Financial product information structured data extraction method and data subject standardization system

Abstract

The invention relates to the technical field of financial data processing, and discloses a financial product information structured data extraction method and a data subject standardization system. A data acquisition module of the system obtains original information from a plurality of financial data sources and analyzes the original information to obtain structured data fragments; the feature extraction module collects key features based on the structured data fragments to obtain financial attribute features and subject coding features; the data processing module performs standardized learning on the features through the structured data fragments to generate a standard subject mapping array; the subject standardization module is combined with a standard subject system and a standard subject mapping array to control a subject matching process of a data standardization engine; the adjusting module collects context parameters of the data fragments to determine a dynamic subject deviation value, and corrects the matching direction of the data standardization engine. According to the system, efficient extraction and accurate standardization of multi-source financial product information can be realized, data processing standardization and accuracy are improved, and data application requirements of the financial industry are met.

Description

Technical Field

[0001] This invention relates to the field of financial data processing technology, specifically to a method for extracting structured data of financial product information and a data subject standardization system. Background Technology

[0002] In the process of digital transformation in the financial industry, financial products are becoming increasingly diverse, encompassing multiple categories such as bank wealth management products, fund products, insurance products, and bond products. The information storage formats, data dimensions, and subject definitions of different categories of financial products vary significantly. Currently, the data sources from which financial institutions obtain financial product information are extremely fragmented. These include structured data stored in internal core business systems, such as transaction records and key terms of product contracts, as well as semi-structured and unstructured data from external public channels, such as product filing documents issued by financial regulatory agencies, product brochures provided by third-party data platforms, and product introductions on financial information websites. This raw financial product information lacks standardized data formats; some data sources store data in tabular form, some present information in document form, and there is even a large amount of unstructured content existing in the form of images or scanned documents, significantly increasing the difficulty of data parsing.

[0003] In actual data processing, due to inconsistent definition standards for financial product attributes among different financial institutions, the same attribute may have multiple expressions. For example, "expected rate of return" may be expressed as "expected annualized return" or "expected annualized rate of return." Furthermore, different attributes may have similar names, making it difficult to form a unified attribute feature system after data extraction. At the same time, regarding account coding, each financial institution often formulates its own internal coding rules based on its business needs, resulting in significant differences in code length and coding logic. For example, some institutions use 6-digit numeric codes, while others use 8-digit alphanumeric codes. Moreover, the codes for the same account can be completely different in different institutions, while different accounts may have similar codes in different institutions, further exacerbating the difficulty of data standardization.

[0004] Data extraction for financial product information often relies on single-template matching, which can only process data sources with specific formats. When faced with diverse data sources, the accuracy drops significantly, and much unstructured data cannot even be effectively parsed, leading to incomplete data collection. In the data standardization stage, traditional methods depend heavily on manually defining mapping rules to match extracted data with standard subjects. This is not only costly in terms of manpower and time, but also difficult to cover all data scenarios. When financial product information is updated or new data sources are added, the rules need repeated adjustments, resulting in poor adaptability. Furthermore, current technologies lack consideration for data context. Relying solely on a single data field during subject matching easily leads to mismatches. For example, when a data field value is in a fuzzy range, it cannot be comprehensively judged by considering contextual parameters such as the corresponding product type, issuing institution, and term. This makes it difficult to guarantee the accuracy of standardized data, impacting subsequent financial product analysis, risk assessment, and regulatory reporting, and failing to meet the financial industry's needs for efficient data processing and accurate application. Summary of the Invention

[0005] The purpose of this invention is to provide a method for extracting structured data of financial product information and a data subject standardization system to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a system for extracting structured data and standardizing data categories of financial product information, the system comprising: The data acquisition module is used to obtain raw financial product information from multiple financial data sources, and to parse the raw financial product information to obtain structured data fragments. The feature extraction module is used to collect key features based on the structured data fragments to obtain the financial attribute features and account code features corresponding to each data fragment. The data processing module is used to perform standardized learning on the various financial attribute features and subject coding features of the structured data fragments to obtain a standard subject mapping array; The subject standardization module is used to obtain a standard subject system and, based on the mapping values ​​of each standard subject in the standard subject mapping array and their corresponding data fragments, control the processing flow of the data standardization engine when matching subjects. The adjustment module is used to collect context parameters of the current data segment during the standardization process to determine the dynamic subject deviation value, and to correct the matching direction of the data standardization engine based on the dynamic subject deviation value and the standard subject mapping value of the current data segment.

[0007] Preferably, the feature extraction module specifically includes: A region segmentation unit is used to acquire the structured data fragments and perform region segmentation based on the information density in the structured data fragments to obtain multiple data regions. The attribute acquisition unit is used to obtain the data classification direction, select an initial data region for financial attribute feature acquisition based on the data classification direction, select the next data region for financial attribute feature acquisition based on the data classification direction, and so on until the financial attribute features corresponding to each data region are determined. The encoding acquisition unit is used to acquire encoding rule vectors, reselect an initial data region based on the encoding rule vectors to acquire subject encoding features, select the next data region based on the encoding rule vectors to acquire subject encoding features, and so on, until the subject encoding features corresponding to each data region are determined.

[0008] Preferably, the feature extraction module further includes: obtaining the data classification direction and encoding rule vector at the current moment through the big data center of the financial data platform.

[0009] Preferably, the data processing module specifically includes: A depth determination unit is used to determine the depth of the comparison region based on the structured data fragment; The first learning unit is used to acquire the financial attribute features and subject coding features corresponding to the first data region, and select a comparison data region for standardized learning based on the location coordinates of the first data region and the depth of the comparison region to obtain the standard subject mapping value corresponding to the first data region. The array filling unit is used to obtain the financial attribute features and account coding features corresponding to other data areas, determine the standard account mapping values ​​corresponding to each data area in the same way, and fill the array according to the structured data fragments to obtain the standard account mapping array.

[0010] Preferably, the subject standardization module specifically includes: The mapping value acquisition unit is used to acquire the mapping values ​​of each standard subject in the standard subject mapping array and their corresponding data areas. The path acquisition unit is used to acquire all processable path segments of the data standardization engine. The adjacent region acquisition unit is used to acquire each processable path segment and the adjacent data region corresponding to the processable path segment; The trajectory determination unit is used to obtain the starting point of the processing trajectory of the data standardization engine, weight and summarize the standard subject mapping values ​​of all adjacent data areas corresponding to the processable path segment as the priority score corresponding to the processable path segment, and use a graph search algorithm to determine the processing trajectory of the data standardization engine when performing subject matching from the starting point of the processing trajectory.

