Long document content review method based on artificial intelligence

Abstract

The invention relates to a long document content review method based on artificial intelligence, and belongs to the field of artificial intelligence and information security. According to the method, for a long document or a super-long document, firstly, differential analysis is carried out according to document types, contents such as texts and pictures are extracted, and semantic analysis and text extraction are carried out on the pictures by utilizing a multi-modal large model; and for the text, performing slicing processing and extracting an external link, and realizing multi-level detection and abstract generation of external webpages, documents and pictures by combining a recursive analysis mechanism. Sensitive data detection, spelling and grammar verification and semantic consistency analysis are sequentially carried out in the review analysis module, context multi-modal correlation analysis including image-text correlation and external link correlation is executed through a large model, and a structured review result document is generated. According to the method, the automatic, high-efficiency and recursive content review of the complex long document can be realized, and the intelligent level of content compliance detection, risk identification and information security management and control is greatly improved.

Description

Technical Field

[0001] This invention relates to a method for reviewing the content of long documents based on artificial intelligence. Background Technology

[0002] In today's context of rapid information and digital development, document data in scenarios such as office work, scientific research, and compliance supervision is characterized by large volume, complex formats, and multimodal content. Long or extremely long documents, in particular, often contain multi-source information such as text, images, and external links, making manual review extremely inefficient and prone to oversights. Existing document detection methods generally suffer from the following shortcomings: limited processing capabilities for extremely long documents, unable to effectively segment and analyze them; lack of a unified parsing and review mechanism for multimodal information such as images and external links; traditional detection methods rely on rule matching, lacking sufficient intelligence to perform contextual semantic analysis and cross-modal information correlation analysis; and lack of automated risk alerts and recursive external resource analysis capabilities in risk control and compliance review. Therefore, there is an urgent need for a new document detection and review method that integrates artificial intelligence technology, multimodal parsing capabilities, and automated risk identification mechanisms to improve the efficiency and accuracy of document content compliance and security detection. Summary of the Invention

[0003] The purpose of this invention is to provide an artificial intelligence-based method for reviewing long document content. This method uses a multimodal large-scale model to uniformly parse document text, images, and external links, and employs slicing and recursive mechanisms to achieve multi-level detection of long documents and external resources. In the review and analysis phase, sensitive data detection, spelling and grammar verification, and semantic consistency analysis are performed sequentially, and the large-scale model is used to review the relevance of text, images, and external links. Finally, a structured review result document is output by the result integration module. This invention enables automated, multimodal, and intelligent review of document content, improving the efficiency and accuracy of compliance detection and risk identification.

[0004] To achieve the above objectives, the technical solution of the present invention is: a method for reviewing long document content based on artificial intelligence, comprising the following steps:

[0005] Step S1: Prepare the documents to be tested;

[0006] Step S2: Based on the document type, use differentiated data parsing methods to parse the document content, and input the parsing results into the text processing module and the image processing module;

[0007] Step S3: Utilize a multimodal large model to detect and parse the images I={I1,I2,…,Im} contained in the document content, outputting the text information T1 or visual features F1 contained in the images, and use LLM to generate a content summary S of the images.I The text information T1 is input into the text processing module, and the visual features F1 and content summary S are input into the text processing module. I Input to the review and analysis module; where:

[0008]

[0009] Step S4: Perform slicing processing on the data received by the text processing module, extract external links from the data, and transmit the external links to the external link processing module. Simultaneously, send the sliced ​​text data to the review and analysis module. The slicing processing method is represented as follows:

[0010]

[0011] Where text represents the complete text content, L is the standard slice length, ΔL is the overlap length, and P... start (i) and P end (i) represents the start and end positions of the i-th slice, and slice i represents the text content after slicing.

[0012] Step S5: In the external link processing module, perform accessibility checks on the external links: if they are inaccessible, generate a risk warning; if they are accessible, perform recursive parsing based on the type of external link. In the recursive process, introduce a depth parameter d and set a maximum depth d_max. Continue expanding when d ≤ d_max, and terminate the recursion and output the current result when d_max is exceeded.

