Finance report analysis method, device and equipment based on multi-source heterogeneous data processing

Abstract

The invention relates to the technical field of artificial intelligence, and discloses a financial report analysis method, and the method comprises the steps: carrying out the data preprocessing of historical financial association data and historical business operation data, and obtaining to-be-analyzed historical data; performing multi-order feature engineering processing on the to-be-analyzed historical data to obtain historical feature vector data; constructing an initial financial analysis large model, and training the initial financial analysis large model by using the historical feature vector data to obtain a trained financial analysis large model; accessing the trained large financial analysis model into a target system, and triggering the large financial analysis model to operate; and driving the large financial analysis model to perform multi-dimensional semantic analysis and quantitative reasoning on the to-be-analyzed financial report data to obtain visual financial report analysis result data. The method can be applied to internal financial statements of enterprises with businesses of science and technology finance, medical health, old-age care and the like, and the financial statement analysis efficiency and comprehensiveness can be improved through the multi-modal data fusion and knowledge enhancement large model architecture technology.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to a financial report analysis method and device based on multi-source heterogeneous data processing and equipment. BACKGROUND

[0002] In the process of industry financial report analysis, the existing technology mainly relies on manual processing of financial data, which has significant technical bottlenecks: manual statistical analysis of financial reports is prone to data omission, resulting in incomplete analysis conclusions; limited by manpower, the analysis dimension is single, and it is difficult to cover the deep correlation analysis of key indicators such as income growth rate and profit margin; in the face of massive business data, financial transaction data and other situations, manual processing cannot realize large-scale real-time processing, and the response efficiency is low; financial report note information generation and investment suggestions rely on subjective judgment, lack of intelligent analysis model support, and it is difficult to meet the real-time and accuracy needs of investors. The inefficiency and limitations of traditional manual mode result in insufficient timeliness of internal business analysis and external information disclosure. SUMMARY

[0003] The technical problem solved by the embodiments of the present application is that the traditional financial report analysis method is inefficient and not comprehensive.

[0004] To solve the above technical problems, the first technical solution adopted by the embodiments of the present application is to provide a financial report analysis method based on multi-source heterogeneous data processing, comprising: acquiring historical financial correlation data and historical business operation data of a preset type, performing data preprocessing on the historical financial correlation data and the historical business operation data through a preset rule engine to obtain historical data to be analyzed; based on a preset industry characteristic mapping rule, performing a preset multi-stage feature engineering processing on the historical data to be analyzed to obtain constructed historical feature vector data; constructing an initial financial analysis large model through a preset machine learning algorithm, training the initial financial analysis large model using the historical feature vector data to obtain a trained financial analysis large model; connecting the trained financial analysis large model to a target system, triggering the financial analysis large model to run through natural language questioning or index retrieval; generating an analysis instruction based on a prompt engineering, and using the analysis instruction to drive the financial analysis large model to perform multi-dimensional semantic analysis and quantitative reasoning on the financial report data to be analyzed to obtain visualized financial report analysis result data of a preset specification.

[0005] Optionally, the step of performing data preprocessing on the historical financial related data and the historical business operation data through a preset rule engine to obtain the historical data to be analyzed includes: performing format standardization operations on the historical financial related data and the historical business operation data, converting unstructured text data into structured fields through a natural language processing model, unifying the data units and storage formats of heterogeneous numerical data from different data sources, and obtaining standardized data; performing data cleaning on the standardized data, identifying and removing duplicate data, outliers and invalid records based on a preset cleaning rule library, wherein the cleaning rule library includes financial indicator threshold rules and business data logic rules, and obtaining cleaned data; filling missing data in the cleaned data using a multiple filling method or a prediction model based on machine learning, wherein numerical missing data is filled by prediction using a random forest regression model, and text missing data is completed by generating content using a generative pre-training model, and obtaining the historical data to be analyzed.

[0006] Optionally, the step of performing preset multi-order feature engineering processing on the historical data to be analyzed based on preset industry feature mapping rules to obtain constructed historical feature vector data includes: identifying and extracting financial indicator feature data and business operation feature data from the historical data to be analyzed according to the preset industry feature mapping rules; establishing a basic feature data set based on the financial indicator feature data and the business operation feature data; performing feature transformation operations on the basic feature data set, analyzing the correlation between features of the basic feature data, mapping the high-dimensional feature space of the basic feature data to a low-dimensional feature space, and obtaining a reduced-dimensional feature data set; performing feature intersection and combination operations on the reduced-dimensional feature data set, generating high-order composite features by constructing product terms and quotient terms between features; and performing preset data standardization processing on the high-order composite features to construct the structured historical feature vector data.

[0007] Optionally, the step of constructing an initial financial analysis large model using a preset machine learning algorithm, training the initial financial analysis large model using the historical feature vector data, and obtaining a trained financial analysis large model includes: constructing the initial financial analysis large model based on a preset neural network architecture, the initial financial analysis large model including a feature encoding module and an analysis result generation module, wherein the feature encoding module is used to perform feature extraction and representation learning on the input historical feature vector data; dividing the historical feature vector data into a training data set, a validation data set, and a test data set according to a preset ratio; using the training data set to iteratively train the initial financial analysis large model, and in each training cycle, adjusting the model parameters using a preset parameter optimization algorithm so that a preset loss function converges; using the validation data set to perform performance evaluation on the initial financial analysis large model during the training process, and adjusting the model training parameters based on the evaluation results, the training parameters including the learning rate, batch data size, and number of training rounds; using the test data set to perform performance testing on the trained initial financial analysis large model, calculating the performance score of the large model under preset evaluation indicators, and obtaining the trained financial analysis large model when the performance score is within a preset threshold range.

[0008] Optionally, the step of generating analysis instructions based on the prompt engineering, using the analysis instructions to drive the financial analysis big model to perform multi-dimensional semantic analysis and quantitative reasoning on the financial report data to be analyzed, and obtaining preset standardized visualized financial report analysis result data includes: based on the natural language question or indicator retrieval condition expression input by the user, generating the analysis instructions for the financial analysis big model through a prompt template generation module, wherein the prompt template includes industry knowledge base data, analysis target guidance data, and reasoning path constraint data; inputting the analysis instructions and the financial report data to be analyzed into the financial analysis big model, performing feature extraction and semantic representation on the financial report data to be analyzed through the feature encoding module of the financial analysis big model, and generating a multi-dimensional feature vector; performing financial semantic analysis and quantitative calculation on the multi-dimensional feature vector through the analysis result generation module of the financial analysis big model, and generating structured analysis data including financial report indicator correlation, time series characteristics, horizontal comparison results, and risk assessment results; inputting the structured analysis data into a visualization generation module, and generating the visualized financial report analysis result data including dynamic data charts, text interpretation data, and decision recommendation data through the visualization generation module according to preset chart type mapping rules.

[0009] Optionally, the analysis result generation module of the financial analysis large model performs financial semantic analysis and quantitative calculation on the multi-dimensional feature vector to generate structured analysis data including financial report indicator correlation, time series characteristics, horizontal comparison results and risk assessment results, including: constructing a semantic association network between financial report indicator data, identifying causal relationship data between financial report indicator data through the semantic association network, and calculating correlation strength data, setting the causal relationship data and the correlation strength data as the correlation relationship; identifying the trend feature data and periodic regularity data of the multi-dimensional feature vector changing with time through a time series analysis algorithm, and setting the trend feature data and periodic regularity data as the time series characteristics; performing horizontal comparison calculation on the multi-dimensional feature vector with pre-calculated industry average data and competitive target data to obtain the horizontal comparison result; calculating various risk indexes of preset data indicators based on a preset risk assessment model, and generating the risk assessment result according to the various risk indexes.

[0010] Optionally, the step of connecting the trained financial analysis big model to the target system and triggering the operation of the financial analysis big model through natural language questions or indicator retrieval includes: deploying the trained financial analysis big model in the target system and configuring the corresponding application interface, wherein the application interface supports receiving analysis requests in natural language form and analysis requests in indicator retrieval form; when receiving a natural language question input by the user, parsing the natural language question through the semantic understanding module, extracting the question intent and key entities, and generating a corresponding structured query statement; when receiving an indicator retrieval condition input by the user, converting the indicator retrieval condition into a query instruction that conforms to the input format that the financial analysis big model can receive; and inputting the structured query statement or the query instruction into the financial analysis big model.