[0011] Preferably, the adjustment module specifically includes: The deviation calculation unit is used to obtain the deviation difference between the dynamic subject deviation value and the standard subject mapping value in the current data area. The threshold comparison unit is used to determine the correction depth based on the context parameters of the current data region when the deviation difference is higher than a preset threshold, and to obtain the dynamic subject deviation value and standard subject mapping value of the adjacent data region based on the correction depth. The correction unit is used to correct the matching direction of the data standardization engine based on the dynamic subject deviation value and standard subject mapping value of the adjacent data area.

[0012] Preferably, the system further includes: The index building module is used to perform correlation analysis between historical structured data fragments and standardized response patterns through the data processing module, and to set a first directional correlation index structure; An index matching library is used to store the first directional association index structure; A bidirectional retrieval engine is used to perform bidirectional traversal of the first directional associated index structure in the index matching library, establish a standardized parameter vector for the subject, and use a prediction model to provide feedback verification of the standardized parameter vector for the subject. The adjustment module is also used to receive feedback verification results and correct the matching direction of the data standardization engine.

[0013] Preferably, the feature extraction module further includes: The keyword extraction unit is used to extract thematic keywords from the structured data fragments, compare the thematic keywords with standard subject keywords based on statistical feature technology, and select matching data fragments as candidate datasets. The scope analysis unit is used to analyze and process the data segments in the candidate dataset according to their respective subjects and data quantity to obtain the retrieval scope value. Based on the retrieval scope value, it determines whether the retrieval scope needs to be reduced. If so, it generates a reduction signal. The reduction unit is used to narrow the search scope based on the reduction signal, the topic keywords, and logical operators, and to obtain new candidate datasets.

[0014] Preferably, the feature extraction module further includes: The vector transformation unit is used to convert data fragments in a new candidate dataset into vector representations using natural language processing techniques, thereby obtaining semantic vectors for the data fragments. The similarity calculation unit is used to calculate the similarity between a data segment and a standard subject using cosine similarity.

[0015] Preferably, the system further includes: The correlation analysis unit is used to acquire the user's contextual feature data and combine it with the feature data of the corresponding new candidate dataset to perform correlation analysis and obtain the correlation value between the user's contextual feature data and the data fragment. The matching degree calculation unit is used to combine the similarity and correlation values ​​between data fragments and standard subjects, process and analyze them to obtain the matching degree, and recommend corresponding data fragments to users based on the matching degree.

[0016] Compared with the prior art, the beneficial effects of the present invention are: This system, through its data acquisition module, can obtain raw financial product information from multiple financial data sources and parse it to obtain structured data fragments. This module overcomes the limitations of traditional single-template matching, enabling adaptive parsing of data sources with different formats. Whether it's unstructured data in the form of tables, documents, or scanned images, it can be transformed into structured data fragments through multi-dimensional parsing algorithms. This effectively solves the problem of incomplete data acquisition caused by the single format of data sources and the difficulty in parsing unstructured data in traditional technologies, ensuring the comprehensiveness of financial product information collection and providing a complete data foundation for subsequent data processing.

[0017] The feature extraction module collects key features from structured data fragments, obtaining the financial attribute features and account coding features corresponding to each data fragment. By constructing a multi-dimensional feature recognition model, this module can automatically identify attribute and coding information in data fragments. Even when faced with attributes with similar expressions such as "expected rate of return" and "expected annualized return," it can classify them as the same financial attribute feature through semantic analysis and feature comparison. For account codes that differ significantly between different institutions, it can extract the essential features of the corresponding accounts through code structure analysis and code logic decomposition, forming a unified account coding feature system. This avoids the feature confusion caused by inconsistent attribute expressions and large differences in coding rules in traditional technologies, providing a standardized feature foundation for subsequent data standardization learning.

[0018] The data processing module performs standardized learning on various financial attribute features and account coding features from structured data fragments to obtain a standard account mapping array. This module employs machine learning algorithms to automatically learn the correlation between attribute features and account coding features from a large number of structured data fragments, as well as the correspondence between these features and standard accounts. This eliminates the need for repeated manual formulation and adjustment of mapping rules, significantly reducing labor and time costs. Furthermore, the machine learning algorithm has self-optimization capabilities. When financial product information is updated or new data sources are added, the learning model can be continuously optimized through new data fragments, constantly improving the standard account mapping array. This allows the mapping rules to dynamically adapt to data changes, solving the problem of poor adaptability of traditional manual rules and improving the flexibility and timeliness of data standardization.

[0019] The subject standardization module acquires the standard subject system and, based on the mapping values ​​of each standard subject in the standard subject mapping array and their corresponding data fragments, controls the processing flow of the data standardization engine during subject matching. This module, through the precise mapping relationships provided by the standard subject mapping array, guides the data standardization engine to efficiently complete subject matching according to predetermined logic, avoiding the blindness of traditional matching processes. During the matching process, a comprehensive judgment is made based on the corresponding data fragments to ensure that the matching results highly match the actual meaning of the data, reducing mismatches caused by single-field judgments, improving the accuracy of subject matching, and enabling the standardized data to accurately correspond to the standard subject system, meeting the data standardization requirements of the financial industry.

[0020] The adjustment module collects contextual parameters of the current data segment during the standardization process to determine the dynamic account deviation value. Based on this dynamic account deviation value and the standard account mapping value, it corrects the matching direction of the data standardization engine. This module fully considers the contextual information of the data, such as the type of financial product, issuing institution, product term, and risk level. By analyzing the deviation between these parameters and the standard account mapping value, it determines the dynamic account deviation value and then corrects the matching direction in real time. When data field values ​​are in a fuzzy range or are ambiguous, the matching logic can be adjusted based on the contextual parameters to ensure that the account matching results remain accurate and reliable even in complex data scenarios. This function further compensates for the matching error problem caused by traditional technologies ignoring data context, significantly improving the quality of standardized data. This better supports the core business work of financial institutions, such as subsequent product analysis, risk assessment, and regulatory reporting, helping financial institutions improve data application efficiency and business decision-making accuracy. Attached Figure Description

[0021] Figure 1 This is a sequence diagram of the financial product information structured data extraction and data subject standardization system described in this invention; Figure 2 This is a schematic diagram illustrating the working principle of the feature extraction module. Figure 3 This is a schematic diagram illustrating the working principle of the data processing module. Detailed Implementation

[0022] 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.

