Intelligent financial analysis method for natural language processing and related device

Abstract

The application discloses a kind of natural language processing intelligent financial analysis method and related equipment, method includes: obtaining historical financial data, generates analysis target by BERT model analysis user demand, and generates account matching rule using T5 model;After pre-processing and intelligent splicing to account data, extract budget execution rate information to construct multidimensional financial knowledge graph in combination with the latest policies and regulations;Then adopt the fusion model of graph neural network and deep reinforcement learning, under the guidance of user target and budget execution analysis data, realize financial risk identification and trend prediction, automatically generate visual analysis report.The application reduces the matching time of each financial statement in account text data, can timely combine the latest policies and regulations to carry out financial dynamic analysis, automatically generate visual analysis report, improve the timeliness and convenience of financial analysis, can be widely applied in data processing technical field.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an intelligent financial analysis method and related equipment using natural language processing. Background Technology

[0002] In today's digital economy era, the financial data generated by organizations and individuals is experiencing explosive growth, exhibiting diverse and complex forms and content. This financial data forms the foundation for financial analysis. Financial analysis is undergoing profound transformation, with cutting-edge technologies such as big data, cloud computing, and artificial intelligence becoming core driving forces. Big data technology, through distributed storage and parallel computing, enables the real-time collection, cleaning, and integration of massive amounts of financial data, solving the problem of mixed structured and unstructured data that traditional methods struggle to handle. Cloud computing provides elastic and scalable computing resources, supporting the efficient operation of financial analysis models, reducing organizational IT costs, providing stronger data support for organizational decision-making, and simultaneously enhancing the personalization of individual financial management.

[0003] However, existing technologies also have some technical shortcomings, specifically as follows: First, existing technologies rely on manual rule configuration of financial statement matching logic, which is cumbersome, error-prone, and has high communication costs. Second, the data processing and analysis process is isolated and static, making it difficult to dynamically adjust to changes in industry policy environments, resulting in delayed analysis results. Furthermore, the entire analysis chain is fragmented; from data preparation and intent understanding to risk prediction and report generation, the automation level of each link is low, heavily relying on professional knowledge and manual intervention, resulting in unsatisfactory timeliness, accuracy, and ease of use in financial analysis. Summary of the Invention

[0004] The main objective of this invention is to provide an intelligent financial analysis method, device, electronic device, storage medium, and program product using natural language processing, aiming to solve at least one problem in the prior art.

[0005] To achieve the above objectives, one aspect of this invention proposes an intelligent financial analysis method using natural language processing, the method comprising:

[0006] Acquire historical financial data; this includes textual account data, user-reported financial analysis requests, and user-reported account matching methods. The BERT model is used to perform language parsing and intent recognition on users' verbal financial analysis needs, and to generate user goals. Based on the T5 model, language parsing is performed on the user's verbal account matching method to generate matching rules; The accounting text data is preprocessed, and the preprocessed accounting text data is deconstructed and reconstructed based on the user's goal, and irrelevant items are filtered out to obtain the filtered accounting text data. Based on the matching rules, the filtered account text data is concatenated and organized through a pointer generator network to obtain the organized account text data. To obtain the latest policies and regulations, based on the organized accounting text data, a cross-attention dual-tower model enhanced with accounting knowledge is used to perform feature association on the policies and regulations, and extract budget execution analysis data. Using the organized accounting text data as the main data, a multi-dimensional financial knowledge graph is constructed through entity extraction and relation modeling. Based on a multidimensional financial knowledge graph, with user goals as the target guide and budget execution analysis data as the correction item, a graph analysis deep computing engine is used to identify financial risks and predict financial risk trends, generating a visualized financial analysis report.

[0007] In some embodiments, obtaining historical financial data includes the following steps: Data was extracted from a multi-source system consisting of personal financial statements, corporate financial systems, and payment platforms to obtain account text data; among which, the account text data is multi-source heterogeneous text data; The first audio data segment of the user's spoken financial analysis needs was collected and transcribed, and then classified and merged to obtain the user's spoken financial analysis needs. The second voice data segment, which collects and transcribes the interrelationships between different data reports and the matching content of key items in the user's dictated account text data, is classified and merged to obtain the user's dictated account matching method.

[0008] In some embodiments, the BERT model is used to perform language parsing and intent recognition on users' verbal financial analysis needs to generate user goals, including the following steps: The BERT model was used to perform word segmentation, part-of-speech tagging, and dependency parsing on users’ verbal financial analysis needs to obtain initial user needs analysis results. Identify keywords that represent the core intent of users based on the initial user needs analysis results; keywords include at least one of cost analysis, budget execution rate, and financial risk analysis. The semantic components of keywords representing the user's core intent are structured to generate quantifiable user goals.

[0009] In some embodiments, language parsing is performed on the user's verbal account matching method based on the T5 model to generate matching rules, including the following steps: The user's verbal account matching method is converted into a standardized account matching text sequence, and then the encoder of the T5 model is used to parse the grammatical structure and identify the matching conditions and logical relationships between different accounts. Based on the matching conditions and logical relationships between different accounts, the T5 model decoder converts the colloquial descriptions in the user's verbal account matching method into logical expressions of attributes-operators-values, resulting in a structured rule sequence. By using the T5 model and a self-attention mechanism, the rule generation strategy of the structured rule sequence is dynamically adjusted to obtain matching rules.

[0010] In some embodiments, the accounting text data is preprocessed, and the preprocessed accounting text data is deconstructed and reconstructed based on user objectives to filter out items irrelevant to user objectives, resulting in filtered accounting text data. This includes the following steps: The accounting text data is preprocessed to obtain preprocessed accounting text data; the preprocessing includes filling in the missing values ​​of financial amount related items and unifying the financial data attribute values; Based on the user's objectives, the preprocessed account text data is deconstructed, redundant data items and noisy records are actively filtered out as items irrelevant to the user's objectives, and key data fields and entities related to the user's objectives are extracted. Use key data fields and entities related to user goals as filtered account text data.

