Construction method and device of audit knowledge base

Abstract

The invention discloses an audit knowledge base construction method and device. The method comprises the steps that basic elements are extracted from audit data through an information structured processing method, and the basic elements comprise text elements, image elements and table elements; performing logic organization on the basic elements by using a content awareness method to form structured data blocks; and constructing an audit knowledge base based on the structured data blocks. The technical problems of low auditing efficiency and poor auditing decision are solved.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and in particular to a method and device for constructing an audit knowledge base. Background Art

[0002] In the audit of accounting firms, the effective application of LLMs depends on the construction and implementation of a high-quality audit knowledge base. The knowledge base integrates resources such as regulations, standards, and cases, and through efficient management, it can quickly retrieve and utilize information and improve work efficiency. At the same time, the knowledge base promotes knowledge sharing and transfer, promotes collaborative work and innovative thinking. In a complex audit environment, the knowledge base helps understand and adapt to various audit scenarios, and improves the efficiency and quality of audit work. However, the audit knowledge base currently used in accounting firms has technical problems such as low audit efficiency and poor audit decision-making due to unprocessed work.

[0003] To address the above-mentioned problems, no effective solution has been proposed yet. Summary of the invention

[0004] The embodiment of the present invention provides a method and device for constructing an audit knowledge base, so as to at least solve the technical problems of low audit efficiency and poor audit decision-making.

[0005] According to one aspect of an embodiment of the present invention, a method for constructing an audit knowledge base is provided, comprising: extracting basic elements from audit data through an information structuring processing method, wherein the basic elements include text elements, image elements and table elements; logically organizing the basic elements using a content-aware method to form structured data blocks; and constructing an audit knowledge base based on the structured data blocks.

[0006] According to another aspect of an embodiment of the present invention, there is also provided a device for constructing an audit knowledge base, comprising: an extraction module, configured to extract basic elements from audit data through an information structuring processing method, wherein the basic elements include text, images and tables; a formation module, configured to logically organize the basic elements using a content-aware method to form structured data blocks; and a construction module, configured to construct an audit knowledge base based on the structured data blocks.

[0007] In the embodiment of the present invention, basic elements are extracted from audit data by using an information structuring processing method, wherein the basic elements include text elements, image elements and table elements; the basic elements are logically organized using a content-aware method to form structured data blocks. The above solution solves the technical problems of low audit efficiency and poor audit decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0009] Figure 1 is a flow chart of a method for constructing an audit knowledge base according to an embodiment of the present invention;

[0010] Figure 2 is a flowchart of another method for constructing an audit knowledge base according to an embodiment of the present invention;

[0011] Figure 3 is a flow chart of an unstructured data processing method according to an embodiment of the present invention;

[0012] Figure 4 is a flowchart of a method for converting text data into an embedding vector through an embedding calculation model according to an embodiment of the present invention;

[0013] Figure 5 is a configuration diagram of a Pinecone vector database according to an embodiment of the present invention;

[0014] Figure 6 is a code for obtaining embedding according to an embodiment of the present invention;

[0015] Figure 7 is a code of a response example according to an embodiment of the present invention;

[0016] Figure 8 is a code of a response example according to an embodiment of the present invention;

[0017] Fig. 9 is a code for inserting vector data into Pinecone according to an embodiment of the present invention;

[0018] Fig.10 is a code for an example of interaction with Weaviate according to an embodiment of the present invention;

[0019] Fig.11 is a structural diagram of another method for constructing an audit knowledge base according to an embodiment of the present invention;

[0020] Fig.12 A schematic diagram of the structure of an electronic device suitable for implementing the embodiments of the present disclosure is shown. DETAILED DESCRIPTION

[0021] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0022] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0023] According to an embodiment of the present invention, a method embodiment of a method for constructing an audit knowledge base is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0024] Figure 1 is a method for constructing an audit knowledge base according to an embodiment of the present invention, such as Figure 1 As shown, the method comprises the following steps:

[0025] Step S102, extracting basic elements from the audit data through an information structuring processing method, wherein the basic elements include text elements, image elements and table elements.

[0026] First, the audit data is preprocessed, for example, the document corresponding to the audit data is converted into an image, the image is resized and standardized, and a preprocessed image is obtained as the preprocessed audit data.