[0011] To solve the above technical problems, the second technical solution adopted by the embodiment of the present application is: to provide a financial report analysis device based on multi-source heterogeneous data processing, including: a historical data acquisition module, used to obtain historical financial related data and historical business operation data of preset types, and perform data preprocessing on the historical financial related data and the historical business operation data through a preset rule engine to obtain historical data to be analyzed; a feature vector data module, used to perform preset multi-order feature engineering processing on the historical data to be analyzed based on preset industry feature mapping rules to obtain constructed historical feature vector data; a financial analysis big model module, used to construct an initial financial analysis big model through a preset machine learning algorithm, and use the historical feature vector data to train the initial financial analysis big model to obtain a trained financial analysis big model; a big model access module, used to access the trained financial analysis big model to the target system, and trigger the operation of the financial analysis big model through natural language questions or indicator retrieval; a financial report analysis result module, used to generate analysis instructions based on prompt engineering, and use the analysis instructions to drive the financial analysis big model to perform multi-dimensional semantic parsing and quantitative reasoning on the financial report data to be analyzed, to obtain preset standardized visual financial report analysis result data.

[0012] In order to solve the above technical problems, the third technical solution adopted in the embodiment of the present application is: to provide an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the financial report analysis method based on multi-source heterogeneous data processing as described above.

[0013] In order to solve the above technical problems, the fourth technical solution adopted in the embodiment of the present application is: to provide a non-volatile computer-readable storage medium, characterized in that the non-volatile computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are executed by an electronic device, the electronic device executes the financial report analysis method based on multi-source heterogeneous data processing as described above.

[0014] Different from the related technologies, this application converts heterogeneous data into standardized data through data preprocessing to solve the quality problem of original data, uses multi-order feature engineering combined with industry rules to optimize features, enhances data expression capabilities, builds models based on machine learning to realize automatic learning of financial data patterns, improves analysis accuracy, connects the model to the system and lowers the operation threshold through a dual-trigger mechanism, prompts engineering-driven model analysis and reasoning and generates visual results, thereby improving the efficiency of financial report analysis and the effectiveness of decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] One or more embodiments are exemplarily illustrated by corresponding drawings, which do not constitute limitations on the embodiments. Elements with the same reference numerals in the drawings are represented as similar elements, and unless otherwise stated, the figures in the drawings do not constitute proportional limitations.

[0016] Figure 1 This is a schematic diagram of the operating environment of the financial report analysis method based on multi-source heterogeneous data processing provided in an embodiment of the present application.

[0017] Figure 2 This is a schematic diagram of the execution flow of the financial report analysis method based on multi-source heterogeneous data processing provided in an embodiment of the present application.

[0018] Figure 3 It is a schematic diagram of the execution flow of the historical feature vector data constructed in the financial report analysis method based on multi-source heterogeneous data processing provided in an embodiment of the present application.

[0019] Figure 4 It is a schematic diagram of the execution flow of the large financial analysis model trained in the financial report analysis method based on multi-source heterogeneous data processing provided in an embodiment of the present application.

[0020] Figure 5 This is a schematic diagram of the system structure of a financial report analysis device based on multi-source heterogeneous data processing provided in an embodiment of the present application.

[0021] Figure 6 This is a hardware structure diagram of an electronic device that performs a financial report analysis method based on multi-source heterogeneous data processing provided in an embodiment of the present application. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0023] It should be noted that, if there is no conflict, the various features in the embodiments of the present application can be combined with each other and are all within the scope of protection of the present application. In addition, although the functional modules are divided in the device schematics and the logical order is shown in the flow charts, in some cases, the steps shown or described can be performed in a different order than the module division in the device schematics or the order in the flow charts.

[0024] Unless otherwise defined, all technical and scientific terms used in this specification have the same meanings as those commonly understood by those skilled in the art to which this application belongs. The terms used in this specification are intended only to describe specific embodiments and are not intended to limit this application. The term "and / or" as used in this specification includes any and all combinations of one or more of the relevant listed items.

[0025] To facilitate understanding of this embodiment, first, a financial report analysis method based on multi-source heterogeneous data processing disclosed in the embodiment of this application is introduced in detail. Figure 1 , Figure 1 Schematic diagram of the operating environment of the financial report analysis method based on multi-source heterogeneous data processing provided by the embodiment of the present application, such as Figure 1 As shown, the execution subject of the financial report analysis method based on multi-source heterogeneous data processing provided in the embodiment of the present application is generally an electronic device with certain computing capabilities, such as a computer device. In some possible implementations, the financial report analysis method based on multi-source heterogeneous data processing can be implemented by a processor calling computer-readable instructions stored in a memory. Figure 1 The computer device in the text can be a server. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. It can be understood that Figure 1 The number of computer devices in the figure is only for reference and can be expanded to any number according to actual needs.

[0026] Please continue reading Figure 2 , Figure 2 FIG. 1 is a schematic diagram of the execution flow of the financial report analysis method based on multi-source heterogeneous data processing provided by the embodiment of the present application. Figure 2 As shown, the following steps are included:

[0027] S1. Obtain historical financial related data and historical business operation data of a preset type, and pre-process the historical financial related data and historical business operation data through a preset rule engine to obtain historical data to be analyzed.

[0028] Among them, step S1, as the starting point of the entire financial report analysis process, focuses on building a high-quality training data foundation. Historical financial related data covers core report data such as the company's balance sheet, income statement, cash flow statement, and derivative data such as taxation, investment and financing that are closely related to finance. The preset type of historical business operation data includes business process data such as sales orders, customer management, and supply chain flows, which together constitute a multi-dimensional analysis data source. The preset rule engine serves as the hub of data preprocessing. Through built-in cleaning, conversion, and verification rules, it performs preliminary processing on the original data, such as format unification, outlier filtering, and missing value marking. It converts heterogeneous data scattered in different systems (for example, ERP, CRM) and different formats (structured tables, semi-structured logs, and unstructured documents) into standardized data sets that meet the requirements of subsequent analysis, providing reliable data input for subsequent feature engineering and model training.

[0029] As an optional implementation, the process of performing data preprocessing in the above step S1 may further specifically include the following steps S11 to S13.

[0030] S11. Perform format standardization operations on historical financial-related data and historical business operation data, convert unstructured text data into structured fields through a natural language processing model, unify the data units and storage formats of heterogeneous numerical data from different data sources, and obtain standardized data.

[0031] For example, unstructured text data (such as the Management Discussion and Analysis section in financial reports) cannot be processed directly by the model. Natural language processing (NLP) models (such as named entity recognition models and text classification models) are required to extract key information and convert it into structured fields (for example, quantifiable or categorizable attributes such as "market share" and "customer complaint rate"). For numerical data, since different systems may use different units (for example, "10,000 yuan" versus "100 million yuan") or storage formats (for example, CSV, JSON), it is necessary to uniformly convert it into model-compatible units (for example, standardizing it to "yuan") and formats (for example, unified two-dimensional table storage) through preset rules to ensure data consistency in physical storage and logical structure.

[0032] S12. Clean the standardized data and identify and remove duplicate data, outliers, and invalid records based on a preset cleaning rule library, where the cleaning rule library includes financial indicator threshold rules and business data logic rules to obtain cleaned data.

[0033] For example, the preset cleaning rule library integrates financial expertise and business logic. For example, the financial indicator threshold rules can define "gross profit margin should be between 0% and 100%", "asset-liability ratio should not exceed the industry warning line", etc., and automatically filter out obvious outliers (for example, records with a gross profit margin of 150%) through the rule engine. Business data logic rules verify the rationality of business processes, such as "order creation time must be earlier than payment time" and "delivery quantity must not exceed order quantity", and identify and delete contradictory records through logical judgment. The identification of duplicate data is usually based on unique identification fields (for example, transaction serial number, customer ID) to ensure the uniqueness of each record and avoid redundant data interfering with model training.

[0034] S13. Missing data in the cleaned data are filled using a multiple imputation method or a machine learning-based prediction model. Numerical missing data are predicted and filled using a random forest regression model, and text missing data are completed by generating content using a generative pre-training model to obtain historical data to be analyzed.

