Document content auditing method and system based on combined artificial intelligence

Abstract

The invention discloses a document content auditing method and system based on combined artificial intelligence, and relates to the technical field of content auditing. According to the method, an original document is converted and standardized into a unified format, a natural language processing technology and a regular expression are applied to extract text fragments, images and meta-information of the text fragments, the images and the meta-information of the text fragments, the images and the meta-information; the method comprises the following steps: inputting a file into a combined artificial intelligence model to respectively detect semantic violation, keyword violation and image violation, generating violation condition description and correction cases through a large language model, marking the auditing result in the file through highlighting to form an auditing report, and providing an auditing conclusion and a modification suggestion for file content auditing.

Description

Technical Field

[0001] The present invention relates to the field of content review, and in particular to a document content review method and system based on combined artificial intelligence. Background Art

[0002] In the era of information explosion and digital transformation, the review of contracts, credits, archives and other documents has become a critical and indispensable process. However, the traditional manual review method is not only time-consuming and costly, but also easily affected by human subjective factors, and it is difficult to meet the needs of modern society for efficient and accurate document processing.

[0003] With the development of machine learning and deep learning technologies, especially the advancement of natural language processing and computer vision technologies, the automation of document review has become possible. In recent years, the vigorous development of large language model technology has also provided basic model support with stronger reasoning capabilities and broader world knowledge.

[0004] Despite this, the existing document review technology still cannot meet the complex review scenarios. In actual document scenarios, the text volume is large, the structure is complex, there are many review items, and different rules have different review objects and scopes. The existing document content review system cannot meet the review requirements. Therefore, how to deal with the problem of low efficiency of document content review in real scenarios and use artificial intelligence technology to complete automated document content review is an urgent problem to be solved. Summary of the invention

[0005] A series of simplified concepts are introduced in the Summary of the Invention, which will be further described in detail in the Detailed Description of the Invention. The Summary of the Invention does not mean to attempt to define the key features and essential technical features of the claimed technical solution, nor does it mean to attempt to determine the scope of protection of the claimed technical solution.

[0006] The present invention mainly solves the problem that the existing technology is not suitable for the low efficiency of document content review in real scenarios, and proposes a document content review method and system based on combined artificial intelligence to achieve the purpose of automated document content review.

[0007] In a first aspect, the present invention provides a document content review method based on combined artificial intelligence, comprising the following steps:

[0008] Step 100, parsing documents in different data formats, standardizing the document structure and content to obtain a document set in PDF (Portable Document Format, a commonly used file format in the computer field) format , extract images and text paragraphs from the document and get the image set and a collection of text paragraphs ;

[0009] Step 200, using a divide-and-conquer strategy, using a combined artificial intelligence model to determine the image set and a collection of text paragraphs The illegal content in the file is recorded, and the illegal location, illegal item and illegal text are recorded to obtain the illegal content set. ;

[0010] Step 300: The illegal content set obtained in step 200 is The violation location, violation item, and violation text recorded in the document are applied to the prompt word template to prompt the large language model to generate explanations and corrections for specific violations and attach them to the violation content set. ;

[0011] Step 400, a set of illegal content including illegal description and correction plan obtained in step 300 The recorded violations are in the document set The illegal information is highlighted in turn to form an audit report of the document.

[0012] Furthermore, in step 100, documents of different data formats are parsed, and the document structure and content are standardized to obtain a document set in PDF format. , extract images and text paragraphs from the document and get the image set and a collection of text paragraphs , including the following steps:

[0013] Step 101: Use the document format parsing library to convert different documents into PDF format to obtain a document set. ;

[0014] Step 102, using regular expressions to match the result tags of the document, obtaining structured meta-information such as paragraphs and chapter divisions in the document, and obtaining paragraph-by-paragraph text sets through the PDF parsing library ;

[0015] Step 103, applying natural language processing technology to the paragraph-by-paragraph text set obtained in step 102 Apply the sentence segmentation model to obtain independent sentences, and attach structural meta-information such as the paragraph and chapter to each sentence to form a text paragraph set. ;

[0016] Step 104, use a PDF parsing tool to extract images from the document, and attach structural meta information such as the paragraph and chapter where each image is located, as well as text context information near the image, to form an image set. .

