generation of support for public security and judicial processing is completed to support transparent analysis of the public security and judicial process.

[0141] The above embodiments are preferred implementation modes of the present invention and are not intended to limit the scope of implementation of the present invention. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be considered as equivalent replacement methods and are covered by the scope of protection of the present invention.

Claims

1. A case factor analysis and processing method based on intelligent semantic understanding, characterized in that: include: Step S1: Preprocessing the multimodal information of public security and judicial documents: using a character recognition module to parse the image or PDF data of input documents such as interrogation records, investigation reports, or confidential materials, and analyze the text content and layout features; using regular expression matching and syntax analyzer to standardize the document structure and generate a structured document containing a title, party information, fact determination, and inquiry content; constructing a public security and judicial knowledge base, integrating a public security and judicial professional dictionary, an entity type library, and a legal interpretation association rule library, wherein the entity type library contains at least one legal element entity label of the criminal subject, subjective aspect, criminal object, objective behavior, criminal tool, evidence type, legal provision, and processing result; and the legal interpretation association rule library covers multi-level legal provision reference relationships; Step S2: Joint extraction of multi-granular case elements. Based on a pre-trained general large-scale language model, domain-adaptive fine-tuning strategies are used to learn public security and judicial semantic features. The fine-tuning data includes a cross-case training set consisting of tens of thousands of annotated interrogation records, investigation reports, or confidential documents. A hierarchical prompt template is designed to extract "subject-behavior-result" triples in the fact description paragraph and perform association extraction between treatment result and legal basis in the treatment reason paragraph. A multi-head attention mechanism is used to fuse text features with knowledge base entity vectors to output a structured element set including the case subject, subjective intent, illegal and criminal means, degree of harm, and treatment result. Step S3: Public security judicial logical reasoning and contradiction prediction, constructing a dynamic public security judicial interpretation knowledge graph, with nodes containing three types of entities: legal provisions, case elements, and evidence types. Edge relationships define the applicable conditions of the provisions, the path of element proof, and the closure rules of the evidence chain; using the graph neural network reasoning mechanism to verify the legality of the extracted elements: verify the integrity of the case constituent elements through subgraph matching, detect missing elements and generate completion suggestions. Apply the rule reasoning module to judge the adaptability of the processing results and legal provisions, output legal application conflict warnings, and calculate the credibility of the evidence chain based on the probability reasoning model. When the support of key evidence is lower than the threshold, a review prompt is triggered; Step S4: Intelligent assisted processing and explainable output: input structured elements into the penalty prediction model, which contains a cascaded MLP classifier, outputs the penalty interval and probability distribution, and visualizes the weights of influencing factors; combined with the similar case recommendation engine, it matches the historical case library through the cosine similarity of the element vectors, returns a comparative analysis table of several similar cases and their key points of processing, and generates an interactive regulatory question-and-answer interface that supports automatic parsing of natural language queries. When returning answers, it simultaneously displays the regulatory basis and related case fragments.

2. The method according to claim 1, characterized in that The character recognition module in step S1 includes the following steps: Step 2.1 Text area detection; An improved network based on YOLOv5 is used to detect and locate tilted text regions in public security and judicial documents such as interrogation transcripts, investigation reports, or confidential materials by optimizing the feature extraction layer and adding a rotated bounding box prediction head. The improved network introduces a multi-scale feature fusion module to enhance the detection capability of text of different sizes and angles, and combines it with a tilt angle loss function to improve the positioning accuracy of the rotated bounding box. This step accurately identifies the boundary coordinates of text regions, addressing common problems such as tilt and low resolution in scanned documents of public security and judicial documents, providing high-quality input data for subsequent text recognition. Step 2.2: Text recognition; Combining a generative adversarial network (GAN) and an attention mechanism, adversarial generative repair is performed on blurred or stamp-covered public security and judicial documents, generating clear text content. The GAN uses a generator to reconstruct blurred text areas, while a discriminator optimizes the authenticity of the repaired text. The attention mechanism focuses on key text areas, prioritizing text segments containing case elements such as the parties' names, time, and location. This step significantly improves text recognition accuracy by addressing common problems such as stamp occlusion and blurred ink in public security and judicial documents, and generates structured text sequences to support subsequent semantic analysis. Step 2.3: Layout analysis; Graph Convolutional Network (GCN) is used to extract coordinate metadata of multimodal elements such as tables, signatures, and seals in public security and judicial documents, and analyze their semantic relationships for subsequent case factor analysis. GCN constructs a topological map of the document layout to identify table boundaries, signature locations, and seal types, and aggregates the contextual information of multimodal elements based on node features. The generated coordinate metadata includes element type, location, and association relationships, providing support for structured extraction and semantic reasoning of case elements.