[0013] Step S6: In the review and analysis module, analyze the sliced ​​text data, visual features F1, and content summary S. I Sensitive data detection and text spelling detection are performed sequentially, and context analysis, image-text correlation analysis, and external link correlation analysis are conducted in conjunction with the large model.

[0014] Step S7: Based on the risk warning in step S5 and the analysis results in step S6, the results integration module generates a structured review result document in the original document order.

[0015] Furthermore, in step S4, external links include URLs and document links.

[0016] Furthermore, in step S5, external link types are divided into four categories, each corresponding to a different processing method, as detailed below:

[0017] URL: Use web crawling tools to capture the page and generate a summary of the page content using LLM;

[0018] Document: Read the document and generate a summary from the LLM, while recursively returning step S3;

[0019] Image: Read the image and recursively return step S3;

[0020] Emails, scripts, and others: Directly generate risk warnings.

[0021] Further details regarding the review and analysis module:

[0022] Step S601: Vectorize the text using the Embedding model, compare its similarity with the sensitive word vector library, quickly identify potentially sensitive content, and mark high-risk segments.

[0023] Step S602: Use a pre-trained language model including BERT to perform spelling and grammar checks on the sliced ​​text data, and identify typos, grammatical errors and abnormal expressions;

[0024] Step S603: Use LLM to perform semantic understanding and analysis on the sliced ​​text data, determine the text theme, sentiment, potential risks and inappropriate expressions, and output structured review results;

[0025] Step S604: By analyzing the contextual relationship between the preceding and following segments of the document, determine the rationality, consistency and potential misleading information of the sliced ​​text data in the overall document, and identify logical conflicts or semantic anomalies.

[0026] Step S604: Use the Score model to evaluate visual features F1 and content summary S. I The text data after the current slice is scored to evaluate the consistency between the external content and the document semantics. When the score is below the threshold σ, the tag relevance is low.

[0027] Step S605: Use the Score model to score the summary content of external links and the text data after the current slice to evaluate the consistency between external content and document semantics. If it is lower than the threshold σ, the relevance is low.

[0028] Furthermore, the method can perform content slicing and progressive processing on ultra-long documents, and use a multimodal large model to perform semantic parsing and text extraction on images embedded in the document, thereby achieving cross-modal document content review.

[0029] Furthermore, the method employs a parallel processing mode during the external link parsing process to identify the type of external links and read their content. It also uses a recursive parsing mechanism to examine and generate summaries of the documents, web pages, and images contained in the links layer by layer, ensuring the consistency and integrity of the external resources with the main document.

[0030] Furthermore, the review and analysis module analyzes the sliced ​​text data, external links, visual features F1, and content summary S. IThe system performs unified and integrated processing, sequentially completing sensitive data detection, spelling and grammar verification, and semantic consistency analysis. Based on a large model, it performs contextual multimodal correlation analysis, and finally outputs a structured review result document that can be used for compliance review, risk control, and security testing.

[0031] This invention also provides an artificial intelligence-based system for reviewing the content of long documents, comprising:

[0032] The data input module prepares the documents to be tested.

[0033] The document parsing module performs structured parsing of documents, extracts text and image content, and outputs text and image data.

[0034] The image processing module utilizes a multimodal large model to detect and parse image data, outputting the textual information T1 or visual features F1 contained in the image data, and using LLM to generate a content summary S for the image. I The text information T1 is input into the text processing module, and the visual features F1 and content summary S are input into the text processing module. I Pass the data to the review and analysis module;

[0035] The text processing module extracts external links based on the text data and text information T1, and transmits the external links to the external link processing module. At the same time, it segments the text according to the paragraph and text length to generate structured text segment data, and sends the text segment data to the review and analysis module.

[0036] The external link processing module performs accessibility checks on external links: if they are inaccessible, a risk warning is generated; if they are accessible, the module performs recursive parsing based on the external link type and summarizes its content using LLM.

