Intelligent system and method for government affairs review based on adaptive vision and cross-modal reasoning

Abstract

The application provides a government affair review intelligent system and method based on adaptive vision and cross-modal reasoning, and belongs to the technical field of government affair informatization. The system comprises a preprocessing module for self-correction enhancement processing to generate standardized images; a difference detection module for detecting visual and text differences of multi-version government materials; a visual text coding module for generating multi-modal embedding representation of a unified vector space; a cross-modal review module for generating compliance scores; a dynamic Token module for dynamically adjusting image patch division to optimize target area perception ability; a bidirectional alignment module for constructing semantic mapping of risk features and regulations; an evidence chain generation module for generating traceable evidence graph structure; a risk early warning module for risk assessment and triggering early warning levels; and a decision module for generating final review conclusions and reason chains. The application realizes high cross-modal review accuracy, low-quality image enhancement recognition, reduced missed review rate, full-link review evidence chain, and automatic high-risk early warning.

Description

Technical Field

[0001] This invention relates to the field of e-government technology, specifically to an intelligent system and method for e-government review based on adaptive vision and cross-modal reasoning. Background Technology

[0002] In government affairs sectors such as market supervision and administration, and administrative approval, the review materials contain a large amount of multimodal data, such as images of certificates and licenses, scene photos, application form texts, and legal clauses. Currently, intelligent government affairs review systems are widely used in departments such as market supervision administration bureaus, administrative approval service bureaus, and emergency management bureaus for certificate and license approval, scene verification, and compliance review, covering sub-scenarios such as enterprise registration, business licenses, and special equipment supervision.

[0003] In existing technologies, government review systems mainly rely on manual verification, which typically suffers from the following problems: First, cross-modal consistency review is difficult, as fields in images (e.g., business licenses) and text (e.g., application forms) are often inconsistent, and manual verification is time-consuming; second, low-quality government images have low recognition rates, as blurry, skewed, or reflective photographs reduce the accuracy of OCR and scene recognition; third, regulations and violations cannot be automatically identified, and scene photos are difficult to automatically determine whether they violate relevant legal provisions; fourth, the review process lacks automation, requiring manual judgment for material understanding, regulation matching, risk identification, and conclusion output; and finally, comparison of multiple versions of materials lacks technical support, as companies submit materials multiple times, easily changing images and text, and existing government review systems cannot accurately identify version differences. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent system and method for government affairs review based on adaptive vision and cross-modal reasoning, in order to solve the technical problems of existing government affairs review systems, such as difficulty in cross-modal consistency review, low recognition rate of low-quality government affairs images, inability to automatically locate regulations and violations, lack of automation in the review process, and lack of technical support for comparison of multiple versions of materials.

[0005] To achieve the above objectives, this invention provides an intelligent system for government affairs review based on adaptive vision and cross-modal reasoning. The intelligent system includes: a preprocessing module for performing self-correction enhancement processing on low-quality government affairs material images to generate standardized images; a difference detection module for detecting visual and textual differences between multiple versions of government affairs materials based on the standardized images and the OCR text stream of the government affairs materials, and generating a structured difference report; a visual text encoding module for generating a multimodal embedding representation in a unified vector space based on the standardized images and the original text of the government affairs materials using multimodal alignment technology; a cross-modal review module for checking the consistency between image fields and text fields based on the multimodal embedding representation and the OCR text stream of the government affairs materials, and generating a compliance score; and a dynamic token module for using the standardized image... The system comprises four modules: a visual feature vector representation based on image and multimodal embeddings; a dynamic adjustment module for image patch segmentation; a bidirectional alignment module for constructing a semantic mapping between risk features and regulatory provisions based on image feature vectors and regulatory vectors using contrastive learning methods; an evidence chain generation module for generating a traceable evidence graph structure using visualization tools based on matching results between image and text fields, compliance scores, semantic mapping results, and multimodal embeddings; a risk warning module for conducting risk assessments and triggering warning levels based on multimodal embeddings, compliance scores, and historical operational data using a rule engine; and a decision-making module for generating the final review conclusion and reasoning chain based on the evidence graph structure, risk assessment results, and regulatory provision retrieval results using a chain-reasoning method.

[0006] Optionally, the step of performing self-correction enhancement processing on low-quality government affairs material images from government affairs review to generate standardized images includes: performing multi-scale deblurring processing on the low-quality government affairs material images, dynamically adjusting the intensity in conjunction with EXIF ​​metadata to generate a deblurred image; based on the deblurred image, identifying text regions and performing super-resolution enhancement using a text detector, and triggering optimization through OCR confidence; performing geometric correction on the enhanced image, correcting image tilt and distortion through a corner regression network and affine transformation; performing exposure compensation on the geometrically corrected image, using a Retinex model to balance brightness, and recording processing parameters.

[0007] Optionally, the step of detecting visual and textual differences between multiple versions of government documents and generating a structured difference report includes: extracting multi-feature visual fingerprints and calculating similarity based on the standardized image to generate a visual difference index; applying a character-level diff algorithm to compare the text versions of government documents based on the OCR text stream, and performing rigorous comparison for target fields; associating the compared visual and textual differences through an attention mechanism to construct an evidence graph and perform time series analysis; generating a structured difference report based on the scores of visual and textual differences, and triggering suggested actions.

[0008] Optionally, the step of generating a multimodal embedding representation in a unified vector space using multimodal alignment technology includes: generating visual vectors by performing patch partitioning and multi-scale feature extraction based on standardized images using VisionTransformer; performing semantic alignment based on visual vectors and the original text of government documents through contrastive learning methods, and supporting cross-modal retrieval; encoding the regulatory text of government documents into vector segments and injecting regulatory structure metadata to construct a vector index; and using the Hungarian matching algorithm to calculate the optimal pairing of image patches and text spans to generate a multimodal embedding representation in a unified vector space.

[0009] Optionally, checking the consistency between the image field and the text field to generate a compliance score includes: extracting text content and layout information based on a standardized image using the LayoutViT model; identifying the target field using a preset government entity recognition model based on an OCR text stream; aligning the image field and the text field by calculating the similarity through a two-layer alignment strategy; and calculating the compliance score based on the alignment result using a preset scoring formula and generating an interpretability result.