3. The method according to claim 1, characterized in that The multi-granularity case factor joint extraction method in step S2 includes the following steps: Step 3.1: Model fine-tuning; Using a pre-trained general-purpose large-scale language model, through a domain-adaptive fine-tuning strategy, and using a training set containing tens of thousands of cross-case annotated documents , learn the semantic features unique to the public security and judicial fields; fine-tune the optimization goal to minimize the loss function ,in is the cross entropy loss, is the domain adaptation item, is the regularization term, and are weight coefficients to improve the model's understanding of public security and judicial terminology and case logic; Step 3.2: Design the prompt template; Designing a layered prompt template ,in The extraction of "subject-behavior-result" triples for fact description paragraphs is defined as , Indicates the subject of the case, Indicates behavioral relationships. Indicates harmful consequences; The processing result used to process the reason paragraph is associated with the legal basis and is defined as , Indicates the processing result. Represents the legal basis, which is used to guide the model to accurately locate case elements of different granularities; Step 3.3: Feature extraction; Based on the fine-tuned language model, combined with the hierarchical prompt template, triple extraction is performed in the fact description paragraph, and the output set , perform association extraction in the processing reason paragraph and output the set , generating a preliminary element set including the subject of the case, subjective intent, illegal and criminal means, degree of harm, and handling results; Step 3.4: Feature fusion and output; Adopt multi-head attention mechanism to fuse text feature vectors and the entity vector of the public security and judicial knowledge base , calculate the fusion features ,in Represents a multi-head attention function to enhance the semantic relevance between elements. Based on fusion features , through the decoding function , combined with the prompt template , mapping the fusion features into a structured feature set ,in Provide a semantic representation basis, It provides explicit output for subsequent semantic reasoning and auxiliary processing.

4. The method according to claim 1, wherein The construction and reasoning method of the dynamic judicial interpretation knowledge graph described in step S3 includes the following steps: Step 4.1: Knowledge graph construction; Knowledge graph construction: building a dynamic public security and judicial interpretation knowledge graph , where the node set Contains regulatory provisions , Case Elements and type of evidence Three types of entities, edge sets Defined as the conditions for application of the provisions , element proof path and the rule of evidence chain closure , extracting "if-then" conditional rules through dependency analysis of legal provisions Establish a mapping relationship between clauses and constituent elements, and use the TransR knowledge representation learning model to embed case elements into the vector space , and set up a dynamic update interface to trigger subgraph reconstruction and association rule regeneration when new public security judicial interpretations are released; Step 4.2 Graph neural network verification; Adopting the graph neural network reasoning mechanism, the extracted case elements are verified for legitimacy, the integrity of the case elements is verified through the subgraph matching algorithm, and the node feature updates are calculated. , detect missing elements and generate completion suggestions ; Step 4.3: Rule reasoning and conflict warning; Apply rule reasoning module to calculate processing results and regulatory basis The fitness function ,when Output regulatory conflict warning , combined with the edge relationship of the public security and judicial knowledge graph and Analyze the causes of conflict; Step 4.4 Probabilistic Reasoning and Review Tips; Combined with the probability reasoning model, calculate the credibility of the evidence chain ,in Indicates the Items of evidence in the evidence subset The conditional probability under Dynamically adjust the weight of evidence based on evidence relevance and new evidence input; set credibility thresholds ,when Trigger key evidence review prompt , prompts include the type of missing evidence and suggestions for completion , through the Bayesian update rule Incorporating new evidence Reassess credibility and support public security and judicial personnel in optimizing the chain of evidence and improving processing accuracy.

5. The method according to claim 1, wherein The intelligent auxiliary processing and interpretable output method in step S4 includes the following steps: Step 5.1 Penalty prediction; The structured elements are input into the penalty prediction model, which contains a cascaded multi-layer perceptron (MLP) classifier. First, the first-level network extracts continuous features such as penalty amount and degree of injury, and calculates the baseline penalty prediction value. Then, the second-level network processes discrete circumstances such as surrender and recidivism to calculate the penalty adjustment range. Finally, the attention mechanism is introduced through the fusion module to dynamically allocate feature weights , output the penalty interval and probability distribution including upper and lower limits , used to assist public security and judicial processing; Step 5.2: Recommend similar cases; Combined with the similar case recommendation engine, by calculating the current element vector Vector with historical case library Cosine similarity of ,match several similar cases from the public security and judicial historical case database, return a comparative analysis table of key processing points, and integrate new processing data into the case database through a dynamic update interface to maintain the timeliness of recommendations; Step 5.3: Visualization and interaction; Generate weight visualization module to present the weight of influencing factors in the form of charts The contribution degree of public security and judicial personnel is supported to intuitively evaluate the basis for handling; at the same time, an interactive legal question-and-answer interface is built to support natural language queries. , automatically parse the query intent through the semantic parsing model and generate answers , and simultaneously display the legal basis provisions and related case fragments to improve the efficiency of public security and judicial processing.

Images