[0037] The review and analysis module analyzes the visual features F1 and content summary S of the text slice data. I The system sequentially performs sensitive data detection, large-scale model semantic analysis, contextual analysis, image-text correlation analysis, and external link correlation analysis to complete multi-level intelligent review.

[0038] The results integration module summarizes the review and analysis results and external link risk warning information in a unified manner, and generates a structured review results document in the original document order.

[0039] The present invention also provides an electronic device, including a memory, a processor, and computer program instructions stored in the memory and executable by the processor, wherein when the processor executes the computer program instructions, it can implement the steps of any of the methods described above.

[0040] The present invention also provides a computer-readable storage medium having stored thereon computer program instructions that can be executed by a processor, wherein when the processor executes the computer program instructions, it can implement the steps of any of the methods described above.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] (1) This invention uses artificial intelligence as its core technology and integrates multimodal parsing and recursive review mechanisms to automatically complete the comprehensive detection of long document text, images and external links, and realize intelligent analysis of compliance, risk and semantic consistency.

[0043] (2) This invention significantly improves the efficiency and accuracy of document review, reduces the workload and risk of omissions in manual review, and can effectively change the traditional content compliance review method based on "rules + manpower". It has the advantages of low cost, high reliability and high intelligence. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the method in an embodiment of the present invention.

[0045] Figure 2 This is a functional block diagram of an embodiment of the present invention.

[0046] Figure 3 This is a general flowchart of an embodiment of the present invention. Detailed Implementation

[0047] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.

[0048] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0049] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0050] This embodiment provides a method for reviewing the content of long documents based on artificial intelligence. The specific implementation method includes the following steps:

[0051] Step S1: Prepare the documents to be tested;

[0052] Step S2: Based on the document type, use differentiated data parsing methods to parse the document content, and input the parsing results into the text processing module and the image processing module;

[0053] Step S3: Utilize a multimodal large model to detect and parse the images I={I1,I2,…,Im} contained in the document content, outputting the text information T1 or visual features F1 contained in the images, and use LLM to generate a content summary S of the images. I The text information T1 is input into the text processing module, and the visual features F1 and content summary S are input into the text processing module. I Input to the review and analysis module; where:

[0054]

[0055] Step S4: Perform slicing processing on the data received by the text processing module, extract external links (including URLs and document links) from the data, and transmit the external links to the external link processing module. Simultaneously, send the sliced ​​text data to the review and analysis module. The slicing processing method is represented as follows:

[0056]

[0057] Where text represents the complete text content, L is the standard slice length, ΔL is the overlap length, and P... start (i) and P end (i) represents the start and end positions of the i-th slice, and slice i represents the text content after slicing.

[0058] Step S5: In the external link processing module, perform accessibility checks on the external links: if they are inaccessible, generate a risk warning; if they are accessible, perform recursive parsing based on the external link type.

[0059] External links are categorized into four types, each with a different processing method, as detailed below:

[0060] URL: Use web crawling tools to capture the page and generate a summary of the page content using LLM;

[0061] Document: Read the document and generate a summary from the LLM, while recursively returning step S3;

[0062] Image: Read the image and recursively return step S3;

[0063] Emails, scripts, and others: Directly generate risk warnings;

[0064] In the recursive process, a depth parameter d is introduced, and a maximum depth d_max is set. When d ≤ d_max, the recursion continues, and when d_max is exceeded, the recursion terminates and the current result is output.

[0065] Step S6: In the review and analysis module, analyze the sliced ​​text data, visual features F1, and content summary S. I Sensitive data detection and text spelling detection are performed sequentially, followed by context analysis, image-text correlation analysis, and external link correlation analysis using a large model; detailed steps of the review and analysis module:

[0066] Step S601: Vectorize the text using the Embedding model, compare its similarity with the sensitive word vector library, quickly identify potentially sensitive content, and mark high-risk segments.

