Multi-source heterogeneous data fusion type financial analysis report generation method and system and medium

Abstract

The application provides a multi-source heterogeneous data fusion type financial analysis report generation method, system and medium, belongs to the technical field of financial analysis report generation, and relates to the technical field of financial analysis report generation.The method comprises the following steps: constructing a financial multi-modal large model, loading weight parameters, and importing a financial knowledge graph; analyzing an analysis target, generating an index demand list based on the financial knowledge graph, and associating data sources; collecting multi-source financial data from the data sources; extracting features from the multi-source financial data and outputting feature vectors; dynamically weighting and fusing the feature vectors through a cross-modal multi-head attention mechanism; based on the fused feature vectors, constructing a logical framework of the report through the financial multi-modal large model, generating a multi-modal financial analysis report, and adjusting the content granularity according to the user role; updating the fusion rules and weights of the feature vectors according to the user feedback behavior, and optimizing the semantic retrieval library matched with the feature vectors.The application breaks the data island, improves the generation efficiency and accuracy of the financial analysis report, and adapts to multi-scene requirements.

Description

Technical Field

[0001] This application belongs to the field of financial analysis report generation technology, specifically involving a method, system, and medium for generating financial analysis reports based on multi-source heterogeneous data fusion. Background Technology

[0002] With the development of enterprise digital transformation, financial analysis has shifted from traditional ex-post accounting to real-time decision support. Traditional financial reporting mainly relies on manual compilation of Excel spreadsheets, while automated tools often generate standardized reports based on fixed templates, such as balance sheets. In recent years, large-scale artificial intelligence models have been gradually applied to the field of financial analysis, but most existing systems are still limited to analyzing single data sources, such as generating profit analysis reports solely based on ERP data. They cannot effectively integrate unstructured data, such as audit reports, and thus struggle to meet the needs of complex financial scenarios.

[0003] Traditional financial analysis report generation suffers from the following shortcomings: First, weak data integration capabilities. Financial data is scattered across multiple systems such as ERP, CRM, and auditing systems, with significant differences in formats such as database fields, PDF tables, and handwritten annotations. Traditional tools lack a unified data access and preprocessing framework. Second, a lack of semantic consistency. Different departments have different definitions of the same financial indicators; for example, whether "cost" includes tax varies across departments. The lack of a unified semantic mapping mechanism leads to conflicting analysis conclusions. Third, low utilization of unstructured data. Key information (such as potential liabilities and risk warnings) contained in unstructured data such as audit reports and contract texts cannot be effectively extracted and utilized. Finally, insufficient dynamic adaptability. When faced with user-defined analysis requirements (such as specific impairment tests), the analysis model needs to be manually reconstructed, and real-time dynamic adjustments cannot be achieved after the response.

[0004] The relevant patents have generated reports using fixed templates, which only support structured data processing or focus on automated calculations from a single data source. They do not involve cross-source semantic fusion and dynamic weighting mechanisms, and lack a multimodal data processing framework specifically for the financial field. Summary of the Invention

[0005] In a first aspect, embodiments of this application provide a method for generating a multi-source heterogeneous data fusion-based financial analysis report, comprising the following steps:

[0006] S1. Integrates a structured processing engine, a semi-structured parsing module, an unstructured understanding module, a data access agent, a feature extraction agent, and an interactive agent to construct a large-scale financial multimodal model;

[0007] S2. Load the weight parameters of the pre-trained financial multimodal large model and import the financial knowledge graph;

[0008] S3. Parse the user's input analysis objectives through an interactive intelligent agent, generate a list of indicator requirements based on the financial knowledge graph, and associate it with the corresponding data source;

[0009] If an indicator in the indicator requirement list cannot be matched with a data source, obtain supplementary information from the user through the interactive intelligent agent;

[0010] S4. Collect multi-source financial data from associated data sources, including structured data, semi-structured data, and unstructured data, through a data access agent;

[0011] The feature extraction agent extracts features from the three types of data respectively and outputs feature vectors.

[0012] S5. The feature vectors from various data sources are mapped to a unified financial ontology semantic space by a feature extraction agent, and dynamic weighted fusion is performed through a cross-modal multi-head attention mechanism to generate a unified fused feature vector.

[0013] S6. Based on the fusion feature vector, the logical framework of the report is constructed through the financial multimodal big model, generating a multimodal financial analysis report that includes indicator analysis, anomaly interpretation and correlation reasoning, and adaptively adjusting the content granularity of the multimodal financial analysis report according to the user role;

[0014] S7. Based on user feedback behavior, update the feature vector fusion rules and weights through the financial multimodal large model, and optimize the semantic retrieval library for feature vector matching.

[0015] Furthermore, the specific steps for loading the weight parameters of the pre-trained financial multimodal large model in step S2 are as follows:

[0016] S21. Construct a dataset using labeled financial-related structured, semi-structured, and unstructured data, and train a large financial multimodal model.

[0017] S22. Import an industry-specific financial knowledge graph, which constructs the relationships between accounting subjects, report items, and business indicators;

[0018] S23. Configure the data access agent of the financial multimodal big model to connect to the financial database via JDBC / ODBC, call the market data interface via API, and recognize PDF reports via the OCR module.

