Financial data management system based on large language model assistance

Abstract

The invention relates to the technical field of financial data management, and particularly discloses a financial data management system based on large language model assistance, which processes a financial report file by using a natural language technology based on deep learning, extracts text content in the financial report file, and performs word granularity semantic coding on the text content to obtain a semantic coding result; the method comprises the following steps of: acquiring a semantic meaning of each word granularity description in a financial report, performing semantic association enhancement based on local context on each word granularity description semantic feature of the financial report to enhance the semantic feature expression capability of each word granularity description, and performing financial term annotation by means of a large language model on the basis. And based on the financial term annotation information, performing automatic financial report structured database recording on the financial report text content. In this way, the financial report processing efficiency and accuracy can be remarkably improved, meanwhile, manual errors are reduced, and a more accurate financial analysis tool is provided for investors and supervision organizations.

Description

Technical Field

[0001] The present application relates to the technical field of financial data management, and more specifically, to a financial data management system assisted by a large language model. Background Art

[0002] With the rapid development of the global economy and the increasing scale of enterprises, financial management plays a vital role in the operation of enterprises. In today's business environment, financial reports, as important records of corporate economic activities, usually contain profound insights into key areas such as corporate performance, market trends, and management discussions and analysis. They are not only the key basis for management decision-making, but also the main window for investors, regulators and the public to understand the financial status of enterprises.

[0003] In order to better manage and analyze financial data, it is usually necessary to organize and report the financial information of an enterprise or organization according to certain standards and formats. However, as an important carrier of a company's financial status and operating results, financial reports usually contain a large number of professional terms and complex sentence structures. Traditionally, when conducting financial data analysis and report compilation, it is often necessary to rely on manual review and entry, which is not only time-consuming and labor-intensive, but also prone to data errors or omissions due to human factors, seriously affecting the efficiency and accuracy of financial management. Some existing intelligent financial data management systems mainly rely on simple text searches or rule-based pattern matching, which are difficult to adapt to the complex context and diverse financial terms in financial reports, resulting in limited system performance.

[0004] Therefore, an optimized financial data management system assisted by a large language model is expected. Summary of the invention

[0005] In order to solve the above technical problems, the present application is proposed. The embodiment of the present application provides a financial data management system assisted by a large language model, which uses natural language technology based on deep learning to process financial report files, extracts the text content therein, and performs word-granular semantic encoding on it to capture the semantic meaning of each word-granular description in the text content of the financial report, and then, by strengthening the semantic association of each word-granular description of the financial report based on the local context, the semantic feature expression ability of each word-granular description is enhanced, and on this basis, financial terms are annotated with the help of a large language model, and then based on the financial terminology annotation information, the financial report text content is automatically recorded in a structured database of financial reports. In this way, the processing efficiency and accuracy of financial reports can be significantly improved, while reducing the subjective influence of manual interpretation, providing investors and regulators with more accurate financial analysis tools.

[0006] Accordingly, according to one aspect of the present application, a financial data management system based on the assistance of a large language model is provided, which includes: A file acquisition module, used to acquire financial report files; A text content extraction module, used to extract text content from the financial report file to obtain financial report text content; A word granularity semantic encoding module, used for performing word granularity semantic encoding on the text content of the financial report to obtain a sequence distribution of a semantic embedding encoding vector describing the word granularity of the text content of the financial report; A semantic association constraint optimization module, used for performing semantic optimization on the sequence distribution of the word granularity description semantic embedding coding vector of the financial report text content based on the medium-granularity semantic association constraint to obtain a sequence of medium-granularity semantically enhanced financial report text content word granularity coding vectors; A financial term labeling module, used for labeling the sequence of medium-granularity semantically enhanced financial report text content word granularity encoding vectors with financial terms to obtain a sequence of financial report text content word granularity term labels; The structured data record module is used to map the financial report text content to a predefined data model based on the sequence of the word granularity term tags of the financial report text content to obtain a financial report structured database record.

[0007] Compared with the prior art, the financial data management system provided by the present application, which is assisted by a large language model, uses natural language technology based on deep learning to process financial report files, extract the text content therein, and perform word-granular semantic encoding on it, so as to capture the semantic meaning of each word-granular description in the text content of the financial report, and then, by strengthening the semantic association of each word-granular description of the financial report based on the local context, the semantic feature expression ability of each word-granular description is enhanced, and on this basis, financial terminology is annotated with the help of a large language model, and then, based on the financial terminology annotation information, the financial report text content is automatically recorded in a structured database of financial reports. In this way, the processing efficiency and accuracy of financial reports can be significantly improved, while reducing the subjective influence of manual interpretation, providing investors and regulators with more accurate financial analysis tools. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0009] Figure 1 It is a block diagram of a financial data management system based on the assistance of a large language model according to an embodiment of the present application.