[0023] Please see Figure 1 This invention provides a method for extracting structured data of financial product information and a data subject standardization system. The system includes: a data acquisition module, a feature extraction module, a data processing module, a subject standardization module, and an adjustment module.

[0024] The data acquisition module obtains raw financial product information from multiple financial data sources, including public market data platforms, internal databases of financial institutions, and third-party data service interfaces. The module periodically crawls data using web scraping technology or API interfaces and parses the raw information. The parsing process involves natural language processing (NLP) to identify structured elements in the text, generating structured data fragments stored in JSON or XML format. The feature extraction module receives these structured data fragments and extracts financial attribute features and account coding features through key feature acquisition operations. Financial attribute features include product type, issuer information, and yield indicators, while account coding features are identified based on industry standard coding systems such as COA (Account Code A). The data processing module uses machine learning algorithms to standardize the financial attribute features and account coding features. The standardization process employs a supervised learning model, using historical data as a training set to generate a standard account mapping array. This array is a multi-dimensional array structure that stores the mapping relationship between each data fragment and a standard account. The subject standardization module integrates a standard subject system, which predefines a standardized subject library from financial regulatory agencies. The module controls the data standardization engine's processing flow based on mapping values ​​in the standard subject mapping array. This data standardization engine is a rule-based engine that prioritizes paths with high matching degrees based on mapping values ​​when performing subject matching operations. The adjustment module monitors the context parameters of data segments in real time during the standardization process. These context parameters include timestamps, data source credibility indicators, and historical matching records. It calculates dynamic subject deviation values ​​to reflect the current matching direction offset and dynamically corrects the data standardization engine's matching direction based on the deviation value and standard subject mapping values ​​to ensure matching accuracy.

[0025] Example 1: See Figure 2 The implementation of the feature extraction module relies on the collaborative work of multiple internal units to achieve in-depth processing of structured data fragments. The region segmentation unit first receives structured data fragments from the data acquisition module. These fragments are usually presented as semi-structured data in a certain hierarchical format. The core task of this unit is to analyze the information distribution density within the fragments. The information density is evaluated based on multiple dimensions such as the coverage of effective data within fields, the diversity of field types, and the frequency of numerical occurrences. By applying clustering algorithms, the unit can identify naturally formed clustered regions in the data fragments. High-density regions usually correspond to field groups containing core financial values ​​(such as interest rates and prices), while low-density regions may contain descriptive text or auxiliary information. The segmentation process not only considers the static distribution of the data but also dynamically adjusts the clustering parameters in conjunction with historical processing patterns to ensure that the segmentation results not only reflect the characteristics of the current data but also form a good connection with subsequent processing steps. Finally, multiple data regions output are assigned unique identifiers, and their position coordinates and density levels in the entire fragment are recorded, providing clear input objects for subsequent feature acquisition.

[0026] The attribute acquisition unit's functionality begins with obtaining external data classification directions. These directional instructions originate from the big data center of the financial data platform, an authoritative data source that aggregates industry consensus and the latest classification rules. Based on the acquired classification direction, the unit selects an initial processing region from the data region set generated by the region segmentation unit. The initial region selection strategy typically prioritizes the region with the highest information density or selects the most representative starting point based on historical experience. For the selected initial region, the unit performs financial attribute feature acquisition operations. This process involves parsing and semantically annotating the region content to identify key attributes such as asset class, risk level, and issuing institution. During the acquisition process, predefined attribute dictionaries and matching rules are invoked to ensure that the extracted feature labels conform to industry-standard terminology. After completing the acquisition of the initial region, the unit automatically determines the next region to be processed based on the processing logic implicit in the data classification direction. This order may be dynamically adjusted based on the region's physical location, logical correlation, or attribute importance. The iterative acquisition process continues until all identified data regions have been processed, ultimately generating a financial attribute feature set covering all regions. Each feature is associated with its source region and includes a confidence index.

[0027] The execution flow of the coding acquisition unit is parallel to that of the attribute acquisition unit, but its focus is on the extraction of subject coding features. This unit first obtains the coding rule vector, which is also sourced in real-time from the big data center of the financial data platform. This vector defines the coding structure specifications, verification algorithms, and version information. Based on the search strategy specified by the coding rule vector, the unit reselects the initial processing region from the data region set. This initial region selection may favor regions containing coding identifiers (such as specific prefixes or fixed-length numeric strings). Within the selected region, the unit performs subject coding feature acquisition. This acquisition process deeply integrates rule matching and pattern recognition technologies, enabling it to handle variant forms and abbreviations of codes. The coding rule vector not only provides matching templates but also specifies verification steps during the acquisition process, such as verifying the validity of codes through checksums or cross-referencing. Based on the sequential logic defined by the coding rule vector, the unit sequentially selects subsequent data regions for acquisition, ensuring that all regions potentially containing coding information are covered. Finally, the output set of subject coding features and the set of financial attribute features correspond at the data region level, together constituting the complete output of the feature extraction module.

[0028] The powerful functionality of the feature extraction module relies in part on its real-time interaction with an external big data center. The financial data platform's big data center, acting as a dynamically updated knowledge base, continuously aggregates data classification standards and coding specifications from multiple authoritative financial institutions. Before initiating the processing flow, the module first requests the big data center to obtain the latest data classification direction file and coding rule vector file. This design allows the module to adapt instantly to adjustments in industry classification standards and updates to coding rules, such as responding to the emergence of new financial products or changes in regulatory policies. The data classification direction is typically a structured configuration file or data stream that defines not only the attribute categories to be collected but may also include processing priority suggestions and region selection hints. The coding rule vector is usually a set of parameterized matching and verification rules that guide the coding and acquisition unit to complete its work efficiently and accurately. Through this external dependency mechanism, the feature extraction module does not need to build in all possible rules and directions, maintaining high flexibility and scalability, and can keep pace with industry development trends.

