Official document verification method and system based on artificial intelligence

Abstract

The invention discloses an official document verification method and system based on artificial intelligence, and relates to the technical field of document verification, and the method comprises the steps: obtaining a to-be-verified official document, converting the to-be-verified official document into a structured format, and recognizing a text segment and an image region; language feature analysis is carried out based on the text segments, paragraph features are constructed, curvature response in document structure evolution is calculated in combination with the paragraph features, abnormal judgment of the text segments is carried out, and an abnormal text segment set is obtained; carrying out backtracking based on the abnormal text fragments, extracting corresponding image areas to carry out image-text consistency verification to obtain an image-text abnormal pair set, and constructing a proposition and responsibility subject set to carry out knowledge affiliation relationship and responsibility conflict detection to obtain an abnormal proposition set; according to the method, a whole-process artificial intelligence verification system of structure detection, image-text verification, responsibility judgment and abnormity traceability is constructed, and the accuracy and reliability of official document verification can be greatly improved.

Description

Technical Field

[0001] This invention relates to the field of document verification technology, and in particular to an artificial intelligence-based method and system for verifying official documents. Background Technology

[0002] With the widespread adoption of information technology in government offices, enterprise and public institution management, and industry standardization, electronic documents have gradually become the mainstream document carrier. In recent years, the rapid development of artificial intelligence technologies, particularly natural language processing, computer vision, document layout analysis, dependency parsing, semantic embedding models, and topic models, has led to significant progress in automated document processing, intelligent proofreading, and text consistency assessment. In the field of document understanding, BERT pre-trained language models, LDA topic models, dependency syntax tree generation tools, and visual Transformers are widely used for tasks such as text structure parsing, semantic understanding, and image recognition. Meanwhile, with the construction of various document management platforms and digital government systems, using APIs to obtain documents to be processed and combining them with DocumentObjectModel for document structure transformation has become one of the industry standard processes. Against this backdrop, both academia and industry are committed to building a unified document parsing framework to achieve automated verification and risk detection of electronic documents through multimodal input, feature extraction, and intelligent reasoning.

[0003] Existing technologies lack the ability to quantify the evolution of document paragraph structure, cannot capture structural breaks and semantic drift, and only perform alignment based on feature similarity without conducting bidirectional alignment and deep correlation analysis of text context semantic embedding and local image region embedding. This makes it difficult to identify tampering, mismatch, or inconsistencies between text and image content in documents. Furthermore, existing technologies rely heavily on manual rules and cannot provide structured propositional expressions for behaviors such as "unit approval" and "personnel signature." They also lack logical inference mechanisms to determine whether the responsible party truly knows or approves the validity of the proposition. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an artificial intelligence-based document verification method and system to solve the problems of existing technologies being unable to capture structural breaks and semantic drift, having difficulty identifying tampering, mismatches, or inconsistencies between text and graphics in documents, and lacking a logical inference mechanism to determine whether the responsible party truly knows or acknowledges the validity of the proposition.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides an artificial intelligence-based document verification method, which includes, The document to be verified is obtained and converted into a structured format, and text fragments and image regions are identified. Based on text fragments, language feature analysis is performed and paragraph features are constructed. Combining the paragraph features, the curvature response in the document structure evolution is calculated to identify anomalies in the text fragments, resulting in a set of anomalous text fragments. Backtracking is performed based on abnormal text fragments, and corresponding image regions are extracted for image-text consistency verification to obtain a set of abnormal image-text pairs. Then, a set of propositions and responsible entities is constructed to detect knowledge attribution relationships and responsibility conflicts, resulting in a set of abnormal propositions. All sets are uniformly constructed into an abnormal node set, and an abnormal source cause-effect graph is generated. After the correlation strength between abnormal nodes in the graph is calculated to confirm the main abnormal source chain, the abnormal score is calculated according to the abnormal node type and the weighted sum is obtained to get the total score. The official document is verified based on the total score, and the verification conclusion is stored.

[0007] As a preferred embodiment of the AI-based document verification method of the present invention, the following steps are performed: Language feature analysis is conducted on text fragments to construct paragraph features; the curvature response in the document structure evolution is calculated based on the paragraph features to determine anomalies in the text fragments, resulting in a set of abnormal text fragments. Specifically, a dependency tree is built on the text fragments using dependency parsing tools, and the average depth of the path from the root node to all leaf nodes in the dependency tree is calculated. A pre-trained BERT-base model is used to obtain the context word vector embeddings of the text fragments, and the lexical alignment between adjacent context word vector embeddings is calculated. A pre-trained LDA model is applied to obtain the topic probability distribution vectors of the current text fragment and the previous text fragment, and KL divergence is used to measure the difference between the topic probability distribution vectors. Then, a paragraph feature function is constructed based on the average depth, lexical alignment, and difference. The paragraph feature values ​​are calculated and sorted in descending order. Using the sorted sequence of paragraph feature values, the curvature response of each text fragment in the document structure evolution is constructed, and anomaly determination of the current text fragment is performed based on the curvature response, resulting in a set of abnormal text fragments.