[0017] Furthermore, the step 200 adopts a divide-and-conquer strategy and uses a combined artificial intelligence model to determine the image set and a collection of text paragraphs The illegal content in the file is recorded, and the illegal location, illegal item and illegal text are recorded to obtain the illegal content set. , including the following steps:

[0018] Step 210: construct a first text violation detection branch, apply named entity recognition technology, and detect the text paragraph set There are semantically illegal items in the set, forming a set of illegal content ;

[0019] Step 220: construct a second text violation detection branch and apply pattern matching technology to detect the text paragraph set. There are projects with keyword violations in the content, forming a collection of illegal content ;

[0020] Step 230: construct a first image violation detection branch, apply multimodal large language model technology, and detect the image set The image violation items in the form of a violation content set ;

[0021] Step 240, merging the set of illegal content obtained in steps 310 to 230 , , , remove duplicates and form a set of illegal content .

[0022] Furthermore, in step 210, a first text violation detection branch is constructed to detect a set of text paragraphs using named entity recognition technology. There are semantically illegal items in the set, forming a set of illegal content , including the following steps:

[0023] Step 211: Use the representation network to represent the text paragraph set Characterize the text in to get the text feature set ;

[0024] Step 212, based on the text feature set obtained in step 211 , apply the Global Pointer predictor (a named entity recognition method in the field of deep learning) to predict the violation location, violation item and violation text, and obtain the violation content set .

[0025] Furthermore, in step 220, a second text violation detection branch is constructed to detect the text paragraph set by applying pattern matching technology. There are projects with keyword violations in the content, forming a collection of illegal content , including the following steps:

[0026] Step 221, compile AC automata (Aho-CorasickAutomata, an algorithm for string pattern matching) according to the violation keyword lists specified by different violation items;

[0027] Step 222, applying the AC automaton obtained in step 221 to match the text in the text paragraph set T, outputting the violation position, violation item and violation text, and obtaining the violation content set .

[0028] Furthermore, in step 230, a first image violation detection branch is constructed, and a multimodal large language model technology is applied to detect the image set. The image violation items in the form of a violation content set , including the following steps:

[0029] Step 231, call the first prompt word template and add the image set The image and image violation standard in the prompt word template are used to obtain the instruction for performing the image violation detection task, and the multimodal large language model is prompted to obtain the instruction execution result. ;

[0030] Step 232, analyzing the instruction execution result obtained in step 231 ,extract The JSON (JavaScript Object Notation, a standard information representation method commonly used in the computer field) structured results in the image violation items are obtained, and the violation content set is obtained. .

[0031] Furthermore, in step 300, the violation position, violation item and violation text recorded in the violation content set C obtained in step 200 are applied to the prompt word template to prompt the large language model to generate an explanation and correction for the specific violation, and append it to the violation content set. , including the following steps:

[0032] Step 301, call the second prompt word template, add the illegal content in the illegal content set and the detection standard of the illegal item to the prompt word template in turn, obtain the instruction to execute the illegal situation description task, and prompt the large language model to obtain the instruction execution result ;

[0033] Step 302: Analyze the command execution result obtained in step 301 ,extract JSON structured result in the result to get the violation description;

[0034] Step 303, calling the third prompt word template, sequentially adding the illegal content in the illegal content set, the detection standard of the illegal item and the illegal situation description obtained in step 302 to the prompt word template, obtaining the instruction to execute the amendment generation task, and prompting the large language model to obtain the instruction execution result ;

[0035] Step 304: Analyze the instruction execution result obtained in step 303 ,extract The JSON structured result in step 302 is used to obtain the violation amendment, which is attached to the violation content set together with the violation description obtained in step 302. middle.

