A method, apparatus, medium, and device for extracting data from financial announcements.

Abstract

This invention discloses a method, apparatus, medium, and device for extracting financial announcement data, belonging to the field of data extraction. This application acquires preprocessed financial announcement data and user query requests, utilizes a multi-path recall mechanism (semantic segmentation, directory navigation, and keyword retrieval) combined with text structure features to filter out relevant contexts matching the query; then, it standardizes the tabular data in the relevant contexts using financial domain synonym mapping; finally, it inputs the standardized context into a large language model trained with supervised fine-tuning on a financial domain dataset, outputting a standardized structured extraction result, thus achieving efficient and accurate extraction and processing of financial announcement data. This application effectively solves the problem that existing technologies cannot accurately and efficiently extract financial announcement data.

Description

Technical Field

[0001] This invention relates to the field of data extraction, and in particular to a method, apparatus, medium and equipment for extracting data from financial announcements. Background Technology

[0002] In fields such as financial investment analysis, risk assessment, and market supervision, financial announcements (such as corporate annual reports, prospectuses, and rating reports) are the core carriers reflecting a company's operating conditions, financial data, and risk information. Accurately and efficiently extracting key information from these announcements is a crucial prerequisite for subsequent data analysis and decision-making. With the expansion of the capital market and increasingly stringent information disclosure requirements, the number of financial announcements has exploded, and they are diverse in format (including PDF files and scanned copies), complex in content, and dense with technical terminology, placing higher demands on the efficiency and accuracy of information extraction.

[0003] Currently, existing technologies for extracting financial announcement data mainly fall into four categories: The first is manual compilation and outsourcing services, where researchers or analysts manually read announcements and extract key information, or this is outsourced to third-party data service providers. This method relies heavily on human resources, is time-consuming and labor-intensive, and has high outsourcing costs, making it difficult to adapt to the extraction needs of massive amounts of announcements. The second is automated extraction based on templates or rules, which uses preset templates or regular expressions to extract specific fields from fixed-format documents. However, this method is rigid; once the format or wording of the announcement is slightly adjusted, the preset rules may become invalid, requiring continuous and significant effort to maintain, and thus has very limited adaptability. The third is traditional OCR and keyword retrieval technology, which first converts scanned images into text using OCR, and then locates information using keyword matching or simple natural language processing technology. However, this method cannot understand the semantics of the text, is prone to errors and omissions in extraction, and has weak processing capabilities for tabular data. The fourth is simple augmented retrieval generation (RAG) technology, which uses a single vector semantic retrieval method to provide context for a large language model and perform question answering. However, it has limited effectiveness in recalling specific terms and precise numbers in the financial field, making it difficult to accurately locate target chapters, resulting in inaccurate contextual information and ultimately affecting the reliability of the extraction results.

[0004] Furthermore, existing technologies generally suffer from a core pain point: a large amount of key data in financial announcements is presented in tabular form, and different announcements lack a unified standard for naming the same type of project (e.g., "government bonds" vs. "national bonds," "policy bank bonds" vs. "policy bank bonds"). Traditional technologies cannot effectively normalize these non-standard terms, making it difficult to directly use the extracted data for cross-enterprise and cross-period comparative analysis, further restricting the application value of the data. At the same time, the limitations of single search strategies and inconsistent formats of extracted results also make it difficult for existing technologies to meet the actual needs of the financial sector for "efficient, accurate, standardized, and integrable" data extraction. These shortcomings prevent existing technologies from accurately and efficiently extracting data from financial announcements. Summary of the Invention

[0005] This invention provides a method, apparatus, medium, and device for extracting data from financial announcements, in order to solve the problem that existing technologies cannot accurately and efficiently extract data from financial announcements.

[0006] Firstly, this application provides a method for extracting data from financial announcements, including: Obtain preprocessed financial announcement data and user query requests; Based on the text structure characteristics of financial announcement data, and combined with a pre-set multi-path recall mechanism that includes semantic segmentation, directory navigation and keyword retrieval, the associated context that matches the user's query request is filtered from the financial announcement data. By using a pre-defined synonym mapping relationship in the financial field, the tabular data in the associated context is subjected to terminology standardization processing; The standardized contextual information is input into a pre-defined large language model so that the large language model outputs structured extraction results that conform to the pre-defined specifications. The large language model is obtained through supervised fine-tuning training based on a dataset from the financial field.