[0011] In some embodiments, based on matching rules, the filtered account text data is concatenated and organized through a pointer generator network to obtain organized account text data, including the following steps: The encoder of the pointer generator network performs deep encoding on the filtered account text, and extracts the semantic features and contextual information of each account record as the encoded account features. An attention mechanism is used to associate the encoded account features with matching rules to identify the encoded features; the encoded features represent the relevant account items to be concatenated and the strength of the association. Using the correlation strength as the attention weight, and based on the relevant account entries, the decoder of the pointer generator network is used to decode and generate an account text sequence as the concatenation result; The system performs logical verification and format standardization on the sequence of account texts, and then evaluates the splicing result through preset verification rules. The splicing result that passes the evaluation is output as the organized account text data.

[0012] In some embodiments, the latest policies and regulations for a given period are obtained. Based on the organized accounting text data, a cross-attention dual-tower model enhanced with accounting knowledge is used to perform feature association on the policies and regulations, and budget execution analysis data is extracted. This includes the following steps: Web crawling technology is used to crawl and clean the latest time period of policies and regulations, and the structured text of the policies and regulations is output; where the latest time period is a preset time period from the current time. The structured text and the organized accounting text data are input into the cross-attention dual-tower model that enhances accounting knowledge. The resulting accounting context vector is then processed and output as a cross-attention key-value pair. Based on the cross-attention key-value pairs, using the cross-attention layer of the cross-attention dual-tower model, and taking the account context vector as the query, the policy and regulation vector is weighted to obtain the financial and accounting integration features. The financial and accounting integration features are input into a multilayer perceptron for decoding, and an investment quantitative benefit score is output. Budget execution analysis data is obtained by tracing the source of investment quantitative benefit scores.

[0013] In some embodiments, the organized accounting text data is used as the main data, and a multidimensional financial knowledge graph is constructed through entity extraction and relation modeling, including the following steps: Identify and extract financial entities from the organized accounting text data; the financial entities include project name, expense account, transaction item and unit information; Based on syntactic analysis and semantic role labeling, financial relationships are established between each financial entity; these relationships include the execution status of financial accounting and budgetary accounting, transaction history, nature of funds, and project linkage. By mapping financial entities and financial relationships to a unified graph model, a multidimensional financial knowledge graph is constructed.

[0014] In some embodiments, the graph analytics deep computing engine is a fusion model of graph neural networks and deep reinforcement learning networks. Based on a multidimensional financial knowledge graph, it uses user goals as the objective guide and budget execution analysis data as the correction term. The graph analytics deep computing engine is used to identify financial risks and predict financial risk trends, generating a visualized financial analysis report, including the following steps: Based on a multidimensional financial knowledge graph, and with user goals as the target guide, we use graph neural networks to extract the structure of related subgraphs, and track the transmission path and trend characteristics of financial risks through node embedding and relationship propagation. Using budget execution data as an environmental correction item, a risk identification model is constructed by simulating regulatory dynamics using deep reinforcement learning networks. By inputting the financial risk transmission path and financial risk trend characteristics into the risk identification model, the results of the graph analysis are obtained and a visual financial analysis report is generated.

[0015] To achieve the above objectives, another aspect of the present invention provides an intelligent financial analysis device using natural language processing, the device comprising: The first module is used to acquire historical financial data; the historical financial data includes textual account data, user-described financial analysis requirements, and user-described account matching methods. The second module is used to perform language parsing and intent recognition on the user's verbal financial analysis needs using the BERT model, and generate user goals. The third module is used to perform language parsing on the user's verbal account matching method based on the T5 model and generate matching rules. The fourth module is used to preprocess the account text data. Based on the user's goal, the preprocessed account text data is deconstructed and reconstructed, and irrelevant items are filtered out to obtain the filtered account text data. The fifth module is used to concatenate and organize the filtered account text data based on matching rules through a pointer generator network to obtain the organized account text data. The sixth module is used to obtain the latest policies and regulations for the current period. Based on the organized accounting text data, it uses a cross-attention dual-tower model enhanced with accounting knowledge to perform feature association on the policies and regulations and extract budget execution analysis data. The seventh module is used to construct a multi-dimensional financial knowledge graph by taking the organized accounting text data as the main data and using entity extraction and relation modeling. The eighth module is used to identify financial risks and predict financial risk trends based on a multidimensional financial knowledge graph, with user goals as the target guide and budget execution analysis data as the correction item. It uses a graph analysis deep computing engine to generate a visualized financial analysis report.

[0016] To achieve the above objectives, another aspect of the present invention provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the aforementioned method.

[0017] To achieve the above objectives, another aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method.

[0018] To achieve the above objectives, another aspect of the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the aforementioned method.