[0036] Further, in step 400, the set of illegal contents including the illegal situation description and the correction plan obtained in step 300 is The recorded violations are in the document set The illegal information is highlighted in turn to form a document audit report, including the following steps:

[0037] Step 401, in the document set The set of illegal content obtained in the matching step 300 Obtain the coordinate position of each illegal text in the PDF document one by one ;

[0038] Step 402, construct a PDF standard annotation object, specify the annotation type as highlight, and use the coordinate position obtained in step 401 Specify the highlight range and specify the annotation content as the set of illegal content The violation items, violation texts, violation descriptions and amendments in the code are used to obtain the annotated object set. ;

[0039] Step 403: The labeled object set obtained in step 402 is Apply to Document Set In the audit, you will get the final audit result report.

[0040] In a second aspect, the present invention provides a document content review system based on combined artificial intelligence, comprising:

[0041] The Web service module obtains the document set to be reviewed by processing the interface request , and returns the result report to the user through a response after the audit is completed;

[0042] The PDF processing module completes the document set Extraction, get the image set and a collection of text paragraphs , and output the illegal content set in the algorithm module After completing the document set The process of highlighting the violation information in turn to form a document audit report;

[0043] The task scheduling module is responsible for dynamically scheduling the number of executors of the algorithm module in the server according to the server status and request load, supporting the operation of the algorithm module;

[0044] The algorithm module is responsible for completing the image set and a collection of text paragraphs Detect illegal content and generate illegal content descriptions and amendments to form illegal content collection process.

[0045] The document content review method and system based on combined artificial intelligence provided by the present invention have the following advantages compared with the prior art:

[0046] 1. It replaces the traditional manual document content review process, avoiding the problems of low efficiency, personal influence and low quality of manual work.

[0047] 2. A divide-and-conquer combined artificial intelligence method was introduced to solve the problem of using artificial intelligence for content review when the review project includes checkpoints in different modalities such as images and texts. It also strengthened the ability to detect violations in various situations such as keyword violations and semantic violations.

[0048] 3. The introduction of a large language model to generate violation descriptions and amendments solves the "black box" problem of traditional AI content review, improves the interpretability of review results, and further ensures the quality of document content review. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a flow chart of a document content review method based on combined artificial intelligence described in the present invention;

[0050] Figure 2 This is a schematic diagram of illegal content detection in step 200 of a document content review method based on combined artificial intelligence according to the present invention;

[0051] Figure 3 It is a schematic diagram of step 300 of the document content review method based on combined artificial intelligence described in the present invention, describing the violation and generating amendments;

[0052] Figure 4 It is a schematic diagram of a document content review method based on combined artificial intelligence according to the present invention; DETAILED DESCRIPTION

[0053] In order to make the technical problems solved by the present invention, the technical solutions adopted and the technical effects achieved clearer, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It is understood that the specific embodiments described herein are only used to explain the present invention, rather than to limit the present invention. It should also be noted that, for the convenience of description, only the parts related to the present invention are shown in the accompanying drawings, rather than all the contents.

[0054] First, as Figure 1 As shown, an embodiment of the present invention provides a document content review method based on combined artificial intelligence, comprising the following steps:

[0055] Step 100: parse documents in different data formats, standardize the document structure and content to obtain a document set in PDF format , extract images and text paragraphs from the document and get the image set and a collection of text paragraphs .

[0056] Step 101: Use the document format parsing library to convert different documents into PDF format to obtain a document set. .

[0057] Step 102, using regular expressions to match the result tags of the document, obtaining structured meta-information such as paragraph and chapter divisions in the document, and obtaining paragraph-by-paragraph text sets through the PDF parsing library .

[0058] Specifically, structured meta information includes but is not limited to paragraphs and chapters. All structured information required to meet the audit project requirements should be considered meta information. For example, if an audit project requires that the product description must include a "Safety Instructions" field, then the field in which the text paragraph is located also belongs to structured meta information.

