A method for extracting enterprise audit information based on artificial intelligence

Abstract

The present invention discloses an enterprise audit information extraction method based on artificial intelligence, which relates to the field of enterprise audit information processing technology. The method comprises the following steps: obtaining original audit documents from an enterprise and performing multimodal conversion processing to generate a standard data set; performing deep semantic analysis on the standard data set to construct an audit knowledge graph; cross-modally aligning the audit knowledge graph with image features, table features, and time series features to construct a joint audit analysis model; dynamically verifying the audit analysis results with a historical case library to form a dynamically evolving knowledge base; and generating a final audit document based on the updated knowledge base. The present invention realizes dynamic correlation analysis between financial anomalies and business flow fluctuations through cross-modal feature vector alignment and attention-driven clause applicability scoring, thereby improving the efficiency of cross-modal clue discovery.

Description

Technical Field

[0001] The present invention relates to the technical field of enterprise audit information processing, and in particular to an enterprise audit information extraction method based on artificial intelligence. Background Art

[0002] In recent years, enterprise audit information technology has rapidly developed, and multimodal data analysis methods based on optical character recognition (OCR) and natural language processing have become mainstream. Existing technologies typically employ a modular processing process: extracting text information through traditional image processing, combining it with rule engines or statistical models to match audit features, and finally relying on manual review to correlate legal provisions. While these methods perform reliably for structured data processing, they have significant limitations in cross-modal correlation and semantic understanding of unstructured data. Traditional OCR relies on fixed threshold segmentation, which can easily lead to character breakage or splicing errors in low-quality scans. Entity recognition based on keyword matching struggles to capture complex semantic combinations such as "inflated revenue," resulting in a high rate of missed audit leads. This problem primarily manifests itself in the areas of dynamic correlation and knowledge fusion. Firstly, most systems lack a unified representation method for multi-source data (text, tables, and time series), leading to a breakdown in the cross-modal chain of evidence. For example, financial statement anomalies and time-series fluctuations in business logs cannot be correlated using simple rules. Secondly, legal clause mapping often relies on hard-coded rules, failing to consider the impact of the enterprise organizational structure on the applicability of clauses. For example, the same "internal control deficiency" may correspond to different regulatory provisions in different departments, but existing technologies have not introduced attention mechanisms to quantify the dynamic correlation between provisions and organizational characteristics, resulting in insufficient targeting of audit recommendations. Summary of the Invention

[0003] In view of the above existing problems, the present invention is proposed.

[0004] Therefore, the present invention provides an enterprise audit information extraction method based on artificial intelligence to solve the problems of missed audit clues and misjudgment caused by poor cross-modal correlation of multimodal audit data and static mapping of legal terms.

[0005] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0006] In a first aspect, the present invention provides an enterprise audit information extraction method based on artificial intelligence, which comprises obtaining original audit documents from the enterprise and performing multimodal conversion processing to generate a standard data set;

[0007] Conduct deep semantic analysis on standard data sets and build an audit knowledge graph;

[0008] Cross-modally align the audit knowledge graph with image features, table features, and time series features to build a joint audit analysis model;

[0009] Dynamically verify the audit reports generated by the joint audit analysis model with the historical case library to form a dynamically evolving knowledge base;

[0010] Generate final audit documents based on the updated knowledge base.

[0011] As a preferred solution of the enterprise audit information extraction method based on artificial intelligence of the present invention, wherein: the multimodal conversion process includes the following steps:

[0012] Grayscale the PDF files of the contract text and the JPG files of the scanned financial statements, apply the Otsu algorithm to calculate the adaptive threshold, generate black and white binary images, and perform noise removal.

[0013] Perform optical character recognition on the denoised binary image, segment it by text line area, and output JSON format text with character coordinates;

[0014] Perform spectral subtraction noise reduction and speech endpoint detection on the WAV files of the meeting minutes recordings, segment the valid speech segments and transcribe them into text;

[0015] Perform sentence segmentation and entity tagging on the CSV file of the business system log to generate a corrected text after dependency syntax analysis.