[0067] Step S602: Use a pre-trained language model including BERT to perform spelling and grammar checks on the sliced ​​text data, and identify typos, grammatical errors and abnormal expressions;

[0068] Step S603: Use LLM to perform semantic understanding and analysis on the sliced ​​text data, determine the text theme, sentiment, potential risks and inappropriate expressions, and output structured review results;

[0069] Step S604: By analyzing the contextual relationship between the preceding and following segments of the document, determine the rationality, consistency and potential misleading information of the sliced ​​text data in the overall document, and identify logical conflicts or semantic anomalies.

[0070] Step S604: Use the Score model to evaluate visual features F1 and content summary S. I The text data after the current slice is scored to evaluate the consistency between the external content and the document semantics. When the score is below the threshold σ, the tag relevance is low.

[0071] Step S605: Use the Score model to score the summary content of external links and the text data after the current slice to evaluate the consistency between external content and document semantics. If it is lower than the threshold σ, the relevance is low.

[0072] Step S7: Based on the risk warning in step S5 and the analysis results in step S6, the results integration module generates a structured review result document in the original document order.

[0073] Furthermore, in step S5,

[0074] Further details regarding the review and analysis module:

[0075] Step S601: Vectorize the text using the Embedding model, compare its similarity with the sensitive word vector library, quickly identify potentially sensitive content, and mark high-risk segments.

[0076] Step S602: Use a pre-trained language model including BERT to perform spelling and grammar checks on the sliced ​​text data, and identify typos, grammatical errors and abnormal expressions;

[0077] Step S603: Use LLM to perform semantic understanding and analysis on the sliced ​​text data, determine the text theme, sentiment, potential risks and inappropriate expressions, and output structured review results;

[0078] Step S604: By analyzing the contextual relationship between the preceding and following segments of the document, determine the rationality, consistency and potential misleading information of the sliced ​​text data in the overall document, and identify logical conflicts or semantic anomalies.

[0079] Step S604: Use the Score model to evaluate visual features F1 and content summary S. I The text data after the current slice is scored to evaluate the consistency between the external content and the document semantics. When the score is below the threshold σ, the tag relevance is low.

[0080] Step S605: Use the Score model to score the summary content of external links and the text data after the current slice to evaluate the consistency between external content and document semantics. If it is lower than the threshold σ, the relevance is low.

[0081] In one embodiment of the present invention, the content review method includes three parts: document slicing processing, image parsing, and recursive analysis of external links. These three types of data are uniformly input into the review and analysis module to achieve multimodal, full-process review of documents.

[0082] In one embodiment of the present invention, the method uses a multimodal large model for unified processing in document reading, external link parsing, image content extraction and text analysis, and achieves layer-by-layer parsing of external resources and generation of content summaries through a recursive mechanism.

[0083] In one embodiment of the present invention, the method can use a parallel processing mode to achieve efficient review of long document content, based on the complexity of the document content, the accuracy requirements of the calculation model, the differences in computing resources, and the requirements for real-time review.

[0084] Preferably, in this embodiment, the document type may include, but is not limited to, text documents, image documents, mixed documents, and very long documents, and the original data may be represented in the form of text, images, external links, etc.

[0085] Preferably, in this embodiment, to enhance the model's review and detection capabilities, in addition to the document text, external resource data can be introduced, including webpage content linked by hyperlinks, external document content, images, or media content. During model execution, static or quasi-static data can also be introduced, including compliance rule bases, sensitive word databases, and trusted domain / IP data. Since users or organizations may introduce potential risks due to lack of experience or process defects when editing documents, the method of this invention can also introduce unstructured data such as processes and specifications as reference information.

[0086] Preferably, in this embodiment, the feature processing of document content includes data preprocessing, slicing, and feature extraction. Specific techniques may include, but are not limited to, word segmentation, keyword extraction, and text vectorization.

[0087] Preferably, the specific implementation logic of text detection in this embodiment is as follows:

[0088] After slicing the document text, extract its features and input them into a large model. Using semantic analysis methods, sensitive data detection methods, and spelling and grammar checks, calculate the matching degree between the text features and the sensitive word library. If the matching value exceeds the threshold, generate a risk warning.