[0029] Data flow and processing coordination between units within the module are managed by an internal scheduler. This scheduler monitors the completion status of the region segmentation unit and then triggers the initialization of the attribute acquisition and encoding acquisition units. The scheduler ensures that the processing progress of the two acquisition units is synchronized at the data region level and merges the results upon completion. All acquired feature data is written to a standardized output interface, which provides clear and structured feature data input for subsequent data processing modules. The entire implementation reflects a progressive processing philosophy from coarse-grained region segmentation to fine-grained feature extraction. Through the combination of dynamic external resource referencing and internal iterative processing mechanisms, efficient and accurate feature extraction of complex financial product information is achieved.

[0030] Example 2: See Figure 3 The depth determination unit performs analysis operations. This unit receives structured data fragments from the feature extraction module and parses their internal structure. Structured data fragments typically contain multiple data regions and their interrelationships. The depth determination unit quantifies the depth of the comparison region by analyzing the nesting level between regions, the strength of field associations, and historical interaction patterns. The comparison region depth is a dynamically calculated parameter that defines the range of neighboring regions that each data region needs to reference during the learning process. Its value is positively correlated with the complexity of the data structure. The calculation process comprehensively considers the region's type attributes, content stability, and logical distance from other regions. The final determined comparison region depth value is appended to the metadata of the structured data fragment, providing crucial computational basis for subsequent learning units.

[0031] The first learning unit then initiates the processing flow. This unit first acquires the financial attribute features and subject code features corresponding to the first data region. These feature data come from the output of the feature extraction module. The unit determines the learning starting point based on the position coordinates of the first data region in the overall structure. The position coordinates may be identified in the form of linear indexes or multi-dimensional coordinates. Combining the comparison region depth values ​​provided by the depth determination unit, the first learning unit selects a set of comparison data regions. These regions are usually located within a range centered on the current region and with the depth value as the radius. The standardization learning process uses a machine learning method based on similarity metrics to compare the distribution of the current region and the comparison regions in the feature space. The learning algorithm calculates feature weights and evaluates the degree of consistency between regions, ultimately outputting a standard subject mapping value representing the mapping confidence. This mapping value may be expressed in various forms such as probability values, similarity scores, or code identifiers, depending on the choice of the learning algorithm model.

[0032] The array filling unit is responsible for integrating the learning results into a complete mapping array. This unit sequentially acquires feature data from all remaining data regions, including financial attribute features and subject coding features. For each region to be processed, the unit adopts the same standardized learning process as the first learning unit, i.e., selecting a set of comparison regions based on the region's location coordinates and the depth of the comparison region. During the learning process, the consistency of algorithm parameters and processing logic is maintained to ensure that the standard subject mapping values ​​of all regions are comparable and consistent. After completing the calculation of the mapping values ​​of all regions, the unit begins to construct the standard subject mapping array. The structure design of the array matches the organization form of the original structured data fragments, and may use a two-dimensional matrix or a high-dimensional tensor for data organization. Each array position stores the standard subject mapping value of the corresponding data region, while retaining the correspondence between the region identifier and the mapping value. The final generated standard subject mapping array will serve as the basic input data for subsequent subject standardization processing.

[0033] The implementation of the subject standardization module relies on the collaborative work of multiple functional units. The mapping value acquisition unit first extracts each mapping value and its corresponding data area information from the standard subject mapping array. This unit parses the array data structure and establishes a correspondence table between mapping values ​​and area identifiers. At the same time, it verifies the validity and completeness of the mapping values ​​and filters out mapping results with low confidence or invalidity. The path acquisition unit accesses the internal structure of the data standardization engine and obtains all processable path segments supported by the engine. These path segments represent different processing strategies or algorithm branches that may be adopted in the subject matching process. Each path has applicable conditions and execution characteristics. The unit generates detailed description information of the path segments, including metadata such as path identifier, processing parameters, and performance characteristics.

[0034] The adjacent region acquisition unit analyzes the data access patterns of each processable path segment, identifying relevant data regions that may be involved in the processing of that path segment. The determination of these adjacent data regions is based on the processing logic and data dependencies of the path segment, and may include sets of spatially adjacent, logically related, or semantically similar regions. The unit establishes a list of adjacent regions for each path segment and records the correlation strength index between regions. Finally, the trajectory determination unit performs optimization calculations on the processing trajectory. This unit first determines the starting point of the data standardization engine's processing trajectory, which is usually the engine's initial state or the endpoint state of the previous processing task. For each processable path segment, the unit collects the standard subject mapping values ​​of all its adjacent data regions and uses a weighted aggregation algorithm to calculate the priority score of the path segment. The weighting is dynamically adjusted based on the confidence level of the mapping values, regional importance, and the real-time processing environment. Based on the priority scores of all path segments, the unit uses a graph search algorithm to construct a processing path graph, where nodes represent processing states and edges represent path segments. The algorithm explores all possible processing trajectories starting from the starting point, selects the optimal path based on the priority score, and finally determines the processing trajectory for the data standardization engine when performing subject matching. This processing trajectory will guide the engine to complete the subject standardization work in the most efficient way.

[0035] Example 3: The implementation of the adjustment module achieves dynamic correction through the collaborative work of the deviation calculation unit, threshold comparison unit, and correction unit. The deviation calculation unit continuously monitors the processing of the data standardization engine, obtaining the dynamic subject deviation value and standard subject mapping value of the currently processed data area. The dynamic subject deviation value is calculated by real-time analysis of context parameters, including indicators such as data timestamp, source data quality score, and historical matching consistency rate. The standard subject mapping value comes from the standard subject mapping array generated by the data processing module, representing the initial learning mapping relationship. The deviation calculation unit uses a difference measurement algorithm to calculate the deviation difference between the two, which reflects the degree of deviation between the actual processing result and the expected mapping. During the calculation process, the type characteristics of the data area and processing environment factors are considered to ensure the accuracy and comparability of the deviation difference.

[0036] After receiving the deviation difference, the threshold comparison unit performs a threshold judgment. The preset threshold is a dynamically adjusted parameter whose benchmark value comes from the statistical analysis of historical processed data and is adaptively adjusted according to changes in the processing environment. When the deviation difference is higher than the preset threshold, the unit determines that a correction process needs to be initiated. At this time, the unit determines the correction depth based on the context parameters of the current data area. The calculation of the correction depth comprehensively considers the magnitude of the deviation difference, the importance of the data area, and the timeliness requirements of processing. Based on the determined correction depth, the unit obtains the dynamic subject deviation values ​​and standard subject mapping values ​​of neighboring data areas within a specified range. The selection of these neighboring areas is based on the dual criteria of spatial proximity and logical correlation. The obtained neighboring area data undergoes validity verification and normalization processing to provide reliable input for subsequent correction.