[0016] As a preferred solution of the enterprise audit information extraction method based on artificial intelligence described in the present invention, wherein: the construction of a structured audit knowledge graph includes the following steps:

[0017] Based on the text modal data in the standard dataset, extracting audit suspicious entities through entity recognition to generate an intermediate dataset;

[0018] The audit doubt entities in the intermediate dataset are input into a sequence labeling model driven by a bidirectional long short-term memory network, matched with a pre-built database of legal and regulatory clauses, and a mapping data table of clause index codes is generated;

[0019] Perform integrity check and conflict detection on the mapping data table, and dynamically complete missing entity attribute fields;

[0020] Based on the completed mapping data table, the BERT-base model is used to identify organizational entities in the text and construct a correlation matrix between audit suspicious entities and organizational structures;

[0021] The association matrix is ​​combined with the clause index data, and the clause applicability score is calculated through the multi-head attention mechanism to generate a verified structured audit knowledge graph.

[0022] As a preferred solution of the enterprise audit information extraction method based on artificial intelligence of the present invention, wherein: the construction of the joint audit analysis model includes the following steps:

[0023] Extract text feature vectors, image feature vectors, and time series feature vectors from the audit knowledge graph;

[0024] Perform cross-modal alignment of text feature vectors, image feature vectors, and time series feature vectors in a shared embedding space and dynamically adjust feature weights;

[0025] The table structure features and audit text features are integrated through the multi-head attention mechanism to generate a joint representation;

[0026] Based on the joint representation training of the fully connected network branches, the audit problem classification, risk level prediction and clause matching results are output in parallel to generate a structured audit report.

[0027] As a preferred solution of the enterprise audit information extraction method based on artificial intelligence of the present invention, wherein: the forming of the dynamically evolving knowledge base includes the following steps:

[0028] Convert the structured audit report into an XML data package, call the pre-trained embedding model to generate semantic vectors, retrieve similar cases from the historical case library, and generate a verification report;

[0029] Based on the difference annotation results in the verification report, receive the correction feedback file, analyze the clause index changes and generate the difference comparison table;

[0030] Use the proximal strategy optimization algorithm to fine-tune the joint audit analysis model, update the training dataset and rebuild the audit knowledge graph;

[0031] The updated audit doubt entity is associated with the clause index node, and cross-modal positioning information is automatically inherited to form a dynamically evolving knowledge base.

[0032] As a preferred solution of the enterprise audit information extraction method based on artificial intelligence of the present invention, wherein: the generating of the final audit document includes the following steps:

[0033] Based on the audit doubt entity codes in the dynamically evolving knowledge base, locate the original voucher screenshot and generate a visual image with a red mark box;

[0034] Retrieve legal clauses through hash index and construct a paragraph framework of the clauses;

[0035] Generate rectification process nodes based on risky wind turbine parameters in the latest version of the audit doubt entity node set;

[0036] Input the visual images, clause contents, and rectification process into the natural language generation system, fill in the semantic templates, reconstruct the sentences, and generate structured documents;

[0037] By associating text entities with multimodal evidence chains through a coordinate binding algorithm, the interactive display of original vouchers, data tables, and time series graphs is triggered to generate the final audit document.

[0038] As a preferred solution of the enterprise audit information extraction method based on artificial intelligence of the present invention, further processing of the grayscaled financial statement scans includes the following steps:

[0039] A table recognition algorithm based on deep learning is used to locate the table bounding box, and projection analysis is used to segment cells and extract row numbers, column numbers, and content.

[0040] Map the table header and data rows to generate two-dimensional structured table data, and integrate it into a PNG format screenshot;

[0041] Associate the two-dimensional structured table data with the time series data of the business system log to generate time series data.

[0042] As a preferred solution of the enterprise audit information extraction method based on artificial intelligence of the present invention, the dynamic completion of missing entity attribute fields includes the following steps:

[0043] Call the extractor to complete the amount, time and subject fields by locating the original audit document;

[0044] Compare the confidence differences of terms mappings for the same entity to generate completion request instructions;

[0045] The relationship strength between audit questionable entities and organizational structure is calculated based on co-occurrence frequency and syntactic distance.

[0046] In a second aspect, the present invention provides a computer device comprising 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 enterprise audit information extraction method based on artificial intelligence as described in the first aspect of the present invention.