[0019] Furthermore, the specific steps in step S3 of parsing the user input analysis target through the interactive intelligent agent are as follows:

[0020] S31. Perform intent recognition on the analysis target input by the user in natural language and extract key analysis dimensions;

[0021] S32. Based on the financial knowledge graph, key analysis dimensions are mapped to a set of standard financial indicators to generate an indicator requirement list;

[0022] S33. Associate known data sources with the indicator data in the indicator requirement list;

[0023] S34. When there is an indicator in the indicator requirement list that cannot be associated with any data source, the interactive agent is activated to generate a query statement to prompt the user to specify the data source of the corresponding indicator, and the new data source mapping relationship fed back by the user is stored in the knowledge base.

[0024] Furthermore, in step S4, the feature extraction agent performs feature extraction on the three types of data respectively. The specific steps are as follows:

[0025] S41. For structured data, use a rule engine and a lightweight MLP model to identify and quantify key accounting items to obtain the feature vector of the structured data;

[0026] S42. For semi-structured data, an adaptive parser is used to identify table hierarchy and formula dependencies, and an improved BiLSTM model that integrates financial term embeddings is used to normalize the diverse indicator names into standard terms, thus obtaining the feature vector of the semi-structured data.

[0027] S43. For unstructured data, a fine-tuned financial BERT model is used for named entity recognition and relation extraction, and causal relationships are identified through dependency parsing to obtain the feature vector of the unstructured data.

[0028] Furthermore, the specific steps of step S5 are as follows:

[0029] S51. Construct a financial ontology model, define the core financial concepts, attributes and relationships, and map the extracted feature vectors to the unified semantic space defined by the financial ontology model to eliminate semantic ambiguity;

[0030] S52. Divide the mapped feature vectors into core financial indicator vectors and non-core financial indicator vectors;

[0031] S53. A cross-modal multi-head attention mechanism is used for fusion, with the core financial indicator vector as the query parameter Q and the non-core financial indicator vector as the key parameter K and value parameter V. The attention weight is calculated using the following formula:

[0032] ;

[0033] in, It is the dimension of the feature vector. It is a normalized activation function;

[0034] S54. The weighted and fused feature vectors are projected into a shared semantic space and optimized by an autoencoder based on contrastive learning to generate fused feature vectors.

[0035] Furthermore, in step S53, a dynamic importance factor is introduced when fusing cross-modal multi-head attention mechanisms;

[0036] The dynamic importance factor is dynamically adjusted based on the data source type and data timeliness. The specific calculation formula is as follows:

[0037]

[0038] in, Let be the dynamic importance factor of the i-th eigenvector. It is a non-linear activation function. and These are learnable parameters. It is an embedded vector of data source type. It is an embedding vector for data timeliness. This represents the concatenation of the embedding vector representing the data source type and the embedding vector representing the data timeliness.

[0039] The formula for calculating attention weights after introducing a dynamic importance factor is adjusted as follows:

[0040]

[0041] in, It is a vector composed of dynamic importance factors corresponding to each feature vector. It is a normalized activation function, where K is the key parameter and V is the value parameter. It is the dimension of the feature vector.

[0042] Furthermore, in step S6, the logical framework of the report is established. The logical framework of the report shall include at least the core indicator analysis, interpretation of abnormal fluctuations, and multi-indicator correlation reasoning.

[0043] Generating multimodal financial analysis reports includes: matching line charts for trend indicators and pie charts for percentage indicators;

[0044] Content is adaptively adjusted based on user role, including:

[0045] The content granularity of the multimodal financial analysis report is dynamically adjusted based on the identified user role.

[0046] Furthermore, the specific steps for updating based on user feedback behavior in step S7 are as follows:

[0047] S71. Collect both explicit and implicit user feedback;

[0048] S72. Based on explicit feedback, the feature extraction or fusion rules are updated through online learning, and based on implicit feedback, the fusion weights of the corresponding indicators are dynamically adjusted.

[0049] S73. The validated fusion logic is transformed into a high-dimensional vector and stored in the retrieval library. When a new analysis scenario is encountered, the historical similarity fusion logic is matched from the retrieval library by calculating the cosine similarity.

[0050] Secondly, embodiments of this application also provide a multi-source heterogeneous data fusion-based financial analysis report generation system, including:

[0051] The financial multimodal large model building unit is used to integrate a structured processing engine, a semi-structured parsing module, an unstructured understanding module, a data access agent, a feature extraction agent, and an interactive agent to build a financial multimodal large model.

[0052] The large model loading unit is used to load the weight parameters of a pre-trained financial multimodal large model and import it into the financial knowledge graph.

[0053] The financial analysis data requirement determination unit is used to parse the analysis objectives input by the user through an interactive intelligent agent, generate a list of indicator requirements based on the financial knowledge graph, and associate it with the corresponding data sources;

[0054] If an indicator in the indicator requirement list cannot be matched with a data source, obtain supplementary information from the user through the interactive intelligent agent;

[0055] The multi-source heterogeneous financial data feature extraction unit is used to collect multi-source financial data from associated data sources through a data access agent, including structured data, semi-structured data and unstructured data.

[0056] The feature extraction agent extracts features from the three types of data respectively and outputs feature vectors.

[0057] The financial feature fusion and semantic alignment unit is used to map the feature vectors of various data sources to a unified financial ontology semantic space through the feature extraction agent, and to perform dynamic weighted fusion through a cross-modal multi-head attention mechanism to generate a unified fused feature vector.