[0059] Step 103, applying natural language processing technology to the paragraph-by-paragraph text set obtained in step 102 Apply the sentence segmentation model to obtain independent sentences, and attach structural meta-information such as the paragraph and chapter to each sentence to form a text paragraph set. .

[0060] For example, the sentence segmentation model provided by the algorithm library spaCy can be used to group text paragraphs into The long text in is divided into several sentences to form a set of text paragraphs In order to ensure the quality of the review of violations in the subsequent steps, the text paragraph set Saved in JSONL format, each text paragraph record contains text, file name, page number, structured metadata, etc.

[0061] It should be noted that the embodiments of the present invention do not limit the sentence segmentation method used, including but not limited to the spaCy algorithm library, the NLTK algorithm library, the stop word-based method or the neural network-based method.

[0062] Step 104, use a PDF parsing tool to extract images from the document, and attach structural meta information such as the paragraph and chapter where each image is located, as well as text context information near the image, to form an image set. .

[0063] Specifically, the algorithm library PyPDF2 is used to parse images from PDF to form an image set ,The image set is saved in JSONL format, and the images extracted from the PDF are encoded into strings using Base64 and saved in the JSON object.,Each image record contains the Base64 encoding of the image, the file name, the page number, the text context of the image neighborhood, and structured meta-information.

[0064] Step 200, using a divide-and-conquer strategy, using a combined artificial intelligence model to determine the image set and a collection of text paragraphs The illegal content in the file is recorded, and the illegal location, illegal item and illegal text are recorded to obtain the illegal content set. .

[0065] Specifically, this step is the core step of document content review. For common keyword violations, semantic violations, and image violations in document review, a divide-and-conquer strategy is adopted to complete the review work using a combined artificial intelligence model, such as Figure 2 shown.

[0066] Step 210: construct a first text violation detection branch, apply named entity recognition technology, and detect the text paragraph set There are semantically illegal items in the set, forming a set of illegal content ;

[0067] This step is mainly aimed at detecting semantic violations. Semantic violations require the perception and judgment of the overall meaning of the text paragraph, and the search for illegal content in combination with the context. The embodiment of the present invention introduces named entity recognition technology to detect semantic violations in the file.

[0068] Step 211: Use the representation network to represent the text paragraph set Characterize the text in to get the text feature set ;

[0069] Specifically, for a set of text paragraphs A text in , using embedding technology to embed text paragraphs The encoding length is sequence , using the DeBERTa (decoupled attention encoding enhanced BERT architecture) backbone network to train the sequence Characterize and obtain the characteristic expression of the text paragraph After this step, the text paragraph set Characterized into text feature set .

[0070] It should be noted that the embodiments of the present invention do not limit the representation network used, including but not limited to the Transformer architecture representation network represented by the DeBERTa and the recurrent neural network architecture representation network represented by BiLSTM (Bidirectional Long ShortTerm Memory).

[0071] Step 212, based on the text feature set obtained in step 211 , apply the Global Pointer (a named entity recognition method) predictor to predict the violation location, violation item and violation text, and obtain the violation content set .

[0072] Specifically, for the text feature set A column of feature tensors in , use the first multi-layer perceptron to predict a column of tensors , the output dimension of the first multilayer perceptron is is set as the type of semantic violation in the audit project. Therefore, each position of the output of the first multilayer perceptron is a column of probability vectors. The bit represents the first It is a place Similarly, use the second multi-layer perceptron to predict a column of tensors , the first column vector of each column in the output tensor of the second multilayer perceptron The position of the sequence is the first Further, we can calculate the probability of the illegal content ending. Position to Subsequence of bits , which belongs to The probability score of a violation can be calculated as .

[0073] In particular, considering the use of sequence start and end boundary information, step 212 introduces RoPE (Rotary Position Embedding) position encoding, and the above probability score is corrected to ,in , and It is Position and Bit position encoding.