[0010] Figure 2 This is a data flow diagram of a financial data management system assisted by a large language model according to an embodiment of the present application.

[0011] Figure 3 It is a block diagram of a word granularity semantic encoding module in a financial data management system assisted by a large language model according to an embodiment of the present application.

[0012] Figure 4 It is a block diagram of a semantic association constraint optimization module in a financial data management system assisted by a large language model according to an embodiment of the present application. DETAILED DESCRIPTION

[0013] Below, the embodiments of the present application will be described in more detail in conjunction with the accompanying drawings, and the above and other purposes, features and advantages of the present application will become more apparent. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the example embodiments described herein.

[0014] In response to the defects of existing financial data management systems, this application proposes a financial data management system assisted by a large language model, which uses natural language technology based on deep learning to process financial report files, extract the text content therein, and perform word-granular semantic encoding on it to capture the semantic meaning of each word-granular description in the text content of the financial report, and then, by strengthening the semantic association of each word-granular description of the financial report based on the local context, the semantic feature expression ability of each word-granular description is enhanced, and on this basis, financial terminology is annotated with the help of a large language model, and then, based on the financial terminology annotation information, the financial report text content is automatically recorded in a structured database of financial reports. In this way, the processing efficiency and accuracy of financial reports can be significantly improved, while reducing the subjective impact of manual interpretation, providing investors and regulators with more accurate financial analysis tools.

[0015] Figure 1 It is a block diagram of a financial data management system based on the assistance of a large language model according to an embodiment of the present application. Figure 2 FIG. 1 is a data flow diagram of a financial data management system based on a large language model according to an embodiment of the present application. Figure 1 and Figure 2As shown, the financial data management system 100 assisted by a large language model includes: a file acquisition module 110, which is used to acquire financial report files; a text content extraction module 120, which is used to extract text content from the financial report file to obtain financial report text content; a word granularity semantic encoding module 130, which is used to perform word granularity semantic encoding on the financial report text content to obtain a sequence distribution of word granularity description semantic embedding encoding vectors of the financial report text content; a semantic association constraint optimization module 140, which is used to perform semantic optimization on the sequence distribution of the word granularity description semantic embedding encoding vectors of the financial report text content based on medium granularity semantic association constraints to obtain a sequence of medium granularity semantically enhanced financial report text content word granularity encoding vectors; a financial term labeling module 150, which is used to perform financial term labeling on the sequence of medium granularity semantically enhanced financial report text content word granularity encoding vectors to obtain a sequence of financial report text content word granularity term labels; a structured data recording module 160, which is used to map the financial report text content to a predefined data model based on the sequence of word granularity term labels of the financial report text content to obtain a financial report structured database record.

[0016] In the above-mentioned financial data management system 100 assisted by a large language model, the file acquisition module 110 is used to acquire financial report files. In the embodiment of the present application, the financial report files include but are not limited to electronic financial statements, annual reports, quarterly reports, etc.

[0017] In the modern enterprise environment, financial report files usually exist in the form of electronic documents, which may be stored on local computers, internal company network servers or cloud services. For many companies, their internal financial systems are the main source of financial report generation. In this case, obtaining financial report files needs to be integrated with the company's internal information system to facilitate automated access and retrieval of these files. Based on this, it is first necessary to understand the data storage structure and interface specifications of the company's internal information system. For example, some companies use well-known enterprise resource planning (ERP) systems, such as SAP, Oracle, etc. These systems usually provide a wealth of application programming interfaces (APIs) that allow external programs to access and obtain the financial data stored therein in an authorized manner. Developers can use these APIs to write specialized program codes to achieve connection with the financial system and extract the required financial report file data from the system's database based on specific query conditions, such as reporting period, report type, etc.

[0018] With the widespread application of cloud computing technology, some companies may store financial report files in cloud storage services. Cloud storage providers such as Amazon S3 and Microsoft Azure Blob Storage provide companies with flexible, scalable and secure file storage solutions. In this case, obtaining financial report files first requires relevant personnel of the company to have credentials to access cloud storage services, usually including access keys and secret keys. By using these credentials, with the help of software development kits (SDKs) or command line tools provided by cloud storage service providers, developers can write programs to access and download financial report files in cloud storage buckets. For example, using Amazon S3's SDK, developers can programmatically obtain the required financial report files from the S3 bucket by specifying information such as the bucket name and file path, and download them to the local computer or other specified storage location for subsequent processing and analysis.