[0010] Optionally, the dynamic adjustment of image patch segmentation and optimization of the perception capability of the target region includes: calculating an importance score based on a standardized image using a weighted fusion formula, and dynamically adjusting a preset threshold according to the quality of the standardized image; dynamically segmenting the patch size based on the importance score using a preset segmentation algorithm, and performing multi-scale feature fusion through a pyramid patching mechanism; merging tokens based on low-importance regions in the patch segmentation results to optimize computational cost; and calculating importance perception loss during the training phase of the segmentation algorithm to optimize the perception capability of the target region and enhance the alignment effect of the target region.

[0011] Optionally, the step of constructing a semantic mapping between risk features and legal provisions through contrastive learning includes: defining a set of risk concepts based on historical cases of government violations, generating visual prototype embedding vectors, and constructing a visual risk feature library; segmenting and encoding the legal texts in the legal knowledge base into legal vectors, and constructing an efficient vector index; training a model through bidirectional contrastive loss to achieve semantic alignment between visual features and legal provisions; detecting risk concepts based on new government material images, combined with the visual risk feature library, retrieving relevant legal provisions, and generating an interpretable chain of evidence.

[0012] Optionally, the step of using visualization tools to generate a traceable evidence graph structure includes: defining node types and instantiation based on the output of the preceding modules, and establishing a node relationship index and node update mechanism; defining edge types according to node types, and calculating edge weights based on the output of the preceding modules to construct logical relationships of the evidence chain; constructing a complete evidence graph structure based on the logical relationships of the evidence chain through evidence anchoring and logical node integration; and generating interactive views and reports through structured queries and visualization tools.

[0013] Optionally, the step of conducting risk assessment and triggering early warning levels through a rule engine includes: calculating anomaly scores for visual and textual modalities based on multimodal embedding representation using Gaussian mixture models and deep energy models, and fusing them into a final anomaly score; generating a comprehensive compliance score based on the final anomaly score, compliance score, and historical operating data using a preset weighted scoring formula; defining display rules based on the comprehensive compliance score and the output results of preceding modules, and conducting risk assessment through the rule engine; prioritizing the risk assessment result when a rule is triggered, otherwise using the comprehensive compliance score, and resolving conflicts between the rule and the Gaussian mixture model and deep energy model; setting risk thresholds based on historical operating data, triggering early warning levels according to the comprehensive compliance score, and conducting online learning through manual review of the results.

[0014] On the other hand, this invention provides an intelligent method for government affairs review based on adaptive vision and cross-modal reasoning. The intelligent method includes: performing self-correction enhancement processing on low-quality government affairs material images to generate standardized images; detecting visual and textual differences between multiple versions of government affairs materials based on the standardized images and the OCR text stream of the government affairs materials, and generating a structured difference report; generating a multimodal embedding representation in a unified vector space based on the standardized images and the original text of the government affairs materials using multimodal alignment technology; checking the consistency between image fields and text fields based on the multimodal embedding representation and the OCR text stream of the government affairs materials, and generating a compliance score; and further... The system dynamically adjusts image patch segmentation based on visual feature vectors and optimizes the perception capability of target areas. Using contrastive learning, it constructs a semantic mapping between risk features and regulatory provisions based on image feature vectors and regulatory vectors with multimodal embedding representations. Based on the matching results of image and text fields, compliance scores, semantic mapping results, and multimodal embedding representations, it generates a traceable evidence graph structure using visualization tools. Based on multimodal embedding representations, compliance scores, and historical operational data, it conducts risk assessments and triggers warning levels through a rule engine. Finally, based on the evidence graph structure, risk assessment results, and regulatory provision retrieval results, it generates the final review conclusion and reasoning chain using a chain-reasoning method.

[0015] The above technical solutions enable automatic cross-modal consistency review with high accuracy, including key fields such as address and business scope, reducing manual verification; low-quality images can also be automatically enhanced to improve OCR and certificate recognition quality by 20-40%; automatic identification of violation scenarios and matching with legal provisions significantly reduces the rate of missed reviews and improves the standardization of law enforcement; it can generate a full-chain review evidence chain to support regulatory record keeping, review, and accountability scenarios; automatic early warning of high-risk applications improves approval efficiency and risk screening capabilities; and accurate identification of tampering behaviors such as "changing images and words" by enterprises strengthens compliance supervision.

[0016] Other features and advantages of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating the intelligent government affairs review system based on adaptive vision and cross-modal reasoning, as described in this invention. Figure 1 ; Figure 2 This is a schematic diagram of the process for generating standardized images in this invention; Figure 3This is a schematic diagram of the process for generating a structured difference report in this invention; Figure 4 This is a schematic diagram of the process for generating a multimodal embedding representation of a unified vector space in this invention; Figure 5 This is a schematic diagram of the process for generating compliance scores in this invention; Figure 6 This is a schematic diagram of the process of dynamically adjusting image patch division in this invention; Figure 7 This is a schematic diagram of the process for constructing a semantic mapping between risk features and regulatory provisions in this invention; Figure 8 This is a schematic diagram of the process for generating a traceable evidence graph structure in this invention; Figure 9 This is a flowchart illustrating the process of risk assessment and triggering early warning levels in this invention; Figure 10 This is a flowchart illustrating the process of generating the final review conclusion and chain of reasons in this invention; Figure 11 This is a flowchart illustrating the intelligent method for government affairs review based on adaptive vision and cross-modal reasoning of the present invention. Figure 12 This is a flowchart illustrating the intelligent government affairs review system based on adaptive vision and cross-modal reasoning, as described in this invention. Figure 2 . Detailed Implementation

[0018] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the scope of the present invention.

[0019] It should be noted that the acquisition, transmission, storage, use, and processing of data in the technical solution of this application all comply with the relevant provisions of national laws and regulations. In the embodiments of this application, certain existing industry solutions such as software, components, and models may be mentioned. These should be considered exemplary, intended only to illustrate the feasibility of implementing the technical solution of this application, and do not imply that the applicant has already used or necessarily used such solutions.