[0035] For example, for missing numerical data, the random forest regression model uses an ensemble learning algorithm and other relevant features (for example, using features such as "operating income" and "cost" to predict missing values ​​of "net profit") to build a prediction model, infer missing values ​​through the correlation between samples, and improve the accuracy of filling. For missing text data, generative pre-training models (for example, GPT-type models) learn language patterns based on large-scale text corpora and generate reasonable content based on contextual semantics (for example, completing missing risk warning paragraphs in financial reports). The multiple filling method generates multiple complete data sets through multiple simulations, comprehensively evaluates the uncertainty of the filling results, further improves the reliability of the data, and ensures the effectiveness of subsequent feature engineering and model training.

[0036] As an example, in the financial sector, banks collect historical financial data from customers, such as credit card spending records, loan repayment history, and wealth management product holdings. They also obtain preset types of historical business operations data, including customer credit assessment reports and customer service communication records. This data is then pre-processed using a preset rule engine, converting unstructured credit assessment text into structured credit score fields. The currency units of transaction data across different accounts are unified, and abnormal spending records and duplicate transactions are removed. Missing repayment dates are then predicted and filled in using machine learning models. Ultimately, the resulting historical data can be used for customer credit risk analysis and personalized financial product recommendations.

[0037] As another example, in healthcare, hospitals can access patients' historical financial data, such as medical expense settlement statements and medical insurance reimbursement records, as well as pre-set historical business operations data, including electronic medical records, examination and test reports, and treatment records. Using a pre-set rule engine, these data are processed, converting unstructured medical record text into structured disease diagnosis codes and symptom labels. Numerical data units generated by different testing equipment are standardized, duplicated test results are removed, and missing medication dosage data is filled in using a multi-pass imputation method. This generates historical data for analysis that can support reasonable medical expense assessments.

[0038] The above steps S11 to S13 form a complete data preprocessing chain through operations such as data format standardization, cleaning and missing data filling. Among them, format standardization converts unstructured text into structured fields and unifies the units and formats of numerical data, solving the problem of heterogeneous data sources, so that complex text and diverse numerical data can be efficiently processed by computers. Data cleaning eliminates duplicate, abnormal and invalid records based on the preset rule library, corrects erroneous information that violates financial logic and business rules, and improves data authenticity and reliability. Missing data filling uses random forest regression models, generative pre-training models and multiple filling methods for numerical and text data respectively to ensure that missing values ​​are reasonably completed and reduce uncertainty. The above series of operations jointly guarantee the standardization, accuracy and completeness of the data, provide a high-quality data foundation for subsequent feature engineering and model training, and significantly improve the efficiency, accuracy and model performance of financial report analysis.

[0039] S2. Based on the preset industry feature mapping rules, the historical data to be analyzed is subjected to preset multi-order feature engineering processing to obtain constructed historical feature vector data.

[0040] Multi-stage feature engineering is the key bridge between raw data and model training. By combining industry feature mapping rules with pre-set algorithms, the historical data to be analyzed is converted into structured feature vectors suitable for processing by machine learning models. This process not only optimizes data dimensions and enhances semantics, but also uses multi-stage processing strategies to uncover deep correlation patterns in the data, providing the model with more expressive input features, significantly improving the accuracy and generalization of financial report analysis.

[0041] As an optional implementation, please continue to refer to Figure 3 , Figure 3 : is a schematic diagram of the execution flow of the historical feature vector data constructed in the financial report analysis method based on multi-source heterogeneous data processing provided by the embodiment of the present application, such as Figure 3 As shown, the process of obtaining the constructed historical feature vector data in the above step S2 may specifically include the following steps S21 to S25.

[0042] S21. According to preset industry feature mapping rules, identify and extract financial indicator feature data and business operation feature data from the historical data to be analyzed.

[0043] The preset industry feature mapping rules are built based on the financial knowledge graph. Using semantic matching algorithms, they automatically identify key financial indicators (e.g., current ratio, net profit growth rate) and business operation indicators (e.g., new customer acquisition rate, supply chain cycle) in the historical data to be analyzed. This step utilizes natural language processing technologies such as named entity recognition and relationship extraction to accurately extract feature data highly relevant to industry analysis from unstructured text and structured tables, ensuring the professionalism and pertinence of feature selection.

[0044] S22. Establish a basic feature data set based on the financial indicator feature data and the business operation feature data.

[0045] The basic feature data set is established using feature indexing technology, which categorizes and stores extracted financial indicator and business operation features, and constructs a feature association matrix. By annotating feature metadata (e.g., data type, business meaning, and calculation logic), a standardized feature warehouse is formed, providing a unified data interface for subsequent feature engineering. This also supports rapid feature retrieval and reuse, improving processing efficiency.

[0046] S23. Perform feature transformation operations on the basic feature data set, analyze the correlation between the features of the basic feature data, map the high-dimensional feature space of the basic feature data to the low-dimensional feature space, and obtain a feature data set after dimensionality reduction.

[0047] The feature transformation operation is based on dimensionality reduction algorithms such as principal component analysis (PCA) and singular value decomposition (SVD). By calculating the Pearson correlation coefficient matrix between features, it identifies clusters of highly correlated features. When projecting the original high-dimensional feature space onto a low-dimensional subspace, principal components whose cumulative variance contribution exceeds a threshold (e.g., 95%) are retained. This significantly reduces the number of features while maximizing data variability and key information, effectively alleviating the curse of dimensionality.

[0048] S24. Perform feature crossover and combination operations on the feature data set after dimensionality reduction, and generate high-order composite features by constructing product terms and quotient terms between features.

[0049] For example, feature crossover and combination operations employ a feature generator network to automatically construct mathematical operations (e.g., multiplication and division) between features to generate higher-order composite features. For example, combining "accounts receivable" with "operating income" generates "accounts receivable turnover," while combining "fixed assets" with "total assets" generates "fixed asset ratio." These composite features capture nonlinear relationships between the original features, revealing deeper financial business logic and enhancing the model's ability to identify complex patterns.

[0050] S25. Perform preset data standardization processing on the high-order composite features to construct structured historical feature vector data.

[0051] For example, data normalization uses the Z-score method, which calculates the mean and standard deviation of each high-order composite feature and maps the feature values ​​to a standard normal distribution with a mean of 0 and a standard deviation of 1. This operation eliminates the impact of dimensional differences in different features on model training, ensuring more efficient convergence of the gradient descent algorithm. It also prevents features with large numerical ranges from receiving excessive weight in the model, improving model stability and predictive accuracy.

[0052] As an example, in the financial sector, when assessing customer credit, banks use pre-defined industry feature mapping rules to identify and extract financial indicator feature data from the historical data to be analyzed. These include the customer's debt-to-asset ratio and current ratio, as well as operational feature data such as credit card purchase frequency and loan repayment rate. This feature data is then classified and integrated to establish a basic feature data set. Feature transformations are then performed on this basic feature data set, analyzing correlations between features. This process maps the high-dimensional feature space to a lower-dimensional space, removing redundant indicators such as those that repeatedly reflect debt repayment ability, resulting in a reduced-dimensional feature data set. This reduced-dimensional feature data set is then cross-referenced and combined to construct higher-order composite features, such as the debt-to-income ratio, to more accurately measure debt repayment pressure. Finally, these high-order composite features are standardized to construct structured historical feature vector data, which helps generate financial reports that more accurately reflect the bank's operating conditions and risk levels.

[0053] As another example, in the healthcare sector, when hospitals conduct disease prediction analysis, they use pre-set industry feature mapping rules to extract financial indicator feature data from the historical data to be analyzed. This includes the patient's total medical expenses, medical insurance reimbursement ratio, and business operation feature data, such as the number of hospital stays and follow-up visits. After establishing a basic feature set from the extracted feature data, feature transformation is performed on it to analyze correlations and reduce dimensionality, removing features such as routine examination items that have weak correlations with the disease. The reduced features are then cross-referenced and combined, for example, combining patient age and illness duration to assess disease trends. Finally, standardization is performed to construct structured historical feature vector data, providing data support for financial reports and accurately presenting the hospital's financial revenue and expenditure and operational efficiency.

[0054] The above steps S21 to S25, through feature engineering processing in a hierarchical and progressive manner, systematically improve the data quality and model adaptability of financial report analysis. Based on the industry feature mapping rules, the financial and business operation feature data are accurately extracted to ensure that the feature selection fits the needs of financial report analysis and avoid interference from irrelevant data. The extracted features are integrated into a basic set, and a standardized data organization form is established to provide a clear framework for subsequent processing. Through dimensionality reduction operations, feature correlation is analyzed and redundancy is eliminated to solve the dimensionality disaster problem, reduce data complexity, and improve model training efficiency. High-order composite features are generated through feature intersection and combination, and the potential relationship between data is deeply explored, the feature expression ability is enhanced, and the model can capture more complex financial business logic. High-order features are standardized to eliminate dimensional differences, avoid imbalance of feature weights, and ensure the stability and convergence speed of model training. The above operations jointly construct structured historical feature vector data, provide high-quality input for the financial analysis model, and significantly improve the accuracy and effectiveness of financial report analysis.