[0058] The financial analysis report generation unit is used to construct the logical framework of the report based on the fused feature vector through a financial multimodal big model, generate a multimodal financial analysis report that includes indicator analysis, anomaly interpretation and correlation reasoning, and adaptively adjust the content granularity of the multimodal financial analysis report according to the user role;

[0059] The dynamic response and optimization unit is used to update the fusion rules and weights of feature vectors through a financial multimodal large model based on user feedback behavior, and to optimize the semantic retrieval library for feature vector matching.

[0060] Thirdly, embodiments of this application also provide a storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the multi-source heterogeneous data fusion financial analysis report generation method as described in the first aspect.

[0061] As can be seen from the above technical solutions, this application has the following advantages:

[0062] The method, system, and medium for generating multi-source heterogeneous data financial analysis reports provided in this application achieve intelligent and dynamic generation of financial analysis reports by constructing a complete multi-source heterogeneous data fusion system and combining semantic understanding of large models and collaboration of multiple intelligent agents. This effectively solves the problems of data silos, semantic inconsistencies, low utilization of unstructured data, and insufficient dynamic adaptability in existing technologies, improves the efficiency and accuracy of financial analysis report generation, and provides more comprehensive and accurate support for financial decision-making. Attached Figure Description

[0063] To more clearly illustrate the technical solution of this application, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0064] Figure 1 This is a flowchart illustrating the method for generating multi-source heterogeneous data fusion-based financial analysis reports according to the present invention.

[0065] Figure 2 This is a schematic diagram of the multi-source heterogeneous data fusion financial analysis report generation system of the present invention. Detailed Implementation

[0066] The various embodiments of this disclosure will be described more fully in the detailed steps of the method for generating multi-source heterogeneous data fusion financial analysis reports described below. This disclosure may have various embodiments, and adjustments and changes may be made therein. However, it should be understood that there is no intention to limit the various embodiments of this disclosure to the specific embodiments disclosed herein, but rather this disclosure should be understood to cover all adjustments, equivalents, and / or alternatives falling within the spirit and scope of the various embodiments of this disclosure.

[0067] For example, as enterprises advance their digital transformation, financial analysis has shifted from traditional ex-post accounting to real-time decision support. However, traditional financial report generation mainly relies on manual compilation of Excel spreadsheets, and automated tools are mostly based on fixed templates to generate standardized reports, such as balance sheets. In recent years, although large-scale artificial intelligence models have begun to be applied to financial analysis, most existing systems are limited to analyzing a single data source, such as generating profit analysis reports based solely on ERP data. They cannot effectively integrate unstructured data, such as audit reports, and are unable to meet the needs of complex financial scenarios.

[0068] Traditional financial analysis report generation suffers from the following problems: First, weak data integration capabilities. Financial data is scattered across multiple systems such as ERP, CRM, and auditing systems, with significant differences in data formats, such as database fields, PDF tables, and handwritten annotations. Traditional tools lack a unified data access and preprocessing framework. Second, lack of semantic consistency. Different departments have different definitions of the same financial indicators. For example, the definition of "cost" including tax differs across departments, lacking a unified semantic mapping mechanism, leading to conflicting analysis conclusions. Third, low utilization rate of unstructured data. Key information (such as potential liabilities and risk warnings) contained in unstructured data such as audit reports and contract texts cannot be effectively extracted and utilized. Fourth, insufficient dynamic adaptability. Faced with user-defined analysis needs (such as special impairment tests), manual reconstruction of the analysis model is required, resulting in a lag in response and an inability to achieve real-time dynamic adjustments.

[0069] Existing related patents either generate reports based on fixed templates, supporting only structured data processing, or focus on automated calculations from a single data source, without involving cross-source semantic fusion and dynamic weighting mechanisms, and lack a multimodal data processing framework specifically for the financial field.

[0070] To address the aforementioned issues, this embodiment provides a method for generating financial analysis reports based on the fusion of multi-source heterogeneous data. By integrating multi-source heterogeneous data and combining large models with multiple agents, it breaks down data silos, achieves efficient integration and in-depth analysis of financial data, and generates comprehensive, accurate, and adaptive financial analysis reports, significantly improving financial decision support capabilities.

[0071] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0072] Please see Figure 1 The diagram shows a flowchart of a method for generating a multi-source heterogeneous data fusion-based financial analysis report in a specific embodiment. The method includes the following steps:

[0073] S1. Integrates a structured processing engine, a semi-structured parsing module, an unstructured understanding module, a data access agent, a feature extraction agent, and an interactive agent to construct a large-scale financial multimodal model;

[0074] It should be noted that the deployed intelligent agent can achieve efficient data access, feature extraction, and user interaction, laying the foundation for subsequent steps;

[0075] S2. Load the weight parameters of the pre-trained financial multimodal large model and import the financial knowledge graph;

[0076] It should be noted that this step provides semantic understanding and data processing capabilities. Through pre-trained large models and financial knowledge graphs, it is possible to quickly understand and process financial data.

[0077] S3. Parse the user's input analysis objectives through an interactive intelligent agent, generate a list of indicator requirements based on the financial knowledge graph, and associate it with the corresponding data source;

[0078] If an indicator in the indicator requirement list cannot be matched with a data source, obtain supplementary information from the user through the interactive intelligent agent;

[0079] It should be noted that the system can accurately understand users' analytical needs, generate a specific list of indicator requirements, and automatically connect to data sources, thereby improving the system's automation level and response speed to user needs; by supplementing information through interactive intelligent agents, the system's adaptability to complex requirements is further improved.