[0020] Please refer to Figure 1 and Figure 12This invention provides an intelligent system for government affairs review based on adaptive vision and cross-modal reasoning. The system includes: a preprocessing module for performing self-correction enhancement processing on low-quality government affairs materials images to generate standardized images; a difference detection module for detecting visual and textual differences between multiple versions of government affairs materials based on the standardized images and OCR text streams of the government affairs materials, and generating a structured difference report; a visual text encoding module for generating a multimodal embedding representation in a unified vector space based on the standardized images and the original text of the government affairs materials using multimodal alignment technology; a cross-modal review module for checking the consistency between image fields and text fields based on the multimodal embedding representation and the OCR text stream of the government affairs materials, and generating a compliance score; and a dynamic token module for using standardized images... The system dynamically adjusts image patch segmentation and optimizes the perception capability of target regions based on visual feature vectors and multimodal embedding representations. A bidirectional alignment module can be used to construct a semantic mapping between risk features and regulatory provisions using contrastive learning methods, based on image feature vectors and regulatory vectors derived from multimodal embedding representations. An evidence chain generation module can generate a traceable evidence graph structure using visualization tools, based on matching results between image and text fields, compliance scores, semantic mapping results, and multimodal embedding representations. A risk warning module can conduct risk assessments and trigger warning levels based on multimodal embedding representations, compliance scores, and historical operational data through a rule engine. A decision-making module can generate the final review conclusion and reasoning chain based on the evidence graph structure, risk assessment results, and regulatory provision retrieval results using a chain-reasoning method.

[0021] Please refer to Figure 2 In this embodiment of the invention, the preprocessing module can be used to perform self-correction and enhancement processing on low-quality government affairs materials images from government affairs review to generate standardized images, and may include: Step S100: Perform multi-scale deblurring on low-quality government affairs material images, dynamically adjust the intensity based on EXIF ​​metadata, and generate a deblurred image.

[0022] In a preferred embodiment of the present invention, based on the original RGB image, a Transformer-based encoder-decoder architecture is adopted. The encoder uses a multi-scale attention mechanism to extract global features (e.g., overall image structure) at low-resolution layers and focus on local details (e.g., small fonts) at high-resolution layers, providing multi-scale feature maps for deblurring. The decoder reconstructs a clear image based on the encoder output through attention weights. The loss function combines pixel-level L1 loss and perceptual loss (e.g., based on the intermediate layer feature distance of a pre-trained VGG network to ensure detail preservation). The training data includes synthetic blur (e.g., motion blur and compression artifacts) and relies on multi-scale features to ensure that the deblurring effect does not impair high-frequency information. During inference, the intensity is dynamically adjusted based on the deblurring result and EXIF ​​metadata (e.g., shutter speed) (e.g., lightweight processing is used at high shutter speeds, otherwise depth deblurring is triggered) to ensure adaptability.

[0023] Step S110: Based on the deblurred image, identify the text region and perform super-resolution enhancement using a text detector, and trigger optimization through OCR confidence.

[0024] In a preferred embodiment of the present invention, a lightweight text detector (e.g., a variant of CTPN or EAST) is used to detect text or stamp regions based on the deblurred output image, generating bounding boxes. The lightweight text detector locates key regions through a sliding window and attention mechanism, directly relying on the deblurring result to improve detection accuracy. For the detected bounding boxes, a patch-level ViT super-resolution network is applied to first segment the bounding boxes into small patches (e.g., 4x4 pixels), learn local texture through a Transformer layer, and then upsample to restore details. The loss function uses perceptual loss and adversarial loss to ensure realism. A pipeline is designed to first perform lightweight enhancement, and then trigger high-precision enhancement based on OCR confidence (e.g., below 90%), optimizing the bounding box quality through a feedback loop.

[0025] Step S120: Perform geometric correction on the enhanced image by using a corner regression network and affine transformation to correct image tilt and distortion.

[0026] In a preferred embodiment of the present invention, a small CNN regression network (e.g., based on a ResNet backbone) is trained based on the enhanced image to predict the coordinates of the four corner points. The loss function uses a smoothed L1 loss to ensure corner accuracy, relying on the output of the bounding box enhancement because the enhanced image is easier to detect corner points. The homography matrix is ​​optimized using the RANSAC algorithm, and affine transformation correction is performed. After correction, the image is rotated to the standard orientation based on a text orientation detector (e.g., Hough transform). The robustness of the geometric correction is ensured based on the corner point output. If corner detection fails (e.g., in an image without corner points), a fallback to a feature matching scheme (e.g., SIFT features) is provided to provide fault tolerance.

[0027] Step S130: Perform exposure compensation on the geometrically corrected image, use the Retinex model to equalize the brightness, and record the processing parameters.

[0028] In a preferred embodiment of the present invention, based on the geometrically corrected image, the Retinex theory is applied to decompose the image into illumination and reflection components. Gaussian filtering is used to estimate the illumination map for brightness equalization, and the exposure processing is directly dependent on the correction output to ensure that it is performed on the correct geometry. A U-Net style network is trained, with the Retinex decomposition result as input and the exposure correction map as output. The loss function combines exposure loss and detail preservation loss to further refine the brightness. All processing steps record the original image and processing parameters, supporting audit rollback.

[0029] Please refer to Figure 3 In this embodiment of the invention, the difference detection module can be used to detect visual and textual differences between multiple versions of government documents based on standardized images and OCR text streams of government documents, and generate a structured difference report, which may include: Step S200: Based on the standardized image, extract multi-feature visual fingerprints and calculate similarity to generate visual difference indicators.

[0030] In a preferred embodiment of the present invention, based on the corrected image output by the preprocessing module, ViT's CLSToken is used as a semantic fingerprint, and perceptual hash and local features (e.g., SIFT descriptors) are calculated simultaneously. This directly depends on the image quality of the preprocessing module, as a clear image can improve feature stability. Cosine similarity and geographic consistency checks (e.g., bounding box position changes) are calculated based on the fingerprint. Local patch comparison uses a sliding window to calculate the SSIM index to ensure the accuracy of difference detection. After the fingerprint is vectorized, it is stored in the database to support fast version comparison and provide a foundation for subsequent text difference detection.

[0031] Step S210: Based on the OCR text stream of government documents, apply a character-level diff algorithm to compare the text versions of the government documents and perform a strict comparison of the target fields.