[0019] The embodiments of this invention include at least the following beneficial effects: This invention provides an intelligent financial analysis method, device, electronic device, storage medium, and program product using natural language processing. The solution first acquires complete historical financial data, including account text data, user-stated requirements, and matching methods. A BERT model is used to perform deep semantic parsing and intent recognition on the user's stated financial analysis requirements, generating quantifiable and executable analysis objectives. Simultaneously, a T5 model is used to parse the user's stated account matching logic, automatically generating executable matching rules, simplifying the manual operation process required for traditional financial statement matching, and reducing the time spent on comparison and communication arising from matching issues between different reports. Preprocessing of the account text data ensures data quality consistency. Based on user objectives, the preprocessed data is deconstructed and reorganized, and irrelevant items are filtered out. A pointer generator network is used to intelligently stitch together the account data, improving data processing efficiency. By introducing the latest policies and regulations as dynamic correction factors, and utilizing a cross-attention dual-tower model enhanced with accounting knowledge, the processed account data is deeply feature-associated with the latest policies and regulations to extract time-sensitive budget execution analysis data. This enables timely responses to dynamic changes in regulation, allowing for corresponding corrections and providing accurate data support for subsequent analysis. A multi-dimensional financial knowledge graph is constructed through entity extraction and relational modeling. A computational engine integrating graph neural networks and deep reinforcement learning is employed, guided by user objectives and combined with budget execution data, to identify financial risks and predict trends. It not only tracks complex financial relationship networks but also continuously optimizes the accuracy of risk prediction through reinforcement learning. Finally, a visualized financial analysis report containing risk identification results and trend predictions is automatically generated, achieving full automation from data collection, processing, and analysis to report generation. This invention addresses the pain points of traditional financial analysis, such as low data processing efficiency, delayed policy response, and high professional knowledge barriers. It reduces the matching time of various financial statements in accounting text data, enables timely integration of the latest policies and regulations for dynamic financial analysis, and automatically generates visualized analysis reports, improving the timeliness and convenience of financial analysis and providing more intelligent and automated financial decision support for both institutional and individual users. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of an implementation environment for the intelligent financial analysis method for natural language processing provided in this embodiment of the invention; Figure 2 This is a flowchart illustrating an intelligent financial analysis method using natural language processing provided in an embodiment of the present invention. Figure 3 This is a schematic diagram of the unfolding process of step S100 provided in the embodiment of the present invention; Figure 4This is a schematic diagram of the unfolding process of step S200 provided in the embodiment of the present invention; Figure 5 This is a schematic diagram of the unfolding process of step S300 provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the unfolding process of step S300 provided in an embodiment of the present invention; Figure 7 This is a schematic diagram illustrating the principle architecture of the intelligent financial analysis method based on natural language processing provided in this embodiment of the invention. Figure 8 This is a schematic diagram illustrating the overall application process of the intelligent financial analysis method using natural language processing provided in this embodiment of the invention. Figure 9 This is a schematic diagram of the structure of an intelligent financial analysis device using natural language processing, provided in an embodiment of the present invention. Figure 10 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the embodiments of this invention; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this invention as detailed in the appended claims.

[0022] It is understood that the terms “first,” “second,” etc., used in this invention may be used herein to describe various concepts, but unless specifically stated otherwise, these concepts are not limited by these terms. These terms are used only to distinguish one concept from another. For example, first information may also be referred to as second information without departing from the scope of embodiments of the invention, and similarly, second information may also be referred to as first information. Depending on the context, the words “if,” “when,” or “in response to determination” as used herein may be interpreted as “when…” or “when…” or “in response to determination.”

[0023] The terms “at least one,” “multiple,” “each,” “any,” etc., used in this invention, “at least one” includes one, two, or more than two; “multiple” includes two or more than two; “each” refers to each of the corresponding multiple; and “any” refers to any one of the multiple.

[0024] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein is for the purpose of describing embodiments of the invention only and is not intended to limit the invention.

[0025] Among related technologies, firstly, existing technologies rely on manual rule configuration of financial statement matching logic, which is cumbersome, error-prone, and has high communication costs. Secondly, the data processing and analysis process is isolated and static, making it difficult to dynamically adjust to changes in industry policy environment, resulting in delayed analysis results. Furthermore, the entire analysis chain is broken; from data preparation and intent understanding to risk prediction and report generation, the automation level of each link is low, heavily relying on professional knowledge and human intervention, resulting in unsatisfactory timeliness, accuracy, and ease of use of financial analysis.

[0026] In view of this, this invention provides an intelligent financial analysis method and related equipment using natural language processing. The solution first acquires complete historical financial data, including account text data, user-stated requirements, and matching methods. A BERT model is used to perform deep semantic parsing and intent recognition on the user's stated financial analysis requirements, generating quantifiable and executable analysis objectives. Simultaneously, a T5 model is used to parse the user's stated account matching logic, automatically generating executable matching rules, simplifying the manual operation process required for traditional financial statement matching, and reducing the time spent on comparison and communication due to matching issues between different reports. Preprocessing of the account text data ensures data quality consistency. Based on user objectives, the preprocessed data is deconstructed and reorganized, and irrelevant items are filtered out. A pointer generator network is used to intelligently stitch together the account data, improving data processing efficiency. By introducing the latest policies and regulations as dynamic correction factors, and using a cross-attention dual-tower model enhanced with accounting knowledge, the processed account data is deeply correlated with the latest policies and regulations to extract timely budget execution analysis data. This enables timely response to dynamic changes in regulation, making corresponding corrections and providing accurate data support for subsequent analysis. This invention constructs a multi-dimensional financial knowledge graph through entity extraction and relational modeling, and employs a computational engine that integrates graph neural networks and deep reinforcement learning. Guided by user objectives, it combines budget execution data to identify financial risks and predict trends. It not only tracks complex financial relationship networks but also continuously optimizes the accuracy of risk prediction through reinforcement learning. Finally, it automatically generates a visualized financial analysis report containing risk identification results and trend predictions, achieving full automation from data collection, processing, and analysis to report generation. This invention addresses the pain points of traditional financial analysis, such as low data processing efficiency, delayed policy response, and high professional knowledge barriers. It reduces the matching time of various financial statements in accounting text data, enables timely integration of the latest policies and regulations for dynamic financial analysis, and automatically generates visualized analysis reports, improving the timeliness and convenience of financial analysis and providing more intelligent and automated financial decision support for both institutional and individual users.