[0008] As a preferred embodiment of the document verification method based on artificial intelligence described in this invention, the following steps are performed: backtracking based on abnormal text fragments, extracting corresponding image regions for image-text consistency verification, obtaining a set of abnormal image-text pairs, constructing a set of propositions and responsible entities for knowledge attribution relationships and responsibility conflicts, obtaining a set of abnormal propositions; backtracking is performed on each abnormal point based on the set of abnormal text fragments, extracting context word vector embeddings and corresponding or adjacent image regions, loading a pre-trained VisionTransformer model in the Timm library, using the image regions as input to the VisionTransformer model, outputting the embedding vectors of the image regions, performing alignment operations in conjunction with context word vector embeddings, calculating semantic similarity, setting a consistency threshold, and performing image-text anomaly labeling in conjunction with semantic similarity, obtaining a set of abnormal image-text pairs; Based on the set of image-text anomalies, text fragments are extracted from the set. Named entity recognition is used to identify entities in the text fragments, including units, personnel, and organizations. Then, dependency parsing is used to analyze the predicates in the text fragments, including approval, signing, and stamping, to construct a set of propositions and a set of responsible entities. The system queries the current text segment for explicit statements signed by the responsible party. If such statements exist, the responsible party is considered aware that the current proposition is true; otherwise, it is considered unknown. The true and unknown are encoded as Boolean values. Based on the Boolean values, a logical formula is established to determine the knowledge attribution relationship between each proposition and the group of responsible parties. If the logical formula for the knowledge attribution relationship between the proposition and any of the responsible parties in the group conflicts, the current proposition is considered an abnormal proposition, and a set of abnormal propositions is generated; otherwise, it is considered normal.

[0009] As a preferred embodiment of the document verification method based on artificial intelligence described in this invention, the step of uniformly constructing each set into an abnormal node set and generating an abnormal causal graph refers to constructing a unified abnormal node set based on the abnormal text fragment set, the image-text abnormal pair set, and the abnormal proposition set. The abnormal node set is traversed, and when there is a pairing between a text fragment and an image region in the set, an edge is established. When the text fragment contains a proposition, an edge is established. When there is a conflict between the proposition and the responsible party, an attribution edge is established. Combined with the abnormal nodes, an abnormal causal graph is generated.

[0010] As a preferred embodiment of the document verification method based on artificial intelligence described in this invention, the following steps are performed: after confirming the main anomaly tracing chain by calculating the correlation strength between abnormal nodes in the graph, anomaly scores are calculated according to the abnormal node types and weighted summation is performed to obtain the total score. Based on the total score, the document verification is performed by traversing the anomaly tracing cause-effect graph, querying the graph from the initial abnormal node, tracing back the cause-effect chain layer by layer along the connection path, generating an abnormal path set, and calculating the correlation strength between abnormal nodes in each abnormal path as the weight of the abnormal path. Based on the weight, the cumulative influence strength of any abnormal path is calculated. The cumulative influence intensity is sorted in descending order to obtain a sorted list. The abnormal path with the largest cumulative influence intensity is marked as the main abnormal source chain. The starting node, the involved graphics and text, and the total number of propositional abnormal nodes are recorded. Based on the main abnormal source chain, the abnormal score of the abnormal node is calculated. Based on the anomaly score, the total score is calculated, and then the document is verified for compliance. If the total score is greater than or equal to the compliance threshold, the document is considered to be a seriously abnormal document, and a main anomaly path graph is generated. The entire document is then marked with a warning label of "structural imbalance - unclear responsibility". Otherwise, the document is considered to have passed the verification.

[0011] As a preferred embodiment of the document verification method based on artificial intelligence described in this invention, the step of storing the obtained verification conclusion refers to transmitting the verification conclusion and the total score to the database via wireless transmission technology, creating a folder for the verification conclusion and the total score in the database, adding the transmitted verification conclusion and the total score to the folder, and giving the folder a name consistent with the current document.

[0012] As a preferred embodiment of the AI-based document verification method of the present invention, the steps of obtaining the document to be verified and converting it into a structured format, and identifying text fragments and image regions refer to using an API interface to obtain the document to be verified from the document management platform, applying the DocumentObject Model method to convert the document to be verified into a structured document object, and using text region detection algorithms and layout analysis algorithms to perform text region detection and image region recognition on the structured document object to obtain text fragments and image regions.

[0013] Secondly, the present invention provides an artificial intelligence-based document verification system, comprising, The document recognition module acquires the document to be verified, converts it into a structured format, and recognizes text fragments and image regions. A judgment module is constructed to perform language feature analysis on text fragments and construct paragraph features. Combining the paragraph features, the curvature response in the document structure evolution is calculated to judge the anomalies of text fragments, and a set of abnormal text fragments is obtained. The backtracking detection module backtracks based on abnormal text fragments, extracts the corresponding image regions for image-text consistency verification, obtains a set of abnormal image-text pairs, constructs a set of propositions and responsible entities, and performs knowledge attribution relationship and responsibility conflict detection to obtain a set of abnormal propositions; The calculation and verification module unifies all sets into an abnormal node set and generates an abnormal source cause-effect graph. After confirming the main abnormal source chain by calculating the correlation strength between abnormal nodes in the graph, it calculates the abnormal score according to the abnormal node type and calculates the weighted sum to obtain the total score. Based on the total score, it verifies the official document and obtains the verification conclusion for storage.