[0047] In a third aspect, 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 enterprise audit information extraction method based on artificial intelligence as described in the first aspect of the present invention.

[0048] The beneficial effects of the present invention are as follows: through multimodal feature alignment and dynamic knowledge graph construction, the accuracy of audit clue discovery and the adaptability of legal clauses are significantly improved. In the data preprocessing stage, the Otsu algorithm is used for adaptive binarization and median filtering noise reduction to solve the problem of character breakage in scanned documents and improve the OCR accuracy; spectral subtraction noise reduction and dual-threshold endpoint detection technology reduce the error rate of speech transcription; through cross-modal feature vector alignment and attention-driven clause applicability scoring, dynamic correlation analysis between financial anomalies and business flow fluctuations is achieved, which improves the efficiency of cross-modal clue discovery; based on the quality verification mechanism of the graph neural network and the fine-tuning of clause matching through proximal strategy optimization, the dynamic evolution capability of the knowledge graph is ensured. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0050] Figure 1 Flowchart of the enterprise audit information extraction method based on artificial intelligence.

[0051] Figure 2 Schematic diagram of multimodal conversion processing.

[0052] Figure 3 Schematic diagram of the audit knowledge graph.

[0053] Figure 4 Schematic diagram constructed for the joint audit analysis model. DETAILED DESCRIPTION

[0054] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0055] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0056] Secondly, the term "one embodiment" or "embodiment" 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 various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0057] Reference Figures 1 to 4 , this embodiment provides an enterprise audit information extraction method based on artificial intelligence, comprising the following steps:

[0058] S1. Obtain the PDF file of the contract text, the JPG file of the scanned financial statement, the CSV file of the business system log, and the WAV file of the meeting minutes recording from the enterprise information system.

[0059] Grayscale the PDF files of contract texts and the JPG files of scanned financial statements.

[0060] Specifically, the weighted average method is used to convert the RGB images of the PDF file of the contract text and the JPG file of the scanned financial statement into grayscale images.

[0061] Binarize the grayscale image.

[0062] Specifically, the Otsu algorithm is applied to calculate the adaptive threshold, calculate the grayscale histogram, iteratively calculate the maximum inter-class variance, determine the segmentation threshold, and generate a black and white binary image.

[0063] Further explanation, the pixel value mapping rule is: the area with pixel value lower than the segmentation threshold is set to black (0), and the area with pixel value higher than the segmentation threshold is set to white (255), generating a black and white binary image of the contract text and a black and white binary image of the financial statement.

[0064] Perform noise removal on the binarized image.

[0065] Specifically, a median filter is used to create a 3×3 pixel sliding window, the 9 pixel values ​​in the window are sorted and the median is taken, and the reflection filling method is used to expand the image edge to remove noise.

[0066] Perform optical character recognition on the denoised binary image.

[0067] Specifically, it divides the text into continuous line areas from left to right and from top to bottom, and performs vertical projection analysis to determine the line spacing (such as 32 pixels for line height). Horizontal projection is used to segment the text lines (such as 28 lines on page 5). Sequence recognition is performed on each text line area, and the image is segmented into [32×600] pixel blocks by line. The character content and its coordinates in the original binary image are output in sequence to generate digital text in JSON format with character coordinates.

[0068] Perform spectral subtraction noise reduction processing on the WAV file of the meeting minutes recording to eliminate the environmental background noise.

[0069] Specifically, through frame processing, background noise with a duration exceeding M is eliminated to generate a noise-reduced audio file.

[0070] Perform speech endpoint detection on the noise-reduced recording files, and segment valid speech segments based on the short-time energy and zero-crossing rate double threshold method; input the valid speech segments into the speech transcription engine based on connection time series classification to generate text records with timestamps; perform sentence processing on the text records generated by the speech transcription engine and the CSV files of the business system logs, using regular expressions to match periods, question marks, exclamation marks, etc. as sentence boundary identifiers; perform entity tagging on the sentence-separated text records and business system log texts, and use a predefined audit entity dictionary for precise string matching.

[0071] Perform dependency parsing on entity-labeled transcripts and business system log text.