[0027] It is understood that the intelligent financial analysis method using natural language processing provided by this invention can be applied to any computer device with data processing and computing capabilities, and this computer device can be various terminals or servers. When the computer device in the embodiment is a server, the server is an independent physical server, or a server cluster or distributed system composed of multiple physical servers, 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 communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms. Optionally, the terminal can be a smartphone, tablet, laptop, or desktop computer, but it is not limited to these.

[0028] like Figure 1 The diagram shown is a schematic representation of an implementation environment provided by an embodiment of the present invention. (Refer to...) Figure 1 The implementation environment includes at least one terminal 102 and a server 101. The terminal 102 and the server 101 can be connected via a network, either wirelessly or via a wired connection, to complete data transmission and exchange.

[0029] Server 101 can be a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, 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 communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms.

[0030] Additionally, server 101 can also be a node server in a blockchain network. Blockchain is a novel application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms.

[0031] Terminal 102 can be a smartphone, tablet computer, laptop computer, desktop computer, smart speaker, smartwatch, etc., but is not limited to these. Terminal 102 and server 101 can be directly or indirectly connected via wired or wireless communication, and this embodiment of the invention does not impose any limitations.

[0032] For example, based on Figure 1 The implementation environment shown in this embodiment of the invention provides an intelligent financial analysis method using natural language processing. The following description uses the application of this intelligent financial analysis method in server 101 as an example. It can be understood that this intelligent financial analysis method using natural language processing can also be applied in terminal 102.

[0033] Reference Figure 2 , Figure 2 This is an optional flowchart of the intelligent financial analysis method using natural language processing provided in the embodiments of the present invention. The executing entity of the intelligent financial analysis method using natural language processing can be any of the aforementioned computer devices (including servers or terminals). Figure 2 The method may include, but is not limited to, steps S100 to S800.

[0034] Step S100: Obtain historical financial data; Historical financial data includes textual account data, user-reported financial analysis needs, and user-reported account matching methods. It should be noted that in some embodiments, such as Figure 3 As shown, step S100 may include the following steps: S110, extracting data from a multi-source system consisting of personal financial statements, company financial systems, and payment platforms to obtain account text data; wherein, the account text data is multi-source heterogeneous text data; S120, collecting and transcribing a first voice data segment of the user's verbal financial analysis needs, classifying and merging it to obtain the user's verbal financial analysis needs; S130, collecting and transcribing a second voice data segment of the interrelationships between different data reports and the matching content of key items in the user's verbal account text data, classifying and merging it to obtain the user's verbal account matching method.

[0035] For example, in some specific implementations, complete historical financial data, including account text data, user verbal requirements, and matching methods, is first obtained as the data basis for subsequent processing steps.

[0036] In some specific application scenarios, for example, a user's expense data can come from personal Excel ledgers, accounting software statements, etc. Furthermore, a user can say via voice, "I want to analyze last month's total household expenses and investment returns" (generating the first voice segment), and "My 'dining and food' expenses in my accounting software correspond to the 'food and beverages' category in my Excel ledger" (generating the second voice segment). The system collects and transcribes these voice recordings, categorizes them, and obtains structured input information.

[0037] Step S200: Use the BERT model to perform language parsing and intent recognition on the user's verbal financial analysis needs, and generate user goals; It should be noted that in some embodiments, such as Figure 4As shown, step S200 may include the following steps: S210, using the BERT model to perform word segmentation, part-of-speech tagging, and dependency parsing on the user's verbal financial analysis needs to obtain initial user needs analysis results; S220, identifying keywords of the user's core intent based on the initial user needs analysis results; keywords include at least one of cost analysis, budget execution rate, and financial risk analysis; S230, structuring the semantic components of the keywords of the user's core intent to generate quantified user goals.

[0038] For example, in some specific implementations, taking the user's verbal request "Compare the marketing expenses of this year and the third quarter of last year?" as an example, the BERT model performs syntactic analysis on it, identifies the core intent keyword as "expense analysis", and structures the semantic components such as "third quarter of this year", "third quarter of last year", "marketing expenses", and "effect" (which can be quantified as "return on investment"), and finally generates a quantified user goal: "Calculate and compare the return on investment of marketing expenses in Q3 of 2023 with that in Q3 of 2022".

[0039] Specifically, this invention utilizes BERT's powerful contextual understanding capabilities to transform vague and colloquial user needs into clear, structured, and quantifiable analytical objectives, effectively avoiding the ambiguity and limitations of traditional keyword- or fixed-template matching. This invention can automate the process of clarifying intent that originally required multiple business-IT communications, thereby directly reducing analytical biases and project costs caused by poor communication.

[0040] Step S300: Based on the T5 model, perform language parsing on the user's verbal account matching method to generate matching rules; It should be noted that in some embodiments, such as Figure 5 As shown, step S300 may include the following steps: S310, converting the user's verbal account matching method into a standardized account matching text sequence, and then using the encoder of the T5 model to parse the grammatical structure and identify the matching conditions and logical relationships between different accounts; S320, based on the matching conditions and logical relationships between different accounts, using the decoder of the T5 model to convert the colloquial description in the user's verbal account matching method into a logical expression of attribute-operator-value, to obtain a structured rule sequence; S330, using the T5 model to dynamically adjust the rule generation strategy of the structured rule sequence through a self-attention mechanism to obtain matching rules.

[0041] For example, in some specific implementations, a user might verbally state: "Match the 'Customer ID' in table A with the 'Customer ID' in table B. The record must be 'Active' and the 'Order Amount' in table A must be greater than the 'Credit Limit' in table B." Based on the aforementioned example, the T5 model can parse and convert this into a structured logical expression: [Matching conditions: A. Customer ID = B. Customer ID AND B. Status = 'Active' AND A. Order Amount > B. Credit Limit].