[0014] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the artificial intelligence-based document verification method described in the first aspect of the present invention.

[0015] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the artificial intelligence-based document verification method described in the first aspect of the present invention.

[0016] The beneficial effects of this invention are as follows: By constructing paragraph feature functions through dependency parsing, contextual word vector embedding, and topic distribution changes, and combining them with generalized fractional curvature response, the structural evolution trend of official documents can be quantified, thereby accurately identifying structural abrupt changes, semantic jumps, and abnormal layout paragraphs, significantly improving the ability to detect structural anomalies. Furthermore, by using BERT semantic embedding and ViT image embedding to calculate the semantic similarity between text fragments and image regions, contradictions or forgeries between image content such as seals, signatures, and closings and the main text description can be effectively identified, achieving highly reliable image-text consistency verification. Secondly, by constructing proposition sets and responsibility subject sets through named entity recognition and responsibility expression parsing, the authenticity and consistency of responsibility statements are judged using logical attribution formulas, and further, the main anomaly tracing chain is inferred through anomaly causal graphs, achieving path-level tracking and risk scoring of anomaly sources. Therefore, this invention not only constructs a full-process artificial intelligence verification system of "structural detection—image-text verification—responsibility judgment—anomaly tracing," but also significantly improves the accuracy and reliability of official document verification. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of the AI-based document verification method in Example 1.

[0019] Figure 2 This is a structural diagram of the AI-based document verification system in Example 1.

[0020] Figure 3 This is a flowchart of the document verification process in Example 1. Detailed Implementation

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0022] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0023] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0024] Example 1, referring to Figures 1-3 This is the first embodiment of the present invention, which provides an artificial intelligence-based document verification method, including the following steps: S1. Obtain the official document to be verified and convert it into a structured format, then identify text fragments and image regions; Specifically, obtaining the document to be verified and converting it into a structured format, and identifying text fragments and image regions refer to using the API interface to obtain the document to be verified from the document management platform, applying the DocumentObjectModel method to convert the document to be verified into a structured document object, and using text region detection algorithms and layout analysis algorithms to perform text region detection and image region recognition on the structured document object to obtain text fragments and image regions.

[0025] By acquiring official documents through API interfaces and combining DOM structure parsing, text region detection, and layout analysis algorithms, structured modeling and multimodal element extraction of the documents were achieved. This processing flow not only improves the system's adaptability to unstructured documents and diverse document formats but also provides a high-quality input foundation for subsequent processing. The construction of structured document objects gives text, images, and other elements clear boundaries and hierarchical relationships, which helps to accurately identify key content such as titles, body text, attachments, and signature areas in the document.