[0019] When the user starts the acquisition process, the system will first check whether there is a pre-set configuration file, which contains the default file path, access permission settings and other necessary parameters. If such a configuration exists, it will be initialized according to its instructions; otherwise, the user will be prompted to manually specify the file source. For local files, users can select the specific file location through the graphical interface, or enter the absolute path through the command line to point to the target document. For resources on remote servers, the system should provide an intuitive and easy-to-use connection interface, allowing users to fill in information such as the host address, port number, user name and password, and download the required files after establishing a secure connection.

[0020] In the process of obtaining financial report files, different file formats also need to be considered. Common file formats for financial reports include PDF, Excel, Word, etc., and each format has its own characteristics and processing methods. Financial report files in PDF format are widely used because of their good typesetting and cross-platform compatibility. However, the content extraction of PDF files is relatively complex and requires the help of specialized PDF parsing tools and libraries. For example, the PyPDF2 library in Python can be used to read the text content of PDF files, extract page information, etc. However, the extraction of tables and complex formats of financial data in PDF may require more advanced technical means, such as using optical character recognition (OCR) technology combined with text parsing algorithms to convert image content in PDF into editable text data, and then perform structured data extraction and collation. For financial report files in Excel format, its advantage lies in the structuring and easy processing of data. You can use the Excel processing library in the programming language, such as Python's openpyxl library, to directly open the Excel file, read the data in the worksheet, and extract the required financial data fields according to pre-defined rules and logic, such as specific data items in each financial statement, formula calculation results, etc. Financial report files in Word format are relatively rarely used for direct data extraction, but in some cases, it may be necessary to extract key financial information from the text description in the Word file. This can be done through a Word document parsing library, such as Python's python-docx library, to read paragraphs, tables and other elements in the file and extract the text content. Then, financial-related information can be identified and organized through text analysis technology.

[0021] It is worth noting that in actual applications, the financial reports of many companies are not publicly accessible resources, but are subject to strict access control protection. Therefore, this system must have a complete authentication mechanism that can legally obtain the private information of authorized users under the premise of complying with laws and regulations. This includes but is not limited to using the HTTPS protocol to encrypt the transmission channel, supporting modern authentication frameworks such as OAuth2.0, and integrating single sign-on (SSO) functions, so that users can seamlessly access financial documents in multiple related systems with just one login.

[0022] In addition to static documents, some dynamically generated web pages may also appear as carriers of financial reports. For such online resources, traditional file acquisition methods are obviously no longer applicable. At this time, the system will enable a specially designed web crawler to simulate the behavior of a real browser to access the target URL, parse the HTML source code, and extract the core content. More importantly, the tool must also have a certain degree of intelligence to deal with the content delay problem of JavaScript loading and wait for the page to be fully rendered before starting the collection work.

[0023] After completing the initial data reading, it is necessary to preprocess it to ensure that it meets the input requirements of subsequent modules. This step mainly includes standardization operations such as removing redundant spaces, unifying date formats, and converting currency symbols. Especially for financial reports in a multilingual environment, the system must be able to identify different language versions and convert them into the target language to facilitate subsequent consistency analysis. In addition, given that some special industries may adopt customized professional term abbreviations or coding rules, the system also reserves an extension interface to allow users to import customized dictionaries to supplement or replace entries in the default vocabulary.

[0024] Most importantly, the entire acquisition process should be performed within a strict security framework. Whether it is local or remote operation, the principle of least privilege should be followed, that is, only the minimum permissions required to complete the task should be granted to prevent potential security risks.

[0025] In the above-mentioned financial data management system 100 assisted by a large language model, the text content extraction module 120 is used to extract text content from the financial report file to obtain the text content of the financial report. It should be understood that, considering that the original financial report file may be a document in PDF, Word or other formats, there are a large number of non-text elements, such as charts, pictures, etc. Therefore, in order to filter out non-text elements that are not related to the financial terminology annotation task, the present application further uses text extraction technology (such as optical character recognition (OCR) technology) to process the financial report file, extract the text content therein, and obtain the financial report text content, thereby providing purer input data for subsequent financial terminology annotation tasks.

[0026] Specifically, considering the format differences of electronic documents such as PDF, Word (.docx), Excel (.xlsx), plain text files (.txt), etc., in order to be able to effectively extract the required text content from these different types of files, powerful file parsing capabilities must be possessed. For example, PDF files often contain rich visual elements, such as charts, pictures, and complex typesetting styles, which makes it difficult to read the text directly; while Word documents may include various embedded objects and formatting tags, which also increase the complexity of processing. In order to cope with this diversity, this application adopts a multi-level parsing strategy, first determining its type based on the file extension, and then calling the corresponding parser to perform specific operations.