[0042] Specifically, embodiments of the present invention can transform the process of relying on technicians to manually write SQL or configure ETL rules into one where the model automatically parses and generates natural language descriptions. The present invention can solve the core defect of the prior art that is "cumbersome and error-prone". Specifically, the sequence-to-sequence generation capability of the T5 model can handle complex, multi-conditional colloquial matching logic and adapt to the different expression habits of different users. Based on this, the generated rules can be closer to the user's original intention.

[0043] Step S400: Preprocess the account text data, deconstruct and reorganize the preprocessed account text data based on the user's target, filter out items irrelevant to the user's target, and obtain the filtered account text data. It should be noted that in some embodiments, step S400 may include the following steps: preprocessing the account text data to obtain preprocessed account text data; wherein, the preprocessing includes completing the financial amount default value related items and unifying the financial data attribute values; according to the user's goal, deconstructing the preprocessed account text data, actively filtering out redundant data items and noisy records as items irrelevant to the user's goal, and extracting key data fields and entities related to the user's goal; using the key data fields and entities related to the user's goal as the filtered account text data.

[0044] For example, in some specific implementations, the user's goal is to perform "travel expense analysis." The system preprocesses the expense text (e.g., unifying "USD 1000" to "RMB 6,800"). Subsequently, the system actively filters out data items unrelated to "travel," such as noisy records like "fixed asset purchases" and "R&D personnel salaries," retaining only key fields such as "transportation expenses," "accommodation expenses," and "travel allowances," as well as related entities (such as employees and projects).

[0045] In some optional implementations, the missing value of financial amount related items refers to the missing value of certain amounts. In this case, a reference value is inferred based on the context to fill the missing value. This process can quickly obtain complete accounting text data and quickly obtain a visualized financial analysis report while ensuring accuracy. The unified financial data attribute value refers to the unification of various names and terms in the financial data.

[0046] Specifically, the embodiments of the present invention actively filter irrelevant and redundant data according to the analysis target, which can improve the efficiency and accuracy of subsequent processing and analysis, thereby avoiding the interference of "data overload" on the model. Specifically, the present invention can solve the problem of inconsistency of multi-source data through preprocessing such as default value completion and attribute value unification, thereby providing clean and standardized data input for downstream tasks.

[0047] Step S500: Based on the matching rules, the filtered account text data is spliced ​​and organized through the pointer generator network to obtain the organized account text data. It should be noted that in some embodiments, step S500 may include the following steps: deep encoding the filtered account text through the encoder of the pointer generator network, extracting the semantic features and contextual information of each account record as encoded account features; using an attention mechanism to associate the encoded account features with matching rules to identify the encoded features; wherein, the encoded features represent the relevant account entries to be spliced ​​and the association strength; using the association strength as the attention weight, based on the relevant account entries, using the decoder of the pointer generator network to decode and generate an account text sequence as the splicing result; performing logical verification and format unification on the account text sequence, and then evaluating the splicing result through preset verification rules, and outputting the splicing result that passes the evaluation as the organized account text data.

[0048] For example, in some implementations, the pointer generator network is an improved sequence-to-sequence model that combines traditional generation mechanisms with pointer networks, enabling the model to flexibly copy words from source text or generate new words from a fixed vocabulary. This model utilizes attention distributions to calculate copy probabilities, addressing the issues of handling out-of-vocabulary and rare words. While maintaining text fluency, it ensures the accuracy of technical terms and key data, making it particularly suitable for tasks requiring precise reproduction of original information, such as financial text summarization and data splicing.

[0049] In some specific application scenarios, taking the matching rule as associating "purchase order" and "inbound order" as an example, the pointer generator network reads all orders and inbound orders. Through the attention mechanism, it finds that "order number PO20231001" and "inbound order RK20231005" have the highest association strength and both point to the same material. Therefore, the information of these two records can be concatenated into a complete "purchase-inbound" flow record, and finally logical verification (such as whether the quantity is consistent) is performed.

[0050] Specifically, this embodiment of the invention utilizes an attention mechanism to dynamically determine the correlation and strength between data entries, enabling accurate and intelligent data splicing and organization. Compared to traditional rigid matching based on fixed key values, this embodiment of the invention offers higher fault tolerance and accuracy. Furthermore, this embodiment of the invention ensures the rationality and reliability of the splicing results through logical verification and preset rule evaluation, laying the foundation for constructing a high-quality knowledge graph.

[0051] Step S600: Obtain the latest policies and regulations for the current period. Based on the organized accounting text data, use the cross-attention dual-tower model enhanced with accounting knowledge to perform feature association on the policies and regulations and extract budget execution analysis data. It should be noted that in some embodiments, step S600 may include the following steps: using web crawling technology to crawl and clean the policies and regulations for the latest time period, and outputting the structured text of the policies and regulations; wherein, the latest time period is a preset time period starting from the current time; inputting the structured text and the organized accounting text data into the cross-attention dual-tower model of accounting knowledge enhancement, processing to obtain the accounting context vector as the output, forming cross-attention key-value pairs; based on the cross-attention key-value pairs, using the cross-attention layer of the cross-attention dual-tower model, using the accounting context vector as the query, weighting the vector of policies and regulations to obtain accounting fusion features; inputting the accounting fusion features into a multilayer perceptron for decoding, and outputting an investment quantitative benefit score; performing source analysis on the investment quantitative benefit score to obtain budget execution analysis data.

[0052] For example, in some specific implementations, the accounting knowledge-enhanced cross-attention dual-tower model is a deep learning model that combines a dual-tower model architecture with a cross-attention mechanism and incorporates knowledge from the financial field. The accounting knowledge-enhanced cross-attention dual-tower model of the present invention is mainly used for tasks such as semantic matching of financial data, risk prediction, or investment recommendation.