[0055] S3. Build an initial financial analysis model using a preset machine learning algorithm, and use historical feature vector data to train the initial financial analysis model to obtain a trained financial analysis model.

[0056] This step involves building and training a large financial analysis model using a machine learning algorithm, transforming historical feature vector data into an intelligent model with predictive and analytical capabilities. The pre-set machine learning algorithm, based on a neural network architecture, optimizes model parameters through multiple rounds of iterative training, enabling it to automatically learn complex patterns and semantic associations within financial data. This process not only transforms data into knowledge but also, through the model's generalization capabilities, provides a reusable intelligent tool for financial report analysis, significantly improving analytical efficiency and accuracy.

[0057] As an optional implementation, please continue to refer to Figure 4 , Figure 4This is a schematic diagram of the execution flow of the financial analysis model trained in the financial report analysis method based on multi-source heterogeneous data processing provided by the embodiment of the present application, such as Figure 4 As shown, the process of obtaining the trained financial analysis model in the above step S3 may specifically include the following steps S31 to S35.

[0058] S31. Construct an initial financial analysis model based on a preset neural network architecture. The initial financial analysis model includes a feature encoding module and an analysis result generation module.

[0059] The feature encoding module is used to extract features and learn representations of the input historical feature vector data. For example, the preset neural network architecture uses the Transformer architecture as its foundation to construct an initial financial analysis model consisting of a feature encoding module and an analysis result generation module. The feature encoding module uses a multi-head self-attention mechanism to perform sequence modeling on the input historical feature vector data, capturing the temporal dependencies and semantic associations between financial indicators and converting high-dimensional and sparse feature vectors into low-dimensional and dense semantic representations. The analysis result generation module, based on the encoded feature representation, generates financial analysis results through a feedforward neural network, achieving a mapping from data features to analytical conclusions.

[0060] S32. Divide the historical feature vector data into a training data set, a validation data set, and a test data set according to a preset ratio.

[0061] For example, data partitioning uses a stratified random sampling strategy, dividing historical feature vector data into training, validation, and test sets according to preset proportions (e.g., 70%, 15%, and 15%). To ensure data distribution consistency, the sample ratios of various financial scenarios (e.g., profits, losses, and risk events) are maintained during the partitioning process to avoid model overfitting due to data skew. Independent validation and test sets are used for parameter tuning during model training and final performance evaluation, respectively, to ensure the model's generalization ability on unseen data.

[0062] S33. Use the training data set to iteratively train the initial financial analysis model. In each training cycle, adjust the model parameters through a preset parameter optimization algorithm so that the preset loss function converges.

[0063] For example, the iterative training process utilizes a variant of stochastic gradient descent (SGD) (e.g., the Adam optimizer). During each training cycle, batches of training data are fed into the model, predictions are calculated via forward propagation, and the gradient of the loss function is calculated via backpropagation. The default loss function combines mean squared error (MSE) and cross-entropy loss, taking into account both the accuracy of financial indicator forecasts and the precision of the classification task. By continuously adjusting model parameters, the loss function gradually converges, enabling the model to learn patterns in the financial data.

[0064] S34. Use the validation dataset to evaluate the performance of the initial financial analysis model during training, and adjust the model's training parameters based on the evaluation results. The training parameters include the learning rate, batch size, and number of training rounds.

[0065] During the performance evaluation phase, the model is monitored in real time using a validation dataset, calculating multiple evaluation metrics (e.g., accuracy, recall, F1 score, and root mean square error). Training parameters are dynamically adjusted based on these evaluation results. If the model shows signs of overfitting, the learning rate is reduced or the regularization strength is increased. If convergence is slow, the batch size is increased or the learning rate is increased. This adaptive adjustment mechanism ensures that the model maintains stable performance improvements during training, avoiding regression into local optimal solutions.

[0066] S35. Use the test data set to perform a performance test on the trained initial financial analysis model, calculate the performance score of the model under the preset evaluation indicators, and when the performance score is within the preset threshold range, the trained financial analysis model is obtained.

[0067] The final performance test uses an independent test dataset to calculate the model's comprehensive score based on pre-set evaluation metrics. These metrics include the accuracy of financial indicator forecasts, the recall rate of risk event identification, and the confidence level of analytical conclusions. When each performance score reaches the preset threshold, the model demonstrates sufficient generalization and stability for practical application. This rigorous testing mechanism ensures that the trained financial analysis model accurately captures the inherent patterns of financial data, providing reliable intelligent support for financial report analysis.

[0068] As an example, in the financial sector, when banks build customer credit risk assessment models, they use the Transformer architecture to construct the initial model. The feature encoding module learns semantic representations of historical customer feature vector data (e.g., debt-to-asset ratio, repayment history), capturing the temporal dependencies between financial indicators. The data is divided into training, validation, and test sets with a 70% / 15% / 15% ratio to ensure a balanced distribution of samples from customers with different credit ratings. The model is iteratively trained using the Adam optimizer, with a loss function combining mean squared error and cross-entropy loss, simultaneously considering the accuracy of default probability prediction and risk classification. During training, the learning rate and batch size are dynamically adjusted using the validation set to prevent overfitting. Model performance is finally evaluated on the test set, and model training is completed when the default prediction accuracy exceeds 90% and the F1 score reaches 0.85.

[0069] As another example, in the medical field, when a hospital builds a medical cost prediction model, it builds an initial model based on a preset neural network architecture. The feature encoding module performs time series modeling on the patient's historical feature vector data (for example, medical treatment items, length of hospitalization), and extracts features related to medical expenses. The data is stratified and sampled at a ratio of 7:1.5:1.5 to keep the sample ratios of different disease types consistent. The model is iteratively trained using a stochastic gradient descent algorithm, and the loss function focuses on the accuracy of medical cost prediction. The model performance is monitored through a validation set, and the regularization parameters and the number of training rounds are dynamically adjusted. When the root mean square error on the test set is lower than the industry threshold and the cost prediction deviation rate is controlled within 15%, the model training is completed. The model can assist in generating financial statements containing cost structure and trend analysis to optimize the allocation of medical resources.

[0070] Steps S31 to S35, through a hierarchical and progressive model construction and training strategy, systematically improve the accuracy and generalization capabilities of the financial analysis model. A feature encoding module based on the Transformer architecture enables deep semantic parsing of financial data. The self-attention mechanism captures complex dependencies between indicators, significantly enhancing the model's understanding of financial semantics. A stratified data sampling strategy ensures consistent data distribution across training, validation, and test sets, preventing model overfitting due to data skew and improving the model's adaptability to different scenarios. Combining the optimization objectives of mean squared error and cross-entropy loss balances the training performance of regression prediction and classification tasks, enabling the model to accurately predict financial indicator values ​​and identify risk event types. A dynamic training parameter adjustment mechanism ensures stable model convergence during training, avoiding local optimal solutions and accelerating model training efficiency. A rigorous multi-metric performance evaluation system ensures that the trained model meets industry standards, accurately captures the inherent patterns of financial data, provides reliable intelligent support for financial report analysis, and ultimately achieves efficient data-to-decision transformation.

[0071] S4. Connect the trained financial analysis model to the target system and trigger the operation of the financial analysis model through natural language questions or indicator retrieval.

[0072] The system integrates trained large financial analysis models into target systems, building a bridge between the model and actual application scenarios, transforming model training results into real business value. Triggering model execution through natural language questions or indicator retrieval eliminates the traditional reliance on specialized knowledge and complex operations for financial analysis. Flexible and interactive methods meet the diverse analytical needs of different users, allowing financial analysis capabilities to be deeply integrated into business processes, improving decision-making efficiency and accuracy.

[0073] As an optional implementation, the process of triggering the operation of the financial analysis model in the above step S4 may specifically include the following steps S41 to S44.

[0074] S41. Deploy the trained financial analysis model in the target system and configure a corresponding application program interface (API), wherein the application program interface supports receiving analysis requests in natural language form and analysis requests in indicator retrieval form.