[0080] S4. Collect multi-source financial data from associated data sources, including structured data, semi-structured data, and unstructured data, through a data access agent;

[0081] The feature extraction agent extracts features from the three types of data respectively and outputs feature vectors.

[0082] It should be noted that this step achieves comprehensive collection of multi-source financial data, covering structured, semi-structured and unstructured data, ensuring the integrity and diversity of the data; by using a feature extraction agent to perform targeted feature extraction on different types of data, valuable information can be extracted, providing high-quality feature vectors for subsequent semantic fusion.

[0083] S5. The feature vectors from various data sources are mapped to a unified financial ontology semantic space by a feature extraction agent, and dynamic weighted fusion is performed through a cross-modal multi-head attention mechanism to generate a unified fused feature vector.

[0084] It should be noted that the financial ontology model eliminates semantic ambiguity between different data sources and achieves semantic alignment of feature vectors; the cross-modal multi-head attention mechanism can dynamically weight and fuse feature vectors, strengthen features with high relevance to the analysis target, suppress redundant or noisy features, and improve the accuracy and adaptability of data fusion.

[0085] S6. Based on the fusion feature vector, the logical framework of the report is constructed through the financial multimodal big model, generating a multimodal financial analysis report that includes indicator analysis, anomaly interpretation and correlation reasoning, and adaptively adjusting the content granularity of the multimodal financial analysis report according to the user role;

[0086] It should be noted that the multimodal financial analysis report generated based on fused feature vectors can comprehensively and accurately reflect the financial situation, including indicator analysis, anomaly interpretation, and correlation reasoning, providing users with rich analytical information; the granularity of the content is dynamically adjusted according to the user role, which improves the relevance and practicality of the report and meets the needs of different users;

[0087] S7. Based on user feedback behavior, update the feature vector fusion rules and weights through the financial multimodal big model, and optimize the semantic retrieval library for feature vector matching;

[0088] It should be noted that by collecting user feedback and dynamically updating the fusion rules and optimizing the semantic retrieval library, the system has achieved self-evolution and optimization. It can be continuously adjusted and improved based on the actual usage of users, and better adapt to new analysis scenarios and user needs.

[0089] This embodiment provides a comprehensive and systematic solution that enables efficient integration and in-depth analysis of multi-source heterogeneous financial data, generating high-quality, adaptive financial analysis reports.

[0090] Furthermore, as a refinement and extension of the specific implementation methods of the above embodiments, in order to fully illustrate the specific implementation process of this embodiment, another method for generating multi-source heterogeneous data fusion financial analysis reports is provided, which includes the following steps:

[0091] S1. Integrates a structured processing engine, a semi-structured parsing module, an unstructured understanding module, a data access agent, a feature extraction agent, and an interactive agent to construct a large-scale financial multimodal model;

[0092] Specifically, the structured processing engine is used to parse the balance sheet and cash flow statement, the semi-structured parsing module is used to identify the formula logic of Excel templates, and the unstructured understanding module is used to extract key information from audit reports and contract texts.

[0093] S2. Load the weight parameters of the pre-trained financial multimodal large model and import the financial knowledge graph;

[0094] Step S2, loading the weight parameters of the pre-trained financial multimodal large model, is detailed below:

[0095] S21. Construct a dataset using labeled financial-related structured, semi-structured, and unstructured data, and train a large financial multimodal model.

[0096] S22. Import an industry-specific financial knowledge graph, which constructs the relationships between accounting subjects, report items, and business indicators;

[0097] S23. Configure the data access agent of the financial multimodal big model to connect to the financial database through JDBC / ODBC, call the market data interface through API, and recognize PDF reports through the OCR module;

[0098] S3. Parse the user's input analysis objectives through an interactive intelligent agent, generate a list of indicator requirements based on the financial knowledge graph, and associate it with the corresponding data source;

[0099] If an indicator in the indicator requirement list cannot be matched with a data source, obtain supplementary information from the user through the interactive intelligent agent;

[0100] The specific steps in step S3, where the interactive agent parses the user-inputted analysis target, are as follows:

[0101] S31. Perform intent recognition on the analysis target input by the user in natural language and extract key analysis dimensions;

[0102] S32. Based on the financial knowledge graph, key analysis dimensions are mapped to a set of standard financial indicators to generate an indicator requirement list;

[0103] S33. Associate known data sources with the indicator data in the indicator requirement list;

[0104] S34. When there is an indicator in the indicator requirement list that cannot be associated with any data source, the interactive agent is activated to generate a query statement to prompt the user to specify the data source of the corresponding indicator, and the new data source mapping relationship fed back by the user is stored in the knowledge base;

[0105] S4. Collect multi-source financial data from associated data sources, including structured data, semi-structured data, and unstructured data, through a data access agent;

[0106] The feature extraction agent extracts features from the three types of data respectively and outputs feature vectors.

[0107] In step S4, the feature extraction agent performs feature extraction on the three types of data respectively. The specific steps are as follows:

[0108] S41. For structured data, use a rule engine (such as verifying the reconciliation relationship of "Assets = Liabilities + Owner's Equity") and a lightweight MLP model to identify and quantify key accounting items to obtain the feature vector of the structured data;

[0109] S42. For semi-structured data, an adaptive parser is used to identify table hierarchy and formula dependencies, and an improved BiLSTM model that integrates financial term embeddings is used to normalize the diverse indicator names into standard terms, thus obtaining the feature vector of the semi-structured data.