[0053] Specifically, the accounting knowledge-enhanced cross-attention dual-tower model is a deep learning architecture specifically designed for financial text analysis. Its core consists of three parts: a pre-trained accounting knowledge model, a dual-tower encoding structure, and a cross-attention fusion module. The accounting knowledge-enhanced cross-attention dual-tower model is a specialized analytical architecture for deeply linking financial accounts with policy texts. The model employs a dual-tower structure, where one tower encodes pre-trained policy and regulatory texts infused with accounting knowledge (such as accounts and standards), and the other tower encodes organized accounting data. Its core lies in the cross-attention mechanism, which allows accounting data to act as a "query," dynamically retrieving and weighting the most relevant compliance requirements and incentive clauses from the "keys" of policy texts to achieve semantic alignment. This process can automatically identify correlations such as "whether specific R&D expenses match the latest tax incentive policies," ultimately decoding and outputting quantitative policy impact analysis (such as budget execution adjustment recommendations or investment benefit scores), enabling static accounts to respond to the dynamic regulatory environment.

[0054] In some specific application scenarios, for example, the system crawls the latest policy on "increasing the deduction ratio for R&D expenses of high-tech enterprises." The dual-tower model correlates the policy text with the company's compiled R&D project accounting data. The model finds a strong correlation between the policy benefits and entities such as "software R&D department" and "hardware innovation projects," ultimately outputting an upward adjustment to the "investment quantitative benefit score" for these projects. Furthermore, it traces the source of the data to derive budget execution analysis data such as "It is estimated that tax savings of XX yuan can be achieved, and it is recommended to increase the budget for related projects."

[0055] Specifically, this invention creatively links external unstructured policies and regulations with internal financial data through deep feature correlation, enabling financial analysis to move beyond a closed and static approach and gain an awareness of the external environment. Specifically, the output budget execution analysis data includes a quantitative assessment of policy impacts, directly guiding the dynamic adjustment and optimization of the budget, thus achieving a "dynamic correction" that is difficult to achieve in existing technologies.

[0056] Step S700: Using the organized accounting text data as the main data, a multi-dimensional financial knowledge graph is constructed through entity extraction and relation modeling. It should be noted that, in some embodiments, step S700 may include the following steps: identifying and extracting financial entities from the organized accounting text data; wherein, the financial entities include project names, expense accounts, transaction items and unit information; establishing financial relationships between each financial entity based on syntactic analysis and semantic role labeling; the financial relationships include the execution status of financial accounting and budget accounting, transaction exchanges, nature of funds and project linkage; mapping the financial entities and financial relationships to a unified graph model to construct a multidimensional financial knowledge graph.

[0057] For example, in some specific implementations, from the organized data, such as "XX R&D Project" (Entity 1), "R&D Personnel Salaries" (Entity 2), and "Company A" (Entity 3), relationships are then established: Entity 1 "occupies" Entity 2 (financial relationship), Entity 2 "pays" to Entity 3 (transactional relationship), and Entity 1 "belongs" to the high-tech product catalog (project affixation). All entities and relationships constitute a network-like multidimensional knowledge graph.

[0058] Specifically, the embodiments of the present invention can transform flat accounting tables into graphs rich in semantic relationships, thereby intuitively revealing the complex network behind cash flow and business flow, providing the best data structure for in-depth analysis; among them, the graph structure is the basis for advanced analysis such as risk transmission analysis and trend prediction, and can solve the limitation of traditional report analysis that "only sees the trees and not the forest".

[0059] Step S800: Based on a multidimensional financial knowledge graph, with user goals as the target guide and budget execution analysis data as the correction item, a graph analysis deep computing engine is used to identify financial risks and predict financial risk trends, generating a visualized financial analysis report. It should be noted that the graph analysis deep computing engine is a fusion model of graph neural networks and deep reinforcement learning networks. In some embodiments, step S800 may include the following steps: based on a multidimensional financial knowledge graph, with user goals as the target guide, extracting the structure of related subgraphs using graph neural networks, and tracking the transmission path and trend characteristics of financial risks through node embedding and relationship propagation; using budget execution data as an environmental correction item, simulating regulatory dynamics using deep reinforcement learning networks, and constructing a risk identification model; inputting the transmission path and trend characteristics of financial risks into the risk identification model, analyzing the graph analysis results, and generating a visualized financial analysis report.

[0060] For example, in some specific implementations, guided by the objective of "financial risk," a graph neural network extracts a risk transmission subgraph path from a knowledge graph: "Customer A defaults on payments → causing cash flow problems for Project B → potentially triggering default on supplier C." Simultaneously, a deep reinforcement learning network incorporates "monetary policy tightening" (as a budget execution analysis data / environmental correction term), simulating the possibility of accelerated risk transmission under this environment. Ultimately, it predicts the company's overall cash flow risk level for the next 90 days and generates a risk report with a time-series trend graph.

[0061] Specifically, by combining the relational reasoning capabilities of graph neural networks with the dynamic environment simulation capabilities of deep reinforcement learning, this invention can not only identify static risk points but also simulate the transmission paths of risks in multi-entity networks and their evolution trends in future environments, demonstrating predictive capabilities far exceeding those of traditional statistical models. The final output is a visualized report that integrates specific risk paths, trend predictions, and the influence of external factors, directly serving management decisions and significantly enhancing the practical value and ease of use of financial analysis.

[0062] To explain in detail the principle of the technical solution of the present invention, the overall process of the present invention will be described below with reference to some specific embodiments. It is easy to understand that the following is an explanation of the technical principle of the present invention and should not be regarded as a limitation of the present invention.

[0063] To address the shortcomings of existing technologies, such as Figure 6 and Figure 7 As shown, this embodiment provides an intelligent financial analysis method using natural language processing, which can be implemented through the following steps: S1. Obtain historical financial data; historical financial data includes: account text data, user-described account matching methods, and user-described financial analysis requirements; the account text data is multi-source heterogeneous text data, specifically including the following steps: S11. Collect and transcribe voice data segments of users' verbal financial analysis needs, and classify and merge them to obtain users' verbal financial analysis needs.