[0032] In a preferred embodiment of the present invention, based on the OCR text stream of the preprocessing module, a character-level diff algorithm (e.g., Myers algorithm) is used to compare different versions of text, and to record insertion, deletion, and replacement operations. The OCR quality affects the accuracy of the text. For fields such as registration number and legal representative, regular expressions and edit distance thresholds are applied for strict comparison, and key parts are specially processed based on the diff results. The differences are classified into "addition / deletion / replacement" and a highlighted view is generated to ensure a structured report.

[0033] Step S220: Using an attention mechanism, correlate the visual and textual differences to construct an evidence map and perform time series analysis.

[0034] In a preferred embodiment of the present invention, based on visual and textual differences, image regions are associated with text fields through attention weights to construct an evidence graph and achieve cross-modal integration; version timestamps are recorded, difference evolution patterns are analyzed, and a time-series dimension is added.

[0035] In a preferred embodiment of the present invention, the difference score can be expressed by the following formula:

[0036] in, , , The weights are derived from historical data and used to generate a comprehensive score.

[0037] Step S230: Based on the score of visual and textual differences, generate a structured difference report and trigger suggested actions.

[0038] In a preferred embodiment of the present invention, a PDF report is automatically generated based on the difference score and evidence, including heatmaps and text highlighting, which directly depends on the cross-modal diff output; actions are triggered according to the scoring threshold (e.g., low scores are automatically passed), and automation is achieved based on the scoring results; metadata is embedded in the report to ensure audit compliance and traceability.

[0039] Please refer to Figure 4 In this embodiment of the invention, the visual text encoding module can be used to generate a multimodal embedding representation in a unified vector space based on the original text of standardized images and government materials, using multimodal alignment technology. This may include: Step S300: Based on the standardized image, use VisionTransformer to perform patch partitioning and multi-scale feature extraction to generate visual vectors.

[0040] In a preferred embodiment of the present invention, based on the image output by the preprocessing module, the image is segmented into variable-size patches (e.g., 16x16 pixels) using VisionTransformer (ViT). Patch Tokens and global CLS Tokens are generated through positional encoding and self-attention to provide a visual representation for subsequent alignment. Pyramid ViT is introduced to extract features at multiple scales, enhance small object detection, and improve representation quality. Relative positional encoding is added to improve sensitivity to the layout of government materials and refine spatial information.

[0041] Step S310: Based on visual vectors and the original text of government documents, semantic alignment is performed through contrastive learning methods, and cross-modal retrieval is supported.

[0042] In a preferred embodiment of the present invention, the CLIP model is fine-tuned using government data, the loss function is InfoNCE loss, and the input includes visual vectors and text embeddings to achieve preliminary semantic mapping; it supports coarse-grained image-to-text retrieval, returns Top-K relevant fragments, and is based on CLIP alignment results; ViT's patch tokens are combined with the CLIP model's global vector to achieve fine-grained alignment and enhance accuracy.

[0043] Step S320: Encode the regulatory text of government documents into vector segments, inject regulatory structure metadata, and build a vector index.

[0044] In a preferred embodiment of the present invention, based on the original regulatory text, LLM (e.g., BERT) is used to segment and encode it into vectors, with each segment having ≤512 tokens, providing a foundation for subsequent RAGs; the vectors are stored in a vector database (e.g., Milvus) to support similarity retrieval and are combined with visual encoding output; the regulatory structure is injected into the vectors as additional features to improve retrieval accuracy.

[0045] Step S330: Using the Hungarian matching algorithm, calculate the optimal pairing of image patch and text span to generate a multimodal embedding representation in a unified vector space.

[0046] In a preferred embodiment of the present invention, when calculating the alignment loss, the Hungarian matching algorithm can be used to calculate the optimal pairing of image patch and text span, and the loss function is the contrastive loss with margin to ensure semantic space uniformity; the bounding box coordinates of image token and the character span of text token can be bound to generate an attention heatmap to provide interpretability.

[0047] Please refer to Figure 5 In this embodiment of the invention, the cross-modal review module can be used to check the consistency between image fields and text fields based on multimodal embedding representation and OCR text stream of government documents, and generate a compliance score, which may include: Step S400: Based on the standardized image, extract the text content and layout information using the LayoutViT model.

[0048] In a preferred embodiment of the present invention, when deploying layout-aware OCR, the LayoutViT model can be used to simultaneously output text content and layout information (e.g., bounding boxes) based on the output image of the preprocessing module, directly relying on image quality; for tabular materials, a graph neural network is used to parse key-value pairs and extract structured fields; the OCR confidence can be calibrated using Bayesian methods to reduce false alarms and improve reliability.

[0049] Step S410: Based on the OCR text stream, a preset government entity recognition model is used to identify the target field.

[0050] In a preferred embodiment of the present invention, a BiLSTM-CRF model can be used to identify entities (e.g., company names) on government corpora, with the input being an OCR text stream; a span-based method can be introduced to improve boundary accuracy and refine entity extraction; identified entities can be linked to a knowledge base to verify authenticity and add external verification.

[0051] Step S420: Calculate the similarity between the image field and the text field using a two-layer alignment strategy, and then align the image field and the text field.

[0052] In a preferred embodiment of the present invention, when implementing a two-layer alignment strategy, global alignment can use CLIP to calculate the overall similarity, while field-level alignment can use Token-level similarity calculation; the optimal pairing is determined by Hungarian attention, and the weight is based on field importance to optimize the matching; when alignment conflicts occur, manual review or a rule engine is triggered to ensure robustness.

[0053] Step S430: Based on the alignment results, calculate the compliance score using a preset scoring formula and generate an interpretability result.

[0054] In a preferred embodiment of the present invention, the compliance score can be represented by the following formula:

[0055] in, These represent the field similarity weight, global semantic consistency weight, and rule confidence weight, respectively, which can be adjusted based on the domain or learned through a small dataset.

[0056] In a preferred embodiment of the present invention, when expanding the loss function, a regulatory relevance loss can be added during training to increase the penalty for samples with inconsistent key fields and enhance supervision; each field returns matching evidence and confidence intervals, providing readable results.

[0057] Please refer to Figure 6 In this embodiment of the invention, the dynamic token module can be used to dynamically adjust the image patch segmentation and optimize the perception capability of the target region based on the visual feature vectors of the standardized image and multimodal embedding representation. This can include: Step S500: Based on the standardized image, use a weighted fusion formula to calculate the importance score, and dynamically adjust the preset threshold according to the quality of the standardized image.