[0037] The correction unit employs a weighted correction algorithm based on neighboring region data. This algorithm comprehensively considers the deviation and mapping confidence of each neighboring region to calculate a correction factor that adjusts the matching direction of the data standardization engine. The correction factor is calculated using the following formula: in: Represents the correction factor. Indicates the number of neighboring areas. Indicates the first The weighting coefficients of each neighboring region Show the first Dynamic subject deviation values ​​of neighboring regions Indicates the first Standard subject mapping values ​​for neighboring regions. Weighting coefficients. The correction factor is dynamically determined based on the correlation between the current region and neighboring regions, as well as data quality. The calculated correction factor is used to adjust the path selection weights of the data normalization engine, thereby achieving precise correction of the matching direction.

[0038] The index building module operates independently and processes and analyzes historical data. This module accesses a historical structured data fragment library and a standardized response pattern database, extracting successful matching cases from past processing records. Interrelation analysis employs an improved association rule mining algorithm to discover potential connections between data fragment features and standardized responses. The analysis process considers multi-dimensional association factors, including time features, data source features, and subject features. The generated first-level directional association index structure uses a multi-level index format, supporting efficient range queries and precise retrieval. The index structure is regularly updated to incorporate the latest processed data, maintaining its timeliness and accuracy. The index matching library, as a centralized storage system, manages multiple directional association index structures and their metadata. The library uses a distributed architecture to achieve high availability and scalability, supporting concurrent access and fast retrieval. The stored index structures are optimized and compressed, reducing storage space requirements while ensuring retrieval efficiency. The library maintains index version information and usage statistics, providing data support for index optimization.

[0039] The bidirectional retrieval engine provides efficient index query services. After receiving a query request, the engine loads the first directional related index structure from the index matching library. The bidirectional traversal adopts a strategy of scanning with both left and right pointers simultaneously. The left pointer traverses from front to back, and the right pointer traverses from back to front, until a match is found or the traversal is completed. During the traversal, the scanning step size and direction are dynamically adjusted to optimize query efficiency. A standardized parameter vector for each subject is established based on the traversal results. This vector contains multi-dimensional parameters such as matching degree score, response time prediction, and resource requirement estimation. The prediction model is trained using machine learning methods, using historical query data and behavior logs as the training set. The model performs feedback verification on the parameter vector and outputs verification results indicating the reliability and expected effect of the matching. The verification results are passed to the adjustment module for further improvement and correction decisions. The adjustment module integrates multi-source information from deviation calculation and feedback verification to form a comprehensive correction strategy. The module maintains a correction strategy library, storing correction parameters and processing rules for different scenarios. Based on the real-time processing status and verification results, the module dynamically selects the optimal correction strategy and adjusts the correction parameters. The correction results are fed back to the data standardization engine, affecting its subsequent path selection and matching decisions. At the same time, correction experience and results are recorded in the historical database, providing new data sources for index construction and predictive model training, forming a closed-loop optimization system.

[0040] Example 4; The keyword extraction unit of the feature extraction module receives structured data fragments from the data acquisition module. All fragments focus on the private equity fund sector, covering types such as securities, equity, and venture capital. For example, a data fragment of a securities-type private equity fund includes the fund name "Xinyuan Youxuan Securities Investment Private Equity Fund," the manager "Xintou Wealth Management Co., Ltd.," and the fundraising scale "5000." The data snippet for an equity private equity fund includes the following: Fund name: "Qihang Innovation Equity Private Equity Fund"; Manager: "Qifu Investment Management Co., Ltd."; Investment areas: "Next-generation information technology, high-end equipment manufacturing"; Duration: "7 years (2-year investment period + 5-year exit period)"; Minimum subscription amount: "1 million yuan"; Management fee: "2% of paid-in capital per year"; Data snippet for a venture capital private equity fund includes the following: Fund name: "Science and Technology Innovation Future Venture Capital Private Equity Fund"; Manager: "Science and Technology Innovation Investment Management Co., Ltd."; Investment stage: "Seed round, Angel round"; Investment direction: "Biomedicine, Artificial intelligence"; Duration: "8 years (3-year investment period + 5-year exit period)"; Profit distribution method: "Price recovered first, then profit sharing, excess returns are distributed at a 2:8 ratio".

[0041] The keyword extraction unit employs text analysis technology to extract core thematic keywords. For securities-related private equity fund segments, keywords are extracted from the name ("securities investment," "preferred choice"), investment scope ("Shanghai and Shenzhen A-shares," "treasury bond reverse repurchase," "public funds"), and lock-up period ("short-term lock-up"). For equity-related segments, keywords are extracted from the investment field ("next-generation information technology," "high-end equipment manufacturing"), duration ("long-term duration," "investment period," "exit period"), and minimum investment threshold ("high threshold"). For venture capital-related segments, keywords are extracted from the investment stage ("seed round," "angel round"), investment direction ("biomedicine," "artificial intelligence"), and profit distribution ("return on investment before profit sharing," "profit sharing on excess returns"). The extraction process combines word frequency statistics and semantic analysis, filtering out common words such as "limited company" and "ten thousand yuan" to ensure that keywords reflect the core attributes of the private equity fund segment.

[0042] Based on statistical feature extraction technology, the keyword extraction unit compares thematic keywords with a standard keyword database for private equity funds. This database is constructed according to the private equity fund subject classification system, including dimensions such as product type, investment target, operating mode, fee structure, and duration. Each subject corresponds to a feature vocabulary. For example, the "Securities Private Equity Funds" subject corresponds to "Securities Investment," "Shanghai and Shenzhen A-shares," "Treasury Bond Reverse Repurchase," and "Short-term Closed-end"; "Equity Private Equity Funds" corresponds to "Equity Investment," "Next-generation Information Technology," "Long-term Duration," and "Investment Period"; and "Venture Capital Private Equity Funds" corresponds to "Venture Capital," "Seed Round," "Biomedicine," and "Return on Investment Before Profit Sharing." During the comparison, in addition to perfect matching, semantic similarity calculations identify synonymous and related words. For example, "Shanghai and Shenzhen A-shares" and "A-shares," "Shanghai and Shenzhen Stocks," and "duration" and "operating cycle" are considered related or synonymous. Segments with matching scores exceeding a threshold are selected to form an initial candidate dataset. For example, segments containing "Securities Investment" and "Shanghai and Shenzhen A-shares" are included in the "Securities Private Equity Funds" candidate set, and segments containing "Equity Investment" and "Next-generation Information Technology" are included in the "Equity Private Equity Funds" candidate set.