[0027] Next, the object detection method based on deep learning is used to classify and extract the document elements in the preprocessed audit data to obtain the basic elements. For example, a convolutional neural network is used to extract features from the preprocessed image to obtain a feature map; a convolution kernel is slid on the feature map through a region proposal network to generate a candidate region proposal, wherein the candidate region proposal includes a bounding box and a corresponding score; the candidate region proposal is converted into a feature vector of a fixed size; based on the feature vector, the category of the target is predicted, and the position and size of the bounding box are adjusted, and a non-maximum suppression algorithm is used to retain the bounding box with the highest confidence and suppress other bounding boxes overlapping with it to obtain the basic elements.

[0028] Step S104: logically organize the basic elements using a content-aware method to form structured data blocks.

[0029] The text elements and the image elements are converted into descriptive texts, and the table elements are converted into HTML format; based on the descriptive texts and the table elements in HTML format, the structured data blocks are formed.

[0030] Step S106, constructing an audit knowledge base based on the structured data blocks.

[0031] First, the original data is converted into a high-dimensional vector representation through the embedding model. The text data is processed by the word segmenter to generate a tag sequence, and special tags are added and the sequence length is adjusted according to the requirements of the embedding model. The processed text is input into the embedding model to generate an embedding vector, which lays the foundation for subsequent data storage and analysis. The selection of the embedding model is crucial, and the model performance, data security and applicable scenarios must be comprehensively considered to meet the audit business needs and long-term expansion goals.

[0032] Next, cloud service solutions or local deployment solutions can be used for different application requirements. Cloud services such as Pinecone combined with OpenAI embedding models provide convenient and efficient processing capabilities, suitable for the rapid start of small and medium-sized data; while local deployment solutions such as integrating Weaviate database based on Docker-Compose provide higher security and customization capabilities for processing large-scale or sensitive data. This flexible configuration method supports the efficient construction and stable operation of the audit knowledge base in different scenarios.

[0033] Finally, in the cloud service solution, advanced embedding models are used to convert text data into high-dimensional vectors, and vector retrieval and matching are performed through cosine similarity. Data storage and query are efficiently completed through Pinecone, realizing the combination of semantic search and deep analysis. The local deployment solution realizes data storage and semantic search through Weaviate's multi-module integration, which not only enhances the scalability of the system, but also meets the needs of data confidentiality and autonomous control in the audit field. Through customized configuration, the efficiency and security of audit database construction are ensured.

[0034] This application, through the selection of embedded models and the flexible configuration of vector databases, together constitutes a complete solution for building a database based on structured audit data, providing technical support for the efficient operation and precise application of the audit knowledge base.

[0035] In the digital age, corporate data exists in a variety of formats, and key data in the audit field is often stored in PDF format. This application uses a layout parsing strategy with optical character recognition (OCR) capabilities, combined with paragraph and list relationship recognition technology to improve the efficiency of solving content similarity problems.

[0036] The construction of the audit knowledge base involves data collection, information structured processing and cleaning, as well as knowledge base management and maintenance. Through data processing specifications, information technology applications and continuous updating mechanisms, the timeliness and effectiveness of the knowledge base can be ensured. Figure 2 As shown, the method comprises the following steps:

[0037] Step S202: data collection.

[0038] The basis for building an audit knowledge base is to accurately screen and summarize key data, such as historical audit records, policies and regulations, audit standards and legal documents. The comprehensiveness and quality of these data lay a solid foundation for the in-depth development of the knowledge base and ensure the efficiency and accuracy of audit activities. Therefore, selecting authoritative data sources is the key to the data collection process.

[0039] The data collection work adopts a systematic strategy. The first stage focuses on the collection of laws and regulations, covering the Constitution, criminal law, civil law and administrative regulations, etc., to provide a legal basis for auditing. Then, the rules and regulations, auditing standards and guidelines, and industry operating standards are integrated to provide clear guidance for auditing practice. At the same time, attention is paid to the hierarchy and regionality of the data to ensure the comprehensiveness and timeliness of the collected content.

[0040] In terms of data storage, relational databases are selected as the preferred solution due to their excellent data processing performance. During the database design phase, a clear data model is established and the storage structure is carefully planned to ensure the logic and retrieval efficiency of the data. At the same time, multiple security measures such as data encryption, access control and regular backup are adopted to ensure the security and integrity of database information.

[0041] To this end, this application provides an audit knowledge base collection system based on Angular+SpringBoot, aiming to provide an efficient and reliable data collection, collation and retrieval platform. The system builds a user-friendly audit knowledge base management tool by leveraging the flexibility of Angular and the efficiency of SpringBoot.