[0072] Specifically, a syntax tree is constructed with discovery as the root node and abnormal growth as the object component; for those without a subject who need to start the verification procedure immediately, the default subject audit group is inserted to generate a correction text (the audit group needs to start the verification procedure immediately).

[0073] Perform table area detection on JPG files of scanned financial statements.

[0074] Specifically, a table recognition algorithm based on deep learning is used to locate the table bounding box; cell segmentation is performed on the detected table area, and the projection analysis method is applied to identify the intersection of row and column lines to generate cell coordinates; the text content of the header cell and the numerical value of the data cell are extracted, and a mapping relationship between the column header and the data row is established to generate two-dimensional structured table data containing row numbers, column numbers and cell contents.

[0075] The digital text in JSON format, the timestamp text records generated by the speech transcription engine, the two-dimensional structured table data, and the CSV time series data of the business system log after dependency syntax analysis are unified in format and converted into UTF-8 encoded TXT text files, PNG format table screenshots, and TIMESTAMP-VALUE (timestamp-value) format time series data to form a standard dataset containing text modality, image modality, table modality, and time series modality; the text modality data in the standard dataset is classified according to audit issue types such as contract disputes, financial anomalies, and process violations, and stored in a folder structure named according to the issue category.

[0076] S2. Perform audit problem feature extraction on the text modal data in the standard dataset through entity recognition, and identify audit question entities containing semantic feature tags by comparing with the pre-built audit feature dictionary library through pattern matching algorithm, and generate an intermediate dataset containing entity type code and text location coordinates.

[0077] Specifically, a pre-built audit feature dictionary library is loaded, and an exact string matching algorithm is used to traverse the text data.

[0078] It is further explained that a multi-modal scan is performed on each sentence, and when a keyword is matched, a triple data containing the entity type code, the starting position and the ending position is generated; for text segments that do not hit an exact match, a rule-based feature expansion module is used to identify the subject-object structure through dependency syntax (such as "falsely increased income" → verb "falsely increased" + noun "income").

[0079] Based on the intermediate data set, legal clause mapping is performed on the semantic unit where the audit doubt entity is located. A sequence labeling model driven by a bidirectional long short-term memory network is used to match the clause number format features in the pre-built legal and regulatory clause database to generate a mapping data table containing the clause index code and the corresponding audit doubt entity association relationship.

[0080] Specifically, the sentence containing the entity is input into a sequence labeling model driven by a bidirectional long short-term memory network. The model structure includes an input layer, a bidirectional long short-term memory network layer, and a conditional random field layer.

[0081] Among them, the input layer receives legal terminology word vectors (such as pre-trained on 500,000 regulatory texts), the bidirectional long short-term memory network captures the context dependency through 128 hidden units, and the conditional random field layer decodes and outputs the legal clause number features.

[0082] The extracted audit suspicious entities are interactively verified with the mapped clause index data, structured completion request instructions are automatically generated for missing entity attributes, and entity attribute fields are dynamically updated by parsing feedback data.

[0083] Specifically, a validation rule engine is established for the extracted audit doubtful entities and the clause index data in the mapping data table.

[0084] To further illustrate, the validation rule engine includes integrity checking and conflict detection.

[0085] Among them, integrity verification refers to checking whether the required attributes of the entity (such as amount, time, and subject) exist, and conflict detection refers to comparing the mapping results of multiple terms for the same entity (such as triggering an alarm when the confidence difference is >0.2).

[0086] Automatically generating structured completion request instructions for missing entity attributes refers to locating the original data file by coordinates and calling a specific extractor to complete the fields.

[0087] Perform organizational feature extraction on audit doubtful entities with complete attributes, identify the audit management organization entities of the target enterprise from text semantic units, and establish a correlation matrix between audit doubtful entities and organizational management structure.

[0088] Specifically, identifying the audit management organization entities of the target enterprise from text semantic units means using the BERT-base fine-tuning model to identify organizational entities in the text (such as the audit committee, finance department, etc.) and setting entity relationship extraction rules. For example, establishing affiliation and regulatory relationships within the range of entity spacing <200 characters.