[0027] For PDF files, since they are usually designed as non-editable final output forms, specialized tools are required for content extraction. Based on this, the system integrates advanced optical character recognition (OCR) technology, which can automatically detect page layout, distinguish between text and non-text areas, and convert scanned images into editable text strings. The OCR engine can not only process static images in black and white or color, but also adapt to different resolutions and quality levels to ensure good recognition even in the face of low-definition scans. In addition, in order to improve recognition accuracy, the OCR engine will also refer to known template style libraries and adjust parameters according to the typesetting habits of specific industries to better adapt to various complex situations. For example, in the financial field, some bank reports may have a fixed table layout. Through pre-trained templates, OCR can more accurately locate and extract key data items.

[0028] When encountering files generated by office software such as Word or Excel, the system will use the API interface provided by Microsoft Office or an open source third-party library to parse the document structure. For Word documents, in addition to simple paragraph text, there may also be formatted text elements such as titles, lists, footnotes, and non-text content such as charts and formulas. The system must not only extract all visible text parts, but also retain the original hierarchical relationships so that subsequent modules can perform semantic analysis based on the context. For Excel files, special attention should be paid to the cell contents in the worksheet, especially those results calculated by formulas. Considering that many companies use Excel for preliminary data collation and calculation when preparing financial reports, the system has a built-in intelligent formula parsing function that can restore the calculation logic without affecting the original data, helping users better understand the business meaning behind the report.

[0029] In addition to the common file formats mentioned above, some dynamically generated web pages may also appear as carriers of financial reports. For such online resources, traditional file acquisition methods are obviously no longer applicable. At this time, the system will enable a specially designed web crawler to simulate the behavior of a real browser to access the target website, parse the HTML source code, and extract the core content. Once the web page content is successfully captured, the system needs to further clean up irrelevant HTML tags, leaving pure text information for subsequent processing.

[0030] After completing the extraction of all the text content in the file, the next step is to conduct a preliminary quality check on the text content. Although the various technical means mentioned above have greatly improved the success rate of extraction, some errors or incomplete fragments will inevitably occur. To this end, the system has established a comprehensive verification process that covers spelling checking, grammatical analysis, logical consistency and other aspects. For example, where the amount of money is involved, the system will check whether there are obvious differences in order of magnitude or unreasonable values; for time series data, it will check whether the date format is consistent, whether the time period overlaps, and other issues. If potential problems are found, the system will try to infer the correct value through the context, or directly mark it out to remind manual review. Doing so not only ensures the authenticity and reliability of the data, but also lays a solid foundation for subsequent in-depth analysis.

[0031] In the financial data management system 100 assisted by a large language model, the word granularity semantic encoding module 130 is used to perform word granularity semantic encoding on the financial report text content to obtain a sequence distribution of the word granularity description semantic embedding encoding vectors of the financial report text content. Figure 3 FIG. 1 is a block diagram of a word granularity semantic encoding module in a financial data management system assisted by a large language model according to an embodiment of the present application. Figure 3 As shown, the word granularity semantic encoding module 130 includes: a word segmentation unit 131, which is used to perform word segmentation processing on the financial report text content to obtain a text sequence of word granularity descriptions of the financial report text content; a semantic embedding encoding unit 132, which is used to perform semantic embedding encoding on each financial report text content word granularity description in the text sequence of the financial report text content word granularity description to obtain a sequence distribution of the semantic embedding encoding vectors of the financial report text content word granularity description.

[0032] Specifically, the word segmentation unit 131 is used to perform word segmentation processing on the text content of the financial report to obtain a text sequence of word granularity description of the text content of the financial report. It should be understood that the vocabulary unit is the most basic semantic carrying unit of the text. Through word segmentation processing, the boundary of each vocabulary unit can be clarified, so as to more accurately capture the semantic meaning carried by each vocabulary unit in the text content of the financial report, so as to facilitate the identification and marking of professional terms in the financial report. Therefore, the present application first performs word segmentation processing on the text content of the financial report to split the continuous text content into a series of vocabulary units to form a text sequence of word granularity description of the text content of the financial report.