[0110] S43. For unstructured data, a fine-tuned financial BERT model is used for named entity recognition and relation extraction, and causal relationships are identified through dependency parsing to obtain the feature vector of the unstructured data.

[0111] Specifically, for structured data (such as database tables), data cleaning is performed, including removing duplicate records and standardizing currency units;

[0112] For semi-structured data (such as Excel), directly read the file content;

[0113] For unstructured data (such as PDF reports), the OCR engine is called to perform text recognition;

[0114] S5. The feature vectors from various data sources are mapped to a unified financial ontology semantic space by a feature extraction agent, and dynamic weighted fusion is performed through a cross-modal multi-head attention mechanism to generate a unified fused feature vector.

[0115] The specific steps of step S5 are as follows:

[0116] S51. Construct a financial ontology model, define the core financial concepts, attributes and relationships, and map the extracted feature vectors to the unified semantic space defined by the financial ontology model to eliminate semantic ambiguity;

[0117] S52. Divide the mapped feature vectors into core financial indicator vectors and non-core financial indicator vectors;

[0118] S53. A cross-modal multi-head attention mechanism is adopted for fusion, using the core financial indicator vector (such as "net profit") as the query parameter Q, and the non-core financial indicator vector as the key parameter K and value parameter V, to calculate the attention weight. The specific calculation formula is as follows:

[0119] ;

[0120] in, It is the dimension of the feature vector. It is a normalized activation function;

[0121] It should be noted that attention mechanisms are used to dynamically strengthen features that are highly correlated with the analysis target, while suppressing redundant or noisy features.

[0122] S54. Project the weighted and fused feature vectors into a shared semantic space, and optimize them using an autoencoder based on contrastive learning to generate fused feature vectors;

[0123] Specifically, the contrastive learning autoencoder brings semantically consistent feature vectors closer together (such as "net profit" in the profit statement and "after-tax profit" in the audit report), pushes away semantically irrelevant vectors, and finally generates a unified fusion feature vector rich in contextual semantic information.

[0124] S6. Based on the fusion feature vector, the logical framework of the report is constructed through the financial multimodal big model, generating a multimodal financial analysis report that includes indicator analysis, anomaly interpretation and correlation reasoning, and adaptively adjusting the content granularity of the multimodal financial analysis report according to the user role;

[0125] In step S6, the logical framework of the report is constructed. The logical framework of the report shall include at least the core indicator analysis, interpretation of abnormal fluctuations, and multi-indicator correlation reasoning.

[0126] Generating multimodal financial analysis reports includes: matching line charts for trend indicators (such as "operating revenue in the past three years") and matching pie charts for percentage indicators (such as "cost composition");

[0127] Content is adaptively adjusted based on user role, including:

[0128] The content granularity of the multimodal financial analysis report is dynamically adjusted based on the identified user roles.

[0129] For example, when addressing management, the focus is on the macroeconomic impact and a risk summary;

[0130] When targeting finance personnel, it provides more detailed data breakdowns and accounting analysis;

[0131] S7. Based on user feedback behavior, update the feature vector fusion rules and weights through the financial multimodal big model, and optimize the semantic retrieval library for feature vector matching;

[0132] The specific steps of step S7 are as follows:

[0133] S71. Collect both explicit and implicit user feedback; specifically, explicit feedback includes direct modifications made by users to the content and analysis logic of the multimodal financial analysis report.

[0134] Implicit feedback includes the frequency and duration of user access to specific modules of the multimodal financial analysis report;

[0135] S72. Based on explicit feedback, update feature extraction or fusion rules through online learning; based on implicit feedback, dynamically adjust the fusion weights of corresponding indicators (e.g., increase the weights of tax-related indicators if the "tax burden analysis" module is accessed frequently).

[0136] S73. The validated fusion logic is transformed into a high-dimensional vector and stored in the retrieval library. When a new analysis scenario is encountered, the cosine similarity is calculated to match the historical similar fusion logic from the retrieval library, thereby achieving fast response and zero-training fusion.

[0137] In some embodiments, unlike the embodiments described above, a dynamic importance factor is introduced in step S53 when the cross-modal multi-head attention mechanism is fused.

[0138] The dynamic importance factor is dynamically adjusted based on the data source type and data timeliness. The specific calculation formula is as follows:

[0139]

[0140] in, Let be the dynamic importance factor of the i-th eigenvector. It is a non-linear activation function. and These are learnable parameters. It is an embedded vector of data source type. It is an embedding vector for data timeliness. This represents the concatenation of the embedding vector representing the data source type and the embedding vector representing the data timeliness.

[0141] The formula for calculating attention weights after introducing a dynamic importance factor is adjusted as follows:

[0142]

[0143] in, It is a vector composed of dynamic importance factors corresponding to each feature vector. It is a normalized activation function, where K is the key parameter and V is the value parameter. It is the dimension of the feature vector.

[0144] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0145] like Figure 2As shown, the following are embodiments of the multi-source heterogeneous data fusion financial analysis report generation system provided in this disclosure. This system and the multi-source heterogeneous data fusion financial analysis report generation method of the above embodiments belong to the same inventive concept. For details not described in detail in the embodiments of the multi-source heterogeneous data fusion financial analysis report generation system, please refer to the embodiments of the multi-source heterogeneous data fusion financial analysis report generation method.

[0146] The system includes:

[0147] The financial multimodal large model building unit is used to integrate a structured processing engine, a semi-structured parsing module, an unstructured understanding module, a data access agent, a feature extraction agent, and an interactive agent to build a financial multimodal large model.