[0007] This application provides a clear input foundation for subsequent processing by acquiring preprocessed financial announcement data and user query requests. Then, leveraging the structural features of the financial announcement text, combined with a multi-path recall mechanism of semantic segmentation, directory navigation, and keyword retrieval, it can accurately filter out relevant contexts matching the user query from massive amounts of data. This process not only improves the comprehensiveness of data retrieval but also reduces the risk of information omission through multi-dimensional retrieval methods. Subsequently, the tabular data in the relevant context undergoes terminology standardization processing based on the synonym mapping relationship in the financial domain, eliminating comprehension barriers caused by terminology diversity and ensuring data consistency and comparability. Finally, the standardized relevant context is input into a large language model trained with supervised fine-tuning on a financial domain dataset, further enhancing the model's ability to understand and extract financial text. This results in structured extraction results that better conform to preset specifications, thereby providing users with accurate, efficient, and standardized financial information extraction services. This technical solution not only optimizes the processing flow of financial announcement data but also significantly improves the overall system performance and user experience through the synergistic effect of each step. This application effectively solves the problem that existing technologies cannot accurately and efficiently extract financial announcement data.

[0008] Furthermore, the acquisition of preprocessed financial announcement data and user query requests specifically includes: Obtain financial announcement data that has undergone format conversion and structure processing; the financial announcement data includes PDF files and scanned copies of annual reports, prospectuses, and rating reports, converted into structured text. Obtain user-initiated query requests; the query requests include single indicator value queries, table extraction queries, and multi-dimensional comprehensive analysis queries.

[0009] This application ensures the quality and usability of input data by acquiring pre-processed financial announcement data. Specifically, it converts complex-formatted PDF files and scanned copies of documents such as annual reports, prospectuses, and rating reports into structured text. This process not only solves the problem of diverse raw data formats and difficulties in direct processing but also provides a standardized data foundation for subsequent information extraction and analysis through format conversion and structuring. Simultaneously, it acquires diverse query requests initiated by users, including single-indicator numerical queries, specific table extraction queries, and multi-dimensional comprehensive analysis queries, enabling the system to meet the diverse needs of different users in financial data analysis. This dual optimization of data and needs not only improves the system's flexibility and adaptability but also ensures data consistency and accuracy through preprocessing, thus providing strong support for subsequent accurate extraction and efficient analysis, significantly improving the overall efficiency of financial announcement data processing and the user experience.

[0010] Furthermore, based on the textual structure features of the financial announcement data, and combined with a pre-defined multi-path retrieval mechanism including semantic segmentation, directory navigation, and keyword retrieval, the relevant context matching the user's query request is filtered from the financial announcement data, specifically as follows: The preprocessed financial announcement data is divided into independent paragraphs. The user query request and each independent paragraph are vectorized and converted. The vector similarity between the user query request and each independent paragraph is calculated. The paragraphs with similarity that meet the preset standard are semantically segmented for recall, and the semantic segmentation recall results are obtained. The intent of the user's query request is parsed, and the intent is matched with the original directory tree of the financial announcement data to locate the most relevant chapter. Semantic recall is performed only within the text scope corresponding to the chapter to obtain the directory navigation recall result. Based on the keyword matching algorithm, financial announcement data is retrieved, paragraphs containing the query keywords and related derivative terms are extracted, and the keyword retrieval results are obtained. The semantic block recall results, directory navigation recall results, and keyword retrieval recall results are merged and deduplicated to obtain a candidate context set; Based on the preset rearrangement model, and combined with the candidate context set, the associated context is obtained.

[0011] This application achieves semantic-based precise recall by segmenting preprocessed financial announcement data into independent paragraphs and combining this with vectorized transformation of user query requests and similarity calculations, ensuring a high relevance between the recalled results and user intent. Secondly, by parsing user query intent and matching it with the original directory tree of the financial announcements, the search scope is further narrowed, improving the targeting and efficiency of the recall and reducing interference from invalid information. Furthermore, keyword matching algorithms can quickly extract paragraphs containing query keywords and their derivatives, supplementing information that may have been missed in semantic recall. After merging and deduplicating the three recall results, a re-ranking model is used for optimization, ultimately yielding high-quality contextual information. This process not only fully utilizes the structural features and semantic information of the text but also ensures the comprehensiveness and accuracy of the recall results through multi-dimensional recall strategies and optimization processes, significantly improving the performance and user experience of the financial announcement data processing system.