[0089] Preferably, the specific implementation logic for image detection in this embodiment is as follows:

[0090] For images embedded in documents or external links, a multimodal large model is used for parsing to extract text information and semantic features from the images and generate image summaries. Then, the text information is merged with the main text for unified detection, and the semantic features are correlated with the image and text context. If there are inconsistencies or risk features, a risk warning is output.

[0091] Preferably, the external link detection implementation logic in this embodiment is as follows:

[0092] Accessibility checks are performed on the extracted external links. If the link is invalid, a risk warning is generated directly. If the link is valid, it is processed according to its type: the webpage content is summarized by the large model; the document content is parsed and recursively returned to the text and image detection stage; the images are recursively returned to the image detection stage; if it is an email or script file, a risk warning is generated directly.

[0093] The following is a detailed implementation process of the present invention.

[0094] like Figure 1 The schematic diagram of the method of the present invention shown includes the following:

[0095] (1) Input the original document, parse the document content, and extract two types of information: text and images.

[0096] (2) Perform OCR and semantic parsing on the images in the document to extract text information and image features.

[0097] (3) Extract external links from the document, perform risk detection on the external links, and obtain risk warning information.

[0098] (4) Extract external links from the document. Depending on the type of external link, use different tools to further parse the text, images and other content in the external link to obtain textual information.

[0099] (5) Integrate the text information parsed from the document body and images, as well as the text information parsed from external links, into a unified and processable text data.

[0100] (6) Based on the text paragraphs and lengths, the merged text is sliced ​​to form multiple smaller "plain text" segments, which are convenient for subsequent large model processing.

[0101] (7) Input all plain text fragments into the sensitive word library and large model to perform multi-dimensional analysis, including sensitive data detection, compliance check, semantic consistency analysis, etc.

[0102] (8) Output standardized review report documents.

[0103] like Figure 2 The functional block diagram of the present invention shown includes the following:

[0104] (1) Document reading module: Supports reading the content of different types of documents.

[0105] (2) Image processing module: extracts text and generates semantic features from images.

[0106] (3) External link processing module: Performs accessibility detection and type recognition on external links, and completes web page crawling, document parsing, etc.

[0107] (4) Text processing module: performs text slicing, feature extraction and spell verification.

[0108] (5) Review and Analysis Module: Performs sensitive data detection, semantic consistency analysis, and multimodal context analysis on text, image summaries, and external link summaries.

[0109] (6) Results Integration Module: Integrates risk alerts and analysis results to generate a structured review results document.

[0110] like Figure 3 The overall flowchart of this invention shown includes the following:

[0111] (1) Data preparation: Input various original documents, including DOC, PDF and other formats, as the data source to be reviewed.

[0112] (2) Document parsing module: performs structured parsing of documents, extracts text and image content, and outputs text and image data.

[0113] (3) Image processing module: The CLIP model is used to determine the probability that the image is plain text. If the probability is lower than the threshold θ, the large visual model is used to extract feature information; otherwise, OCR is performed to extract the content of the image. Finally, LLM is used to summarize the image content or feature information.

[0114] (4) Text Processing Module: Based on the parsed text and image text information, external links are extracted, and the text is sliced ​​according to paragraph and text length to generate structured text slice data. This ensures that the slice content meets the input requirements of subsequent model analysis, and external link information is extracted. Here, `text` represents the complete text content, and slices are defined by length L, overlap length ΔL, and P. start (i) and P end (i) Slice the i-th slice by the start and end positions of the paragraph, where slice i is the text content after slicing;

[0115] The specific formula is as follows:

[0116] (5) External Link Processing Module: Determines the type of external link. If it is a webpage, uses Jina to crawl the webpage; if it is a document, loads the document and parses its text; if it is an image, performs OCR and visual large model recognition; if it is an email / compressed file, generates risk warning information. The recognized content of the types of images, webpages, and documents is recursively processed, and its content is summarized using LLM.