[0148] The large model loading unit is used to load the weight parameters of a pre-trained financial multimodal large model and import it into the financial knowledge graph.

[0149] The financial analysis data requirement determination unit is used to parse the analysis objectives input by the user through an interactive intelligent agent, generate a list of indicator requirements based on the financial knowledge graph, and associate it with the corresponding data sources;

[0150] If an indicator in the indicator requirement list cannot be matched with a data source, obtain supplementary information from the user through the interactive intelligent agent;

[0151] The multi-source heterogeneous financial data feature extraction unit is used to collect multi-source financial data from associated data sources through a data access agent, including structured data, semi-structured data and unstructured data.

[0152] The feature extraction agent extracts features from the three types of data respectively and outputs feature vectors.

[0153] The financial feature fusion and semantic alignment unit is used to map the feature vectors of various data sources to a unified financial ontology semantic space through the feature extraction agent, and to perform dynamic weighted fusion through a cross-modal multi-head attention mechanism to generate a unified fused feature vector.

[0154] The financial analysis report generation unit is used to construct the logical framework of the report based on the fused feature vector through a financial multimodal big model, generate a multimodal financial analysis report that includes indicator analysis, anomaly interpretation and correlation reasoning, and adaptively adjust the content granularity of the multimodal financial analysis report according to the user role;

[0155] The dynamic response and optimization unit is used to update the fusion rules and weights of feature vectors through a financial multimodal large model based on user feedback behavior, and to optimize the semantic retrieval library for feature vector matching.

[0156] This embodiment achieves efficient integration and in-depth analysis of financial data through the interaction and collaboration of various units, generating comprehensive, accurate, and adaptive financial analysis reports, and significantly improving financial decision support capabilities.

[0157] The method for generating multi-source heterogeneous data fusion financial analysis reports provided in this application can be applied to electronic devices. Those skilled in the art will understand that the electronic device structure involved in the embodiments of this invention does not constitute a limitation on the electronic device. An electronic device may include more or fewer components than illustrated, or combine certain components, or have different component arrangements. In the embodiments of this invention, the electronic device includes, but is not limited to, laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the embodiments of this application described and / or claimed herein.

[0158] Electronic devices may include processors, external memory interfaces, internal memory, universal serial bus (USB) interfaces, charging management modules, power management modules, batteries, wireless communication modules, audio modules, speakers, microphones, sensor modules, buttons, cameras, displays, and SIM card interfaces, etc.

[0159] It is understood that the structures illustrated in the embodiments of this application do not constitute a specific limitation on the electronic device. In other embodiments of this application, the electronic device may include more or fewer components than illustrated, or combine some components, or split some components, or have different component arrangements. The illustrated components may be implemented in hardware, software, or a combination of software and hardware.

[0160] A processor may include one or more processing units, such as a central processing unit (CPU), an application processor (AP), a modem processor, a graphics processing unit (GPU), an image signal processor (ISP), a controller, memory, a video codec, a digital signal processor (DSP), a baseband processor, and / or a neural network processing unit (NPU). Different processing units may be independent devices or integrated into one or more processors.

[0161] The processor can serve as the nerve center and command center of an electronic device. The controller can generate operation control signals based on the instruction opcode and timing signals to control the fetching and execution of instructions.

[0162] The processor may also include memory for storing instructions and data. In some embodiments, the memory in the processor is a cache memory. This memory can store instructions or data that the processor has just used or that are used repeatedly. If the processor needs to use the instruction or data again, it can retrieve it directly from this memory. This avoids repeated accesses, reduces processor latency, and thus improves system efficiency.

[0163] The aforementioned electronic device integrates a structured processing engine, a semi-structured parsing module, an unstructured understanding module, a data access agent, a feature extraction agent, and an interactive agent into the multi-source heterogeneous data fusion-based financial analysis report generation method of this application, constructing a large-scale financial multimodal model; loading the weight parameters of the pre-trained large-scale financial multimodal model and importing them into a financial knowledge graph; parsing the user-inputted analysis objectives through the interactive agent, generating an indicator requirement list based on the financial knowledge graph, and associating it with corresponding data sources; if any indicator in the indicator requirement list cannot be matched with a data source, obtaining supplementary information from the user through the interactive agent; collecting multi-source financial data, including structured data, semi-structured data, and unstructured data, from the associated data sources through the data access agent; extracting features from the three types of data through the feature extraction agent, outputting feature vectors; and performing feature extraction... The intelligent agent maps feature vectors from various data sources to a unified financial ontology semantic space and dynamically weights and fuses them through a cross-modal multi-head attention mechanism to generate a unified fused feature vector. Based on the fused feature vector, a logical framework for the report is constructed through a financial multimodal big model, generating a multimodal financial analysis report that includes indicator analysis, anomaly interpretation, and correlation reasoning. The content granularity of the multimodal financial analysis report is adaptively adjusted according to the user's role. Based on user feedback, the fusion rules and weights of the feature vectors are updated through the financial multimodal big model, and the technical solution of the semantic retrieval library for feature vector matching is optimized. This achieves the beneficial effect of breaking down data silos by fusing multi-source heterogeneous data, combining big model and multi-agent technology, realizing efficient integration and in-depth analysis of financial data, generating comprehensive, accurate, and adaptive financial analysis reports, and significantly improving financial decision support capabilities.