[0012] Furthermore, the process of obtaining the associated context based on the preset rearrangement model and the candidate context set specifically involves: The intent type of the user query request is parsed, and the corresponding source confidence coefficient is assigned according to the three paths of semantic block recall, directory navigation recall and keyword retrieval. Based on the pre-defined reordering model, the semantic relevance of each paragraph in the candidate context set is scored, and the final score is calculated by combining the source confidence coefficient of the recall path to which the paragraph belongs. Paragraphs with final scores below a preset threshold are removed, and a preset number of paragraphs with the highest scores are selected from the remaining paragraphs to obtain the associated context.

[0013] This application first parses the intent type of the user's query request and assigns corresponding source confidence coefficients to three paths—semantic chunking recall, directory navigation recall, and keyword retrieval—based on different intent types. This process fully considers the effectiveness and reliability of different recall paths under different query intents, providing a scientific basis for subsequent screening. Next, a pre-defined reordering model is used to score the semantic relevance of each paragraph in the candidate context set, and the final score is calculated by combining the source confidence coefficient of the recall path to which the paragraph belongs. This scoring mechanism that combines confidence can comprehensively consider the semantic relevance of the paragraph and the reliability of the recall path, thereby more accurately assessing the actual matching degree between each paragraph and the user's query request. Finally, by removing paragraphs below a preset threshold and selecting a preset number of paragraphs with the highest scores from the remaining paragraphs as associated contexts, the high quality and high relevance of the output results are further ensured.

[0014] Furthermore, the terminology standardization process for the tabular data in the associated context is performed using a pre-defined financial domain synonym mapping relationship, specifically as follows: Identify tabular data from the associated context and extract the item names from the tabular data; The project names in the tabular data are compared with the preset financial field synonym mapping relationship, and the non-standard project names in the tabular data are replaced with the corresponding standard terms to obtain standardized tabular data. The synonym mapping relationship in the financial field includes standard terms and their corresponding synonyms, abbreviations and variants.

[0015] This application utilizes a synonym mapping system in the financial field to standardize the terminology of tabular data within a context, significantly improving the consistency and readability of the data. First, tabular data and item names are identified and extracted from the context, providing a clear target for subsequent standardization. Next, the extracted item names are compared with a pre-defined synonym mapping system in the financial field, replacing non-standard item names with standard terms. This process not only overcomes the comprehension barriers caused by the diversity of financial terminology (such as synonyms, abbreviations, and variant expressions) but also ensures the consistency of tabular data across different sources and formats.

[0016] Furthermore, the large language model is obtained through supervised fine-tuning training based on a dataset from the financial field, specifically as follows: Obtain multiple sets of training samples; each set of training samples includes a financial query command, a text fragment of a financial announcement related to the query command, and a standard structured output result that responds to the query command based on the text fragment of the financial announcement. Multiple sets of training samples are input into the initial large language model so that the initial large language model can be trained through multiple rounds of iteration. After obtaining the output results that conform to the preset structured specifications, the training stops and the large language model is obtained.

[0017] This application provides rich domain-specific data for model training by acquiring multiple sets of training samples containing financial domain query commands, relevant text fragments, and standard structured output results, ensuring that the model can learn the language characteristics and business logic of the financial domain. These training samples are then input into an initial large language model, and multiple rounds of iterative training are conducted with the goal of strengthening the model's adherence to commands and the stability of output formats. This process not only optimizes the model's understanding of financial queries but also ensures the structured and standardized nature of the output results through supervised learning. Finally, training stops once the model can stably output results conforming to preset standards, resulting in a fine-tuned large language model. This training method specifically for the financial domain enables the model to process financial announcement texts more accurately, improving the accuracy and consistency of information extraction, thereby providing a more reliable tool for automated processing in the financial domain and significantly enhancing the system's practicality and professionalism.

[0018] Secondly, this application provides a data extraction device for financial announcements. The data extraction device for financial announcements includes: The acquisition module is used to acquire preprocessed financial announcement data and user query requests; The filtering module is used to filter out the associated context that matches the user's query request from the financial announcement data based on the text structure features of the financial announcement data and in combination with a preset multi-path recall mechanism that includes semantic segmentation, directory navigation and keyword retrieval. The processing module is used to perform terminology standardization processing on the tabular data in the associated context through a preset financial domain synonym mapping relationship; The extraction module is used to input the standardized context into a preset large language model so that the large language model outputs a structured extraction result that conforms to the preset specifications. The large language model is obtained through supervised fine-tuning training based on a dataset from the financial field.