[0064] S12. Extract data from a multi-source system consisting of personal financial statements, corporate financial systems, and payment platforms to obtain account text data.

[0065] S13. Collect and transcribe audio data segments of the interrelationships between different data reports and the matching content of key items in the user's dictated account text data, and classify and merge them to obtain the user's dictated account matching method.

[0066] S2. Utilize the BERT model to perform language parsing and intent recognition on users' verbal financial analysis needs, and generate user goals. This includes the following steps: S21. Use the BERT model to perform word segmentation, part-of-speech tagging, and dependency parsing on users' verbal financial analysis needs to obtain the initial user needs analysis results.

[0067] S22. Identify keywords representing the core intent of users based on the initial user needs analysis results; keywords should include at least one of the following: cost analysis, budget execution rate, and financial risk analysis.

[0068] S23. Structure the semantic components of keywords representing the user's core intent to generate quantified user goals.

[0069] S3. Based on the T5 model, perform language parsing on the user's verbal account matching method and generate matching rules, specifically including the following steps: S31. The user's verbal account matching method is converted into a standardized account matching text sequence. The encoder of the T5 model parses the grammatical structure and identifies the matching conditions and logical relationships between different accounts.

[0070] S32. Based on the matching conditions and logical relationships between different accounts, the T5 model decoder converts the colloquial description of the user's verbal account matching method into a logical expression of attribute-operator-value, resulting in a structured rule sequence.

[0071] S33. Using the T5 model, the rule generation strategy of the structured rule sequence is dynamically adjusted through a self-attention mechanism to obtain matching rules.

[0072] S4. Preprocess the account text data, and deconstruct and reorganize the preprocessed account text data based on the user's objectives, filtering out items irrelevant to the user's objectives to obtain filtered account text data; the preprocessing includes: filling in the missing financial amount related items and unifying the financial data attribute values.

[0073] Step S4 specifically includes the following steps: S41. Preprocess the account text data to obtain preprocessed account text data.

[0074] S42. Based on the user's objectives, deconstruct the accounting text data, actively filter out redundant data items and noisy records as items irrelevant to the user's objectives, and extract key data fields and entities related to the user's objectives.

[0075] S43. Output the key data fields and entities related to the user's objectives as filtered account text data.

[0076] S5. Based on the matching rules, the filtered account text data is concatenated and organized using a pointer generator network to obtain the organized account text data. This process includes the following steps: S51. The filtered account text is deeply encoded by the encoder of the pointer generator network to extract the semantic features and contextual information of each account record.

[0077] S52. Use the attention mechanism to associate the encoded account features with the matching rules to identify the encoded features; the encoded features are: the relevant account items to be concatenated and the association strength.

[0078] S53. Using the correlation strength as the attention weight, the decoder of the pointer generator network is used to decode and generate the sequence of account text based on the relevant account entries.

[0079] S54. Perform logical verification and format standardization on the account text sequence, evaluate the splicing result through preset verification rules, and output the splicing result that passes the evaluation as the organized account text data.

[0080] S6. Obtain the latest policy and regulations for the current time period, and based on the organized accounting text data, use the cross-attention dual-tower model enhanced with accounting knowledge to perform feature association on the policy and regulations, and extract budget execution analysis data; the latest time period is a preset time period starting from the current time.

[0081] Step S6 specifically includes the following steps: S61. Crawl and clean the latest policies and regulations, and output the structured text of the policies and regulations.

[0082] S62. Input the structured text of policies and regulations and the organized accounting text data into the cross-attention dual-tower model that enhances accounting knowledge, and output the accounting context vector to form cross-attention key-value pairs.

[0083] S63. Based on the cross-attention key-value pairs, using the cross-attention layer of the cross-attention dual-tower model, and taking the account context vector as the query, the policy and regulation vectors are weighted to obtain the financial and accounting integration features.

[0084] S64. Input the financial and accounting integration features into the MLP decoder and output a quantitative investment benefit score.

[0085] S65. Conduct source analysis on the quantitative benefit score of investment to obtain budget execution analysis data.

[0086] S7. Using the organized accounting text data as the main data, construct a multi-dimensional financial knowledge graph through entity extraction and relation modeling, specifically including the following steps: S71. Automatically identify and extract financial entities from the organized accounting text data. The financial entities include at least: project name, expense account, transaction item, and unit information.

[0087] S72. Based on syntactic analysis and semantic role labeling, establish financial relationships between financial entities; financial relationships should include at least: financial accounting, budget accounting execution status, transaction history, nature of funds, and project linkage status.

[0088] S73. Map financial entities and financial relationships to a unified graph model to construct a multidimensional financial knowledge graph.

[0089] S8. Based on a multidimensional financial knowledge graph, guided by user goals and using budget execution analysis data as a correction factor, a graph analysis deep computing engine is employed to identify financial risks and predict financial risk trends, generating a visualized financial analysis report. The specific steps include: S81. Based on a multidimensional financial knowledge graph, with user goals as the target guide, a graph neural network is used to extract the structure of related subgraphs, and the transmission path and trend characteristics of financial risks are tracked through node embedding and relationship propagation.

[0090] S82. Using budget execution data as an environmental correction item, a risk identification model is constructed by simulating regulatory dynamics using deep reinforcement learning networks.

[0091] S83. Input the financial risk transmission path and financial risk trend characteristics into the risk identification model, analyze the results of the graph analysis, and generate a visualized financial analysis report.

[0092] Optionally, the visualized financial analysis report is a comprehensive financial analysis report that includes risk heatmaps, trend curves, and correlation networks.