[0033] Specifically, the semantic embedding coding unit 132 is used to perform semantic embedding coding on each financial report text content word granularity description in the text sequence of the financial report text content word granularity description to obtain a sequence distribution of the semantic embedding coding vector of the financial report text content word granularity description. In a specific example of the present application, the semantic embedding coding unit 132 is used to: input each financial report text content word granularity description in the text sequence of the financial report text content word granularity description into a word semantic embedding encoder based on the Bert model to obtain a sequence distribution of the semantic embedding coding vector of the financial report text content word granularity description. A person of ordinary skill in the art should know that Bert (Bidirectional Encoder Representations from Transformers) is a pre-trained language model that has learned rich language patterns and rules by performing bidirectional context understanding training on a large-scale corpus, and can effectively capture the deep semantic information of each financial report text content word granularity description, thereby providing a more accurate semantic basis for subsequent financial term annotation.

[0034] In the above-mentioned financial data management system 100 assisted by a large language model, the semantic association constraint optimization module 140 is used to perform semantic optimization based on medium-granularity semantic association constraints on the sequence distribution of the semantic embedded coding vectors of the word granularity description of the text content of the financial report to obtain a sequence of medium-granularity semantically enhanced word granularity coding vectors of the text content of the financial report. It should be understood that, considering that the semantic meaning of each word granularity description in the text content of the financial report may be affected by its contextual environment, therefore, in order to more accurately capture the semantic features of each word granularity description, the present application proposes a semantic optimization method based on local context semantic association constraints, which semantically optimizes each word granularity description within the local context range, thereby maintaining semantic consistency and coherence over a larger semantic span, and enhancing the semantic expression ability of each word granularity description in a specific context.

[0035] Figure 4 FIG. 1 is a block diagram of a semantic association constraint optimization module in a financial data management system assisted by a large language model according to an embodiment of the present application. Figure 4 As shown, the semantic association constraint optimization module 140 includes: a semantic association window search unit 141, which is used to determine the medium-granularity semantic association window of each financial report text content word granularity description semantic embedding coding vector based on the grammatical depth of each financial report text content word granularity description semantic embedding coding vector in the sequence distribution of the financial report text content word granularity description semantic embedding coding vector; a context semantic association reinforcement unit 142, which is used to perform medium-granularity context semantic reinforcement on each financial report text content word granularity description semantic embedding coding vector based on the medium-granularity semantic association window of each financial report text content word granularity description semantic embedding coding vector to obtain a sequence of the medium-granularity semantically reinforced financial report text content word granularity coding vector.

[0036] Specifically, the semantic association window search unit 141 includes: a hyperbolic space mapping subunit, which is used to map each financial report text content word granularity description semantic embedding coding vector in the sequence distribution of the financial report text content word granularity description semantic embedding coding vector to the hyperbolic space to obtain a sequence of hyperbolic space financial report text content word granularity description semantic embedding coding vectors, which is expressed by the formula:

[0037]

[0038] in, represents the sequence distribution of the semantic embedding encoding vectors describing the word granularity of the content of the financial report text, , , , They represent the first, second, and third words in the sequence distribution of the semantic embedding encoding vector of the word granularity description of the financial report text content. , The word-granular description of the financial report text content is the semantic embedding encoding vector, is the hyperbolic space mapping function, and denote the first mapping weight matrix and the second mapping weight matrix respectively, Indicates the The corresponding hyperbolic space financial report text content word granularity description semantic embedding encoding vector.

[0039] That is, considering that in text data, there are obvious hierarchical relationships between words, such as hyponymy, synonyms and antonyms, etc. Therefore, in order to better represent the hierarchical structure of the text content of the financial report, the present application further maps the semantic embedding encoding vector of each word granularity description of the text content of the financial report to the hyperbolic space, so as to utilize the special properties of hyperbolic geometry to better represent the hierarchical structure of the text content of the financial report, and reveal the semantic hierarchy and correlation between each word granularity description.

[0040] Furthermore, the semantic association window search unit 141 further includes: a grammatical depth measurement subunit, which is used to calculate the grammatical depth implicit representation value of each hyperbolic space financial report text content word granularity description semantic embedding coding vector in the sequence of the hyperbolic space financial report text content word granularity description semantic embedding coding vector to obtain the sequence distribution of the grammatical depth implicit representation value of the financial report text content word granularity description, which is expressed by the formula:

[0041] in, represents the square of the norm of a vector, represents the logarithmic function with base 2, represents the grammatical depth measurement function, Indicates the The corresponding financial report text content word granularity describes the grammatical depth implicit representation value.