[0075] Deploying a large financial analysis model in the target system essentially integrates the trained algorithm and parameters into the actual operating environment, enabling it to respond to external requests. Configuring the Application Programming Interface (API) establishes a standardized channel for interaction between the system and the model. By defining input and output formats, request protocols, and data transmission specifications, this API supports both natural language analysis requests and indicator retrieval. This ensures that the large model can seamlessly integrate with different systems (e.g., enterprise ERP and financial management software), providing a unified entry point for users to invoke the model's capabilities.

[0076] S42. When a natural language question input by a user is received, the natural language question is parsed by a semantic understanding module, the question intent and key entities are extracted, and a corresponding structured query statement is generated.

[0077] For example, when receiving a natural language question from a user, the semantic understanding module performs an in-depth analysis of the text based on natural language processing (NLP) technology. First, the text is broken down into word units through basic processing such as word segmentation and part-of-speech tagging. Then, named entity recognition (NER) technology is used to extract key entities (for example, "net profit", "2023 Q3"), and finally, semantic role labeling and intent classification algorithms are combined to clarify the core demands of the user's question (for example, "analyze the reasons for the changes in net profit in a certain quarter"), and natural language semantics are converted into structured query statements for easy model understanding and processing.

[0078] S43. When receiving the indicator search condition input by the user, convert the indicator search condition into a query instruction that conforms to the input format that can be received by the financial analysis model.

[0079] The system needs to adapt the format and perform semantic conversion for the user-entered indicator search criteria. Because the user-entered indicator search criteria may be expressed in a variety of ways (for example, "Query the debt-to-asset ratio for the past three years"), the system must use regular expression matching, keyword mapping, and other technologies to convert them into a standard format acceptable to the financial analysis model (for example, specific data field combinations and query syntax structures). This ensures that the search instructions accurately convey user needs while complying with the model input specifications, avoiding analysis errors caused by format incompatibility.

[0080] S44: Input the structured query statement or the query instruction into the financial analysis model.

[0081] Inputting structured query statements or converted query instructions into the financial analysis model is a key step in triggering the model to execute analysis tasks. This process passes data into the model through the API interface. The model's feature encoding module and analysis result generation module are activated in sequence, performing feature extraction, semantic understanding, logical reasoning, and quantitative calculation on the input information, ultimately outputting financial report analysis results that meet user needs.

[0082] As an example, in the financial sector, securities trading platforms connect trained financial analysis models to intelligent investment research systems, supporting natural language and indicator retrieval requests through standardized API interfaces. When an analyst inputs a natural language question such as "Interpret the reasons for the abnormal cash flow fluctuations in the 2023 annual report of a new energy vehicle company," the semantic understanding module identifies key entities and question intent, and generates multi-dimensional analysis instructions. If the indicator retrieval condition "Screen technology companies with a net profit growth rate >30% and an R&D investment ratio >15%" is entered, the system automatically converts it into a query expression that meets the input specifications of the large model. The processed request is sent to the model through the API. The large model integrates corporate financial reports, industry data, and market sentiment, and outputs an in-depth analysis report containing financial health scores and risk warning labels to provide quantitative support for investment decisions.

[0083] As another example, in the healthcare sector, a chain of medical institutions connected a large financial analysis model to the group's financial control platform and deployed an API gateway that supports bimodal requests. When the CFO posed a natural language question, "Analyze the impact of equipment procurement on profit margins at each branch hospital in Q1 2024," the semantic understanding module parsed the question and constructed a multivariate analysis framework. If the indicator search condition "Correlation between the proportion of drug revenue and medical insurance reimbursement rates by department" is entered, the system performs dimensional mapping and format conversion. After receiving the request, the large model combines medical business data with financial indicators to output a visual report containing cost driver analysis and resource allocation optimization suggestions.

[0084] The above steps S41 to S44 connect the large financial analysis model to the target system and trigger it to run in multiple ways, which significantly improves the flexibility and practicality of financial analysis. Deploying application programming interfaces breaks down system barriers, realizes seamless integration of models with various business systems, and deeply embeds financial analysis capabilities into daily workflows. The semantic understanding module converts natural language into structured queries, lowers the usage threshold, and enables non-professionals to easily initiate complex analysis requests, while the format conversion of indicator retrieval conditions meets the needs of professional users for accurate data queries. Through the two aforementioned triggering methods, the model can quickly respond to diverse analysis needs, process massive financial data in real time, and output in-depth analysis reports containing trend interpretation, risk warnings, and decision-making recommendations. It efficiently converts data value into business insights, provides strong support for scenarios such as financial investment decisions and medical resource allocation, and greatly enhances the efficiency and scientific nature of financial decision-making for enterprises and institutions.

[0085] S5. Generate analysis instructions based on the prompt project, and use the analysis instructions to drive the financial analysis model to perform multi-dimensional semantic analysis and quantitative reasoning on the financial report data to be analyzed, and obtain the preset standardized visual financial report analysis result data.

[0086] This step enhances the semantic understanding and reasoning capabilities of the large financial analysis model through prompt engineering technology, transforming user needs into executable analytical instructions and driving the model to conduct in-depth analysis of financial report data. Industry knowledge base data within the prompt template provides professional domain knowledge support, analytical objectives guide data to clarify the analysis direction, and reasoning paths constrain data to standardize calculation logic, collectively improving the accuracy and professionalism of model analysis. The combination of multi-dimensional semantic analysis and quantitative reasoning enables the model to not only identify surface features of data but also explore deep connections between indicators, generating analytical results that meet financial professional standards. Ultimately, visual presentation enhances the intuitiveness and effectiveness of decision support.

[0087] As an optional implementation, the process of obtaining the preset standard visualized financial report analysis result data in the above step S5 may specifically include the following steps S51 to S54.

[0088] S51. Based on the natural language question or indicator retrieval condition expression input by the user, the analysis instruction for the financial analysis large model is generated through the prompt template generation module, wherein the prompt template includes industry knowledge base data, analysis target guidance data and reasoning path constraint data.

[0089] Among them, the prompt template generation module generates analysis instructions based on the natural language questions or indicator retrieval conditions input by the user, combined with the preset prompt template structure. The industry knowledge base data contains prior knowledge such as standard values ​​of financial ratios and industry development trends, which are used to constrain the scope of model analysis. The analysis target guidance data clarifies the type of analysis that the user wants to obtain (for example, risk assessment, trend forecasting) through keyword matching and intent classification. The reasoning path constraint data stipulates the logical rules that should be followed during the analysis process (for example, the DuPont analysis system, the cash flow discount model). Combined with the formed prompt instructions, the user's fuzzy needs are converted into structured task descriptions that the model can understand, significantly improving the pertinence of the model analysis.

[0090] S52: Input the analysis instruction and the financial report data to be analyzed into the financial analysis model, perform feature extraction and semantic representation on the financial report data to be analyzed through the feature coding module of the financial analysis model, and generate a multi-dimensional feature vector.

[0091] For example, the feature encoding module uses the Transformer architecture to process the input financial report data to be analyzed, converting the financial text into a vector representation through the word embedding layer, and then using the multi-head self-attention mechanism to capture the temporal dependencies and semantic associations between different indicators. The multi-dimensional feature vector not only contains traditional financial indicators (such as net profit and debt-to-asset ratio), but also incorporates text semantic features (such as risk warnings in management discussions) and temporal features (such as quarterly data trends), forming a three-dimensional representation of the financial report data. The aforementioned representation method retains the original semantic information of the data while converting it into a mathematical expression suitable for model processing.

[0092] S53. Perform financial semantic analysis and quantitative calculation on the multi-dimensional feature vector through the analysis result generation module of the financial analysis large model to generate structured analysis data including correlation relationships of financial report indicators, time series characteristics, horizontal comparison results and risk assessment results.

[0093] The analysis result generation module performs financial semantic parsing and quantitative calculations based on the multi-dimensional feature vectors output by the feature encoding module. Knowledge graph technology is used to identify correlations between indicators (for example, the correlation between gross profit margin and market share). Time series analysis is used to explore data variation patterns. A horizontal comparison algorithm is combined to assess the company's relative position within the industry. Risk assessment models are used to identify potential financial risks. The generated structured analysis data is stored as a graph, with nodes representing financial indicators and edges indicating the strength of relationships between indicators. This fully preserves the logical chain and calculation basis of the analysis process.