[0089] Establishing an association matrix between audit questionable entities and organizational management structures means setting the row dimension to the audit questionable entity ID, the column dimension to the organizational entity code, and the matrix value to the relationship strength (0-1 standardized value) calculated based on co-occurrence frequency and syntactic distance.

[0090] The clause index data is correlated with the extracted organizational structure features, and the attention mechanism is used to calculate the semantic correlation between the specific regulatory clause library and the target enterprise's audit decision records to generate quantitative evaluation parameters for the clause applicability.

[0091] Specifically, when using the attention mechanism to calculate the semantic correlation between a specific regulatory clause library and the target enterprise's audit decision records, it is necessary to first build an attention calculation module and then perform multi-head attention calculation.

[0092] To further illustrate, building an attention calculation module includes query, key and value.

[0093] The query refers to the BERT embedding vector of the clause text, the key refers to the TF-IDF feature vector of the enterprise audit decision record, and the value refers to the association tag between the decision record and the clause.

[0094] Multi-head attention calculation is to calculate the attention distribution of clause features and enterprise records by each head separately, and combine all heads to generate the clause applicability score, which is expressed as:

[0095] ;

[0096] in, Rate the terms for suitability. is the total number of attention heads, used to average multi-head output, is the attention head index, Represents normalization of similarity, generating attention weights (sum is 1), and highlighting key matching items. For the The query matrix of attention heads represents the audit suspicious features that need to be paid attention to (such as inflated income). For the The key matrix of the attention heads represents the characteristics of the enterprise audit decision records to be matched (such as historical violation cases), is the transpose of the key matrix, converting the row vector into a column vector for matrix multiplication. is the dimension of the key vector to prevent the dot product value from being too large. is the dimension index, For the The value matrix of each attention head carries the specific content of the enterprise audit decision record (such as case amount, time and other details).

[0097] Combining the audit doubt entity characteristics, clause index data, and organizational structure characteristics, multi-dimensional association nodes are constructed according to the graph structure data specification. The audit doubt entity node contains the original data file location information, and the clause index node contains the global unique identifier of the regulatory library and the clause text characteristics. The constructed graph structure data is quality checked as follows:

[0098] Based on the graph structure data analysis of historical audit cases, the association path complexity threshold is set by statistically analyzing the average complexity of valid association paths in the compliance knowledge graph. When the entity association path complexity exceeds the association path complexity threshold, the automatic review mechanism is triggered, and the verified association relationship confidence parameters are updated to the graph structure data to generate a structured audit knowledge graph containing multi-dimensional feature nodes, verified clause index association parameters and cross-modal data positioning information.

[0099] To further explain, the construction of the audit knowledge graph includes node creation and edge relationship definition.

[0100] Specifically, node creation includes audit suspicion nodes, clause nodes, and organization nodes; edge relationship definitions include entity-clause edges, entity-organization edges, and clause-organization edges; quality verification refers to path complexity detection.

[0101] Among them, the audit doubt node includes ID, type code and original file path, the clause node includes GUID, regulatory library index number and clause text summary, and the organization node includes department code and hierarchical path (such as / headquarters / audit department).

[0102] S3. Extract the semantic confidence index output by the semantic analysis of the audit text from the audit doubt entity nodes in the structured audit knowledge graph to form a text feature vector containing the entity type code and confidence weight; extract the table structure features of the scanned financial statements through the convolutional neural network and convert them into image feature vectors of fixed dimensions; convert the financial data association relationship modeled by the graph neural network into a table feature vector in the form of an adjacency matrix; and slice the business flow fluctuation pattern processed by the long short-term memory network into a time window to generate a time series feature vector.

[0103] Cosine similarity calculation is performed on text feature vectors, image feature vectors, table feature vectors and time series feature vectors in the shared embedding space to establish a cross-modal feature alignment matrix, where the initial alignment weight of the text feature vector is weighted using the semantic confidence index of the audit document semantic analysis, and the alignment weight of the image feature vector and the text feature vector is dynamically adjusted based on the matching results of the table number field in the scanned financial statement in the text knowledge graph.