[0117] (6) Review and Analysis Module: The text slices are sequentially processed through sensitive data detection, text spelling detection, large-scale model semantic analysis, context analysis, image-text correlation analysis, and external link correlation analysis to complete multi-level intelligent review. Among them:

[0118] Sensitive data detection: By vectorizing text using the Embedding model and comparing its similarity with a sensitive word vector library, potential sensitive content can be quickly identified and high-risk segments can be marked.

[0119] Text spelling detection: Using pre-trained language models such as BERT, the spelling and grammar of sliced ​​text are checked to identify typos, grammatical errors and abnormal expressions, thereby improving the standardization of the text.

[0120] Large Model Semantic Analysis: Using LLM to perform semantic understanding and analysis on sliced ​​text, determine the text topic, sentiment, potential risks and inappropriate expressions, and output structured review results.

[0121] Contextual analysis: By analyzing the contextual relationships between segments of a document, the rationality, consistency, and potential misleading information of the segmented content in the overall document are determined, and logical conflicts or semantic anomalies are identified.

[0122] Image-text correlation analysis and external link correlation analysis: The Score model is used to score the summary content of images and external links and the current text slice to evaluate the consistency between external content and document semantics. When the score is below the threshold σ, the correlation is marked as low.

[0123] (7) Results Integration Module: The review analysis results and external link risk warning information are summarized in a unified manner, the final review result data is output, and a standardized review report document is generated.

[0124] The present invention also provides an electronic device, including a memory, a processor, and computer program instructions stored in the memory and executable by the processor, wherein when the processor executes the computer program instructions, it can implement the steps of any of the methods described above.

[0125] The present invention also provides a computer-readable storage medium having stored thereon computer program instructions that can be executed by a processor, wherein when the processor executes the computer program instructions, it can implement the steps of any of the methods described above.

[0126] The above are preferred embodiments of the present invention. Any changes made to the technical solution of the present invention that do not exceed the scope of the technical solution of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for reviewing the content of long documents based on artificial intelligence, characterized in that, Includes the following steps: Step S1: Prepare the documents to be tested; Step S2: Based on the document type, use differentiated data parsing methods to parse the document content, and input the parsing results into the text processing module and the image processing module; Step S3: Utilize a multimodal large model to detect and parse the images I={I1,I2,…,Im} contained in the document content, outputting the text information T1 or visual features F1 contained in the images, and using LLM to generate a content summary S of the images. I The text information T1 is input into the text processing module, and the visual features F1 and content summary S are input into the text processing module. I Input to the review and analysis module; where: Step S4: Perform slicing processing on the data received by the text processing module, extract external links from the data, and transmit the external links to the external link processing module. Simultaneously, send the sliced ​​text data to the review and analysis module. The slicing processing method is represented as follows: Where text represents the complete text content, L is the standard slice length, ΔL is the overlap length, and P... start (i) and P end (i) represents the start and end positions of the i-th slice, and slice i represents the text content after slicing. Step S5: In the external link processing module, perform accessibility checks on the external links: if they are inaccessible, generate a risk warning; if they are accessible, perform recursive parsing based on the type of external link. In the recursive process, introduce a depth parameter d and set a maximum depth d_max. Continue expanding when d ≤ d_max, and terminate the recursion and output the current result when d_max is exceeded. Step S6: In the review and analysis module, analyze the sliced ​​text data, visual features F1, and content summary S. I Sensitive data detection and text spelling detection are performed sequentially, and context analysis, image-text correlation analysis, and external link correlation analysis are conducted in conjunction with the large model. Step S7: Based on the risk warning in step S5 and the analysis results in step S6, the results integration module generates a structured review result document in the original document order.

2. The method for reviewing long document content based on artificial intelligence according to claim 1, characterized in that, In step S4, external links include URLs and document links.

3. The method for reviewing long document content based on artificial intelligence according to claim 1, characterized in that, In step S5, external links are categorized into four types, each with a different processing method, as detailed below: URL: Use web crawling tools to capture the page and generate a summary of the page content using LLM; Document: Read the document and generate a summary from the LLM, while recursively returning step S3; Image: Read the image and recursively return step S3; Emails, scripts, and others: Directly generate risk warnings.