[0058] In a preferred embodiment of the present invention, based on the enhanced image output by the preprocessing module, multiple signal sources are acquired, such as text detection scores, seal detection scores, attention map estimation, and gradient sensitivity, which directly depend on image quality; a lightweight CNN network is used to calculate the importance map in real time, and the threshold is dynamically adjusted according to the overall image quality (e.g., the threshold is reduced for low-quality images) to ensure adaptability.

[0059] In a preferred embodiment of the present invention, the importance score can be expressed by the following formula:

[0060] in, , , , These represent weights, and an importance graph is generated through grid search optimization.

[0061] Step S510: Based on the importance score, dynamically divide the patch size using a preset partitioning algorithm, and perform multi-scale feature fusion through a pyramid patching mechanism.

[0062] In a preferred embodiment of the present invention, during dynamic patch size allocation, small patches (e.g., 8x8 pixels) can be allocated to high-importance regions (e.g., scores > 0.8) and large patches (e.g., 32x32 pixels) can be allocated to low-importance regions (e.g., scores < 0.3), using a sliding window algorithm combined with the patch embedding layer of ViT. During multi-scale feature fusion, a pyramid patching mechanism can be introduced to divide the image into multi-scale patches, and features of different scales can be fused through the multi-head attention layer of Transformer to enhance representation capabilities. A FLOPs budget (e.g., 1G FLOPs) can be set to dynamically adjust the patch size distribution. If the budget is exceeded, it will fall back to uniform patch division to balance accuracy and efficiency.

[0063] Step S520: Based on the low-importance regions in the patch partitioning results, perform token merging to optimize computational costs.

[0064] In a preferred embodiment of the present invention, adjacent patches in low-importance regions can be merged based on the patch partitioning results. For example, Hard merging directly merges them into a single token (saving computation), while Soft merging retains the original token information through attention reweighting. The selection strategy depends on the device capabilities. For Soft merging, a sigmoid gating function can be used to calculate the weight of each token, based on the importance score, to achieve dynamic reweighting. The merging threshold can be selected experimentally (e.g., merging when the importance score is <0.2) to ensure that the accuracy loss is <5% under the target latency, providing a configurable strategy.

[0065] Step S530: During the training phase of the partitioning algorithm, the importance-aware loss is calculated to optimize the perception capability of the target region and enhance the alignment effect of the target region.

[0066] In a preferred embodiment of the present invention, when calculating the importance-aware loss, during the training phase, the loss function can be combined with the standard cross-entropy loss and the importance-aware loss to force the model to focus on key regions; a regularization term can be added to make the alignment loss weight of visual tokens and text tokens in key regions higher, thereby enhancing cross-modal consistency; the key field recognition rate can be used as an evaluation metric to feed back into the importance evaluation, forming a closed-loop optimization to ensure continuous improvement of the module.

[0067] Please refer to Figure 7 In this embodiment of the invention, the bidirectional alignment module can be used to construct a semantic mapping between risk features and regulatory provisions based on image feature vectors and regulatory vectors using a contrastive learning method, and may include: Step S600: Based on historical cases of government violations, define a set of risk concepts, generate visual prototype embedding vectors, and construct a visual risk feature library.

[0068] In a preferred embodiment of the present invention, a set of risk concepts (e.g., "no fire extinguisher", "open flame operation") can be defined based on historical cases of government violations. Positive and negative sample images are collected for each concept. The sample sources include enhanced images output by the preprocessing module and public datasets. When generating prototypical embeddings, the CLIP or ViT model of the visual encoding module can be used to generate prototypical embeddings (class center vectors) for the sample images of each risk concept. The calculation method is the average value of the sample embeddings to ensure feature consistency. Online learning can be supported, and the feature library can be automatically updated when new violation cases appear to ensure the timeliness of the library.

[0069] Step S610: Encode the legal texts in the legal knowledge base into legal vectors and construct an efficient vector index.

[0070] In a preferred embodiment of the present invention, the legal knowledge base can be segmented into “article-item-explanation” sections, and each text segment can be generated into a vector representation using the LLM encoder of the legal text encoding module to ensure vector quality; the legal structure information (e.g., clause number, penalty level) can be encoded as an additional vector dimension and concatenated with the text vector to improve retrieval accuracy; the vectors can be stored using HNSW (Hierarchical Navigation Small World) graph index to support efficient similarity retrieval and prepare for subsequent alignment.

[0071] Step S620: Train the model using bidirectional contrastive loss to achieve semantic alignment between visual features and legal provisions.

[0072] In a preferred embodiment of the present invention, manually labeled pairs (e.g., image risk features, regulatory fragments) can be used as training data, with positive samples being relevant pairs and negative samples being random combinations. The input includes visual features and regulatory vectors. When calculating the bidirectional contrast loss, the loss function adopts InfoNCE loss, and the retrieval in both the image → regulation and regulation → image directions is optimized. By calculating cosine similarity and applying softmax, the distance between positive samples is shortened and the distance between negative samples is pushed away, ensuring semantic space alignment. Difficult negative sample mining (e.g., selecting negative samples with high similarity) can be introduced to improve the robustness of the model through iterative optimization.

[0073] Step S630: Based on the images of new government materials, detect risk concepts using the visual risk feature database, retrieve relevant legal provisions, and generate an interpretable chain of evidence.

[0074] In a preferred embodiment of the invention, for a new input image from the preprocessing module, risk concepts are first detected using a feature library, then Top-K relevant clauses are retrieved from the regulatory vector space using a k-NN algorithm based on cosine similarity; visual evidence (e.g., bounding boxes), cited clause paragraphs, and natural language reasons are output, with reason generation incorporating LLM, the input being the retrieved clauses and visual features, providing interpretability; Platt scaling is used to calibrate the matching confidence, and an alert is triggered based on a threshold (e.g., confidence > 0.7) to ensure reliability.

[0075] Please refer to Figure 8 In this embodiment of the invention, the evidence chain generation module can be used to generate a traceable evidence graph structure using visualization tools based on the matching results of image and text fields, compliance scores, semantic mapping results, and multimodal embedding representations. This structure may include: Step S700: Based on the output of the preceding module, define the node type and instantiation, and establish a node relationship index and node update mechanism.