[0164] The storage medium provided in this application stores a program product capable of generating a financial analysis report that integrates multi-source heterogeneous data.

[0165] The method for generating multi-source heterogeneous data fusion-based financial analysis reports includes: integrating a structured processing engine, a semi-structured parsing module, an unstructured understanding module, a data access agent, a feature extraction agent, and an interactive agent to construct a large-scale financial multimodal model; loading the weight parameters of the pre-trained large-scale financial multimodal model and importing them into a financial knowledge graph; parsing the user-inputted analysis objectives through the interactive agent, generating a list of indicator requirements based on the financial knowledge graph, and associating it with corresponding data sources; if any indicator in the indicator requirement list cannot be matched with a data source, obtaining supplementary information from the user through the interactive agent; and collecting multi-source financial data, including structured data and semi-structured data, from the associated data sources through the data access agent. Based on unstructured data, a feature extraction agent is used to extract features from the three types of data, outputting feature vectors. These feature vectors from each data source are mapped to a unified financial ontology semantic space, and dynamically weighted and fused using a cross-modal multi-head attention mechanism to generate a unified fused feature vector. Based on the fused feature vector, a logical framework for the report is constructed using a large-scale financial multimodal model, generating a multimodal financial analysis report that includes indicator analysis, anomaly interpretation, and correlation reasoning. The content granularity of the multimodal financial analysis report is adaptively adjusted according to user roles. Based on user feedback, the fusion rules and weights of the feature vectors are updated using the large-scale financial multimodal model, and the semantic retrieval library for feature vector matching is optimized.

[0166] In some possible implementations, the method for generating multi-source heterogeneous data fusion financial analysis reports disclosed herein can be implemented as a program product comprising program code that, when the program product is run on a terminal device, causes the terminal device to perform the steps described in the "Exemplary Methods" section of this specification according to various exemplary embodiments of this disclosure.

[0167] The storage medium disclosed herein may be any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.

[0168] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for generating financial analysis reports based on the fusion of multi-source heterogeneous data, characterized in that, Includes the following steps: S1. Integrates a structured processing engine, a semi-structured parsing module, an unstructured understanding module, a data access agent, a feature extraction agent, and an interactive agent to construct a large-scale financial multimodal model; The structured processing engine is used to parse balance sheets and cash flow statements, the semi-structured parsing module is used to identify Excel template formula logic, and the unstructured understanding module is used to extract key information from audit reports and contract texts. S2. Load the weight parameters of the pre-trained financial multimodal large model and import the financial knowledge graph; S3. Parse the user's input analysis objectives through an interactive intelligent agent, generate a list of indicator requirements based on the financial knowledge graph, and associate it with the corresponding data source; If an indicator in the indicator requirement list cannot be matched with a data source, obtain supplementary information from the user through the interactive intelligent agent; S4. Collect multi-source financial data from associated data sources, including structured data, semi-structured data, and unstructured data, through a data access agent; The feature extraction agent extracts features from the three types of data respectively and outputs feature vectors. S5. The feature vectors from various data sources are mapped to a unified financial ontology semantic space by a feature extraction agent, and dynamic weighted fusion is performed through a cross-modal multi-head attention mechanism to generate a unified fused feature vector. Step S5 includes the following steps: S51. Construct a financial ontology model, define the core financial concepts, attributes and relationships, and map the extracted feature vectors to the unified semantic space defined by the financial ontology model to eliminate semantic ambiguity; S6. Based on the fusion feature vector, the logical framework of the report is constructed through the financial multimodal big model, generating a multimodal financial analysis report that includes indicator analysis, anomaly interpretation and correlation reasoning, and adaptively adjusting the content granularity of the multimodal financial analysis report according to the user role; S7. Based on user feedback behavior, update the feature vector fusion rules and weights through the financial multimodal large model, and optimize the semantic retrieval library for feature vector matching.

2. The method for generating multi-source heterogeneous data fusion-based financial analysis reports according to claim 1, characterized in that, The specific steps for loading the weight parameters of the pre-trained financial multimodal large model in step S2 are as follows: S21. Construct a dataset using labeled financial-related structured, semi-structured, and unstructured data, and train a large financial multimodal model. S22. Import an industry-specific financial knowledge graph, which constructs the relationships between accounting subjects, report items, and business indicators; S23. Configure the data access agent of the financial multimodal big model to connect to the financial database through JDBC / ODBC, call the market data interface through API, and recognize PDF reports through the OCR module.

3. The method for generating multi-source heterogeneous data fusion-based financial analysis reports according to claim 1, characterized in that, The specific steps in step S3, where the interactive agent parses the user-inputted analysis target, are as follows: S31. Perform intent recognition on the analysis target input by the user in natural language and extract key analysis dimensions; S32. Based on the financial knowledge graph, key analysis dimensions are mapped to a set of standard financial indicators to generate an indicator requirement list; S33. Associate known data sources with the indicator data in the indicator requirement list; S34. When there is an indicator in the indicator requirement list that cannot be associated with any data source, the interactive agent is activated to generate a query statement to prompt the user to specify the data source of the corresponding indicator, and the new data source mapping relationship fed back by the user is stored in the knowledge base.