[0026] S2. Perform language feature analysis on text fragments and construct paragraph features. Combine the paragraph features to calculate the curvature response in the document structure evolution to judge the anomalies of text fragments and obtain a set of abnormal text fragments. Specifically, based on text fragments, language feature analysis is performed to construct paragraph features. Combining these paragraph features, the curvature response in document structure evolution is calculated to identify anomalies in the text fragments, resulting in a set of anomalous text fragments. Furthermore, based on the text fragments, a dependency tree is built using a dependency parsing tool (Stanza dependency parser), and the average depth of the path from the root node to all leaf nodes is calculated. A pre-trained BERT-base model is used to obtain the context word embeddings of the text fragments, and the lexical alignment between adjacent context word embeddings is calculated. A pre-trained LDA (Latent Dirichlet Allocation) model is applied to obtain the topic probability distribution vectors of the current text fragment and the previous text fragment. KL divergence is used to measure the difference between the topic probability distribution vectors. Then, based on the average depth, lexical alignment, and difference, a paragraph feature function is constructed. The paragraph feature values ​​are calculated and sorted in descending order. Using the sorted sequence of paragraph feature values, the curvature response of each text fragment in document structure evolution is constructed, and anomaly detection of the current text fragment is performed based on the curvature response, resulting in a set of anomalous text fragments. The average depth of the path from the root node to all leaf nodes in the dependency tree is calculated using the following formula: In the formula, Represents a text fragment average depth, Represents a text fragment The total number of paths from the root node to each leaf node. Indicates the first The depth value of the path; The formula for calculating the lexical alignment between adjacent context word vector embeddings is as follows: In the formula, Represents a text fragment The Middle Each context word vector Represents a text fragment The Middle Each context word vector Describing the L2 norm, Represents a text fragment Lexical alignment Indicates the total number of words; The KL divergence metric is used to measure the differences between the probability distribution vectors of topics, and the formula is as follows: In the formula, Represents a text fragment Compared to the previous text fragment Differences, This represents the total number of topic probability distribution vectors. Represents a text fragment In the The probability distribution vector over each topic Represents the natural logarithm function. Indicates the previous text segment In the Probability distribution vectors over each topic; The formula for calculating the paragraph feature value is as follows: In the formula, Represents a text fragment In sequence position Paragraph feature values, The weight representing the average depth, Weights representing lexical alignment. Weights representing differences; , , The range of values ​​can be set by using grid search and cross-validation in the official document dataset. For example, , , average depth Syntax mutations can serve as the primary trigger for structural mutations, and they best reflect structural discontinuities. Primarily used to capture changes in sentence order. It can supplement the disclosure of cross-paragraph transitions, but it will not determine the anomaly on its own. After taking the value, you need to ensure that the sum of the weights is 1. The formula for constructing the curvature response of each text fragment in the document structure evolution is as follows: In the formula, Represents a text fragment Curvature response in document structure evolution Indicates a text fragment In sequence position First derivative of the above, Represents the memory kernel function, Represents the integral variable; The specific form of the memory kernel function is as follows: In the formula, Represents the time difference variable. This represents the power exponent (the power exponent formula ensures convergence of the integral and moderate memory for changes in document structure, therefore...). The larger the value, the stronger the influence of past content. The smaller the size, the smaller the impact; however, changes in document structure have a significant local delay, therefore... The value of can be set experimentally as follows: , and When the value is less than 0.3, the impact of abnormal segments decays rapidly, leading to a decrease in the recognition rate. A value greater than 0.8 can cause interference from irrelevant paragraphs. The base of the natural logarithm. The exponential decay factor (which can be experimentally set to ensure non-local memory control of the execution item, therefore it must be greater than 0) represents the exponential decay factor. Since the average document structure spans 2-20 paragraphs, the decay must meet two conditions: first, the influence weight of adjacent paragraphs is approximately 60-80%; second, the influence of paragraphs separated by more than 10 paragraphs decreases to 5-10%. Under these constraints, the range of values ​​for the exponential decay factor can be set as follows: ); The method of performing anomaly judgment on the current text segment based on curvature response refers to setting an anomaly threshold through experiments and relevant book knowledge. When the curvature response is greater than the anomaly threshold, it indicates that the text segment has significant structural discontinuity and is represented as an anomaly point; otherwise, it is marked as a normal point.

[0027] This invention constructs a mapping channel from the grammatical layer to the logical structure layer by introducing an average depth metric. This metric is not a shallow statistical feature, but a highly structure-sensitive logical flow surface quantification metric, capturing "potential syntactic mutations" rather than explicit errors. Unlike existing "grammatical error detectors" or "rule engines," this design establishes a basic spatial manifold for subsequent structural mutation responses. Furthermore, lexical alignment is not merely the dot product of context word vectors; it reflects the dynamic continuity of intra-sentence ordering tension. Current NLP models generally ignore "tension mutations caused by word order perturbations within the same segment." This invention constructs this metric through continuous vector field changes, providing a differentiable metric basis for determining "abnormal intra-sentence recombination." Secondly, the KL divergence between topic probability distributions... The introduction of this feature allows it to be incorporated as a parameter of the paragraph evolution function into the curvature response model, rather than being used alone or for topic classification. This design combines three-dimensional information of "syntax-word order-semantics" in structural logic mutation detection, which is significantly different from traditional general modeling. The generation of the structural evolution curvature response function simulates the dynamic delayed propagation of "potential influence of preceding text on following text" in official document writing—a behavior with a strong realistic basis in government affairs, approval, and review documents, representing "non-Markovianness in text composition." Furthermore, this step differs from traditional LSTM or Transformer structures: the latter can only handle "current context" or "attention weight" relationships, but cannot construct a mutation response index under the macroscopic structural evolution trajectory. This invention, by constructing a mechanism of "integral curvature + differential derivation + non-local memory regulation," realizes the location and quantification of structural curvature mutation points in documents. Moreover, the curvature response function contains a memory kernel function, which has the biological neural synaptic decremental memory characteristics, ensuring high responsiveness to neighboring paragraphs while avoiding long-distance interference.