4. The method for reviewing long document content based on artificial intelligence according to claim 1, characterized in that, Detailed steps for reviewing and analyzing the module: Step S601: Vectorize the text using the Embedding model, compare its similarity with the sensitive word vector library, quickly identify potentially sensitive content, and mark high-risk segments. Step S602: Use a pre-trained language model including BERT to perform spelling and grammar checks on the sliced ​​text data, and identify typos, grammatical errors and abnormal expressions; Step S603: Use LLM to perform semantic understanding and analysis on the sliced ​​text data, determine the text theme, sentiment, potential risks and inappropriate expressions, and output structured review results; Step S604: By analyzing the contextual relationship between the preceding and following segments of the document, determine the rationality, consistency and potential misleading information of the sliced ​​text data in the overall document, and identify logical conflicts or semantic anomalies. Step S604: Use the Score model to evaluate visual features F1 and content summary S. I The text data after the current slice is scored to evaluate the consistency between the external content and the document semantics. When the score is below the threshold σ, the tag relevance is low. Step S605: Use the Score model to score the summary content of external links and the text data after the current slice to evaluate the consistency between external content and document semantics. If it is lower than the threshold σ, the relevance is low.

5. The method for reviewing long document content based on artificial intelligence according to claim 1, characterized in that, The method can perform content slicing and progressive processing on ultra-long documents, and use a multimodal large model to perform semantic parsing and text extraction on images embedded in the document, thereby achieving cross-modal document content review.

6. The method for reviewing long document content based on artificial intelligence according to claim 1, characterized in that, The method employs a parallel processing mode during external link parsing to identify the type of external links and read their content. It also uses a recursive parsing mechanism to examine and summarize the documents, web pages, and images contained in the links layer by layer, ensuring the consistency and integrity of the external resources with the main document.

7. The method for reviewing long document content based on artificial intelligence according to claim 1, characterized in that, The review and analysis module analyzes the sliced ​​text data, external links, visual features F1, and content summary S. I The system performs unified and integrated processing, sequentially completing sensitive data detection, spelling and grammar verification, and semantic consistency analysis. Based on a large model, it performs contextual multimodal correlation analysis, and finally outputs a structured review result document that can be used for compliance review, risk control, and security testing.

8. A long document content review system based on artificial intelligence, characterized in that, include: The data input module prepares the documents to be tested. The document parsing module performs structured parsing of documents, extracts text and image content, and outputs text and image data. The image processing module utilizes a multimodal large model to detect and parse image data, outputting the textual information T1 or visual features F1 contained in the image data, and using LLM to generate a content summary S for the image. I The text information T1 is input into the text processing module, and the visual features F1 and content summary S are input into the text processing module. I Pass the data to the review and analysis module; The text processing module extracts external links based on the text data and text information T1, and transmits the external links to the external link processing module. At the same time, it segments the text according to the paragraph and text length to generate structured text segment data, and sends the text segment data to the review and analysis module. The external link processing module performs accessibility checks on external links; if an link is inaccessible, a risk warning is generated. If accessible, recursively parse the external link based on its type and summarize its content using LLM; The review and analysis module analyzes the visual features F1 and content summary S of the text slice data. I The system sequentially performs sensitive data detection, large-scale model semantic analysis, contextual analysis, image-text correlation analysis, and external link correlation analysis to complete multi-level intelligent review. The results integration module summarizes the review and analysis results and external link risk warning information in a unified manner, and generates a structured review results document in the original document order.

9. An electronic device comprising a memory, a processor, and computer program instructions stored in the memory and executable by the processor, wherein when the processor executes the computer program instructions, it is able to implement the steps of the method as described in any one of claims 1-7.

10. A computer-readable storage medium having stored thereon computer program instructions executable by a processor, wherein when the processor executes the computer program instructions, it is able to implement the steps of the method as described in any one of claims 1-7.

Images