[0043] The scope analysis unit statistically analyzes the subject distribution and data volume of the initial candidate dataset. It assumes that "Securities" has 80 segments, "Equity" has 60, and "Venture Capital" has 40. An additional 20 segments may fall under multiple subjects due to keyword ambiguity, resulting in a broad search scope. The unit calculates the search scope value, generating a reduction signal when it exceeds a preset threshold.

[0044] The reduced-cell search units construct precise search criteria based on thematic keywords combined with logical operators. The "Securities" candidate set uses "AND" to connect "securities investment" with "Shanghai and Shenzhen A-shares" or "reverse repurchase of treasury bonds," filtering out segments containing only "securities investment" without a clear underlying asset. The "Equity" candidate set uses "AND" to connect "equity investment" with "investment period" and "exit period," excluding segments confused with other funds. For segments with ambiguous attribution, "NOT" is used to exclude contradictory keywords; for example, segments containing "securities investment" and "long-term existence" are explicitly classified as "securities" using "NOT (long-term existence AND no lock-up period)." The final generated candidate dataset contains 65 entries for "securities," 50 for "equity," and 35 for "venture capital," with no ambiguous attribution segments and higher relevance.

[0045] The vector transformation unit uses natural language processing techniques to input fragments from the new candidate set into a pre-trained financial language model, converting them into high-dimensional semantic vectors. Security-related fragment vectors show high values ​​in the dimensions of "Investment Target - Equity" and "Investment Target - Fixed Income"; equity-related fragments are prominent in the dimensions of "Investment Field - Technology" and "Investment Field - Manufacturing"; and venture capital-related fragments have characteristics in the dimensions of "Investment Stage - Early Stage" and "Profit Distribution - Profit Sharing". The vector dimensions are unified, facilitating subsequent calculations.

[0046] The similarity calculation unit calculates the cosine similarity between the semantic vector of a segment and the semantic vector of its corresponding standard subject. The standard subject vectors are generated using the same technology. The "Securities" vector has significant values ​​in the dimensions of "Product Type - Securities Investment" and "Operating Cycle - Short to Medium Term"; the "Equity" vector is clearly defined in the dimensions of "Product Type - Equity Investment" and "Operating Cycle - Long Term"; and the "Venture Capital" vector is significant in the dimensions of "Product Type - Venture Capital" and "Investment Stage - Early Stage". The cosine similarity score ranges from 0 to 1, with a higher similarity score closer to 1. For example, a securities segment has a similarity of 0.92 to its corresponding standard subject, while an equity segment has a similarity of 0.88, achieving a precise match.

[0047] Example 5: Introducing correlation analysis between user scenario feature data and data fragments enables more personalized financial product information matching and recommendation. The correlation analysis unit first obtains user scenario feature data from the system user interface. This data originates from user registration information, historical operation records, and explicit preference settings. For example, a risk-averse individual investor's scenario feature data might include low risk preference, short investment horizon preference, and a history of primarily purchasing fixed-income products. Simultaneously, the unit obtains a new candidate dataset processed in Example 4. Each data fragment in this dataset has been converted into a semantic vector and its similarity to standard subjects has been calculated. The unit performs multi-dimensional correlation analysis on user scenario features and data fragment features. The analysis process employs a collaborative filtering-based recommendation algorithm to calculate the correlation between user feature vectors and data fragment feature vectors across various dimensions. For the risk preference dimension, the system compares the user's risk tolerance level with the product's risk rating. For the investment horizon dimension, it compares the user's liquidity needs with the product's maturity structure. The analysis also considers auxiliary features such as the user's investment experience level and asset size, and a weighted comprehensive calculation yields a quantitative correlation value. This correlation value reflects the degree of matching between the data fragment and the user's needs.

[0048] The matching score calculation unit receives the association value from the association analysis unit and the data segment similarity (standard subject similarity) from the similarity calculation unit. The unit uses a multi-factor fusion algorithm to comprehensively process these two numerical indicators. The algorithm assigns dynamic weights to the association value and similarity, based on the user's current operational context and system configuration strategy. For example, when a user explicitly performs a subject query, the weight of similarity may be appropriately increased, while when the user browses products, the weight of the association value may dominate. The matching score generated by the fusion calculation is a standardized value between 0 and 1, with a high score indicating that the data segment both meets the subject standard and satisfies the user's personalized needs. Based on the calculated matching score, the system sorts and filters the data segments in the candidate dataset. The filtering threshold is dynamically adjusted based on real-time factors such as user device type and network environment. For mobile users, the system may select the top few results with the highest matching scores for recommendation, while for desktop users, it may provide a richer recommendation list. The final recommendation results are presented in a structured manner through the user interface, with each recommendation item displaying key information and matching scores, allowing users to make further decisions based on this information.

[0049] The system maintains real-time monitoring of user feedback throughout the entire process. When users interact with the recommendation results, such as clicking to view details or making a purchase, this feedback data is recorded and used to update the user's contextual feature data. The updated feature data influences subsequent association analysis calculations, forming a continuously optimized personalized recommendation cycle. This dynamic adjustment mechanism enables the system to adapt to changes in user needs and provide more accurate information services. Data security and privacy protection are fully considered throughout the implementation process. User contextual feature data is stored and transmitted in encrypted form, and anonymization technology is used to remove personally identifiable information during data processing. The system provides transparency control functions, allowing users to view and modify the collected feature data, ensuring users' control over their personal information. All data processing operations comply with relevant financial data protection regulations, ensuring user data security while providing personalized services. For system performance optimization, a distributed computing architecture is adopted. Association analysis and matching degree calculation tasks are distributed to multiple computing nodes for parallel execution. Computing nodes dynamically scale according to load, ensuring low response latency even under high concurrency access. Caching mechanisms are widely used to store intermediate calculation results, such as relatively stable data like user feature vectors and data fragment vectors, reducing redundant calculation overhead.