[0104] Based on the cross-modal alignment matrix, the cross-attention scores of text feature vectors and image feature vectors are calculated through the multi-head attention mechanism to generate a joint representation of the fusion table structure features and audit text features. At the same time, the self-attention distribution of the table feature vector and the time series feature vector is calculated to form a correlation pattern map between financial data fluctuations and business flow changes.

[0105] The joint representation of table structure features and audit text features is hierarchically fused with the associated pattern map, and a gating mechanism is used to control the proportion of each modal feature. The initial value of the text modal gating weight adopts the normalized semantic confidence index, and the image modal weight is dynamically adjusted according to the accuracy of financial statement table structure recognition.

[0106] Based on the joint representation of the fusion of table structure features and audit text features, three fully connected network branches of the joint audit analysis model are constructed in parallel:

[0107] The first branch receives fused features and outputs the classification probability distribution of audit issue types such as financial fraud and internal control deficiencies. The second branch receives the same input and generates probability prediction values ​​for high-risk, medium-risk, and low-risk levels. The third branch receives the concatenated input of fused features and legal clause text features and generates a cosine similarity-ranked list of the best matching clauses.

[0108] The output results of the three branches are trained alternately to optimize the cross-entropy loss of the audit question classifier, the mean square error loss of risk prediction, and the contrast loss of clause matching in turn, while keeping the feature extraction parameters frozen, to obtain the structured audit report output by the trained joint audit analysis model.

[0109] Furthermore, the structured audit report includes the entity coding of the audit doubts in the problem list, the clause index number of the legal advice, and the level parameters of the risk warning.

[0110] S4. Convert the structured audit report output by the joint audit analysis model into an XML format data package compatible with the audit work platform; after importing the XML format data package into the audit work platform, call the pre-trained embedding model to generate a 768-dimensional semantic vector for the audit doubt entity description text, use the cosine similarity algorithm to retrieve the most similar case in the historical case library, and generate a verification report document containing similar case numbers, matching scores, and difference point annotations.

[0111] Furthermore, the XML format data packet compatible with the audit work platform includes the file path of the audit doubt entity in the original multimodal data set, the global unique identifier link of the clause index database and the risk level probability distribution vector.

[0112] Based on the difference annotation results in the verification report document, audit suspicious entities with legal clause reference differences are marked as items to be verified; the correction feedback file is received through the online interface of the audit work platform, and the audit suspicious entity codes in the correction feedback file are parsed and matched with the original structured audit report at the field level. Entity pairs with changed clause index numbers are extracted to form a difference comparison table.

[0113] Furthermore, the correction feedback file contains a CSV file of manually confirmed audit doubt entity codes, revised clause index numbers, and risk level adjustment values.

[0114] The audit doubt entity code in the difference comparison table and the revised clause index number are combined into a training sample pair containing the entity code and the new clause index. The proximal policy optimization algorithm is used to calculate the policy gradient of the legal clause matching branch in the joint audit analysis model, as follows:

[0115] While keeping the text feature extraction parameters frozen, the weight matrix of the fully connected layer is dynamically fine-tuned (for example, the learning rate is set to 3e-5 and the batch size is set to 32). During the iterative update process, the cosine similarity loss function between the clause index number and the audit suspicious entity is optimized first. After the update, cross-validation is performed using similar cases in the verification report document.

[0116] The verified correct audit doubt entity code and clause index number data pairs are converted into structured records that conform to the format of the training dataset.

[0117] Specifically, the structured record contains the original text fragment, entity location coordinates, clause index number, and risk level label.

[0118] Perform deduplication verification on the newly added data, append the data records that pass the verification to the end of the TXT text file of the training dataset, and update the corresponding PNG format table screenshots and the index mapping table of TIMESTAMP-VALUE time series data; re-execute the audit knowledge graph construction based on the updated training dataset, perform dependency syntax analysis and entity relationship verification on the newly added audit suspicious entities, establish bidirectional association edges between the verified entity nodes and the clause index nodes (edge ​​attributes include manual confirmation flags and automatic matching confidence), and automatically inherit cross-modal data positioning information when inserting new nodes into the original graph structure data to form a dynamically evolving knowledge base.