[0019] The acquisition module of this application ensures the accurate input of preprocessed financial announcement data and user query requests, providing high-quality foundational data for subsequent processing. Next, the filtering module utilizes a multi-path recall mechanism (semantic segmentation, directory navigation, and keyword retrieval) combined with text structure features to accurately filter out relevant contexts matching the user query. This process significantly improves the comprehensiveness and accuracy of data retrieval. The processing module standardizes the terminology of the tabular data through a financial domain synonym mapping relationship, eliminating comprehension barriers caused by terminology diversity and ensuring data consistency and comparability. Finally, the extraction module inputs the standardized relevant contexts into a large language model trained with supervised fine-tuning in the financial domain, further enhancing the model's ability to understand and extract financial text, ensuring that the output conforms to the preset structured specifications. This modular design not only optimizes the functionality of each stage but also forms an efficient, accurate, and standardized data extraction process through collaborative work, significantly improving the overall performance and user experience of financial announcement data processing.

[0020] Thirdly, this application provides a computer-readable storage medium comprising a stored computer program, wherein, when the computer program is executed, it controls the device containing the computer-readable storage medium to perform a data extraction method for financial announcements as described above. Its beneficial effects are the same as those of the data extraction method for financial announcements provided in the first aspect of this application.

[0021] Fourthly, this application provides a terminal device including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement any of the financial announcement data extraction methods described in the first aspect. Attached Figure Description

[0022] Figure 1 : A schematic flowchart of an embodiment of the data extraction method for financial announcements provided in this application; Figure 2 : A schematic diagram illustrating an embodiment of the multi-path recall RAG architecture provided in this application; Figure 3 : A schematic diagram of an embodiment of the financial announcement data extraction device provided in this application. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Example 1 Please refer to Figure 1 In order to solve the problem that existing technologies cannot accurately and efficiently extract data from financial announcements, this invention provides a method for extracting data from financial announcements, including steps S01-S04.

[0025] S01: Obtain preprocessed financial announcement data and user query requests.

[0026] In a preferred embodiment of this invention, the acquisition of the preprocessed financial announcement data and the user query request specifically involves: First, financial announcements disclosed by various entities in the capital market are collected. These announcements cover various types, including corporate annual reports, prospectuses, and credit rating reports, and are in various formats, including editable PDFs, non-editable PDFs, and scanned copies of paper announcements. For these raw data in different formats, a unified preprocessing engine is used. This engine integrates optical character recognition (OCR), document parsing, and a structured conversion module. It first converts unstructured image data such as scanned documents into text data, and then converts all text data into Markdown-formatted structured text, ensuring data format consistency. Subsequently, the converted structured text is segmented into multiple independent paragraphs with complete semantics. A pre-defined language model (such as Qwen2-72B, which is related to intent recognition) is used to vectorize each paragraph, generating vector data for subsequent retrieval. Furthermore, native directory information is extracted from the structured text to construct a complete directory tree. This directory tree is then mapped to the segmented structured text and vector data to form preprocessed financial announcement data, facilitating rapid location of target chapters. Regarding the acquisition of user query requests, the system receives various information extraction requests initiated by users through their terminals. These query requests explicitly include a description of the target information to be extracted. Specifically, they can be categorized into single-indicator numerical queries (such as "What is the provision coverage ratio of a company at the end of 2023?"), specific table extraction queries (such as "Extract the detailed table of a company's trading financial assets"), and multi-dimensional comprehensive analysis queries (such as "Summarize the core financial indicators and changes of a company over the past three years"), ensuring that the subsequent data extraction process can accurately respond to the user's intent.

[0027] S02: Based on the text structure features of financial announcement data, and combined with a pre-set multi-path recall mechanism that includes semantic segmentation, directory navigation and keyword retrieval, the associated context matching the user's query request is filtered from the financial announcement data.