[0028] S3. Based on the abnormal text fragments, backtrack and extract the corresponding image regions for image-text consistency verification to obtain a set of abnormal image-text pairs. Construct a set of propositions and responsible entities to detect knowledge attribution relationships and responsibility conflicts, and obtain a set of abnormal propositions. Specifically, backtracking is performed based on abnormal text fragments, corresponding image regions are extracted for image-text consistency verification, resulting in a set of abnormal image-text pairs. A set of propositions and responsible entities is constructed for knowledge attribution and responsibility conflict detection, resulting in a set of abnormal propositions. Then, based on the set of abnormal text fragments, backtracking is performed on each abnormal point, extracting context word vector embeddings and corresponding or adjacent image regions. A pre-trained VisionTransformer (ViT) model is loaded into the Timm library, with the image regions used as input. The output is the embedding vector of the image regions. After performing alignment operations with the context word vector embeddings, semantic similarity is calculated. A consistency threshold is set based on empirical knowledge, and image-text anomaly labeling is performed based on semantic similarity, resulting in a set of abnormal image-text pairs. The formula for calculating semantic similarity is: In the formula, Indicates the first The context word vector embedding and the first The semantic similarity of each embedded vector. Describing the L2 norm, Indicates the first Context word vector embeddings, Indicates the first Embedded vectors; The method of combining semantic similarity to perform image-text anomaly labeling means that when the semantic similarity is less than the consistency threshold, the current embedding vector and the context word vector embedding are treated as a data pair and labeled as an image-text anomaly pair; otherwise, they are labeled as normal. Based on the set of text-image anomalies, text fragments are extracted from the set. Named Entity Recognition (NER) is used to identify entities in the text fragments, including units, personnel, and organizations. Then, dependency parsing is used to analyze the predicates in the text fragments, including approval, signing, and stamping. A set of propositions (each proposition is a responsibility statement such as approval, stamping, and issuance expressed by units, personnel, or organizations, such as "This document has been approved by the unit" or "It has been signed by the person in charge") and a set of responsible entities (each responsible entity is the implementing entity mentioned in the document, including the writer, the person who stamps, and the competent unit, etc.) are constructed. If a clearly stated statement signed by the responsible party appears in the current text fragment, it is assumed that the responsible party is aware that the current proposition is true; otherwise, it is considered unknown. The true and unknown are encoded as Boolean values ​​(i.e., true is represented by a Boolean value of 1, and unknown is represented by a Boolean value of 0). Examples of explicit statements are shown in Table 1: Table 1: Examples of Explicitly Expressed Statements Based on Boolean values, a logical formula is established for the knowledge attribution relationship between each proposition and the responsible subjects in the group. If the logical formula for the knowledge attribution relationship between the proposition and any responsible subject in the group conflicts (i.e. the proposition is claimed to be true, but the logical formula is 0), then the current proposition is regarded as an abnormal proposition and a set of abnormal propositions is generated; otherwise, it is regarded as normal. The logical formula for establishing the knowledge attribution relationship between each proposition and the group of responsible subjects is as follows: In the formula, Indicating the subject of collective responsibility (i.e., the set of responsible parties) knows the first The validity of this proposition Represents logical operators. Indicates the total number of responsible parties. Indicates the first Did the responsible party know about the first...? A Boolean value representing the validity of a proposition.

[0029] This invention employs a double-nested semantic space alignment mechanism, combining contextual word vectors (such as BERT embeddings) with a visual VisionTransformer (ViT) model to generate image semantic vectors. It judges image-text consistency based on continuous spatial semantic similarity rather than text overlap or keyword matching. This step enables robust identification of image-text mismatches even in scenarios with blurred seals, distorted signature images, and ambiguous text descriptions. Furthermore, this invention combines propositional structure modeling with responsible subject reasoning, proposing a formal mechanism for a "semantic proposition-responsibility awareness" logical Boolean graph. Specifically, it constructs a proposition set by extracting subject-verb-object structures using dependency syntax, and then combines this with the subject set identified by NER. A logical formula is used to determine whether "the group is aware that the proposition is true." This step introduces a group Boolean consistency constraint for knowledge attribution and automatically pairs it with explicitly expressed statements in the text. Unlike conventional entity recognition, which only extracts "whether it is mentioned," this invention focuses on "whether it constitutes an intention to confirm responsibility." For example, the relationship between the NER entities "This document was signed by the ×× office" and "×× office" is not a simple occurrence, but rather forms a three-element behavioral chain of "subject-responsibility verb-object". If this chain lacks explicit expression, it forms a cognitive vulnerability-type anomalous proposition. Therefore, by superimposing three types of features—text-image anomalies, proposition-responsibility conflicts, and logical Boolean relationships—the system ultimately achieves a leap from "format and content verification" to "semantic consistency and responsibility closed-loop reasoning" in document verification.

[0030] S4. Construct all sets into a unified set of abnormal nodes and generate an abnormal source cause-effect graph. Calculate the correlation strength between abnormal nodes in the graph to confirm the main abnormal source chain. Calculate the abnormal score according to the abnormal node type and sum the scores by weight to obtain the total score. Verify the official documents based on the total score and store the verification conclusion. Specifically, each set is uniformly constructed into an abnormal node set, and an abnormal causal graph is generated. Based on the abnormal text fragment set, the image-text abnormal pair set, and the abnormal proposition set, the abnormal data in each set is constructed into a unified abnormal node set. The abnormal node set is traversed. When there is a pair between a text fragment and an image region in the set, an edge is established. When the text fragment contains a proposition, an edge is established. When there is a conflict between the proposition and the responsible party, an attribution edge is established. Combined with the abnormal nodes, an abnormal causal graph is generated.