[0050] Example 6: Taking the scenario of a securities-type private fund manager (hereinafter referred to as "Manager A") processing the valuation tables of 10 private securities investment funds under its management as an example, Manager A entrusts three institutions, namely Bank E, Securities Firm F, and Third-Party Custodian G, to be responsible for fund valuation and accounting. Every day, Manager A needs to receive the valuation table attachments sent by the three institutions via email, and after completing the data extraction and subject standardization, synchronize them to the internal investment research system and risk control system.

[0051] During the email access process, staff member A entered the email accounts of three custodian institutions (Tencent Enterprise Email for Bank E, NetEase VIP Email for Securities Firm F, and Alibaba Email for Third-Party Custody G) and authorization codes into the system. The system automatically identified the email type and matched the server address. For Bank E's Tencent Enterprise Email, the system first attempted a POP3 connection with SSL verification disabled. After failure, it switched to a POP3 connection with SSL verification enabled, successfully establishing communication. For Securities Firm F's NetEase VIP Email, both POP3 connections with and without SSL verification failed, and the connection was ultimately established via an IMAP server with SSL verification enabled. For Third-Party Custody G's Alibaba Email, the connection was established directly via an IMAP server with SSL verification enabled. The system was configured with an XXJOB scheduling task to retrieve emails daily at 8:30 AM. The initial retrieval recorded the initial cursor position, and subsequent retrievals started from the last cursor position to avoid duplicate retrievals. When an email subject contained keywords such as "valuation table" or "fund valuation daily report," and the sending account was the official email address of one of the three custodian institutions, attachments were automatically downloaded. Among them, Bank E sends an Excel file (filename: "Star No. 1 Private Equity Investment Fund Valuation Table - 20250901.xlsx"), Brokerage F sends a compressed file (filename: "Manager A Fund Valuation Table 20250901.zip"), the system calls the sevenzipjbinding package to decompress and obtain the Excel file, and the third-party hosting G sends an xls file. The system saves all the original data of the attachments to the database and records metadata such as email sending time and attachment format.

[0052] Subject identification phase: The system first processes the Excel file of Bank E to extract the product name. After obtaining the file name, it filters out interference words such as "valuation table" and "private equity securities investment fund", applies the date regular matching rule to eliminate "20250901", and locates the starting point of the product name as "Star No. 1" and the ending point as before "fund" through the keyword "fund", and extracts "Star No. 1 Private Equity Securities Investment Fund". When extracting the product code, first eliminate the date data. If the regular expression "[a-zA-Z][0-9a-zA-Z]{5}+" fails to match, then eliminate the 8-digit pure digital date, and extract the 6-digit code "XC0001" from the custody remarks through the regular expression "\d{6}+". Subsequently, start the multi-data source fusion matching, load the configuration mapping data for reverse fuzzy matching (standard_name LIKE '%Star No. 1%'), and query in the system-maintained product data and the fund association filing data according to "fundRecordNumber='XC0001'" and "productFullName like '%Star No. 1 Private Equity Securities Investment Fund%'", and match the unique parent fund data and the information of Manager A. The extraction of the product code of a certain fund valuation table of Brokerage F fails and is marked for manual intervention. After manually supplementing the code, it is updated to the configuration mapping data, and subsequent similar products can be automatically matched.

[0053] After completing the product information identification, the system calls the Apache POI library to read the first sheet of the Excel file, identify the table boundary, header, and row and column distributions, and convert them into a dictionary structure containing row labels, column labels, and cell values. Scan the header to locate the column of the subject code in the 2nd column through "subject code", scan down 5 columns (including 2 merged cells), and detect subject codes such as "1021" and "1031" in the 10th row through the regular expression "^1[0-2]\d{2}", and confirm that the details of the valuation table start from the 10th row. Locate the positions of each column according to the item keywords such as "subject code", "subject name", "quantity", "cost local currency", and "market value local currency", for example, "subject name" is in the 3rd column and "market value local currency" is in the 9th column; at the same time, find the summary data row at the end of the table through keywords such as "unit net value" and "net asset value", and record that "unit net value" is in the 5th column of the 150th row and "net asset value" is in the 5th column of the 151st row. Finally, clean the subject name text, remove the space in "bank deposit - fixed - accrued interest", convert the percentage sign in "market value ratio (%)" to a numerical format, and unify "2025.09.01" to "2025-09-01", and save the processed data in the database.

[0054] In the account mapping and normalization process, the direct mapping module is triggered for "102152-Securities Settlement Reserve Fund" in the valuation table of Bank E. The system queries the mapping rule base and matches the standard account code "M102111" and name "Settlement Reserve Fund_Securities Funds Account" using "Original Account Code = 102152" and "Original Account Name = Securities Settlement Reserve Fund", thus completing the mapping. In the valuation table of Brokerage F, the conditional mapping module is activated for "2221050201-Value-Added Tax Payable (Financial Services)". The rule base query finds that "IF Custodian Institution = Brokerage F and Account Contains '22210502' -> M222105-Value-Added Tax Payable". This account meets the condition and is mapped to "M222105-Value-Added Tax Payable".

[0055] In the third-party custodian valuation table, detailed accounts such as "10310601-China Merchants Securities" and "10310602-Guotai Junan Securities" are mapped automatically by the system. The parent account is configured as "M103106-Margin Deposits_Securities Firms" (a third-level account). The original account names are obtained, and after removing invalid characters, "China Merchants Securities" and "Guotai Junan Securities" are extracted. If the keyword "securities" is detected, the last part containing the keyword is taken as the core identifier. If no existing configuration is found, a unified account name is generated according to the rules: "M103106ZD" is encrypted with SHA256, then Base36 converted and truncated to 8 bits. "10310601-China Merchants Securities" is mapped to "M103106ZD78AB12CD", and "10310602-Guotai Junan Securities" is mapped to "M103106ZD90EF34GH". During the process, the third-party custodian G added a new subject "106101-Cross-border Derivative Valuation" without a matching rule, marked it as "Pending Review" and pushed it to the manual review interface. The reviewer manually mapped it to "M106101-Derivative Valuation_Cross-border" and updated it to the mapping rule library.

[0056] In the standardized data generation and output phase, the system assembles standard subject codes (such as M102111, M222105, etc.), raw values ​​(securities settlement reserve fund of 800,000 yuan, VAT payable of 50,000 yuan, etc.), and key metadata (valuation date 2025-09-01, currency CNY, portfolio ID XC0001, etc.) into JSON format data, for example, “subjectCode”:“M102111”, “subjectName”:“settlement reserve fund_securities fund account”, “amount”:800000, “valuationDate”:“2025-09-01”, “currency”:“CNY”, “fundId”:“XC0001”. This data is pushed to the investment research and risk control systems of Manager A through a standardized API interface, and update notifications are sent via ActiveMQ message queues to ensure that downstream systems obtain data in a timely manner for investment analysis and risk monitoring.