[0042] That is, by calculating the grammatical depth implicit representation value of the semantic embedding encoding vector of each word-granular description of the text content of the financial report after hyperbolic space mapping, we can analyze its depth in the syntax tree, reveal the grammatical function carried by each word-granular description in the specific context, and then reflect the importance of each word-granular description in the sentence structure and its relationship with other components. For example, nouns as subjects usually carry more sentence information than adjectives as attributives. In this way, it helps to use the grammatical information of each word-granular description to further strengthen the understanding of its true meaning.

[0043] Furthermore, the semantic association window search unit 141 also includes: a semantic association window determination subunit, which is used to determine the medium-granularity semantic association window of the semantic embedding coding vector of each financial report text content word granularity description based on the sequence distribution of the grammatical depth implicit representation value of the financial report text content word granularity description. More specifically, the semantic association window determination subunit is used to: extract the grammatical depth implicit representation value of the first financial report text content word granularity description semantic embedding coding vector from the sequence distribution of the financial report text content word granularity description semantic embedding coding vector as the starting position of the medium-granularity semantic association window; along the sequence distribution of the grammatical depth implicit representation value of the financial report text content word granularity description, find the grammatical depth implicit representation value of the financial report text content word granularity description semantic embedding coding vector that is closest to the grammatical depth implicit representation value of the first financial report text content word granularity description as the ending position of the medium-granularity semantic association window; based on the starting position of the medium-granularity semantic association window and the ending position of the medium-granularity semantic association window, determine the medium-granularity semantic association window of the first financial report text content word granularity description semantic embedding coding vector from the sequence distribution of the financial report text content word granularity description semantic embedding coding vector.

[0044] The above process is expressed as follows:

[0045] in, Indicates taking the absolute value, Indicates the index corresponding to the minimum value. Indicates the The medium-grained semantic association window, is the The position index of is the same as The grammatical depth implicit representation value is closest to the financial report text content word granularity describing the position index of the grammatical depth implicit representation value.

[0046] It should be understood that by setting a reasonable semantic association window, it can be ensured that when strengthening the semantic features of each word granularity description, neither important contextual information will be missed nor too much irrelevant information interference will be introduced. It helps to capture the medium-scale semantic connection between each word granularity description and other words in a specific context, thereby improving the model's ability to understand complex semantic structures while reducing computational complexity.

[0047] Specifically, the contextual semantic association enhancement unit 142 is used to: calculate the semantic association scores of each financial report text content word granularity description semantic embedding coding vector in the medium-granularity semantic association window of the first financial report text content word granularity description semantic embedding coding vector relative to the first financial report text content word granularity description semantic embedding coding vector to obtain a sequence of semantic association scores; normalize the sequence of semantic association scores to obtain a sequence of normalized semantic association scores; based on the sequence of normalized semantic association scores, calculate the position-weighted sum between each financial report text content word granularity description semantic embedding coding vector in the medium-granularity semantic association window of the first financial report text content word granularity description semantic embedding coding vector to obtain the medium-granularity semantically enhanced financial report text content word granularity description semantic embedding coding vector corresponding to the first financial report text content word granularity description semantic embedding coding vector.

[0048] The above process is expressed as follows:

[0049] in, The first word in the sequence distribution of the semantic embedding encoding vector representing the word granularity description of the financial report text content The word-granular description of the financial report text content is the semantic embedding encoding vector, represents the matrix multiplication operation, is the scoring reference vector, is the scoring reference weight matrix, is the rating bias vector, is the natural exponential function, Indicates the The corresponding medium-granularity semantically enhanced financial report text content word granularity encoding vector.

[0050] That is, based on the determined medium-granularity semantic association window, the contextual information in the window is used to improve the semantic representation of each word-granularity description, so as to realize the medium-granularity contextual semantic enhancement of the semantic embedding encoding vector of each word-granularity description of the text content of the financial report. In this process, based on the semantic association and grammatical relationship between each word-granularity description and other words in a specific window, the feature representation of the word-granularity description is adjusted to better reflect the actual meaning of each word-granularity description in the current context, thereby obtaining a sequence of medium-granularity semantically enhanced word-granularity encoding vectors of the text content of the financial report. In this way, the perception of long-distance semantic dependencies in the text content of the financial report can be effectively improved, thereby enhancing the accuracy of the semantic representation of each word-granularity description.