[0042] Step S204, performing structured processing on the audit data.

[0043] In order to effectively process unstructured data containing rich text, images and table information, this application adopts an innovative information structuring processing strategy, covering layout analysis with comprehensive OCR capabilities, including headers, footers, columns, lists, cross-page spreads and other information.

[0044] Specifically, this application uses information structuring processing strategies to extract basic elements such as text, images, and tables, and filters out non-critical elements such as headers and footers. Next, the processed elements are logically organized using content-aware methods to form structured data blocks to facilitate subsequent analysis and application. Specifically, the unstructured data processing process is as follows: Figure 3 As shown, including:

[0045] Step S2042, document segmentation.

[0046] Step S2044, element extraction.

[0047] In the element extraction process, this application uses a deep learning-based target detection method, Detectron2, to achieve detailed classification and extraction of elements in the document. The specific steps are as follows:

[0048] (1) Data preprocessing.

[0049] Convert documents to image format and perform preprocessing such as resizing and normalization for input into deep learning models.

[0050] (2) Feature extraction

[0051] Convolutional Neural Network (CNN) is used to extract features from preprocessed images and automatically learn the hierarchical feature representation of images. Formula (4.1) describes this process, where is the input image and is the extracted feature map.

[0052] FeatureMap=CNN(I)(4.1)

[0053] (3) Regional proposals.

[0054] The convolution kernel is slid on the feature map through the Region Proposal Network (RPN) to generate potential bounding boxes and corresponding scores. Formula (4.2) describes this process, where Proposals represents the generated candidate region proposals.

[0055] Proposals=RPN(Feature Map)(4.2)

[0056] (4)ROI Align.

[0057] The proposals generated by RPN are converted into fixed-size feature vectors to reduce quantization errors and accurately locate features. Formula (4.3) describes this process, where ROI Features represents the extracted feature vector.

[0058] ROI Features=ROIAlign(FeatureMap,Proposals)(4.3)

[0059] (5) Classification and regression.

[0060] The fully connected layer (FC) is used to classify (predict the category of the target, such as text, image, table, etc.) and regress (finely adjust the position and size of the bounding box) the feature vector output by ROI Align. Formulas (4.4) and (4.5) describe this process, where Class Scores is the classification score and BBox Regression is the regression value of the bounding box.

[0061] Class Scores=FC(ROI Features)(4.4)

[0062] BBox Regression=FC(ROI Features)(4.5)

[0063] (6) Non-Maximum Suppression (NMS).

[0064] In order to remove overlapping detection frames, a non-maximum suppression algorithm is usually used to retain the detection frame with the highest confidence and suppress other detection frames that overlap with it.

[0065] Through these steps, the deep learning-based object detection method can carefully classify and extract elements within the document, generate accurate bounding boxes for each element, and thus effectively extract information from the document.

[0066] Step S2046, element combination.

[0067] The processed elements need to be further cleaned and structured, including standardizing and cleaning text elements, converting image elements into descriptive text, and converting table elements into HTML format. After the data is normalized into a unified format, it is further decomposed through block components, which helps to accurately locate relevant paragraphs and increases the possibility of finding the target content. In response to the more detailed requirements of recall, the documents can be parsed and labeled in more detail to refine the entire system of the knowledge base.

[0068] Ultimately, this application successfully transforms scattered, unstructured data that is difficult to use directly into structured, clear and easy to process information, significantly improving the value and practicality of the audit knowledge base, providing rich and reliable data support for audit practice, and laying a solid foundation for in-depth analysis and knowledge extraction. It effectively improves the efficiency and quality of audit work and provides strong information support for research and practice in the audit field.

[0069] Step S206, building a database based on the structured audit data.

[0070] The method of building a database based on structured audit data will be described in detail below.

[0071] In the process of building the audit knowledge base, the selection of embedded models and vector databases plays a core role in the accuracy and effectiveness of information management. Faced with the dramatic increase in the amount and diversity of audit data, traditional data processing methods are becoming insufficient. Therefore, the use of advanced embedded technology and vector databases has become the only way to improve data processing efficiency and deepen analysis.

[0072] Embedding models refer to the use of machine learning algorithms, especially deep learning techniques, to transform raw data into points in a high-dimensional vector space. This process involves training the original input data (text, images, audio, etc.) to learn the mapping of data points to high-dimensional vectors. Vector databases are databases designed specifically for storing and querying these high-dimensional vector collections. They play an increasingly important role in many fields such as natural language processing, computer vision, and recommendation systems.