[0057] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0058] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A system for extracting structured data and standardizing data categories of financial product information, characterized in that, include: The data acquisition module is used to obtain raw financial product information from multiple financial data sources, and to parse the raw financial product information to obtain structured data fragments. The feature extraction module is used to collect key features based on the structured data fragments to obtain the financial attribute features and account code features corresponding to each data fragment. The data processing module is used to perform standardized learning on the various financial attribute features and subject coding features of the structured data fragments to obtain a standard subject mapping array; The subject standardization module is used to obtain a standard subject system and, based on the mapping values ​​of each standard subject in the standard subject mapping array and their corresponding data fragments, control the processing flow of the data standardization engine when matching subjects. The adjustment module is used to collect context parameters of the current data segment during the standardization process to determine the dynamic subject deviation value, and to correct the matching direction of the data standardization engine based on the dynamic subject deviation value and the standard subject mapping value of the current data segment.

2. The financial product information structured data extraction and data subject standardization system as described in claim 1, characterized in that, The feature extraction module specifically includes: A region segmentation unit is used to acquire the structured data fragments and perform region segmentation based on the information density in the structured data fragments to obtain multiple data regions. The attribute acquisition unit is used to obtain the data classification direction, select an initial data region for financial attribute feature acquisition based on the data classification direction, select the next data region for financial attribute feature acquisition based on the data classification direction, and so on until the financial attribute features corresponding to each data region are determined. The encoding acquisition unit is used to acquire encoding rule vectors, reselect an initial data region based on the encoding rule vectors to acquire subject encoding features, select the next data region based on the encoding rule vectors to acquire subject encoding features, and so on, until the subject encoding features corresponding to each data region are determined.

3. The financial product information structured data extraction and data subject standardization system as described in claim 1, characterized in that, The feature extraction module also includes: obtaining the data classification direction and encoding rule vector at the current moment through the big data center of the financial data platform.

4. The financial product information structured data extraction and data subject standardization system as described in claim 1, characterized in that, The data processing module specifically includes: A depth determination unit is used to determine the depth of the comparison region based on the structured data fragment; The first learning unit is used to acquire the financial attribute features and subject coding features corresponding to the first data region, and select a comparison data region for standardized learning based on the location coordinates of the first data region and the depth of the comparison region to obtain the standard subject mapping value corresponding to the first data region. The array filling unit is used to obtain the financial attribute features and account coding features corresponding to other data areas, determine the standard account mapping values ​​corresponding to each data area in the same way, and fill the array according to the structured data fragments to obtain the standard account mapping array.

5. The financial product information structured data extraction and data subject standardization system as described in claim 1, characterized in that, The subject standardization module specifically includes: The mapping value acquisition unit is used to acquire the mapping values ​​of each standard subject in the standard subject mapping array and their corresponding data areas. The path acquisition unit is used to acquire all processable path segments of the data standardization engine. The adjacent region acquisition unit is used to acquire each processable path segment and the adjacent data region corresponding to the processable path segment; The trajectory determination unit is used to obtain the starting point of the processing trajectory of the data standardization engine, weight and summarize the standard subject mapping values ​​of all adjacent data areas corresponding to the processable path segment as the priority score corresponding to the processable path segment, and use a graph search algorithm to determine the processing trajectory of the data standardization engine when performing subject matching from the starting point of the processing trajectory.

6. The financial product information structured data extraction and data subject standardization system as described in claim 1, characterized in that, The adjustment module specifically includes: The deviation calculation unit is used to obtain the deviation difference between the dynamic subject deviation value and the standard subject mapping value in the current data area; The threshold comparison unit is used to determine the correction depth based on the context parameters of the current data region when the deviation difference is higher than a preset threshold, and to obtain the dynamic subject deviation value and standard subject mapping value of the adjacent data region based on the correction depth. The correction unit is used to correct the matching direction of the data standardization engine based on the dynamic subject deviation value and standard subject mapping value of the adjacent data area.

7. The financial product information structured data extraction and data subject standardization system as described in claim 1, characterized in that, Also includes: The index building module is used to perform correlation analysis between historical structured data fragments and standardized response patterns through the data processing module, and to set a first directional correlation index structure; An index matching library is used to store the first directional association index structure; A bidirectional retrieval engine is used to perform bidirectional traversal of the first directional associated index structure in the index matching library, establish a standardized parameter vector for the subject, and use a prediction model to provide feedback verification of the standardized parameter vector for the subject. The adjustment module is also used to receive feedback verification results and correct the matching direction of the data standardization engine.

8. The financial product information structured data extraction and data subject standardization system as described in claim 1, characterized in that, The feature extraction module further includes: The keyword extraction unit is used to extract thematic keywords from the structured data fragments, compare the thematic keywords with standard subject keywords based on statistical feature technology, and select matching data fragments as candidate datasets. The scope analysis unit is used to analyze and process the data segments in the candidate dataset according to their respective subjects and data quantity to obtain the retrieval scope value. Based on the retrieval scope value, it determines whether the retrieval scope needs to be reduced. If so, it generates a reduction signal. The reduction unit is used to narrow the search scope based on the reduction signal, the topic keywords, and logical operators, and to obtain new candidate datasets.

9. The financial product information structured data extraction and data subject standardization system as described in claim 8, characterized in that, The feature extraction module further includes: The vector transformation unit is used to convert data fragments in a new candidate dataset into vector representations using natural language processing techniques, thereby obtaining semantic vectors for the data fragments. The similarity calculation unit is used to calculate the similarity between a data segment and a standard subject using cosine similarity.

10. The financial product information structured data extraction and data subject standardization system as described in claim 9, characterized in that, Also includes: The correlation analysis unit is used to acquire the user's contextual feature data and combine it with the feature data of the corresponding new candidate dataset to perform correlation analysis and obtain the correlation value between the user's contextual feature data and the data fragment. The matching degree calculation unit is used to combine the similarity and correlation values ​​between data fragments and standard subjects, process and analyze them to obtain the matching degree, and recommend corresponding data fragments to users based on the matching degree.

Images