[0028] In a preferred embodiment of this invention, the text structure features of the financial announcement data, combined with a pre-defined multi-path retrieval mechanism including semantic segmentation, directory navigation, and keyword retrieval, are used to filter out the associated context matching the user's query request from the financial announcement data. Specifically: First, a semantic segmentation recall operation is performed. Based on the segmentation results completed in the preprocessing stage, the structured text of the financial announcement is divided into multiple independent paragraphs that maintain semantic integrity. A preset language model (Qwen2-72B, which is from the same source as intent recognition) is used to vectorize the user query request and each independent paragraph. By calculating the similarity value in the vector space of the two, paragraphs with similarity reaching the preset standard are selected, and a preset number of high similarity paragraphs (e.g., 20) are initially recalled to form the semantic segmentation recall result. This method can effectively capture the deep relationship between query intent and text semantics. It is suitable for single-point fact query scenarios where information points are scattered and there is no specific chapter pointing to. For example, when a user queries "What is the company's provision coverage ratio?", relevant information fragments can be quickly located from the entire text. Secondly, directory navigation recall is implemented. First, the core intent of the user's query request is analyzed using Qwen2-72B (semantic understanding model). For example, when a user queries "financial investment details", the model can identify that the user's intent is to obtain specific financial statement notes. Then, the core intent is accurately matched with the original directory tree of the financial announcement, locating the most relevant chapter, "Financial Statement Notes". Semantic recall is performed only within the text scope corresponding to this chapter, initially recalling a preset number of relevant paragraphs (e.g., 20), forming the directory navigation recall results. This method greatly improves the accuracy of table queries and specific chapter data queries, and can effectively eliminate noise interference from other irrelevant chapters. Simultaneously, keyword retrieval is initiated, employing mature keyword matching algorithms (such as the BM25 algorithm) to extract core keywords and related derivative terms from the user's query request. Through keyword matching (such as TF-IDF), a full-domain search is performed on the structured text of the financial announcement, filtering out paragraphs containing the target terms and initially recalling a preset number of relevant paragraphs (e.g., 20), thus forming the keyword retrieval recall results. This method, as a supplement to classic text retrieval algorithms, can effectively compensate for the potential recall deficiencies that semantic models may encounter with certain specific terms, abbreviations, or precise numbers.

[0029] After obtaining the three recall results, intelligent merging and rearrangement are performed according to the following steps: First, candidate pool merging and deduplication: collect all candidate paragraphs from the semantic block recall results, directory navigation recall results and keyword retrieval recall results, include them in a unified list, remove duplicate paragraphs by text fingerprint comparison, and construct a candidate context set. The second step is intent recognition and weight allocation: The Qwen2-72B (semantic understanding model) is used to identify the intent type of the user's query request. Based on the intent type, the "source confidence coefficient" corresponding to each recall path is obtained from a pre-defined strategy table. Specifically, if the intent is "table query" (requiring precise chapter location), a coefficient of 1.2 is assigned to the directory navigation recall path, and a coefficient of 1.0 is assigned to both the semantic block recall and keyword retrieval paths. If the intent is "precise numerical query" (requiring matching specific terms), a coefficient of 1.1 is assigned to the keyword retrieval path, and a coefficient of 1.0 is assigned to the other two paths. If the intent is "comprehensive analysis query," a coefficient of 1.0 is assigned to all three recall paths to maintain a balanced weight. If a candidate paragraph is recalled by multiple paths simultaneously, the highest source confidence coefficient among all paths is used, or the coefficients are combined to enhance the overall score. The third step is semantic scoring and weighted ranking: A lightweight re-ranking model based on Cross-Encoder is used to calculate the basic semantic relevance score of each paragraph in the original user query and the candidate context set. Combined with the source confidence coefficient of the recall path to which the paragraph belongs, the final score of each paragraph is calculated by multiplying the basic score by the source confidence coefficient. All candidate paragraphs are then sorted in descending order according to their final scores. The fourth step is two-layer threshold filtering: First, a relevance score threshold (e.g., 0.8) is set, discarding all paragraphs with a final score below this threshold to filter out noisy data. Then, a final quantity threshold (e.g., 5) is set, selecting the highest-scoring pre-defined number of paragraphs from those that passed the first layer of filtering to form a highly relevant context that matches the user query request. The specific values ​​of the relevance score threshold and the final quantity threshold are determined through multiple rounds of optimization on a financial domain-specific validation set to achieve a balance between recall and precision.

[0030] S03: Standardize the terminology of the tabular data in the associated context by using a preset financial domain synonym mapping relationship.

[0031] In a preferred embodiment of this invention, the step of standardizing the terminology of the tabular data in the associated context using a preset financial domain synonym mapping relationship specifically involves: When a user's query request is a table query (e.g., "extract 'trading financial assets' details"), the system first accurately locates the target table from the relevant context according to the user's instruction. Then, it extracts the complete table data through a text structure recognition algorithm. This table data is presented in Markdown source code format, ensuring that the row and column structure is consistent with the original announcement. When processing the table data, the system first extracts the item names (such as financial asset types, financial indicator names, etc.) and corresponding numerical information to ensure the completeness and accuracy of the table data. Subsequently, the system loads a pre-defined synonym mapping relationship in the financial field. This mapping relationship is based on standardized financial industry terminology, commonly used market expressions, and common non-standard terms found in historical announcements. It can be dynamically maintained by business experts and includes a series of standard terms and their corresponding synonyms, abbreviations, and variant expressions. For example, the non-standard terms corresponding to the standard term "government bonds" include "national debt," "state bonds," "government bonds," and "local government bonds"; the non-standard terms corresponding to the standard term "interbank certificates of deposit" include "certificates of deposit" and "interbank certificates of deposit"; and the non-standard terms corresponding to the standard term "policy bank bonds" include "policy bank bonds" and "policy and financial institution bonds," etc.