[0094] S54: Input the structured analysis data into a visualization generation module, and generate the visualized financial report analysis result data including dynamic data charts, text interpretation data, and decision-making recommendation data through the visualization generation module according to preset chart type mapping rules.

[0095] The visualization generation module converts structured analysis data into intuitive visualizations based on preset chart type mapping rules. For time series data, it automatically matches line charts or bar charts to display trend changes. For correlation analysis, it generates heat maps or radar charts to present the relationships between indicators. For risk assessment results, it uses dashboards to display key risk points and safety thresholds. Text interpretation data uses natural language generation technology to convert data analysis results into professional, easy-to-understand text descriptions. Decision recommendation data is based on a preset business rule library and automatically generates executable action plans based on the analysis results. The resulting visualized financial report analysis results are presented to users through an interactive interface, supporting drill-down and filtering operations to meet the needs of different decision-making scenarios.

[0096] As another optional implementation, the above step S54 may further specifically include the following steps S541 to S544.

[0097] S541. Construct a semantic association network between financial report indicator data, identify causal relationship data between financial report indicator data through the semantic association network, calculate correlation strength data, and set the causal relationship data and the correlation strength data as the correlation relationship.

[0098] Among them, this step constructs a semantic association network between financial report indicators, which is essentially to model the relationship between financial data through a graph neural network (GNN). Financial report indicators are used as nodes, and algorithms such as co-occurrence frequency and Granger causality test are used to identify the causal relationship between indicators, such as the causal link between "increase in operating income" and "increase in sales expenses". The strength of the association is calculated through conditional probability, for example, the degree of impact of "increase in debt-to-asset ratio" on "increase in financial expenses". This semantic association network can not only capture explicit financial relationships (for example, net profit = total revenue - total cost), but also explore implicit associations (for example, the lag effect between R&D investment and market share), providing structured knowledge graph support for subsequent analysis and enhancing the model's understanding of financial logic.

[0099] S542. Identify trend feature data and periodic regularity data of the multi-dimensional feature vector that change over time through a time series analysis algorithm, and set the trend feature data and periodic regularity data as the time series feature.

[0100] Time series analysis uses a long short-term memory (LSTM) network combined with a wavelet transform algorithm to mine time series patterns in multi-dimensional feature vectors. LSTM captures long-term dependencies in data, for example, identifying a trend of declining net profit growth for three consecutive years. Wavelet transforms decompose the frequency components of a time series to extract the cyclical patterns of seasonal fluctuations (for example, the Q4 revenue peak for medical companies). The significance of cyclical patterns is verified using Fourier transforms and autocorrelation functions, ultimately using quantitative indicators such as trend slope and cycle length as time series features. These features not only reflect the historical evolution of financial data but also provide semantic information about the time dimension for future forecasts.

[0101] S543: Perform a horizontal comparison calculation on the multi-dimensional feature vector, the pre-calculated industry average data, and the competitive target data to obtain the horizontal comparison result.

[0102] Among them, horizontal comparison calculations are achieved by constructing a corporate financial portrait vector space, and cosine similarity calculations are performed on the multi-dimensional feature vectors of the company to be analyzed, the industry mean vector, and the competitive target vector. The industry mean data comes from a regularly updated database of listed company financial reports, and the competitive target data is extracted from industry research reports through text mining. The comparison results not only include the difference values ​​of traditional financial ratios (such as ROE and gross profit margin), but also calculate the company's comprehensive ranking score in the industry through principal component analysis. This comparison method breaks through the limitations of a single indicator, evaluates the relative position of a company in the industry from a multi-dimensional perspective, and discovers competitive advantages and gaps.

[0103] S544. Based on a preset risk assessment model, various risk indices of preset data indicators are calculated, and the risk assessment results are generated according to the various risk indices.

[0104] Among them, the preset data indicators include liquidity risk (for example, current ratio), debt repayment risk (for example, interest coverage ratio), market risk (for example, beta coefficient), etc. The risk contribution of each indicator is determined by expert weighting method and hierarchical analysis method. The risk index calculation adopts the Z-score model combined with the machine learning anomaly detection algorithm to identify outliers that deviate from the industry benchmark. The final risk assessment result not only includes a comprehensive risk score, but also clarifies the main risk drivers through attribution analysis (for example, "excessive accounts receivable turnover days lead to increased liquidity risk"), providing a quantitative basis for risk warning and control.

[0105] As an example, in the financial sector, when an investment bank analyzed the financial report of a new energy vehicle company, it prompted the template generation module to generate analysis instructions based on the user question "Assess the company's cash flow risk in 2023," combining industry knowledge bases (e.g., new energy industry subsidy policies) and reasoning path constraints (e.g., free cash flow discounting models). The feature encoding module converted the financial report data into a multi-dimensional feature vector containing information such as the cash cycle and operating cash flow. The analysis result generation module used a semantic association network to discover the causal relationship between extended accounts receivable turnover days and cash flow constraints. It used time series analysis to identify seasonal collection patterns and, by comparing it with the industry average, found that the company's cash flow management efficiency was lower than the industry median. The risk assessment model calculated a liquidity risk index of 72 (e.g., the industry threshold is 80). The visualization generation module integrated these results into a dynamic dashboard displaying cash flow trends, risk warning labels, and improvement suggestions.

[0106] As another example, in the healthcare sector, when a hospital group analyzed the financial reports of its oncology specialty hospitals, the prompting engineering team generated analysis instructions based on the user-entered metric search criteria, "Analyze the impact of equipment purchases on profit margins in Q1 2024," encompassing information such as medical equipment depreciation policies and industry profit margin benchmarks. The feature encoding module processed the financial report data, generating feature vectors containing information such as the equipment-to-revenue ratio and depreciation expense ratio. The analysis result generation module constructed a semantic association network between equipment investment and medical service revenue, identifying a strong correlation between equipment utilization and profit margin (e.g., correlation strength 0.82). Time series analysis revealed a significant increase in profit margins three months after equipment purchase (e.g., trend slope 0.15). A horizontal comparison revealed that the hospital's equipment efficiency was 12% higher than the industry average, and the risk assessment model indicated an equipment idle risk index of 45 (e.g., a safety threshold of 50). The visualization generation module transformed the analysis results into interactive charts, showcasing the equipment investment payback period, departmental performance comparisons, and resource optimization recommendations.

[0107] The financial report analysis method based on multi-source heterogeneous data processing provided in the embodiment of the present application constructs an intelligent analysis system through multi-stage technology. Natural language processing and missing filling technology are used in the data preprocessing stage to convert heterogeneous financial and business data into standardized data and establish a high-quality data foundation. Multi-order feature engineering combines industry knowledge graphs and dimensionality reduction algorithms to extract and generate high-order features, optimize data dimensions while mining deep associations, and improve feature expression capabilities. The model based on the Transformer architecture learns complex patterns of financial data and enhances analysis accuracy through stratified sampling and dynamic parameter adjustment. After the model is connected to the system, it is triggered by natural language and indicator retrieval, and the use threshold is lowered with the help of semantic understanding and format conversion. The prompt engineering drives the model to perform multi-dimensional analytical reasoning, combines knowledge graphs and visualization technology, and outputs structured analysis results to assist decision-making. The entire set of methods realizes the intelligence of the entire process, improves the efficiency, accuracy and decision-making support capabilities of financial report analysis, and provides reliable tools for multiple industries.

[0108] Please continue reading Figure 5 , Figure 5 : is a schematic diagram of the system structure of the financial report analysis device based on multi-source heterogeneous data processing provided by the embodiment of the present application, such as Figure 5 As shown, the financial report analysis device 50 based on multi-source heterogeneous data processing includes: a historical data acquisition module 51, a feature vector data module 52, a financial analysis large model module 53, a large model access module 54 and a financial report analysis result module 55.

[0109] The historical data acquisition module 51 is specifically used to acquire historical financial related data and historical business operation data of a preset type, and pre-process the historical financial related data and the historical business operation data through a preset rule engine to obtain historical data to be analyzed.

[0110] The feature vector data module 52 is specifically configured to perform preset multi-stage feature engineering processing on the historical data to be analyzed based on preset industry feature mapping rules to obtain constructed historical feature vector data.

[0111] The financial analysis large model module 53 is specifically used to construct an initial financial analysis large model through a preset machine learning algorithm, and use the historical feature vector data to train the initial financial analysis large model to obtain a trained financial analysis large model.

[0112] The large model access module 54 is specifically used to access the trained financial analysis large model to the target system, and trigger the operation of the financial analysis large model through natural language questions or indicator retrieval.