[0076] In a preferred embodiment of the present invention, five types of nodes are defined, such as field nodes (e.g., "registration address"), text nodes (e.g., field value), image nodes (e.g., ROI or patch), regulatory nodes (e.g., clause fragments), and event nodes (e.g., "high-risk warning"). Each node contains an ID, timestamp, confidence level, and source file attributes. Nodes can be stored using a graph database (e.g., Neo4j) and indexed to support fast queries and ensure scalability. A node update mechanism can be used to dynamically add or update nodes when new evidence is generated (e.g., multi-version difference detection) to keep the graph up-to-date.

[0077] Step S710: Define the edge type according to the node type, calculate the edge weight based on the output of the preceding module, and construct the logical relationship of the evidence chain.

[0078] In a preferred embodiment of the present invention, edge types are defined, such as support / refutation (e.g., representing evidence logic), located (e.g., anchoring a text span to an image ROI), mapped to (e.g., the association between visual risk features and regulatory clauses), and derived (e.g., evidence deduction conclusions). Each edge carries a weight value (e.g., 0-1). When calculating edge weights, the weights are based on confidence levels. For example, the matching confidence level of the cross-modal review module is directly used as the edge weight to ensure the accuracy of the logical relationship. Edge weights can be adjusted according to new evidence, such as updating relevant edges when the conclusion is modified by manual review, to maintain real-time performance.

[0079] Step S720: Based on the logical relationship of the evidence chain, construct a complete evidence graph structure through evidence anchoring and logical node integration.

[0080] In a preferred embodiment of the present invention, attention weights of the visual text encoding module can be used to anchor text entities to the image ROI (e.g., calculate IoU and attention values ​​to determine anchoring strength) to ensure accurate positioning; the field comparison results (e.g., consistent / inconsistent) of the cross-modal review module are used as binary logical nodes and connected to relevant evidence nodes to construct a basic logical chain; during full graph construction and optimization, all nodes and edges can be integrated, and graph optimization algorithms (e.g., community detection) can be applied to simplify the structure and form a complete evidence chain.

[0081] Step S730: Generate interactive views and reports using structured queries and visualization tools.

[0082] In a preferred embodiment of the present invention, structured queries (e.g., Cypher query language) can be supported, allowing the retrieval of evidence chains by legal provisions, fields, or confidence levels, providing flexible access; it can be implemented through visualization tools to generate interactive views, such as heatmaps displaying evidence density, timelines showing multiple version evolutions, rendered using D3.js or similar libraries; it can be exported as machine-readable JSON and human-readable PDF, and automatically archived into government systems to ensure audit compliance.

[0083] Please refer to Figure 9 In this embodiment of the invention, the risk warning module can be used to perform risk assessment and trigger warning levels based on multimodal embedded representation, compliance scoring, and historical operational data through a rule engine, and may include: Step S800: Based on multimodal embedding representation, calculate the anomaly scores of visual and textual modalities through Gaussian mixture model and deep energy model, and fuse them into the final anomaly score.

[0084] In a preferred embodiment of the present invention, the density distribution can be estimated using a Gaussian mixture model (GMM) based on the embedding output by the visual text encoding module. For example, samples in low-density regions are marked as anomalies, which directly depends on the vector quality of the encoding module. A deep energy model is trained to calculate the energy score of each sample (for example, samples with high energy scores are considered anomalies). The model input is the concatenation of multimodal embeddings, providing a supplementary detection mechanism. The anomaly scores of visual and text modalities can be combined, and the maximum value can be taken as the final anomaly score to improve detection robustness.

[0085] Step S810: Based on the final anomaly score, compliance score, and historical operating data, generate a comprehensive compliance score using a preset weighted scoring formula.

[0086] In a preferred embodiment of the present invention, when extracting scoring factors, field consistency scores can be obtained from the cross-modal review module, visual risk confidence scores can be obtained from the violation alignment module, behavioral scores (e.g., modification frequency) and text similarity (e.g., matching with fraud templates) can be obtained from the historical database; the contribution of each factor can be decomposed using SHAP values ​​to output a readable report, which is convenient for manual review and audit traceability.

[0087] In a preferred embodiment of the present invention, the comprehensive compliance score can be expressed by the following formula:

[0088] in, , , , These represent weights, which are generated from 0 to 100 points by learning from historical data to ensure dynamic adaptability.

[0089] Step S820: Based on the comprehensive compliance score and the output results of the preceding modules, define the display rules and conduct a risk assessment through the rule engine.

[0090] In a preferred embodiment of the present invention, explicit rules can be defined (e.g., "business scope does not match registered industry"). When the rule is triggered, the risk level is directly increased. If the rule engine is triggered, the risk level is immediately upgraded (e.g., from medium risk to high risk). Otherwise, the model output is relied upon to ensure decision priority. Through a conflict resolution mechanism, when the model and the rule conflict, the rule result is adopted first and the log is recorded to balance automation and reliability.

[0091] Step S830: When a rule is triggered, the result of the risk assessment shall be used first; otherwise, the comprehensive compliance score shall be used, and the conflict between the rule and the Gaussian mixture model and the deep energy model shall be resolved.

[0092] In a preferred embodiment of the present invention, the results of the rule engine cover the model output, but retain the original score as a reference. In case of conflict, the confidence level is compared (e.g., if the rule confidence level is >0.9, the rule takes precedence). All conflict cases record detailed context (e.g., including model version and input data hash) to support post-analysis and model optimization. The weights of rules and models can be dynamically adjusted according to the conflict frequency (e.g., if the conflict rate is high, the rule weight is reduced) to achieve long-term balance.

[0093] Step S840: Set risk thresholds based on historical operational data, trigger warning levels based on comprehensive compliance scores, and conduct online learning based on manual review results.

[0094] In a preferred embodiment of the present invention, when setting the threshold optimization, historical manual judgment data can be used to select the risk threshold through ROC curve to ensure that the false positive rate is <5% when the high-risk recall rate is >95%; through an online learning mechanism, the manual review results are fed back as new labels to incrementally update the anomaly detection model and scoring weights to achieve continuous optimization; during operation monitoring and auditing, the model performance (e.g., concept drift detection) can be evaluated regularly to avoid system degradation.