[0032] After extracting the table data and loading the synonym mapping, the system compares and matches each item name in the table data with the non-standard terms in the synonym mapping. If an item name in the table data is detected to belong to a set of non-standard terms corresponding to a certain standard term, the item name is replaced with the corresponding standard term; otherwise, the item name remains unchanged. Throughout the replacement process, the numerical information and row and column structure in the table remain the same; only the item names are normalized. The final output is standardized table data with uniform item names and a standardized format. This data is presented in the form of a structured JSON array, ensuring the naming consistency of the same type of data in different financial announcements. This provides reliable support for subsequent cross-announcement and cross-enterprise data comparison and analysis, as well as downstream system integration.

[0033] S04: Input the standardized context into the preset large language model so that the large language model outputs a structured extraction result that conforms to the preset specifications; The large language model is obtained through supervised fine-tuning training based on a dataset from the financial field.

[0034] In a preferred embodiment of this invention, the step of inputting the standardized context into a preset large language model so that the large language model outputs a structured extraction result conforming to a preset specification specifically involves: The standardized context, including normalized tabular data and precisely matched text paragraphs, is input into a pre-defined generative large language model (e.g., Qwen2-72B). This model is specifically adapted for information extraction scenarios in the financial field, accurately identifying the user's query intent based on the input context and performing targeted processing according to the query task type: for single-indicator queries (e.g., "What is the provision coverage ratio?"), the model directly locates the target value and related information (such as reporting period, calculation method) from the context, and processes it according to "indicator name - value". The fixed field "-Reporting Period-Unit" outputs results in JSON format. For example, when the input context is "As of the end of 2023, the bank's provision coverage ratio... was 205.14%", the output result is {"Indicator Name":"Provision Coverage Ratio","Value":"205.14%","Reporting Period":"End of 2023"}. For table queries, the model integrates standardized table data and outputs it in the form of a structured array, ensuring field alignment and clear hierarchy. For comprehensive analysis queries, the model extracts the core information from the context and integrates it according to logical categories, outputting a well-organized structured summary.

[0035] The large language model is trained using supervised fine-tuning (SFT) on a dataset specific to the financial domain. The core purpose of fine-tuning is not to inject financial domain knowledge (which is provided by the context retrieved earlier), but to enhance the accuracy of the model in following instructions and the stability of the output format. The training dataset consists of multiple sets of high-quality training samples. Each set of samples adopts the format of "instruction + context, standard JSON output". For example, the input of a training sample is "[query]:'What is the provision coverage ratio?'[context]:'As of the end of 2023, the bank's provision coverage ratio... is 205.14%'", and the corresponding standard output is {"indicator name":"provision coverage ratio","value":"205.14%","reporting period":"end of 2023"}. During training, multiple sets of training samples are batch-input into the initial large language model. Through multiple rounds of iteration, the model parameters are adjusted, allowing the model to gradually learn the rules for understanding financial terminology, data extraction logic, and structured output format requirements, until the model output consistently conforms to preset structured specifications (such as field completeness, format uniformity, and data accuracy). Fine-tuning training is then completed, and the model parameters are fixed, forming a large language model adapted to the scenario described in this application. The entire output process strictly adheres to preset specifications, ensuring a uniform result format that can be directly parsed and integrated by downstream systems, meeting the practicality and reliability requirements of data processing in the financial field.

[0036] like Figure 2 As shown, Figure 2The core process for extracting financial announcement data in this application is as follows: After a user initiates a query request, the query intent is first identified using Qwen2-72B. Then, three-way recall operations are executed in parallel: LLM block semantic recall, LLM directory navigation recall, and BM25 keyword retrieval, to obtain the corresponding relevant text chunks. Next, the text chunks retrieved from the three paths are merged and deduplicated, and then a lightweight rearrangement model based on Cross-Encoder is used to complete the rearrangement and filtering to obtain highly relevant contextual information. Finally, this contextual information is input into Qwen2-72B to generate a structured extraction result that conforms to the specifications. The result feedback chain is also included, fully demonstrating the entire technical process from query intent identification, multi-way recall, merging and rearrangement to result output.