[0031] Traditional document verification often focuses on detecting local text or image-text matching errors. This invention introduces an "anomaly graph model," which integrates atomic-level anomalies into a global chain of risks, significantly improving the ability to analyze the logical consistency of documents. Through the directional edge design in the graph structure, the system can trace the causal chain of anomalies layer by layer, accurately pinpointing the "initial anomaly occurrence point" and the "final point of responsibility," providing an interpretable basis for accountability and audit compliance. Because spatial location information, entity coreference identification results, and logical conflict judgment logic are incorporated into the graph construction process, it can handle documents with complex structures such as complex layouts, multi-page text and images, and overlapping seals and signatures. The anomaly graph structure provides an information topology foundation for subsequent anomaly node scoring and path scoring, allowing scoring to go beyond local scoring and to "propagate anomaly weights" based on structural influence and path propagation, making the scoring results more risk-sensitive. The anomaly graph itself is a natural visual representation, which can be used in conjunction with the review system to output graphical paths, assisting manual personnel in quickly locating anomaly chains and improving document review efficiency.

[0032] Furthermore, after confirming the main anomaly tracing chain by calculating the correlation strength between anomalous nodes in the graph, anomaly scores are calculated based on the anomalous node types and weighted summed to obtain the total score. Based on the total score, the official document is verified by traversing the anomaly tracing cause-effect graph. Starting from the initially occurring anomalous node, the cause-effect chain is traced back layer by layer along the connection path, generating a set of anomalous paths. The correlation strength between anomalous nodes in each anomalous path is calculated as the weight of the anomalous path, using the following formula: In the formula, Indicates abnormal nodes and The strength of the association, This represents the Sigmoid activation function. This represents a multilayer perceptron. This indicates a splicing operation. and Indicates abnormal nodes and The corresponding vector, Indicates abnormal nodes and The display structure characteristics between nodes (i.e., whether two nodes are on the same page in the document, encoded as a boolean value: 1 for yes, 0 for no); The multilayer perceptron uses the cross-entropy loss function as the loss function (the label of the cross-entropy loss function can be based on expert annotation of the real causal path and the definition of whether the edge is generated by strong correlation), and uses the Adam optimizer for iterative optimization. When the decrease value of the loss function no longer decreases significantly during continuous iteration, the iteration stops and the trained multilayer perceptron is output. Based on the weights, the cumulative influence strength of any abnormal path is calculated using the following formula: In the formula, Indicates the first abnormal path The cumulative strength of influence, Indicates belonging to, Indicates the first An abnormal path; The cumulative influence intensity is sorted in descending order to obtain a sorted list. After marking the abnormal path with the largest cumulative influence intensity as the main abnormal source chain, the starting node, the involved graphics and text, and the total number of propositional abnormal nodes are recorded. Based on the main anomaly tracing chain, the anomaly score of the anomaly node is calculated using the following formula: In the formula, Indicates abnormal nodes Abnormal scores, Indicates abnormal nodes For text fragments Abnormal scores at that time and Let represent the mean and standard deviation of the curvature response, respectively. Indicates abnormal nodes For the first Anomaly score for each image / text anomaly. Indicates abnormal nodes This refers to abnormal scores during the question setting process. Indicates abnormal nodes When proposing a proposition, the Boolean value 1 is taken from the unknown code. This indicates the operation of finding the minimum value. This indicates the operation of retrieving the maximum value; Based on the anomaly score, the total score is calculated, and then the document is verified for compliance. When the total score is greater than or equal to the compliance threshold (the compliance threshold can be set through experience), the current document is considered to be a seriously abnormal document, and a main anomaly path map is generated. Then, the entire document is marked with a warning label of "structural imbalance - unclear responsibility". Otherwise, the current document is considered to have passed the verification. The formula for calculating the total score is as follows: In the formula, This represents the total score of the main anomaly attribution chain. This indicates the total number of abnormal nodes in the main anomaly tracing chain.

[0033] By employing graph neural network logic modeling to traverse and reason about the "anomaly tracing causal graph," an anomaly path weight modeling mechanism for nodes with mixed structural, semantic, and logical characteristics is provided. Among these, the multilayer perceptron (MLP), as a feedforward neural network structure, can effectively model the nonlinear relationships between node vectors. In this invention, the MLP is used to model the semantic and structural coupling between anomaly nodes, and the "Sigmoid" activation function is used to normalize the MLP output to the [0,1] interval, reflecting the "association strength" between anomaly nodes. Structural Features By using Boolean encoding (whether on the same page) to represent the visual adjacency between nodes, a lightweight "visual field consistency" judgment criterion is formed, greatly enhancing the model's adaptability to the logical structure of official document layouts. Based on this, expert-annotated real causal paths are used as supervision signals. An MLP model is trained using a cross-entropy loss function and combined with an Adam optimizer for efficient iteration, generating a generalizable node connection strength estimation mechanism. This mechanism not only improves the model's ability to fit logical paths between anomalous nodes but also adaptively discovers the most meaningful anomalous propagation chains without explicitly defining rules. Secondly, this invention calculates the cumulative influence strength of each anomalous path based on the strength between nodes and uses the path with the highest value as the "main anomalous source chain." The formation of this chain not only has diagnostic significance but also provides a basis for path-level verification decisions. Based on the main anomaly chain, a multi-source anomaly scoring function was designed. The score for structural anomaly nodes is calculated by normalizing with the mean and standard deviation, exhibiting good stability and discriminative power. The score for image-text anomaly pairs is based on a reverse transformation of semantic similarity. Anomalies related to propositional responsibility are directly assigned a Boolean constant value of 1, strengthening their weight in the scoring and highlighting the severity of logical conflicts of responsibility. This overall process ultimately constitutes the core mechanism for document verification decision-making. The total score is calculated by averaging the scores on the main anomaly chain, ensuring the scoring system possesses both path dependence and global aggregation characteristics. When the total score exceeds a set threshold, the system marks the document as "structural imbalance—unclear responsibility" and outputs a main anomaly path graph for regulatory or review use, achieving a closed-loop process from detection to analysis, and from judgment to location.