[0074] Furthermore, by introducing a predictor, according to the calculation rules of the violation probability score, and A scoring matrix , each position in the score matrix Representatives from Position to The subsequence of bits violates The probability of class auditing items. Setting violation thresholds The values ​​of the probability matrix are filtered to obtain a binary matrix, and the results greater than the violation threshold will be considered as violations.

[0075] According to the binary matrix, the location and category of the illegal items are collected, and the text subsequence is obtained to construct a JSON object to complete the illegal content set 's construction.

[0076] Step 220: construct a second text violation detection branch and apply pattern matching technology to detect the text paragraph set. There are projects with keyword violations in the content, forming a collection of illegal content ;

[0077] This step is mainly aimed at detecting keyword violations. Keyword violations require accurate detection of the offending words and cannot be missed. The first text violation detection branch can detect keyword violations. However, the method based on deep learning is affected by probability and has the risk of omission. The embodiment of the present invention introduces pattern matching technology to construct the second text violation detection branch to accurately identify keyword violations in the file.

[0078] Step 221, compile AC automaton according to the violation keyword list specified by different violation items;

[0079] Specifically, a configuration file is loaded to import the violation keyword lists specified by different violation items, and for each violation item corresponding to the keyword list, an AC automaton object for matching the pattern is compiled to form an AC automaton AC.

[0080] Specifically, for a keyword list corresponding to a violation item, a Deterministic Finite Automata (DFA) that can match the pattern is constructed. Using the data structure of a prefix tree (also known as a Trie tree or dictionary tree), each keyword in the keyword list is inserted in turn to form a DFA automaton that can match the pattern.

[0081] Furthermore, a new state transition condition is added to the DFA automaton, called mismatch transition. For each state on the DFA automaton, when the automaton fails to perform pattern matching in the state, it is transferred to the state with the longest strict suffix in the prefix tree. The reconstruction process is completed using the breadth-first search algorithm to obtain an AC automaton that can match the specified pattern.

[0082] Step 222, applying the AC automaton obtained in step 221 to the text paragraph set Match the text in the output, output the violation position, violation item and violation text, and get the violation content set .

[0083] Specifically, for a set of text paragraphs A text in , using automaton AC to match text from left to right , when the next node of the prefix tree can satisfy the current state, execute the state transfer according to the prefix tree, otherwise, execute the mismatch transfer until returning to the root node of the prefix tree. Output all the text patterns matched in the process. Call the lookup table, map the violation items corresponding to each text pattern, build a JSON object, record the violation location, violation item and violation text, and complete the violation content set 's construction.

[0084] Step 230: construct a first image violation detection branch, apply multimodal large language model technology, and detect the image set The image violation items in the form of a violation content set ;

[0085] This step is mainly aimed at the detection of image violations. Image violations require the artificial intelligence responsible for detection to have the ability to understand the two modes of the image to be detected and the rule text. The embodiment of the present invention introduces a multimodal large language model, and uses the command obedience ability and context learning ability of the multimodal large language model to complete the detection of image violations in the file.

[0086] Step 231, call the first prompt word template and add the image set The image and image violation standard in the prompt word template are used to obtain the instruction for performing the image violation detection task, and the multimodal large language model is prompted to obtain the instruction execution result. ;

[0087] Specifically, a prompt project is carried out according to the image violation review items, and the prompt word design is written and optimized in combination with the few-sample prompt technology, so that the model has the ability to review the input image according to the review rule items, and return the review results in JSON format to form the first prompt word template.

[0088] For image sets A batch of Base64-encoded image samples in the , use the sample encoding results to fill the placeholder in the first prompt word template, call the multimodal large language model to perform violation detection, and obtain the instruction execution results generated by the model .

[0089] In some embodiments, since there are many types of audit items, one prompt word template cannot optimize all audit items. Multiple prompt word templates can be grouped and constructed to form a group of templates that together constitute the first prompt word template.

[0090] Step 232, analyzing the instruction execution result obtained in step 231 ,extract JSON structured results in, get the image violation items, get the violation content set .