[0073] This application aims to deeply explore the key technical decisions in the process of embedding model selection and vector database configuration, and verify the feasibility and effectiveness of different technical routes through specific experimental cases. Figure 4 It describes the entire process of converting text data into embedded vectors through an embedding computation model, and then efficiently dumping these vectors into a vector database, showing the complete path from raw text to vector representation and then to vector data storage.

[0074] First, the text data is converted into tokens through the tokenization process. This step uses the tokenizer corresponding to the selected model to decompose the input text into a series of tokens. Different Transformer models may adopt different tokenization strategies, so it is necessary to select a suitable tokenizer based on the specific selected model. Subsequently, these tokens are converted to the input format of the model. This includes adding special tags according to the requirements of the model and adjusting the sequence length to match the maximum input size of the model. Finally, the processed text is input into the embedding model to obtain a vector representation of the text. These vectors can then be used for further analysis and processing, such as being stored in a vector database to support efficient retrieval. Formula (4.6) describes the process of converting text data into an embedding vector through an embedding calculation model, where Embedding Vector is the embedding vector obtained from the text data, Tokenized Text is the text token processed by the tokenizer, and Embedding Model is an embedding model used to convert tokens into high-dimensional vectors.

[0075] Embedding Vector=Embedding Model(Tokenized Text)(4.6)

[0076] Through these technical decisions and experimental verification, the importance of the selection of embedding models and vector databases for building an audit knowledge base is demonstrated, and a solid foundation is provided for subsequent data analysis.

[0077] Specifically, Figure 5 As shown, the following steps are included:

[0078] Step S502: Select an embedding model.

[0079] In the critical process of building an audit knowledge base, it is crucial to select an appropriate embedding model because it directly affects the efficiency and cost of data processing, as well as the quality and security of the final data. The selection considerations should not only cover current technical requirements, but also focus on future scalability and sustainable development capabilities.

[0080] The existing technologies currently provide a variety of embedding model options, such as text2vec-transformers and img2vec-neural, each of which has its own specific application scenarios and advantages. For example, the text2vec-transformers model is good at processing complex text data, while the img2vec-neural model focuses on the vectorization of image data. These technologies convert raw text into vector form, which not only greatly improves the efficiency of semantic search, but also lays a solid foundation for further data analysis and machine learning tasks, thereby significantly improving the functionality and application breadth of the audit knowledge base.

[0081] Therefore, when selecting an embedded model, audit institutions need to comprehensively consider factors such as cost, data privacy and security, model performance, and scope of application in order to select the solution that best suits their business needs, data characteristics, and long-term strategies. As technology evolves and audit requirements change, the selection of embedded models will continue to be a core part of the knowledge base construction and development process.

[0082] Step S504, configuring a vector database.

[0083] In the process of building an audit knowledge base, it is crucial to choose a suitable vector database configuration solution, which not only determines the speed and flexibility of data processing, but is also the key to data security and privacy protection.

[0084] For users who want quick startup and easy management, cloud-based services such as Pinecone and Weaviate provide convenient vector database solutions. These cloud services make it easier to handle small and medium-sized data due to their good scalability and low maintenance requirements. However, hosting data in the cloud may introduce concerns about data security and privacy, especially when dealing with data containing sensitive information. It may also bring budget uncertainties due to the usage-based billing model.

[0085] On the other hand, choosing to deploy the vector database independently, such as using solutions such as Weaviate or Faiss, can provide users with a higher level of data control and security. By processing data in the internal network, the risk of data leakage is greatly reduced, which is particularly suitable for those occasions that process sensitive data that require strict legal protection. In addition, although independent deployment requires higher initial investment and technical maintenance, it provides the possibility of performance optimization and customized services for specific business needs.

[0086] In summary, the configuration of the vector database needs to comprehensively consider the data scale, security requirements, budget constraints, and technical support capabilities. For those projects in the initial stage of exploration or with small data volumes, cloud services may provide an efficient and economical path. For applications that need to process large amounts of or sensitive data, despite the large initial investment, self-deployment solutions can bring better performance and stricter security in the long run. With the continuous advancement of technology and the growth of audit needs, the selection and configuration of vector databases will continue to evolve to better adapt to the continuous development of the audit industry.