[0095] Please refer to Figure 10 In this embodiment of the invention, the decision-making module can be used to generate a final review conclusion and reasoning chain based on the evidence graph structure, risk assessment results, and retrieval results of legal provisions, using a chain reasoning method. This may include: Step S900: Based on the output of the preceding module, the inference chain is generated in a structured manner through a preset process, and natural language text is generated using an LLM model.

[0096] In a preferred embodiment of the present invention, the agent generates a reasoning chain according to a fixed process, such as OCR text extraction → field consistency review → visual risk identification → regulatory retrieval → final recommendation. Each step references a node of the evidence chain module (e.g., field matching result) to ensure logical coherence. When generating natural language text, LLM (e.g., GPT-style model) can be used to generate readable reasoning text, with prompts designed in RAG style and contextualized. ViT's attention map can be used as a visual basis to inject prompts, enhancing the accuracy of the generated text, relying on the attention output of the visual encoding module.

[0097] Step S910: Retrieve relevant clause fragments from the legal knowledge base, connect the relevant clause fragments with the evidence chain in context, and add generation constraints.

[0098] In a preferred embodiment of the present invention, relevant clauses can be retrieved from a regulatory knowledge base, and Top-K clauses can be retrieved based on cosine similarity using a vector database (e.g., Milvus). The retrieved clause fragments are concatenated with the evidence chain to form an input prompt for the LLM, and generation constraints (e.g., clause numbers must be cited) are added to ensure that the output is compliant. Common query results are cached to reduce latency and improve real-time performance.

[0099] Step S920: Optimize the quality of government review recommendations through reward models and reinforcement learning strategies, and update the reward model and reinforcement learning strategy network based on the results of manual review to achieve a feedback loop.

[0100] In a preferred embodiment of the present invention, during the training of the reward model, the reward model can be trained based on the review suggestion data with expert scores, and the output suggestion quality score (e.g., 0-1) can be generated. The training data includes chain-of-thought output and human scores. During the optimization of the reinforcement learning policy, the PPO algorithm can be used to optimize the policy network to maximize the reward score. The input of the policy network is a multimodal embedding, and the output is the review suggestion. The quality is improved through iteration. The results of human review are fed back to update the reward model and the policy network, forming online learning and ensuring that the system continues to evolve.

[0101] Step S930: Utilize a threshold trigger mechanism to trigger manual review, record all decisions in a complete log, and provide an interface for integration.

[0102] In a preferred embodiment of the present invention, a threshold triggering mechanism is used to set a confidence threshold (e.g., <85%) and a risk threshold (e.g., >70%). When uncertainty arises, manual review is automatically triggered. The thresholds are based on the output of the risk warning module to ensure prudent decision-making. All decisions are recorded in a complete log, including model version, input data hash, and output conclusions, supporting end-to-end accountability. A RESTful API can be provided to support asynchronous queries, and a reviewer UI interface (e.g., displaying the reasoning chain and evidence chain) can be built. Reports are automatically archived to the government system to achieve an end-to-end closed loop.

[0103] Please refer to Figure 11 The present invention also provides an intelligent method for government affairs review based on adaptive vision and cross-modal reasoning, which may include: Step S1: Based on the low-quality government documents images from the government review, perform self-correction enhancement processing to generate standardized images.

[0104] Step S2: Based on standardized images and OCR text streams of government documents, detect visual and textual differences between multiple versions of government documents and generate a structured difference report.

[0105] Step S3: Based on the original text of standardized images and government materials, multimodal alignment technology is used to generate a multimodal embedding representation in a unified vector space.

[0106] Step S4: Based on the multimodal embedding representation and the OCR text stream of government documents, check the consistency between image fields and text fields, and generate a compliance score.

[0107] Step S5: Dynamically adjust the image patch segmentation and optimize the perception capability of the target region based on the visual feature vectors of the standardized image and multimodal embedding representation.

[0108] Step S6: Based on the image feature vector and regulatory vector of multimodal embedding representation, construct the semantic mapping between risk features and regulatory provisions through contrastive learning methods.

[0109] Step S7: Based on the matching results of the image field and the text field, the compliance score, the semantic mapping results, and the multimodal embedding representation, use visualization tools to generate a traceable evidence graph structure.

[0110] Step S8: Based on multimodal embedding representation, compliance score and historical operation data, conduct risk assessment and trigger warning levels through the rule engine.

[0111] Step S9: Based on the evidence diagram structure, risk assessment results, and the search results of legal provisions, generate the final review conclusion and reasoning chain through chain reasoning.

[0112] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0113] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0114] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0115] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0116] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0117] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0118] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0119] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0120] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A government affairs review intelligent system based on adaptive vision and cross-modal reasoning, characterized in that, The intelligent system for government review includes: The preprocessing module is used to perform self-correction and enhancement processing on low-quality government documents images from government review, and generate standardized images. The difference detection module is used to detect visual and textual differences between multiple versions of government documents based on standardized images and OCR text streams of government documents, and generate a structured difference report. The visual text encoding module is used to generate a multimodal embedding representation in a unified vector space based on the original text of standardized images and government materials using multimodal alignment technology. The cross-modal review module is used to check the consistency between image fields and text fields based on multimodal embedded representations and OCR text streams of government documents, and generate a compliance score. The Dynamic Token module is used to dynamically adjust the image patch segmentation and optimize the perception capability of the target region based on the visual feature vectors of the standardized image and multimodal embedding representation. The bidirectional alignment module is used to construct a semantic mapping between risk features and regulatory provisions based on image feature vectors and regulatory vectors using a contrastive learning method. The evidence chain generation module is used to generate a traceable evidence graph structure based on the matching results of image and text fields, compliance scores, semantic mapping results, and multimodal embedding representations, using visualization tools. The risk warning module is used to conduct risk assessment and trigger warning levels based on multimodal embedded representation, compliance score and historical operation data through the rule engine; The decision-making module is used to generate the final review conclusion and reasoning chain based on the evidence graph structure, risk assessment results, and the search results of legal provisions, using a chain reasoning method.