[0091] Specifically, the return of the multimodal large language model is in text form. In order to effectively parse the results from the text generated by the multimodal large language model, the Markdown syntax is used to specify the output format, and the few-sample prompts are used to constrain the output format of the multimodal large language model.

[0092] Furthermore, when the instruction execution result is obtained After that, parse the code paragraph surrounded by "```", build a regular expression to match the JSON object or JSON array in it, objectify the matched text, access the illegal project field from it, and get the illegal content set .

[0093] Step 240, merging the set of illegal content obtained in steps 310 to 230 , , , remove duplicates and form a set of illegal content .

[0094] Since the audit targets of the first text violation detection branch and the second text violation detection branch are both text segment sets The text in the detection result may be repeated. The embodiment of the present invention introduces step 240 to remove duplicates of illegal content.

[0095] For example, for a set of text paragraphs A text in , initialize a column of labeled vectors ,in For Text Length, The number of illegal review opinions. , , Get the set of illegal content . Traverse For each illegal content starting with Ends at Violation of the The illegal content of the illegal items is detected in the tag vector Are all the bits of the , otherwise discard. After the traversal is completed, a set of illegal content is formed .

[0096] like Figure 3 As shown, step 300, the illegal content set obtained in step 200 is The violation location, violation item, and violation text recorded in the document are applied to the prompt word template to prompt the large language model to generate explanations and corrections for specific violations and attach them to the violation content set. .

[0097] Step 301, call the second prompt word template, add the illegal content in the illegal content set and the detection standard of the illegal item to the prompt word template in turn, obtain the instruction to execute the illegal situation description task, and prompt the large language model to obtain the instruction execution result ;

[0098] Specifically, we carry out prompt engineering, write and optimize the prompt word design, so that the model has the ability to explain the reasons why the content is in violation according to the rules of the violation project, the violation content, and the context of the violation content, and return the review results in JSON format to form the second prompt word template.

[0099] Similar to step 231, fill in the prompt word template, call the large language model to explain the execution reason, and obtain the instruction execution result generated by the model .

[0100] Step 302: Analyze the command execution result obtained in step 301 ,extract JSON structured result in the result to get the violation description;

[0101] Step 303, calling the third prompt word template, sequentially adding the illegal content in the illegal content set, the detection standard of the illegal item and the illegal situation description obtained in step 302 to the prompt word template, obtaining the instruction to execute the amendment generation task, and prompting the large language model to obtain the instruction execution result ;

[0102] Specifically, we carry out prompt engineering, write and optimize the prompt word design, so that the model has the ability to generate modification suggestions based on the rules of the violation items, the violation content, the violation content context, and the violation reasons, and return the review results in JSON format to form the third prompt word template.

[0103] Similar to step 231, fill in the prompt word template, call the large language model to execute the modification suggestion generation, and obtain the instruction execution result generated by the model .

[0104] Step 304: Analyze the instruction execution result obtained in step 303 ,extract The JSON structured result in step 302 is used to obtain the violation amendment, which is attached to the violation content set together with the violation description obtained in step 302. middle.

[0105] Step 400, a set of illegal content including illegal description and correction plan obtained in step 300 The recorded violations are in the document set The illegal information is highlighted in turn to form an audit report of the document.

[0106] Step 401, in the document set The set of illegal content obtained in the matching step 300 Obtain the coordinate position of each illegal text in the PDF document one by one ;

[0107] To complete the audit report, the violation content and violation details need to be recorded in the audit results. The annotation of the PDF document needs to specify the values ​​of the Rect and QuadPoints fields in the standard coordinate format. Therefore, it is necessary to match the violation content with the PDF page to calculate the coordinate position to be annotated. .