[0113] The financial report analysis result module 55 is specifically configured to generate an analysis instruction based on a prompt engineering, drive the financial analysis large model to perform multi-dimensional semantic analysis and quantitative reasoning on the financial report data to be analyzed using the analysis instruction, and obtain visualized financial report analysis result data in a preset specification.

[0114] As an optional implementation, the historical data acquisition module 51 is further specifically configured to perform a format standardization operation on the historical financial correlation data and the historical business operation data, convert unstructured text data into structured fields through a natural language processing model, unify data units and storage formats of heterogeneous numerical data from different data sources, and obtain standardized data; perform data cleaning on the standardized data, identify and remove duplicate data, outliers, and invalid records based on a preset cleaning rule library, wherein the cleaning rule library includes financial indicator threshold rules and business data logic rules, and obtain cleaned data; fill in missing data in the cleaned data using a multiple filling method or a prediction model based on machine learning, wherein numerical missing data is filled in through a random forest regression model, and text missing data is completed by generating content through a generative pre-training model, and obtain the historical data to be analyzed.

[0115] As an optional implementation, the feature vector data module 52 is further specifically configured to identify and extract financial indicator feature data and business operation feature data from the historical data to be analyzed according to the preset industry feature mapping rule; establish a basic feature data set according to the financial indicator feature data and the business operation feature data; perform feature transformation on the basic feature data set, analyze the correlation between the basic feature data, map the high-dimensional feature space of the basic feature data to a low-dimensional feature space, and obtain a reduced feature data set; perform feature cross and combination operations on the reduced feature data set, generate high-order compound features by constructing product terms and quotient terms between features; and perform a preset data standardization process on the high-order compound features to construct the structured historical feature vector data.

[0116] As an optional implementation, the financial analysis big model module 53 is also specifically used to construct the initial financial analysis big model based on a preset neural network architecture, and the initial financial analysis big model includes a feature encoding module and an analysis result generation module, wherein the feature encoding module is used to perform feature extraction and representation learning on the input historical feature vector data; the historical feature vector data is divided into a training data set, a verification data set and a test data set according to a preset ratio; the initial financial analysis big model is iteratively trained using the training data set, and in each training cycle, the model parameters are adjusted by a preset parameter optimization algorithm so that the preset loss function converges; the initial financial analysis big model during the training process is performance evaluated using the verification data set, and the training parameters of the model are adjusted based on the evaluation results, and the training parameters include learning rate, batch data volume and number of training rounds; the performance test of the trained initial financial analysis big model is performed using the test data set, and the performance score of the big model under the preset evaluation index is calculated. When the performance score is within the preset threshold range, the trained financial analysis big model is obtained.

[0117] As an optional implementation, the financial report analysis result module 55 is also specifically used to generate the analysis instructions for the financial analysis big model based on the natural language questions or indicator retrieval condition expressions input by the user through the prompt template generation module, wherein the prompt template includes industry knowledge base data, analysis target guidance data and reasoning path constraint data; the analysis instructions and the financial report data to be analyzed are input into the financial analysis big model, and the feature encoding module of the financial analysis big model is used to perform feature extraction and semantic representation on the financial report data to be analyzed to generate a multi-dimensional feature vector; the analysis result generation module of the financial analysis big model is used to perform financial semantic analysis and quantitative calculation on the multi-dimensional feature vector to generate structured analysis data including financial report indicator correlation, time series characteristics, horizontal comparison results and risk assessment results; the structured analysis data is input into the visualization generation module, and the visualization generation module is used to generate the visualized financial report analysis result data including dynamic data charts, text interpretation data and decision recommendation data according to preset chart type mapping rules.

[0118] As an optional implementation, the financial report analysis result module 55 is also specifically used to construct a semantic association network between financial report indicator data, identify the causal relationship data between the financial report indicator data through the semantic association network, and calculate the correlation strength data, and set the causal relationship data and the correlation strength data as the correlation relationship; identify the trend feature data and periodic regularity data of the multi-dimensional feature vector changing over time through a time series analysis algorithm, and set the trend feature data and periodic regularity data as the time series feature; perform a horizontal comparison calculation on the multi-dimensional feature vector with the pre-calculated industry average data and the competitive target data to obtain the horizontal comparison result; based on a preset risk assessment model, calculate various risk indexes of the preset data indicators, and generate the risk assessment result according to the various risk indexes.

[0119] As an optional implementation, the big model access module 54 is also specifically used to deploy the trained financial analysis big model in the target system and configure the corresponding application interface, wherein the application interface supports receiving analysis requests in natural language form and analysis requests in indicator retrieval form; when receiving a natural language question input by a user, the natural language question is parsed by the semantic understanding module, the question intent and key entities are extracted, and a corresponding structured query statement is generated; when receiving an indicator retrieval condition input by the user, the indicator retrieval condition is converted into a query instruction that conforms to the input format that the financial analysis big model can receive; and the structured query statement or the query instruction is input into the financial analysis big model.

[0120] It should be noted that the aforementioned financial report analysis device based on multi-source heterogeneous data processing can execute the financial report analysis method based on multi-source heterogeneous data processing provided in the embodiments of this application, and possesses the corresponding functional modules and beneficial effects of executing the method. For technical details not fully described in the embodiments of the financial report analysis device based on multi-source heterogeneous data processing, please refer to the financial report analysis method based on multi-source heterogeneous data processing provided in the embodiments of this application.

[0121] Figure 6 is a schematic diagram of the hardware structure of an electronic device for executing a financial report analysis method based on multi-source heterogeneous data processing provided by an embodiment of the present application, such as Figure 6 As shown, the electronic device 600 includes:

[0122] One or more processors 610 and memory 620, Figure 6 A processor 610 is taken as an example.

[0123] The processor 610 and the memory 620 may be connected via a bus or other means. Figure 6 The bus connection is taken as an example.

[0124] Memory 620, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules, such as the program instructions / modules corresponding to the financial report analysis method based on multi-source heterogeneous data processing in the embodiments of the present application. Processor 610 executes the non-volatile software programs, instructions, and modules stored in memory 620 to execute various server functional applications and data processing, thereby implementing the financial report analysis method based on multi-source heterogeneous data processing in the aforementioned method embodiment.

[0125] The memory 620 may include a program storage area and a data storage area, wherein the program storage area may store an operating system and application programs required for at least one function; the data storage area may store data created based on the use of the financial report analysis device based on multi-source heterogeneous data processing, etc. In addition, the memory 620 may include a high-speed random access memory and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory 620 may optionally include a memory remotely located relative to the processor 610, and these remote memories may be connected to the financial report analysis device based on multi-source heterogeneous data processing via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0126] The one or more modules are stored in the memory 620, and when executed by the one or more processors 610, perform the financial report analysis method based on multi-source heterogeneous data processing in any of the above method embodiments, for example, perform the above described Figure 2 Steps S1 to S5 of the method, Figure 3 Steps S31 to S35 of the method, Figure 4 Steps S41 to S45 of the method are implemented Figure 5 The functions of modules 51-55 in.

[0127] The above-mentioned product can execute the method provided in the embodiment of this application, and has the functional modules and beneficial effects corresponding to the execution method. For technical details not fully described in this embodiment, please refer to the method provided in the embodiment of this application.

[0128] An embodiment of the present application provides a non-volatile computer-readable storage medium, wherein the non-volatile computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are executed by one or more processors, for example Figure 6 A processor 610 in the embodiment may enable the one or more processors to execute the financial report analysis method based on multi-source heterogeneous data processing in any of the above method embodiments, for example, executing the above described Figure 2Steps S1 to S5 of the method, Figure 3 Steps S31 to S35 of the method, Figure 4 Steps S41 to S45 of the method are implemented Figure 5 The functions of modules 51-55 in.

[0129] The present application provides a computer program product, which includes a computer program stored on a non-volatile computer-readable storage medium. The computer program includes program instructions. When the program instructions are executed by the electronic device, the electronic device is enabled to perform the financial report analysis method based on multi-source heterogeneous data processing in any of the above method embodiments, for example, performing the above-described Figure 2 Steps S1 to S5 of the method, Figure 3 Steps S31 to S35 of the method, Figure 4 Steps S41 to S45 of the method are implemented Figure 5 The functions of modules 51-55 in.

[0130] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of this embodiment.