[0087] Step S506, implementing the cloud service solution.

[0088] This application uses the cloud service Pinecone combined with the API embedding model provided by OpenAI to build an audit knowledge base. By adopting OpenAI's advanced embedding model text-embedding-3-small, the model improves the semantic search capability of text data with 1536-dimensional vectorization processing. Based on cosine similarity as a standard for measuring vector similarity, an index that meets high efficiency requirements is configured in Pinecone, and the key parameters are set to METRIC:cosine and DIMENSIONS:1536.

[0089] The specific process is as follows:

[0090] (1) Configuration of Pinecone vector database: Create a vector database in the Pinecone cloud service and configure the necessary parameters as required to facilitate the storage and retrieval of vector data, such as Figure 5 .

[0091] (2) Obtaining embedding vectors: Use the OpenAI API to convert text data into embedding vectors. This step is the initial stage of the data processing process, such as Figure 6 As shown in the example code.

[0092] (3) Embedded response reception: Extract the embedded vector from the response received from the OpenAI API. This vector will be used for subsequent data processing and analysis. See the response example for details. Figure 7 .

[0093] (4) Initialize Pinecone: Establish a connection with the Pinecone vector database and prepare for data insertion. For specific initialization steps, see Figure 8 .

[0094] (5) Insert vector data. Insert the processed embedded vector into the Pinecone database to complete the data storage process, such as Fig. 9 The operation of inserting vector data is shown.

[0095] In this application, cloud services provide convenience and ease of use, simplify the knowledge base construction process, accelerate the vectorization processing and storage of text data, and provide strong support for deep data analysis.

[0096] Step S508: Implement the local deployment solution.

[0097] Docker-Compose is used as a deployment tool, and its flexibility and convenience make it easy to define and run multi-container Docker applications. Through configuration and orchestration, Weaviate integrates with vectorization and reordering modules to maintain system scalability and maintainability. This local deployment strategy meets the audit field's needs for data confidentiality and processing controllability, and provides a secure and autonomous data processing solution.

[0098] The local deployment of the Weaviate service uses the official Docker image, whose main function is data storage and semantic search. To ensure its stable operation, it is achieved by configuring environment variables. At the same time, the vectorizer (text2vec-transformers) and reranker (reranker-transformers) services are used as independent services. The purpose of this is to enhance the flexibility and customization of the data processing process. In terms of configuration and orchestration, Weaviate is tightly integrated with the vectorization and reranking modules, which not only maintains the scalability of the system, but also makes the system easier to maintain. In addition, the local deployment strategy also meets the requirements of the audit field for data confidentiality and processing controllability, providing a secure and autonomous data processing solution. Fig.10 Demonstrates the interaction flow using Weaviate.

[0099] Practice has proved that the method of using Docker-Compose to implement local deployment of vector databases is not only effective, but also successfully solves specific audit data processing tasks through customized configuration. This method supports highly customized data processing processes and ensures data security and privacy throughout the data processing process. It is an ideal solution for specific audit needs. With the advancement of technology, this scalable and secure local deployment strategy is expected to play a more critical role in future audit data processing applications.

[0100] The present application also provides a method for constructing an audit knowledge base. Fig.11 As shown, including:

[0101] Step S1102, extracting basic elements from the audit data through an information structuring processing method.

[0102] Basic elements include raw data in the form of text, images, and tables. Specifically, audit data is classified and identified, and the data is decomposed into basic units that can be processed independently to ensure the accuracy and consistency of subsequent operations. For example, in text processing, specific semantic elements such as titles, paragraphs, and tags need to be identified; for images, visual information related to the audit content needs to be segmented; and tables need to be decomposed into cell contents and their associated row and column relationships. Among them, the classification of structured data blocks by classifiers includes the following: First, after converting unstructured financial statements into text data, multiple independent structured data blocks in the text data are linearly transformed based on the index and multi-scale feature mapping of the current network layer. Specifically, the structured data block features output by the previous layer (i.e., the input of the current layer) are linearly mapped by using the weight matrix of the current layer, and the linear transformation results of each structured data block are generated by combining the bias vector of the current scale. Subsequently, the linear transformation results are nonlinearly mapped using the activation function to extract feature information from different levels and scales to generate nonlinear mapping results. Next, in the last layer of the network, based on the weight matrix of the last layer and the output features of the previous layer, the high-level features are mapped to multiple classification category spaces to generate category scores for each structured data block. Through the Softmax activation function, these category scores are converted into probability values ​​and a probability distribution is generated. In order to improve the flexibility and accuracy of classification, the correlation weights between categories are introduced, and the smoothness of the output is adjusted in combination with the temperature parameter. The temperature parameter can control the degree of exploration of different categories by the model during the training process, and the model decision is made more accurate by lowering the temperature during the inference stage. Finally, the generated category probability distribution is classified by maximizing the likelihood method. Specifically, the category with the highest probability value is selected as the final classification result.