[0108] Specifically, for illegal content collection , read the page number recorded in the meta information. Extract the character object of the page corresponding to the page number in the PDF, and build a mapping table to record the characters and coordinate positions. Match the illegal text in the extracted text and obtain the coordinate positions of all characters. Obtain the Rect field area by calculating the coordinates of the enclosing rectangle, and adjust the QuadPoints point coordinates in a fine-grained manner by calculating whether to split the line.

[0109] In particular, it is necessary to determine whether the illegal text crosses pages. For cross-page violations, two annotations need to be created and the coordinates calculated separately.

[0110] Step 402, construct a PDF standard annotation object, specify the annotation type as highlight, and use the coordinate position obtained in step 401 Specify the highlight range and specify the annotation content as the set of illegal content The violation items, violation texts, violation descriptions and amendments in the code are used to obtain the annotated object set. ;

[0111] Step 403: The labeled object set obtained in step 402 is Apply to Document Set In the process of completing the audit, you will get the final audit result report.

[0112] Second, as Figure 4 As shown, an embodiment of the present invention provides a document content review system based on combined artificial intelligence, including the following modules:

[0113] The Web service module obtains the document set to be reviewed by processing the interface request , and returns the result report to the user through a response after the audit is completed;

[0114] The PDF processing module completes the document set Extraction, get the image set and a collection of text paragraphs , and output the illegal content set in the algorithm module After completing the document set The process of highlighting the violation information in turn to form a document audit report;

[0115] The task scheduling module is responsible for dynamically scheduling the number of executors of the algorithm module in the server according to the server status and request load, supporting the operation of the algorithm module;

[0116] The algorithm module is responsible for completing the image set and a collection of text paragraphs Detect illegal content and generate illegal content descriptions and amendments to form illegal content collection process.

[0117] After testing, it was found that a document content review method and system based on combined artificial intelligence provided by an embodiment of the present invention can efficiently and accurately detect various common violation types in file review scenarios, such as semantic violations, keyword violations, image violations, etc., with excellent accuracy and recall rate, and can significantly reduce the operation time of file review and effectively ensure the operation quality of file review.

[0118] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that modifying the technical solutions described in the aforementioned embodiments, or replacing some or all of the technical features therein by equivalents, does not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A document content review method based on combined artificial intelligence, characterized in that: The following steps are involved: Step 100: parse documents in different data formats, standardize the document structure and content to obtain a document set in PDF format , extract images and text paragraphs from the document and get the image set and a collection of text paragraphs ; Step 200, using a divide-and-conquer strategy, using a combined artificial intelligence model to determine the image set and a collection of text paragraphs The illegal content in the file is recorded, and the illegal location, illegal item and illegal text are recorded to obtain the illegal content set. ; Step 300: The illegal content set obtained in step 200 is The violation location, violation item, and violation text recorded in the document are applied to the prompt word template to prompt the large language model to generate explanations and corrections for specific violations and attach them to the violation content set. ; Step 400, a set of illegal content including illegal description and correction plan obtained in step 300 The recorded violations are in the document set The illegal information is highlighted in turn to form an audit report of the document.

2. A document content review method based on combined artificial intelligence according to claim 1, characterized in that: The step 100 comprises the following steps: Step 101: Use the document format parsing library to convert different documents into PDF format to obtain a document set. ; Step 102, using regular expressions to match the result tags of the document, obtaining the paragraph and chapter division structured meta information in the document, and obtaining the paragraph-by-paragraph text set through the PDF parsing library ; Step 103, applying natural language processing technology to the paragraph-by-paragraph text set obtained in step 102 Apply the sentence segmentation model to obtain independent sentences, and attach the paragraph and chapter structured meta information to each sentence to form a text paragraph set. ; Step 104, use a PDF parsing tool to extract images from the document, and attach paragraph and chapter structured meta information and text context information near the image to each image to form an image set. .