[0037] In summary, this application provides a clear input foundation for subsequent processing by acquiring preprocessed financial announcement data and user query requests. Next, leveraging the structural features of the financial announcement text and combining semantic segmentation, directory navigation, and keyword retrieval with a multi-path recall mechanism, it can accurately filter relevant contexts matching user queries from massive amounts of data. This process not only improves the comprehensiveness of data retrieval but also reduces the risk of information omission through multi-dimensional retrieval methods. Subsequently, the tabular data in the relevant context undergoes terminology standardization based on the mapping relationship of synonyms in the financial domain, eliminating comprehension barriers caused by terminology diversity and ensuring data consistency and comparability. Finally, the standardized relevant context is input into a large language model trained with supervised fine-tuning on a financial domain dataset, further enhancing the model's ability to understand and extract financial text. This results in structured extraction results that better conform to preset specifications, providing users with accurate, efficient, and standardized financial information extraction services. This technical solution not only optimizes the processing flow of financial announcement data but also significantly improves the overall system performance and user experience through the synergistic effect of each step. This application effectively solves the problem that existing technologies cannot accurately and efficiently extract financial announcement data.

[0038] Example 2 Please refer to Figure 3 This is a data extraction device for financial announcements provided in the embodiments of this application.

[0039] In this embodiment, the data extraction device for financial announcements includes an acquisition module 10, a filtering module 20, a processing module 30, and an extraction module 40.

[0040] Module 10 is used to acquire preprocessed financial announcement data and user query requests; The filtering module 20 is used to filter out the associated context that matches the user's query request from the financial announcement data based on the text structure features of the financial announcement data and in combination with a preset multi-way recall mechanism that includes semantic segmentation, directory navigation and keyword retrieval. Processing module 30 is used to perform terminology standardization processing on the tabular data in the associated context through a preset financial domain synonym mapping relationship; Extraction module 40 is used to input the standardized associated context into a preset large language model so that the large language model outputs a structured extraction result that conforms to the preset specifications. The large language model is obtained through supervised fine-tuning training based on a dataset from the financial field.

[0041] For ease of description and brevity, the embodiments of the device of the present invention include all the implementation methods in the above-described embodiments of the data extraction method for financial announcements, and will not be repeated here.

[0042] Example 3 This application provides a computer-readable storage medium, which includes a stored computer program, wherein the computer program, when running, controls the device where the computer-readable storage medium is located to execute the aforementioned method for extracting data from financial announcements. The method for extracting financial announcement data, when implemented as a software functional unit and used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0043] Example 4 This embodiment provides a terminal device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements a data extraction method for any of the financial announcements described in Embodiment 1.

[0044] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for extracting data from financial announcements, characterized in that, include: Obtain preprocessed financial announcement data and user query requests; Based on the text structure characteristics of financial announcement data, and combined with a pre-set multi-path recall mechanism that includes semantic segmentation, directory navigation and keyword retrieval, the associated context that matches the user's query request is filtered from the financial announcement data. By using a pre-defined synonym mapping relationship in the financial field, the tabular data in the associated context is subjected to terminology standardization processing; The standardized contextual information is input into a pre-defined large language model so that the large language model outputs structured extraction results that conform to the pre-defined specifications. The large language model is obtained through supervised fine-tuning training based on a dataset from the financial field.

2. The data extraction method for financial announcements according to claim 1, characterized in that, The specific steps for obtaining the preprocessed financial announcement data and user query requests are as follows: Obtain financial announcement data that has undergone format conversion and structure processing; the financial announcement data includes PDF files and scanned copies of annual reports, prospectuses, and rating reports, converted into structured text. Obtain user-initiated query requests; the query requests include single indicator value queries, table extraction queries, and multi-dimensional comprehensive analysis queries.