[0103] Step S1104: forming a structured data block.

[0104] The content-aware method is used to logically organize the extracted basic elements to form structured data blocks. For example, specific text is matched with the corresponding image or table to ensure the integrity and accuracy of the data block in terms of semantics and content. The content-aware method usually combines natural language processing, image recognition and pattern matching technology to automatically generate structured data blocks by analyzing the contextual relationship between elements, so that it can be directly applied to subsequent audit analysis or storage.

[0105] Step S1106, generating an audit knowledge base.

[0106] Based on the key factors in the financial field, keywords are extracted from the knowledge dictionary of the financial field, and the keywords are combined with the structured data block vectors to generate financial element feature vectors containing contextual dependencies. Deep hidden layer features are extracted layer by layer through a multi-layer neural network. The leading layer captures local dependencies, the middle layer extracts long-distance dependency features, and the top layer obtains overall financial information through a global attention mechanism. Subsequently, based on the dynamic attention mechanism, the deep hidden layer features are fused with the structured data block vectors to generate a fused feature vector containing financial element keywords, which is stored in the vector library to build an audit knowledge base based on financial elements.

[0107] Specifically, after the structured data blocks are generated, they are first vectorized. Through an embedded model or a specific vectorization algorithm, the text content, image features, and table information in the data blocks are converted into multidimensional vector representations. Vectors can capture the semantic information, structural characteristics, and contextual meaning of the data blocks. In this way, a standardized, easy-to-process data representation is provided for subsequent calculations and analysis, allowing complex financial data to be modeled and compared in a unified high-dimensional space.

[0108] Next, we extract key features based on domain knowledge to meet specific needs in the financial field. We use the knowledge dictionary in the financial field to extract keywords that are closely related to the current analysis task. Keywords include core concepts in financial statements, audit focus, or specific business terms.

[0109] Subsequently, these keywords are combined with the vectors of structured data blocks to generate financial element feature vectors containing contextual dependencies. This combination not only retains the basic information of the data block, but also introduces the supplement of domain knowledge, making the generated feature vector more adaptable to the financial context.

[0110] In order to further optimize the feature representation, the feature vectors of financial elements are deeply processed by constructing a multi-layer neural network. The front layer of the network focuses on capturing local dependencies, mainly modeling short-distance associations between elements within the data block, such as the logical relationship between a report field and its adjacent fields. The middle layer focuses on the extraction of long-distance dependency features, which are used to identify more complex associations between different parts of the data block, such as the connection between the report title and the detail items. The top layer integrates features from all levels through a global attention mechanism to extract overall financial information. The global mechanism can dynamically adjust weights to highlight the importance of key financial elements in the global context, thereby more accurately reflecting the overall meaning of the data.

[0111] After deep feature extraction, the generated deep hidden features are fused with the structured data block vector using the dynamic attention mechanism. The dynamic attention mechanism can flexibly adjust the weighting method of features according to the context and the needs of specific tasks, making the fused feature vector more targeted.

[0112] Finally, the generated fused feature vector not only contains the core information of the original data block, but also integrates the financial element keywords and deep semantic features, making it more valuable for analysis. These fused feature vectors are stored in the vector library and used to build an audit knowledge base based on financial elements. This knowledge base provides efficient query and comparison support for audit analysis by organizing and storing vectorized and optimized financial data. With this knowledge base, auditors can quickly retrieve relevant data, discover potential financial anomalies, and generate more accurate audit reports, providing strong support for the analysis and decision-making of financial data.

[0113] This application vectorizes structured data blocks and combines key factors and contextual information in the financial field to generate financial element feature vectors with deep semantic expression, which significantly improves the understanding and analysis capabilities of financial data. A multi-layer neural network is used to extract local, long-distance and global features layer by layer, and a dynamic attention mechanism is combined to achieve efficient fusion of deep hidden features and structured data block vectors, ensuring the semantic integrity and accuracy of the feature vector. In addition, the multi-scale feature mapping and efficient classification of structured data blocks by the classifier not only improves the flexibility and accuracy of classification, but also improves the robustness of the model in the training and reasoning stages through the introduction of temperature parameter adjustment and category correlation weights.