3. According to claim 1, a document content review method based on combined artificial intelligence is characterized by: The step 200 comprises the following steps: Step 210: construct a first text violation detection branch, apply named entity recognition technology, and detect the text paragraph set There are semantically illegal items in the set, forming a set of illegal content ; Step 220: construct a second text violation detection branch and apply pattern matching technology to detect the text paragraph set. There are projects with keyword violations in the content, forming a collection of illegal content ; Step 230: construct a first image violation detection branch, apply multimodal large language model technology, and detect the image set The image violation items in the form of a violation content set ; Step 240, merging the set of illegal content obtained in steps 310 to 230 , , , remove duplicates and form a set of illegal content .

4. According to claim 3, a document content review method based on combined artificial intelligence is characterized by: The step 210 includes the following steps: Step 211, using the representation network to represent the text paragraph set Characterize the text in to get the text feature set ; Step 212, based on the text feature set obtained in step 211 , apply the Global Pointer predictor to predict the violation location, violation item and violation text, and obtain the violation content set .

5. According to claim 3, a document content review method based on combined artificial intelligence is characterized by: The step 220 includes the following steps: Step 221, compile AC automaton according to the violation keyword list specified by different violation items; Step 222, applying the AC automaton obtained in step 221 to the text paragraph set Match the text in the output, output the violation position, violation item and violation text, and get the violation content set .

6. According to claim 3, a document content review method based on combined artificial intelligence is characterized by: The step 230 includes the following steps: Step 231, call the first prompt word template and add the image set The image and image violation standard in the prompt word template are used to obtain the instruction for performing the image violation detection task, and the multimodal large language model is prompted to obtain the instruction execution result. ; Step 232, analyzing the instruction execution result obtained in step 231 ,extract JSON structured results in, get the image violation items, get the violation content set .

7. According to claim 1, a document content review method based on combined artificial intelligence is characterized by: The step 300 includes the following steps: Step 301, call the second prompt word template, add the illegal content in the illegal content set and the detection standard of the illegal item to the prompt word template in turn, obtain the instruction to execute the illegal situation description task, and prompt the large language model to obtain the instruction execution result ; Step 302: Analyze the command execution result obtained in step 301 ,extract JSON structured result in the result to get the violation description; Step 303, calling the third prompt word template, sequentially adding the illegal content in the illegal content set, the detection standard of the illegal item and the illegal situation description obtained in step 302 to the prompt word template, obtaining the instruction to execute the amendment generation task, and prompting the large language model to obtain the instruction execution result ; Step 304: Analyze the instruction execution result obtained in step 303 ,extract The JSON structured result in step 302 is used to obtain the violation amendment, which is attached to the violation content set together with the violation description obtained in step 302. middle.

8. According to claim 1, a document content review method based on combined artificial intelligence is characterized by: The step 400 includes the following steps: Step 401, in the document set The set of illegal content obtained in the matching step 300 Obtain the coordinate position of each illegal text in the PDF document one by one ; Step 402, construct a PDF standard annotation object, specify the annotation type as highlight, and use the coordinate position obtained in step 401 Specify the highlight range and specify the annotation content as the set of illegal content The violation items, violation texts, violation descriptions and amendments in the code are used to obtain the annotated object set. ; Step 403: The labeled object set obtained in step 402 is Apply to Document Set In the process of completing the audit, you will get the final audit result report.

9. A document content review system based on combined artificial intelligence, implementing a document content review method based on combined artificial intelligence as described in any one of claims 1 to 8, characterized in that: Including Web service module, PDF processing module, task scheduling module and algorithm module; The Web service module obtains the document set to be reviewed by processing the request of the interface , and returns the result report to the user through a response after the audit is completed; The PDF processing module completes the document set Extraction, get the image set and a collection of text paragraphs , and output the illegal content set in the algorithm module After completing the document set The process of highlighting the violation information in turn to form a document audit report; The task scheduling module is responsible for dynamically scheduling the number of executors of the algorithm module in the server according to the server status and request load, and supporting the operation of the algorithm module; The algorithm module is responsible for completing the image set and a collection of text paragraphs Detect illegal content and generate illegal content descriptions and amendments to form illegal content collection process.

Images