[0034] Furthermore, storing the verification results involves transmitting the verification results and total score to the database via wireless transmission technology, creating a folder for the verification results and total score in the database, adding the transmitted verification results and total score to the folder, and giving the folder a name consistent with the current document.

[0035] By introducing wireless transmission and a structured database archiving and naming mechanism, a highly efficient and traceable management system for official document verification results has been established. Verification conclusions and total scores can be transmitted to the database wirelessly in real time, improving data synchronization efficiency and system flexibility, and adapting to multi-terminal, distributed office scenarios. The system automatically creates folders in the database with names consistent with the original documents and categorizes and stores the verification results, ensuring one-to-one data correspondence, clear logic, and easy retrieval and historical review.

[0036] This embodiment also provides an artificial intelligence-based document verification system, including: The document recognition module acquires the document to be verified, converts it into a structured format, and recognizes text fragments and image regions. A judgment module is constructed to perform language feature analysis on text fragments and construct paragraph features. Combining the paragraph features, the curvature response in the document structure evolution is calculated to judge the anomalies of text fragments, and a set of abnormal text fragments is obtained. The backtracking detection module backtracks based on abnormal text fragments, extracts the corresponding image regions for image-text consistency verification, obtains a set of abnormal image-text pairs, constructs a set of propositions and responsible entities, and performs knowledge attribution relationship and responsibility conflict detection to obtain a set of abnormal propositions; The calculation and verification module unifies all sets into an abnormal node set and generates an abnormal source cause-effect graph. After confirming the main abnormal source chain by calculating the correlation strength between abnormal nodes in the graph, it calculates the abnormal score according to the abnormal node type and calculates the weighted sum to obtain the total score. Based on the total score, it verifies the official document and obtains the verification conclusion for storage.

[0037] This embodiment also provides a computer device applicable to the artificial intelligence-based document verification method, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the artificial intelligence-based document verification method proposed in the above embodiment. The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0038] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the document verification method based on artificial intelligence as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0039] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A document verification method based on artificial intelligence, characterized in that: include, The document to be verified is obtained and converted into a structured format, and text fragments and image regions are identified. Based on text fragments, language feature analysis is performed and paragraph features are constructed. Combining the paragraph features, the curvature response in the document structure evolution is calculated to identify anomalies in the text fragments, resulting in a set of anomalous text fragments. Backtracking is performed based on abnormal text fragments, and corresponding image regions are extracted for image-text consistency verification to obtain a set of abnormal image-text pairs. Then, a set of propositions and responsible entities is constructed to detect knowledge attribution relationships and responsibility conflicts, resulting in a set of abnormal propositions. All sets are uniformly constructed into an abnormal node set, and an abnormal source cause-effect graph is generated. After the correlation strength between abnormal nodes in the graph is calculated to confirm the main abnormal source chain, the abnormal score is calculated according to the abnormal node type and the weighted sum is obtained to get the total score. The official document is verified based on the total score, and the verification conclusion is stored.

2. The document verification method based on artificial intelligence as described in claim 1, characterized in that: The process involves analyzing language features based on text fragments and constructing paragraph features. Combining these paragraph features with the curvature response in document structure evolution is then used to determine anomalies in the text fragments, resulting in a set of anomalous text fragments. Specifically, based on the text fragments, a dependency parsing tool is used to build a dependency tree, and the average depth of the path from the root node to all leaf nodes is calculated. A pre-trained BERT-base model is used to obtain the context word embeddings of the text fragments, and the lexical alignment between adjacent context word embeddings is calculated. A pre-trained LDA model is applied to obtain the topic probability distribution vectors of the current and previous text fragments, and KL divergence is used to measure the difference between the topic probability distribution vectors. A paragraph feature function is then constructed based on the average depth, lexical alignment, and difference. The paragraph feature values ​​are calculated and sorted in descending order. Using the sorted sequence of paragraph feature values, the curvature response of each text fragment in document structure evolution is constructed, and anomaly determination of the current text fragment is performed based on the curvature response, resulting in a set of anomalous text fragments.