3. The data extraction method for financial announcements according to claim 1, characterized in that, Based on the text structure features of the financial announcement data, and combined with a pre-set multi-path recall mechanism including semantic segmentation, directory navigation, and keyword retrieval, the relevant context matching the user's query request is filtered from the financial announcement data, specifically as follows: The preprocessed financial announcement data is divided into independent paragraphs. The user query request and each independent paragraph are vectorized and converted. The vector similarity between the user query request and each independent paragraph is calculated. The paragraphs with similarity that meet the preset standard are semantically segmented for recall, and the semantic segmentation recall results are obtained. The intent of the user's query request is parsed, and the intent is matched with the original directory tree of the financial announcement data to locate the most relevant chapter. Semantic recall is performed only within the text scope corresponding to the chapter to obtain the directory navigation recall result. Based on the keyword matching algorithm, financial announcement data is retrieved, paragraphs containing the query keywords and related derivative terms are extracted, and the keyword retrieval results are obtained. The semantic block recall results, directory navigation recall results, and keyword retrieval recall results are merged and deduplicated to obtain a candidate context set; Based on the preset rearrangement model, and combined with the candidate context set, the associated context is obtained.

4. The data extraction method for financial announcements according to claim 3, characterized in that, The associated context is obtained by combining the preset rearrangement model with the candidate context set, specifically as follows: The intent type of the user query request is parsed, and the corresponding source confidence coefficient is assigned according to the three paths of semantic block recall, directory navigation recall and keyword retrieval. Based on the pre-defined reordering model, the semantic relevance of each paragraph in the candidate context set is scored, and the final score is calculated by combining the source confidence coefficient of the recall path to which the paragraph belongs. Paragraphs with final scores below a preset threshold are removed, and a preset number of paragraphs with the highest scores are selected from the remaining paragraphs to obtain the associated context.

5. The data extraction method for financial announcements according to claim 1, characterized in that, The terminology standardization process for the tabular data in the associated context is performed using a pre-defined financial domain synonym mapping relationship, specifically as follows: Identify tabular data from the associated context and extract the item names from the tabular data; The project names in the tabular data are compared with the preset financial field synonym mapping relationship, and the non-standard project names in the tabular data are replaced with the corresponding standard terms to obtain standardized tabular data. The synonym mapping relationship in the financial field includes standard terms and their corresponding synonyms, abbreviations and variants.

6. The data extraction method for financial announcements according to claim 1, characterized in that, The large language model is obtained through supervised fine-tuning training based on a dataset from the financial field, specifically as follows: Obtain multiple sets of training samples; each set of training samples includes a financial query command, a text fragment of a financial announcement related to the query command, and a standard structured output result that responds to the query command based on the text fragment of the financial announcement. Multiple sets of training samples are input into the initial large language model so that the initial large language model can be trained through multiple rounds of iteration. After obtaining the output results that conform to the preset structured specifications, the training stops and the large language model is obtained.

7. A data extraction device for financial announcements, characterized in that, include: The acquisition module is used to acquire preprocessed financial announcement data and user query requests; The filtering module is used to filter out the associated context that matches the user's query request from the financial announcement data based on the text structure features of the financial announcement data and in combination with a preset multi-path recall mechanism that includes semantic segmentation, directory navigation and keyword retrieval. The processing module is used to perform terminology standardization processing on the tabular data in the associated context through a preset financial domain synonym mapping relationship; The extraction module is used to input the standardized context into a preset large language model so that the large language model outputs a structured extraction result that conforms to the preset specifications. The large language model is obtained through supervised fine-tuning training based on a dataset from the financial field.

8. The data extraction device for financial announcements according to claim 7, characterized in that, Based on the text structure features of the financial announcement data, and combined with a pre-set multi-path recall mechanism including semantic segmentation, directory navigation, and keyword retrieval, the relevant context matching the user's query request is filtered from the financial announcement data, specifically as follows: The preprocessed financial announcement data is divided into independent paragraphs. The user query request and each independent paragraph are vectorized and converted. The vector similarity between the user query request and each independent paragraph is calculated. The paragraphs with similarity that meet the preset standard are semantically segmented for recall, and the semantic segmentation recall results are obtained. The intent of the user's query request is parsed, and the intent is matched with the original directory tree of the financial announcement data to locate the most relevant chapter. Semantic recall is performed only within the text scope corresponding to the chapter to obtain the directory navigation recall result. Based on the keyword matching algorithm, financial announcement data is retrieved, paragraphs containing the query keywords and related derivative terms are extracted, and the keyword retrieval results are obtained. The semantic block recall results, directory navigation recall results, and keyword retrieval recall results are merged and deduplicated to obtain a candidate context set; Based on the preset rearrangement model, and combined with the candidate context set, the associated context is obtained.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device on which the computer-readable storage medium is located to perform the data extraction method for financial announcements as described in any one of claims 1 to 6.

10. A terminal device, characterized in that, The system includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements the data extraction method for financial announcements as described in any one of claims 1 to 6.

Images