2. The intelligent system for government affairs review according to claim 1, characterized in that, The process of performing self-correction and enhancement processing on low-quality government documents from government review to generate standardized images includes: Low-quality government documents images are deblurred at multiple scales, and the intensity is dynamically adjusted by combining EXIF ​​metadata to generate deblurred images; Based on the deblurred image, a text detector is used to identify text regions and perform super-resolution enhancement, and optimization is triggered by OCR confidence. The enhanced image is geometrically corrected by using a corner regression network and affine transformation to correct image tilt and distortion. Exposure compensation was performed on the geometrically corrected image, and the brightness was balanced using the Retinex model. The processing parameters were then recorded.

3. The intelligent system for government affairs review according to claim 2, characterized in that, The method for detecting visual and textual differences between multiple versions of government documents and generating a structured difference report includes: Based on the standardized image, multi-feature visual fingerprints are extracted and similarity is calculated to generate visual difference indicators; Based on the OCR text stream of government documents, a character-level diff algorithm is applied to compare the text versions of government documents and to strictly compare the target fields. By using an attention mechanism, visual and textual differences are correlated to construct an evidence graph and perform time-series analysis. Based on the scores for visual and textual differences, a structured difference report is generated, and suggested actions are triggered.

4. The intelligent system for government affairs review according to claim 1, characterized in that, The method of generating a multimodal embedding representation in a unified vector space using multimodal alignment techniques includes: Based on standardized images, VisionTransformer is used to perform patch partitioning and multi-scale feature extraction to generate visual vectors; Based on visual vectors and the original text of government documents, semantic alignment is performed through contrastive learning methods, and cross-modal retrieval is supported. The legal texts in government documents are segmented and encoded into vectors, and then injected with legal structure metadata to build a vector index; Using the Hungarian matching algorithm, the optimal pairing of image patches and text spans is calculated to generate a multimodal embedding representation in a unified vector space.

5. The intelligent system for government affairs review according to claim 1, characterized in that, The process of checking the consistency between the image and text fields and generating a compliance score includes: Based on standardized images, the LayoutViT model is used to extract text content and layout information; Based on OCR text stream, a preset government entity recognition model is used to identify target fields; By employing a two-layer alignment strategy, the similarity between the image field and the text field is calculated, and the image field and text field are aligned accordingly. Based on the alignment results, a compliance score is calculated using a pre-defined scoring formula, and an interpretability result is generated.

6. The intelligent system for government affairs review according to claim 1, characterized in that, The dynamic adjustment of image patch segmentation and optimization of target region perception capabilities include: Based on standardized images, a weighted fusion formula is used to calculate importance scores, and a preset threshold is dynamically adjusted according to the quality of the standardized images. Based on the importance score, the patch size is dynamically divided using a preset partitioning algorithm, and multi-scale feature fusion is performed through a pyramid patching mechanism. Token merging is performed based on low-importance regions in the patch partitioning results to optimize computational costs; During the training phase of the partitioning algorithm, importance-aware loss is calculated to optimize the perception capability of the target region and enhance the alignment effect of the target region.

7. The intelligent system for government affairs review according to claim 1, characterized in that, The method of constructing a semantic mapping between risk characteristics and regulatory provisions through contrastive learning includes: Based on historical cases of government violations, a set of risk concepts is defined, and visual prototype embedding vectors are generated to construct a visual risk feature library. The legal texts in the legal knowledge base are segmented and encoded into legal vectors, and an efficient vector index is constructed. By training the model using bidirectional contrastive loss, semantic alignment between visual features and legal provisions is achieved. Based on images of new government documents, risk concepts are detected using a visual risk feature database, relevant legal provisions are retrieved, and an interpretable chain of evidence is generated.

8. The intelligent system for government affairs review according to claim 7, characterized in that, The process of generating a traceable evidence graph structure using visualization tools includes: Based on the output of the preceding module, define the node type and instantiation, and establish a node relationship index and node update mechanism; Define edge types based on node types, calculate edge weights based on the output of the preceding modules, and construct the logical relationship of the evidence chain. Based on the logical relationship of the evidence chain, a complete evidence graph structure is constructed through evidence anchoring and logical node integration; Generate interactive views and reports using structured queries and visualization tools.

9. The intelligent system for government affairs review according to claim 8, characterized in that, The process of using a rules engine to perform risk assessment and trigger alert levels includes: Based on multimodal embedding representation, anomaly scores of visual and textual modalities are calculated using Gaussian mixture models and deep energy models, and then fused into a final anomaly score. Based on the final anomaly score, compliance score, and historical operational data, a comprehensive compliance score is generated using a pre-defined weighted scoring formula. Based on the comprehensive compliance score and the output results of the preceding modules, display rules are defined, and risk assessment is conducted through the rule engine; When a rule is triggered, the result of the risk assessment is used first; otherwise, the comprehensive compliance score is used, and conflicts between the rule and the Gaussian mixture model and the deep energy model are resolved. Risk thresholds are set based on historical operational data, warning levels are triggered based on comprehensive compliance scores, and online learning is conducted based on the results of manual review.

10. A government affairs review intelligent method based on adaptive vision and cross-modal reasoning, characterized in that, The intelligent methods for government review include: Based on the low-quality government documents images from government review, self-correction and enhancement processing is performed to generate standardized images; Based on standardized images and OCR text streams of government documents, visual and textual differences between multiple versions of government documents are detected, and a structured difference report is generated. Based on standardized images and original texts from government documents, multimodal alignment techniques are used to generate multimodal embedding representations in a unified vector space. Based on multimodal embedding representation and OCR text flow of government documents, check the consistency between image fields and text fields to generate a compliance score; Based on the visual feature vectors of standardized images and multimodal embedding representations, the image patch segmentation is dynamically adjusted, and the perception capability of the target region is optimized. Based on image feature vectors and regulatory vectors using multimodal embedding representations, a semantic mapping between risk features and regulatory provisions is constructed through a contrastive learning method. Based on the matching results of image and text fields, compliance scores, semantic mapping results, and multimodal embedding representations, a traceable evidence graph structure is generated using visualization tools. Based on multimodal embedded representation, compliance scoring, and historical operational data, a rules engine is used to conduct risk assessments and trigger warning levels. Based on the evidence diagram structure, risk assessment results, and retrieval results of legal provisions, a chain reasoning method is used to generate the final review conclusion and chain of reasons.

Images