[0119] S5. Extract the latest version of the audit doubt entity node set from the dynamically evolving knowledge base, where each node contains the audit doubt entity code, the associated clause index number, the risk level parameter, and the cross-modal data positioning information; locate the original voucher screenshot path in the audit doubt entity code positioning standard dataset, load the PDF file of the contract text and take a screenshot of the image area after grayscale processing; the screenshot range is based on the coordinate parameters determined by the character coordinate positioning algorithm to generate a visual image of the audit evidence with a red rectangular annotation box.

[0120] The clause index numbers in the dynamically evolving knowledge base are accurately matched with the legal and regulatory clause database, the complete text content of the corresponding clauses is retrieved through the hash index mechanism, and the paragraph framework of the clause content is constructed in accordance with the qualitative format requirements in the standard audit document specifications.

[0121] Based on the risk level probability distribution vector, a corrective measures process is automatically generated for high-risk audit suspicious entities.

[0122] Among them, the process nodes include problem confirmation, responsibility allocation, solution design, execution monitoring and effect verification. The connecting lines of each process node are marked with time constraints and work deliverable standards.

[0123] The audit evidence visualization images, clause contents and rectification process are input into the natural language generation system, and the semantic template is filled in according to the entity attributes verified by the dynamically evolving knowledge base. The order of sentence components is reconstructed through the dependency syntax analysis engine, and structured documents are generated in combination with the standard audit document format specification requirements.

[0124] Establish a spatial mapping relationship between structured documents and multimodal evidence chains, intelligently associate text entities with visual elements through coordinate binding algorithms, trigger the multimodal evidence display protocol when clicking on audit suspicious entities, and synchronously display original voucher screenshots, related data tables and time series analysis graphs optimized by image processing algorithms on the interactive interface to generate the final audit document.

[0125] This embodiment also provides a computer device, which is suitable for the case of an enterprise audit information extraction method based on artificial intelligence, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the enterprise audit information extraction method based on artificial intelligence proposed in the above embodiment.

[0126] The computer device may be a terminal, comprising a processor, memory, a communication interface, a display, and an input device 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 an operating system and computer programs. The internal memory provides an environment for the operating system and computer programs stored in the non-volatile storage media. The communication interface of the computer device is used to communicate with external terminals via wired or wireless communication. Wireless communication may be achieved via Wi-Fi, a carrier network, NFC (near-field communication), or other technologies. The display of the computer device may be a liquid crystal display or an electronic ink display. The input device may be a touchscreen overlay on the display, buttons, a trackball, or a touchpad on the computer device housing, or an external keyboard, touchpad, or mouse.

[0127] This embodiment also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the enterprise audit information extraction method based on artificial intelligence proposed in the above embodiment; 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 read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0128] In summary, this invention significantly improves the accuracy of audit clue discovery and the adaptability of legal clauses through multimodal feature alignment and dynamic knowledge graph construction. In the data preprocessing stage, the Otsu algorithm is used for adaptive binarization and median filtering noise reduction to solve the problem of character breakage in scanned documents and improve the OCR accuracy. Spectral subtraction noise reduction and dual-threshold endpoint detection technology reduce the error rate of speech transcription. Through cross-modal feature vector alignment and attention-driven clause applicability scoring, dynamic correlation analysis between financial anomalies and business flow fluctuations is achieved, which improves the efficiency of cross-modal clue discovery. The quality verification mechanism based on graph neural network and clause matching fine-tuning based on proximal strategy optimization ensure the dynamic evolution capability of the knowledge graph.