[0131] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a general hardware platform, or of course by hardware. Those skilled in the art can understand that all or part of the processes in the above embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above methods. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0132] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Based on the concept of the present application, the technical features in the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations in different aspects of the present application as described above. For the sake of simplicity, they are not provided in detail. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A financial report analysis method based on multi-source heterogeneous data processing, characterized in that: include: Acquire historical financial related data and historical business operation data of a preset type, and perform data preprocessing on the historical financial related data and the historical business operation data using a preset rule engine to obtain historical data to be analyzed; Based on preset industry feature mapping rules, the historical data to be analyzed is subjected to preset multi-order feature engineering processing to obtain constructed historical feature vector data; Constructing an initial financial analysis model using a preset machine learning algorithm, and training the initial financial analysis model using the historical feature vector data to obtain a trained financial analysis model; Connect the trained financial analysis model to the target system and trigger the operation of the financial analysis model through natural language questions or indicator retrieval; Generate analysis instructions based on the prompt project, use the analysis instructions to drive the financial analysis model to perform multi-dimensional semantic analysis and quantitative reasoning on the financial report data to be analyzed, and obtain preset standardized visualized financial report analysis result data.

2. The financial report analysis method based on multi-source heterogeneous data processing according to claim 1 is characterized in that: The step of pre-processing the historical financial related data and the historical business operation data by using a preset rule engine to obtain the historical data to be analyzed includes: Performing format standardization operations on the historical financial related data and the historical business operation data, converting unstructured text data into structured fields through a natural language processing model, unifying the data units and storage formats of heterogeneous numerical data from different data sources, and obtaining standardized data; Cleaning the standardized data to identify and remove duplicate data, outliers, and invalid records based on a preset cleaning rule library, wherein the cleaning rule library includes financial indicator threshold rules and business data logic rules, thereby obtaining cleaned data; The missing data in the cleaned data are filled using a multiple filling method or a prediction model based on machine learning, wherein the numerical missing data are filled by prediction using a random forest regression model, and the text missing data are completed by generating content using a generative pre-training model to obtain the historical data to be analyzed.

3. The financial report analysis method based on multi-source heterogeneous data processing according to claim 1 is characterized in that: The step of performing preset multi-stage feature engineering processing on the historical data to be analyzed based on preset industry feature mapping rules to obtain constructed historical feature vector data includes: Identifying and extracting financial indicator feature data and business operation feature data from the historical data to be analyzed based on the preset industry feature mapping rules; Establishing a basic feature data set based on the financial indicator feature data and the business operation feature data; Performing a feature transformation operation on the basic feature data set, analyzing the correlation between the features of the basic feature data, mapping the high-dimensional feature space of the basic feature data to a low-dimensional feature space, and obtaining a feature data set after dimensionality reduction; Performing feature crossover and combination operations on the feature data set after dimensionality reduction, and generating high-order composite features by constructing product terms and quotient terms between features; The high-order composite features are subjected to a preset data standardization process to construct the structured historical feature vector data.

4. The financial report analysis method based on multi-source heterogeneous data processing according to claim 1 is characterized in that: The steps of constructing an initial financial analysis model using a preset machine learning algorithm, training the initial financial analysis model using the historical feature vector data, and obtaining a trained financial analysis model include: Constructing the initial financial analysis model based on a preset neural network architecture, the initial financial analysis model comprising a feature encoding module and an analysis result generation module, wherein the feature encoding module is used to perform feature extraction and representation learning on the input historical feature vector data; Dividing the historical feature vector data into a training data set, a validation data set, and a test data set according to a preset ratio; Iteratively training the initial financial analysis model using the training data set, and in each training cycle, adjusting the model parameters using a preset parameter optimization algorithm so that a preset loss function converges; Using the validation data set to evaluate the performance of the initial financial analysis model during training, and adjusting the training parameters of the model based on the evaluation results, the training parameters including the learning rate, batch size, and number of training rounds; The performance of the trained initial financial analysis model is tested using the test data set, and the performance score of the model under the preset evaluation index is calculated. When the performance score is within the preset threshold range, the trained financial analysis model is obtained.

5. The financial report analysis method based on multi-source heterogeneous data processing according to claim 1 is characterized in that: The step of generating analysis instructions based on the prompt engineering, using the analysis instructions to drive the financial analysis model to perform multi-dimensional semantic analysis and quantitative reasoning on the financial report data to be analyzed, and obtaining preset standardized visualized financial report analysis result data includes: Based on the natural language question or indicator search condition expression input by the user, the analysis instruction for the financial analysis model is generated by the prompt template generation module, wherein the prompt template includes industry knowledge base data, analysis target guidance data and reasoning path constraint data; Inputting the analysis instruction and the financial report data to be analyzed into the financial analysis model, performing feature extraction and semantic representation on the financial report data to be analyzed by the feature coding module of the financial analysis model to generate a multi-dimensional feature vector; The analysis result generation module of the financial analysis model performs financial semantic analysis and quantitative calculation on the multi-dimensional feature vector to generate structured analysis data including the correlation relationship of financial report indicators, time series characteristics, horizontal comparison results and risk assessment results; The structured analysis data is input into a visualization generation module, and the visualization generation module generates the visualized financial report analysis result data including dynamic data charts, text interpretation data and decision-making recommendation data according to preset chart type mapping rules.

6. The financial report analysis method based on multi-source heterogeneous data processing according to claim 5 is characterized in that: The step of performing financial semantic analysis and quantitative calculation on the multi-dimensional feature vector by the analysis result generation module of the financial analysis large model to generate structured analysis data including financial report indicator correlation, time series characteristics, horizontal comparison results and risk assessment results includes: Constructing a semantic association network between the financial report indicator data, identifying causal relationship data between the financial report indicator data through the semantic association network, calculating correlation strength data, and setting the causal relationship data and the correlation strength data as the correlation relationship; Identifying trend feature data and periodic regularity data of the multi-dimensional feature vector that changes over time through a time series analysis algorithm, and setting the trend feature data and periodic regularity data as the time series features; Performing a horizontal comparison calculation on the multi-dimensional feature vector, the pre-calculated industry average data, and the competitive target data to obtain the horizontal comparison result; Based on a preset risk assessment model, various risk indices of preset data indicators are calculated, and the risk assessment results are generated according to the various risk indices.

7. The financial report analysis method based on multi-source heterogeneous data processing according to claim 1 is characterized in that: The step of connecting the trained financial analysis model to the target system and triggering the operation of the financial analysis model through natural language questions or indicator retrieval includes: Deploy the trained financial analysis model in the target system and configure a corresponding application programming interface (API), wherein the API supports receiving analysis requests in natural language format and analysis requests in indicator retrieval format; When receiving a natural language question input by the user, the semantic understanding module parses the natural language question, extracts the question intent and key entities, and generates a corresponding structured query statement; When receiving the index search condition input by the user, converting the index search condition into a query instruction that conforms to the input format that the financial analysis model can receive; The structured query statement or the query instruction is input into the financial analysis model.

8. A financial report analysis device based on multi-source heterogeneous data processing, characterized in that: include: A historical data acquisition module is used to acquire historical financial related data and historical business operation data of a preset type, and pre-process the historical financial related data and the historical business operation data using a preset rule engine to obtain historical data to be analyzed; A feature vector data module is used to perform preset multi-order feature engineering processing on the historical data to be analyzed based on preset industry feature mapping rules to obtain constructed historical feature vector data; A financial analysis model module is used to construct an initial financial analysis model using a preset machine learning algorithm, and train the initial financial analysis model using the historical feature vector data to obtain a trained financial analysis model; A large model access module is used to connect the trained financial analysis large model to the target system and trigger the operation of the financial analysis large model through natural language questions or indicator retrieval; The financial report analysis result module is used to generate analysis instructions based on the prompt project, and use the analysis instructions to drive the financial analysis model to perform multi-dimensional semantic analysis and quantitative reasoning on the financial report data to be analyzed, so as to obtain visual financial report analysis result data with preset specifications.

9. An electronic device, characterized in that: include: at least one processor; as well as, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the financial report analysis method based on multi-source heterogeneous data processing as described in any one of claims 1-7.

10. A non-volatile computer-readable storage medium, characterized in that: The non-volatile computer-readable storage medium stores computer-executable instructions. When the computer-executable instructions are executed by an electronic device, the electronic device executes the financial report analysis method based on multi-source heterogeneous data processing according to any one of claims 1 to 7.

Images