[0114] The present application also provides a device for constructing an audit knowledge base, comprising: an extraction module, configured to extract basic elements from audit data through an information structuring processing method, wherein the basic elements include text, images and tables; a formation module, configured to logically organize the basic elements using a content-aware method to form structured data blocks; and a construction module, configured to construct an audit knowledge base based on the structured data blocks.

[0115] It should be noted that the construction device of the audit knowledge base provided in the above embodiment is only illustrated by the division of the above functional modules. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. In addition, the construction device of the audit knowledge base provided in the above embodiment and the construction method embodiment of the audit knowledge base belong to the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0116] Fig.12FIG. 1 shows a schematic diagram of the structure of an electronic device suitable for implementing the embodiment of the present disclosure. It should be noted that: Fig.12 The electronic device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present disclosure.

[0117] like Fig.12 As shown, the electronic device includes a central processing unit (CPU) 1001, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage part 1008 into a random access memory (RAM) 1003. In the RAM 1003, various programs and data required for system operation are also stored. The CPU 1001, the ROM 1002, and the RAM 1003 are connected to each other via a bus 1004. An input / output (I / O) interface 1005 is also connected to the bus 1004.

[0118] The following components are connected to the I / O interface 1005: an input section 1006 including a keyboard, a mouse, etc.; an output section 1007 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 1008 including a hard disk, etc.; and a communication section 1009 including a network interface card such as a LAN card, a modem, etc. The communication section 1009 performs communication processing via a network such as the Internet. A drive 1010 is also connected to the I / O interface 1005 as needed. A removable medium 1011, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 1010 as needed, so that a computer program read therefrom is installed into the storage section 1008 as needed.

[0119] The above is only a preferred implementation of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A method for constructing an audit knowledge base, characterized in that: include: Extracting basic elements from the audit data by using an information structuring processing method, wherein the basic elements include text elements, image elements and table elements; logically organizing the basic elements using a content-aware method to form structured data blocks; An audit knowledge base is constructed based on the structured data blocks.

2. The method according to claim 1, characterized in that Extract basic elements from audit data through information structured processing methods, including: Preprocessing the audit data; The target detection method based on deep learning classifies and extracts the document elements in the preprocessed audit data to obtain the basic elements.

3. The method according to claim 2, characterized in that Preprocessing the audit data includes: Converting the document corresponding to the audit data into an image; The image is resized and standardized to obtain a preprocessed image as the preprocessed audit data.

4. The method according to claim 3, characterized in that The target detection method based on deep learning classifies and extracts the document elements in the preprocessed audit data to obtain the basic elements, including: Use convolutional neural network to extract features from the preprocessed image to obtain feature map; Sliding a convolution kernel on the feature map through a region proposal network to generate a candidate region proposal, wherein the candidate region proposal includes a bounding box and a corresponding score; Convert the candidate region proposal into a feature vector of fixed size; Based on the feature vector, the category of the target is predicted, and the position and size of the bounding box are adjusted to obtain the basic element.

5. The method according to claim 4, characterized in that After adjusting the position and size of the bounding box, the method further includes: using a non-maximum suppression algorithm to retain the bounding box with the highest confidence and suppress other bounding boxes overlapping with the bounding box.

6. The method according to claim 1, characterized in that The basic elements are logically organized using a content-aware method to form structured data blocks, including: Convert the text element and the image element into descriptive text, and convert the table element into HTML format; The structured data block is formed based on the descriptive text and the table elements in the HTML format.

7. A device for constructing an audit knowledge base, characterized in that: include: An extraction module is configured to extract basic elements from the audit data through an information structuring processing method, wherein the basic elements include text, images and tables; A forming module configured to logically organize the basic elements using a content-aware method to form a structured data block; The building module is configured to build an audit knowledge base based on the structured data blocks.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored program, wherein when the program is executed, the device where the computer-readable storage medium is located is controlled to execute the method according to any one of claims 1 to 6.

9. A computer device, characterized in that: include: Memory and processor, The memory stores a computer program; The processor is used to execute the computer program stored in the memory, and when the computer program is run, the processor is enabled to execute the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images