[0051] In the above-mentioned financial data management system 100 assisted by a large language model, the financial terminology annotation module 150 is used to annotate the sequence of the medium-granularity semantically enhanced financial report text content word granularity encoding vectors with financial terms to obtain a sequence of financial report text content word granularity term labels. In a specific example of the present application, the financial terminology annotation module 150 is used to: add a prompt word to the tail of each medium-granularity semantically enhanced financial report text content word granularity encoding vector in the sequence of the medium-granularity semantically enhanced financial report text content word granularity encoding vector, and then input it into the financial term annotator based on the large language model to obtain the sequence of financial report text content word granularity term labels, and the prompt word is "based on word semantic information, financial term labeling". That is, the deep learning ability of the large language model is used to identify and annotate the financial terms for each word granularity description. It should be understood that since the large language model has been pre-trained on a large-scale text data set, it can understand complex language patterns and industry terms. Therefore, after inputting the prompt word (Prompt), the large language model can perform in-depth semantic analysis and identification and labeling of financial terms on each medium-granularity semantically enhanced financial report text content word granularity encoding vector according to the indication information of the prompt word. Here, the prompt word is used to provide additional information or instructions to the large language model without changing the semantic information of the original data, limit the scope of the large language model's behavior, and thus guide it to mark each medium-granularity semantically enhanced financial report text content word granularity encoding vector with financial terms. In an embodiment of the present application, the large language model is a GPT-3 model.

[0052] In the above-mentioned financial data management system 100 assisted by a large language model, the structured data recording module 160 is used to map the text content of the financial report to a predefined data model based on a sequence of term tags of the word granularity of the text content of the financial report to obtain a structured database record of the financial report. Here, the predefined data model provides a standard framework for converting the text content in the financial report into a structured data format, which generally includes a series of predefined fields, such as income, expenditure, balance sheet, income statement, cash flow statement, etc., which correspond to key information points in the financial report. By mapping the text content of the financial report to the corresponding fields, the automatic extraction and organization of the financial data can be achieved, thereby improving the efficiency and accuracy of data processing.

[0053] During the mapping process, the system will automatically identify and match the various term tags in the text content of the financial report and place them in the corresponding structured fields of the data model. For example, if the term "net profit" appears in the text content of the financial report, it will be identified and mapped to the "net profit" field of the data model, thereby converting the unstructured financial report text into structured financial data records stored in the database for efficient query, analysis and utilization. In addition, this method helps to eliminate differences between financial reports from different sources or formats, ensuring data consistency and standardization, which is particularly important for cross-departmental and cross-organizational data sharing.

[0054] In summary, according to the embodiment of the present application, a financial data management system assisted by a large language model is explained, which uses natural language technology based on deep learning to process financial report files, extract the text content therein, and perform word-granular semantic encoding on it to capture the semantic meaning of each word-granular description in the text content of the financial report. Then, the semantic association of each word-granular description of the financial report is strengthened based on the local context to enhance the semantic feature expression ability of each word-granular description, and on this basis, financial terminology is annotated with the help of a large language model, and then the financial report text content is automatically recorded in a structured database based on the financial terminology annotation information. In this way, the processing efficiency and accuracy of financial reports can be significantly improved, while reducing the subjective impact of manual interpretation, providing investors and regulators with more accurate financial analysis tools. The basic principle of the present invention is described above in conjunction with specific embodiments. However, it should be pointed out that the advantages, strengths, effects, etc. mentioned in the present invention are only examples and not limitations, and it cannot be considered that these advantages, strengths, effects, etc. must be possessed by each embodiment of the present invention. In addition, the specific details of the above embodiments are only for the purpose of illustration and facilitation of understanding, rather than limitation, and the above details do not limit the present invention to be implemented by adopting the above specific details.

Claims

1. A financial data management system based on a large language model, characterized in that: include: A file acquisition module, used to acquire financial report files; A text content extraction module, used to extract text content from the financial report file to obtain financial report text content; A word granularity semantic encoding module, used for performing word granularity semantic encoding on the text content of the financial report to obtain a sequence distribution of a semantic embedding encoding vector describing the word granularity of the text content of the financial report; A semantic association constraint optimization module, used for performing semantic optimization on the sequence distribution of the word granularity description semantic embedding coding vector of the financial report text content based on the medium-granularity semantic association constraint to obtain a sequence of medium-granularity semantically enhanced financial report text content word granularity coding vectors; A financial term labeling module, used for labeling the sequence of medium-granularity semantically enhanced financial report text content word granularity encoding vectors with financial terms to obtain a sequence of financial report text content word granularity term labels; The structured data record module is used to map the financial report text content to a predefined data model based on the sequence of the word granularity term tags of the financial report text content to obtain a financial report structured database record.

2. The financial data management system based on the assistance of a large language model according to claim 1 is characterized in that: The word granularity semantic encoding module includes: A word segmentation unit is used to perform word segmentation processing on the text content of the financial report to obtain a text sequence describing the word granularity of the text content of the financial report; The semantic embedding coding unit is used to perform semantic embedding coding on each financial report text content word granularity description in the text sequence of the financial report text content word granularity description to obtain a sequence distribution of the semantic embedding coding vector of the financial report text content word granularity description.