3. The document verification method based on artificial intelligence as described in claim 2, characterized in that: The process involves backtracking based on abnormal text fragments, extracting corresponding image regions for image-text consistency verification, obtaining a set of abnormal image-text pairs, constructing a set of propositions and responsible entities for knowledge attribution and responsibility conflict detection, and obtaining a set of abnormal propositions. Based on the set of abnormal text fragments, backtracking is performed on each abnormal point, extracting context word vector embeddings and corresponding or adjacent image regions. A pre-trained VisionTransformer model is loaded into the Timm library, using the image regions as input to the VisionTransformer model, outputting the image region embedding vectors. After performing alignment operations with the context word vector embeddings, semantic similarity is calculated, a consistency threshold is set, and image-text anomaly labeling is performed based on semantic similarity, resulting in a set of abnormal image-text pairs. Based on the set of image-text anomalies, text fragments are extracted from the set. Named entity recognition is used to identify entities in the text fragments, including units, personnel, and organizations. Then, dependency parsing is used to analyze the predicates in the text fragments, including approval, signing, and stamping, to construct a set of propositions and a set of responsible entities. The system queries the current text segment for explicit statements signed by the responsible party. If such statements exist, the responsible party is considered aware that the current proposition is true; otherwise, it is considered unknown. The true and unknown are encoded as Boolean values. Based on the Boolean values, a logical formula is established to determine the knowledge attribution relationship between each proposition and the group of responsible parties. If the logical formula for the knowledge attribution relationship between the proposition and any of the responsible parties in the group conflicts, the current proposition is considered an abnormal proposition, and a set of abnormal propositions is generated; otherwise, it is considered normal.

4. The document verification method based on artificial intelligence as described in claim 3, characterized in that: The process of unifying all sets into an anomaly node set and generating an anomaly tracing cause-effect graph refers to constructing a unified anomaly node set based on the anomaly text fragment set, the image-text anomaly pair set, and the anomaly proposition set. The anomaly node set is traversed, and if there is a pairing between a text fragment and an image region in the set, an edge is established. If the text fragment contains a proposition, an edge is established. If there is a conflict between the proposition and the responsible party, an attribution edge is established. Combined with the anomaly nodes, an anomaly tracing cause-effect graph is generated.

5. The document verification method based on artificial intelligence as described in claim 4, characterized in that: After confirming the main anomaly tracing chain by the correlation strength between abnormal nodes in the computation graph, the anomaly score is calculated according to the anomaly node type and the total score is obtained by weighted summation. Based on the total score, the document verification is performed by traversing the anomaly tracing cause-effect graph. Starting from the initial abnormal node, the cause-effect chain is traced back layer by layer along the connection path to generate an anomaly path set. The correlation strength between abnormal nodes in each anomaly path is calculated as the weight of the anomaly path. Based on the weight, the cumulative influence strength of any anomaly path is calculated. The cumulative influence intensity is sorted in descending order to obtain a sorted list. The abnormal path with the largest cumulative influence intensity is marked as the main abnormal source chain. The starting node, the involved graphics and text, and the total number of propositional abnormal nodes are recorded. Based on the main abnormal source chain, the abnormal score of the abnormal node is calculated. Based on the anomaly score, the total score is calculated, and then the document is verified for compliance. If the total score is greater than or equal to the compliance threshold, the document is considered a seriously abnormal document, and a main anomaly path graph is generated. The entire document is then marked with a warning label of "structural imbalance - unclear responsibility". Otherwise, the document is considered to have passed the verification.

6. The document verification method based on artificial intelligence as described in claim 5, characterized in that: The process of storing the verification conclusion refers to transmitting the verification conclusion and the total score to the database via wireless transmission technology, creating a folder for the verification conclusion and the total score in the database, adding the transmitted verification conclusion and the total score to the folder, and giving the folder a name consistent with the current document.

7. The document verification method based on artificial intelligence as described in claim 6, characterized in that: The process of obtaining the document to be verified and converting it into a structured format, and identifying text fragments and image regions, refers to using an API interface to obtain the document to be verified from the document management platform, applying the DocumentObject Model method to convert the document to be verified into a structured document object, and using text region detection algorithms and layout analysis algorithms to perform text region detection and image region recognition on the structured document object to obtain text fragments and image regions.

8. An artificial intelligence-based document verification system, based on the artificial intelligence-based document verification method according to any one of claims 1 to 7, characterized in that: include, The document recognition module acquires the document to be verified, converts it into a structured format, and recognizes text fragments and image regions. A judgment module is constructed to perform language feature analysis on text fragments and construct paragraph features. Combining the paragraph features, the curvature response in the document structure evolution is calculated to judge the anomalies of text fragments, and a set of abnormal text fragments is obtained. The backtracking detection module backtracks based on abnormal text fragments, extracts the corresponding image regions for image-text consistency verification, obtains a set of abnormal image-text pairs, constructs a set of propositions and responsible entities, and performs knowledge attribution relationship and responsibility conflict detection to obtain a set of abnormal propositions; The calculation and verification module unifies all sets into an abnormal node set and generates an abnormal source cause-effect graph. After confirming the main abnormal source chain by calculating the correlation strength between abnormal nodes in the graph, it calculates the abnormal score according to the abnormal node type and calculates the weighted sum to obtain the total score. Based on the total score, it verifies the official document and obtains the verification conclusion for storage.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the document verification method based on artificial intelligence as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the document verification method based on artificial intelligence as described in any one of claims 1 to 7.

Images