4. The method for generating multi-source heterogeneous data fusion-based financial analysis reports according to claim 1, characterized in that, In step S4, the feature extraction agent performs feature extraction on the three types of data respectively. The specific steps are as follows: S41. For structured data, use a rule engine and a lightweight MLP model to identify and quantify key accounting items to obtain the feature vector of the structured data; S42. For semi-structured data, an adaptive parser is used to identify table hierarchy and formula dependencies, and an improved BiLSTM model that integrates financial term embeddings is used to normalize the diverse indicator names into standard terms, thus obtaining the feature vector of the semi-structured data. S43. For unstructured data, a fine-tuned financial BERT model is used for named entity recognition and relation extraction, and causal relationships are identified through dependency parsing to obtain the feature vector of the unstructured data.

5. The method for generating multi-source heterogeneous data fusion financial analysis reports according to claim 1, characterized in that, Step S5 also includes the following steps: S52. Divide the mapped feature vectors into core financial indicator vectors and non-core financial indicator vectors; S53. A cross-modal multi-head attention mechanism is used for fusion, with the core financial indicator vector as the query parameter Q and the non-core financial indicator vector as the key parameter K and value parameter V. The attention weight is calculated using the following formula: ； in, It is the dimension of the feature vector. It is a normalized activation function; S54. Project the weighted and fused feature vectors into a shared semantic space, and optimize them using an autoencoder based on contrastive learning to generate fused feature vectors.

6. The method for generating multi-source heterogeneous data fusion-based financial analysis reports according to claim 5, characterized in that, In step S53, a dynamic importance factor is introduced when fusing cross-modal multi-head attention mechanisms; The dynamic importance factor is dynamically adjusted based on the data source type and data timeliness. The specific calculation formula is as follows: in, Let be the dynamic importance factor of the i-th eigenvector. It is a non-linear activation function. and These are learnable parameters. It is an embedded vector of data source type. It is an embedding vector for data timeliness. This represents the concatenation of the embedding vector representing the data source type and the embedding vector representing the data timeliness. The formula for calculating attention weights after introducing a dynamic importance factor is adjusted as follows: in, It is a vector composed of dynamic importance factors corresponding to each feature vector. It is a normalized activation function, where K is the key parameter and V is the value parameter. It is the dimension of the feature vector.

7. The method for generating multi-source heterogeneous data fusion financial analysis reports according to claim 1, characterized in that, In step S6, the logical framework of the report is constructed. The logical framework of the report shall include at least the core indicator analysis, interpretation of abnormal fluctuations, and multi-indicator correlation reasoning. Generating multimodal financial analysis reports includes: matching line charts for trend indicators and pie charts for percentage indicators; Content is adaptively adjusted based on user role, including: The content granularity of the multimodal financial analysis report is dynamically adjusted based on the identified user role.

8. The method for generating multi-source heterogeneous data fusion-based financial analysis reports according to claim 1, characterized in that, The specific steps for updating based on user feedback in step S7 are as follows: S71. Collect both explicit and implicit user feedback; S72. Based on explicit feedback, the feature extraction or fusion rules are updated through online learning, and based on implicit feedback, the fusion weights of the corresponding indicators are dynamically adjusted. S73. The validated fusion logic is transformed into a high-dimensional vector and stored in the retrieval library. When a new analysis scenario is encountered, the historical similarity fusion logic is matched from the retrieval library by calculating the cosine similarity.

9. A multi-source heterogeneous data fusion-based financial analysis report generation system, characterized in that, include: The financial multimodal large model building unit is used to integrate a structured processing engine, a semi-structured parsing module, an unstructured understanding module, a data access agent, a feature extraction agent, and an interactive agent to build a financial multimodal large model. The structured processing engine is used to parse balance sheets and cash flow statements, the semi-structured parsing module is used to identify Excel template formula logic, and the unstructured understanding module is used to extract key information from audit reports and contract texts. The large model loading unit is used to load the weight parameters of a pre-trained financial multimodal large model and import it into the financial knowledge graph. The financial analysis data requirement determination unit is used to parse the analysis objectives input by the user through an interactive intelligent agent, generate a list of indicator requirements based on the financial knowledge graph, and associate it with the corresponding data sources; If an indicator in the indicator requirement list cannot be matched with a data source, obtain supplementary information from the user through the interactive intelligent agent; The multi-source heterogeneous financial data feature extraction unit is used to collect multi-source financial data from associated data sources through a data access agent, including structured data, semi-structured data and unstructured data. The feature extraction agent extracts features from the three types of data respectively and outputs feature vectors. The financial feature fusion and semantic alignment unit is used to map the feature vectors of various data sources to a unified financial ontology semantic space through the feature extraction agent, and to perform dynamic weighted fusion through a cross-modal multi-head attention mechanism to generate a unified fused feature vector. The feature extraction agent maps the feature vectors of various data sources to a unified financial ontology semantic space as follows: a financial ontology model is constructed, the core financial concepts, attributes and relationships are defined, and the extracted feature vectors are mapped to the unified semantic space defined by the financial ontology model to eliminate semantic ambiguity. The financial analysis report generation unit is used to construct the logical framework of the report based on the fused feature vector through a financial multimodal big model, generate a multimodal financial analysis report that includes indicator analysis, anomaly interpretation and correlation reasoning, and adaptively adjust the content granularity of the multimodal financial analysis report according to the user role; The dynamic response and optimization unit is used to update the fusion rules and weights of feature vectors through a financial multimodal large model based on user feedback behavior, and to optimize the semantic retrieval library for feature vector matching.

10. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for generating multi-source heterogeneous data fusion financial analysis reports as described in any one of claims 1 to 8.

Images