3. The financial data management system based on the assistance of a large language model according to claim 2 is characterized in that: The semantic embedding encoding unit is used to: Each financial report text content word granularity description in the text sequence of the financial report text content word granularity description is input into a word semantic embedding encoder based on the Bert model to obtain a sequence distribution of the financial report text content word granularity description semantic embedding coding vector.

4. The financial data management system based on the assistance of a large language model according to claim 3 is characterized in that: The semantic association constraint optimization module includes: A semantic association window search unit, used to determine the medium-granularity semantic association window of each financial report text content word granularity description semantic embedding coding vector based on the grammatical depth of each financial report text content word granularity description semantic embedding coding vector in the sequence distribution of the financial report text content word granularity description semantic embedding coding vector; A context semantic association reinforcement unit is used to perform medium-granularity context semantic reinforcement on each of the financial report text content word granularity description semantic embedding coding vectors based on the medium-granularity semantic association window of the semantic embedding coding vector of the text content word granularity description of each financial report to obtain a sequence of the medium-granularity semantically reinforced financial report text content word granularity coding vectors.

5. The financial data management system based on the assistance of a large language model according to claim 4 is characterized in that: The semantic association window search unit comprises: A hyperbolic space mapping subunit, used to map each financial report text content word granularity description semantic embedding coding vector in the sequence distribution of the financial report text content word granularity description semantic embedding coding vector to the hyperbolic space to obtain a sequence of hyperbolic space financial report text content word granularity description semantic embedding coding vectors; A grammatical depth measurement subunit is used to calculate the grammatical depth implicit representation value of each hyperbolic space financial report text content word granularity description semantic embedding coding vector in the sequence of the hyperbolic space financial report text content word granularity description semantic embedding coding vector to obtain the sequence distribution of the grammatical depth implicit representation value of the financial report text content word granularity description; The semantic association window determination subunit is used to determine the medium-granularity semantic association window of the semantic embedding coding vector of each financial report text content word granularity description based on the sequence distribution of the grammatical depth implicit representation value of the financial report text content word granularity description.

6. The financial data management system based on the assistance of a large language model according to claim 5 is characterized in that: The semantic association window determination subunit is used to: Extracting the grammatical depth implicit representation value of the first financial report text content word granularity description semantic embedding coding vector in the sequence distribution of the financial report text content word granularity description semantic embedding coding vector as the starting position of the medium-granularity semantic association window; Along the sequence distribution of the grammatical depth implicit representation values ​​of the granularity description of the financial report text content words, find the grammatical depth implicit representation value of the granularity description of the financial report text content words that is closest to the grammatical depth implicit representation value of the semantic embedding coding vector of the granularity description of the first financial report text content words as the termination position of the medium-granularity semantic association window; Based on the starting position of the medium-granularity semantic association window and the ending position of the medium-granularity semantic association window, the medium-granularity semantic association window of the first financial report text content word granularity description semantic embedding coding vector is determined from the sequence distribution of the financial report text content word granularity description semantic embedding coding vector.

7. The financial data management system based on the assistance of a large language model according to claim 6 is characterized in that: The context semantic association strengthening unit is used to: Calculating the semantic association scores of each financial report text content word granularity description semantic embedding coding vector in the medium-granularity semantic association window of the first financial report text content word granularity description semantic embedding coding vector relative to the first financial report text content word granularity description semantic embedding coding vector to obtain a sequence of semantic association scores; Normalizing the sequence of semantic association scores to obtain a sequence of normalized semantic association scores; Based on the sequence of normalized semantic association scores, the position-weighted sum between the semantic embedding coding vectors of the granularity description of each financial report text content word in the medium-granularity semantic association window of the semantic embedding coding vector of the granularity description of the first financial report text content word is calculated to obtain the medium-granularity semantically enhanced financial report text content word granularity coding vector corresponding to the semantic embedding coding vector of the granularity description of the first financial report text content word.

8. The financial data management system based on the assistance of a large language model according to claim 7 is characterized in that: The financial term annotation module is used to: After adding a prompt word to the end of each medium-granularity semantically enhanced financial report text content word granularity encoding vector in the sequence of medium-granularity semantically enhanced financial report text content word granularity encoding vectors, the prompt word is input into a financial term tagger based on a large language model to obtain a sequence of financial report text content word granularity term labels, wherein the prompt word is "financial term tagging based on word semantic information".

Images