[0129] 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 the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. An artificial intelligence-based enterprise audit information extraction method, characterized by: include, Obtain original audit documents from enterprises and perform multimodal conversion processing to generate standard data sets; Conduct deep semantic analysis on standard data sets and build an audit knowledge graph; Cross-modal alignment of the audit knowledge graph with image features, table features, and time series features is performed to build a joint audit analysis model. The specific steps include: Extract text feature vectors, image feature vectors, and time series feature vectors from the audit knowledge graph; Perform cross-modal alignment of text feature vectors, image feature vectors, and time series feature vectors in a shared embedding space and dynamically adjust feature weights; The table structure features and audit text features are integrated through the multi-head attention mechanism to generate a joint representation; Based on the joint representation training of the fully connected network branches, the audit problem classification, risk level prediction and clause matching results are output in parallel to generate a structured audit report; The audit report generated by the joint audit analysis model is dynamically verified with the historical case library to form a dynamically evolving knowledge base. The specific steps include: Convert the structured audit report into an XML data package, call the pre-trained embedding model to generate semantic vectors, retrieve similar cases from the historical case library, and generate a verification report; Based on the difference annotation results in the verification report, receive the correction feedback file, analyze the clause index changes and generate the difference comparison table; Use the proximal strategy optimization algorithm to fine-tune the joint audit analysis model, update the training dataset and rebuild the audit knowledge graph; Associate the updated audit doubt entity with the clause index node, automatically inherit cross-modal positioning information, and form a dynamically evolving knowledge base; Generate final audit documents based on the updated knowledge base.

2. The method for extracting enterprise audit information based on artificial intelligence according to claim 1, characterized in that: The multimodal conversion process includes the following steps: Grayscale the PDF files of the contract text and the JPG files of the scanned financial statements, apply the Otsu algorithm to calculate the adaptive threshold, generate black and white binary images, and perform noise removal. Perform optical character recognition on the denoised binary image, segment it by text line area, and output JSON format text with character coordinates; Perform spectral subtraction noise reduction and speech endpoint detection on the WAV files of the meeting minutes recordings, segment the valid speech segments and transcribe them into text; Perform sentence segmentation and entity tagging on the CSV file of the business system log to generate a corrected text after dependency syntax analysis.

3. The method for extracting enterprise audit information based on artificial intelligence according to claim 1, characterized in that: The construction of the audit knowledge graph includes the following steps: Based on the text modal data in the standard dataset, extracting audit suspicious entities through entity recognition to generate an intermediate dataset; The audit doubt entities in the intermediate dataset are input into a sequence labeling model driven by a bidirectional long short-term memory network, matched with a pre-built database of legal and regulatory clauses, and a mapping data table of clause index codes is generated; Perform integrity check and conflict detection on the mapping data table, and dynamically complete missing entity attribute fields; Based on the completed mapping data table, the BERT-base model is used to identify organizational entities in the text and construct a correlation matrix between audit suspicious entities and organizational structures; The association matrix is ​​combined with the clause index data, and the clause applicability score is calculated through the multi-head attention mechanism to generate a verified structured audit knowledge graph.

4. The method for extracting enterprise audit information based on artificial intelligence according to claim 1, characterized in that: The generation of the final audit document includes the following steps: Based on the audit doubt entity codes in the dynamically evolving knowledge base, locate the original voucher screenshot and generate a visual image with a red mark box; Retrieve legal clauses through hash index and construct a paragraph framework of the clauses; Generate rectification process nodes based on the risk level parameters in the latest version of the audit doubt entity node set; Input the visual images, clause contents, and rectification process into the natural language generation system, fill in the semantic templates, reconstruct the sentences, and generate structured documents; By associating text entities with multimodal evidence chains through a coordinate binding algorithm, the interactive display of original vouchers, data tables, and time series graphs is triggered to generate the final audit document.

5. The method for extracting enterprise audit information based on artificial intelligence according to claim 2, characterized in that: The further processing of the grayscaled financial statement scans includes the following steps: A table recognition algorithm based on deep learning is used to locate the table bounding box. Projection analysis is used to segment cells and extract row numbers, column numbers, and content to generate two-dimensional structured table data. Associate the two-dimensional structured table data with the time series data of the business system log to generate time series data.

6. The method for extracting enterprise audit information based on artificial intelligence according to claim 3, characterized in that: The dynamic completion of missing entity attribute fields includes the following steps: The extractor is called to complete the amount, time, and subject fields by locating the original audit data file; the confidence difference of the clause mapping of the same entity is compared to generate the completion request instruction; The relationship strength between audit questionable entities and organizational structure is calculated based on co-occurrence frequency and syntactic distance.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the enterprise audit information extraction method based on artificial intelligence according to any one of claims 1 to 6 are implemented.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the enterprise audit information extraction method based on artificial intelligence according to any one of claims 1 to 6 are implemented.

Images