[0093] In practical applications, such as Figure 8 As shown, the graph neural network (GNN) acts as the "perception engine," deeply understanding the complex relationships within the financial knowledge graph and generating an information-rich "feature vector" for each analytical object. Then, deep reinforcement learning (DRL) acts as the "decision-making brain," using the user's specific goals as action guidelines and referencing external budget execution data in real time to analyze, judge, and correct the features provided by the GNN. Ultimately, this fusion model outputs dynamically evaluated financial risk identification results and financial trend prediction conclusions, generating a comprehensive, accurate, and personalized visualized financial analysis report.

[0094] Based on the above content, the following example will be used to illustrate the needs of public institution Z, namely, "to analyze the budget execution of each project in the previous quarter and, based on the requirements of the superior department on the budget execution, to make arrangements for the budget execution tasks in the next quarter".

[0095] Step 1: Collect the user's voice requests and transcribe them into text. The BERT model analyzes the text and identifies the core intent keywords "budget execution status" and "budget execution task assignment." The intent is then structured and quantified into clear user goals.

[0096] 1) Calculate the ratio of actual expenditure to budget for all projects in the previous quarter (execution rate).

[0097] 2) Identify abnormal projects with excessively high execution rates (potentially due to rushed spending) or excessively low execution rates (idle budget).

[0098] 3) Generate budget adjustment recommendations for the next quarter that align with policy guidelines.

[0099] Step Two: The user verbally describes the data matching method (e.g., "Link the 'planned amount' in the budget sheet with the 'actual expenditure' in the reimbursement system using the project number"). The T5 model parses this into a structured matching rule: [Budget Sheet.Project Number] = [Reimbursement Sheet.Project Number] AND [Budget Sheet.Expense Item] ≈ [Reimbursement Sheet.Expense Category]. Simultaneously, the system extracts account text from multiple sources, including the financial system and reimbursement platform, and performs missing value completion (e.g., adding invoice dates for transactions without dates) and attribute value unification (e.g., unifying "R&D" and "Research and Development Expenditure" into "Research and Development Costs").

[0100] Step 3: Based on the matching rules generated in Step 2, the pointer generator network is enabled. Its encoder deeply understands each budget entry and expenditure record, the attention mechanism calculates the correlation strength according to the rules, and the decoder intelligently "concatenates" the budget values ​​and actual expenditure values ​​of the same project from different tables into a complete "project budget execution record", outputting as well-organized account text data.

[0101] Step Four: Automatically crawl the latest policy and regulatory texts from higher-level departments (such as the Ministry of Finance and the State-owned Assets Supervision and Administration Commission) regarding "budget execution rate assessment requirements" and "revitalizing existing funds." The accounting knowledge-enhanced cross-attention dual-tower model begins operation: one tower encodes the policy text (e.g., "requiring accelerated budget execution"), and the other tower encodes the project execution data output in Step Three. Through the cross-attention mechanism, the model can discover a high correlation between "a project's execution rate of only 30%" and the policy clause on "recovering long-term idle funds," thus outputting budget execution analysis data—specifically, a "policy compliance score" and early warning prompts for each project.

[0102] Step 5: Extract financial entities such as "Project A," "R&D Expenses," and "Department B," as well as financial relationships such as "Overspending," "Belongs to," and "Surplus" from the assembled data to construct a multi-dimensional financial knowledge graph. Based on this, the Graph Neural Network (GNN) automatically tracks the flow of funds and the transmission path of risks. For example, it discovers that the overspending funds of "Project C" (execution rate 150%) mainly flowed to "Project D" under the same person in charge, forming an abnormal fund flow closed loop, and initially identifying the transmission path of financial risks.

[0103] Step Six: With the goal of "optimizing budget allocation and controlling risks," the policy warnings obtained in Step Four (such as "Project A's low execution rate faces the risk of fund recovery") are used as environmental correction items. A Deep Reinforcement Learning (DRL) agent simulates decisions within the graph environment (such as "cutting the budget of Project C and partially transferring it to Project A"). Through multiple rounds of trial and error, the system finds the optimal budget adjustment plan that simultaneously improves overall policy compliance and execution efficiency. Finally, the system integrates the graph analysis results and the DRL simulation plan to automatically generate a visualized financial analysis report that includes a ranking of the previous quarter's execution rate, a diagram of abnormal project risks, and a budget allocation suggestion table for the next quarter.

[0104] The example of financial analysis of public institution Z in this invention demonstrates the complete process from verbal requirements to intelligent analysis reports, which shortens the traditional financial analysis work that takes several days to complete to the hour level. It can also dynamically combine budget execution to provide decision-making suggestions, significantly improving the efficiency and intelligence level of financial analysis.

[0105] In summary, this invention constructs an intelligent financial analysis method for intelligent decision-making from multi-source heterogeneous data. First, it acquires complete historical financial data, including account text data, user-described needs, and matching methods. The BERT model is used to perform deep semantic parsing and intent recognition on the user-described financial analysis needs, generating quantifiable and executable analysis objectives. Simultaneously, the T5 model is used to parse the user-described account matching logic, automatically generating executable matching rules, simplifying the manual operation process required for traditional financial statement matching, and reducing the time spent on comparison and communication due to matching issues between different reports. Preprocessing of the account text data ensures data quality consistency. Based on user objectives, the preprocessed data is deconstructed and reorganized, and irrelevant items are filtered out. A pointer generator network is used to intelligently stitch the account data together, improving data processing efficiency. By introducing the latest policies and regulations as dynamic correction factors, and using a cross-attention dual-tower model enhanced with accounting knowledge, the processed account data is deeply feature-associated with the latest policies and regulations to extract time-sensitive budget execution analysis data. This enables timely responses to regulatory changes and corresponding corrections, providing accurate data support for subsequent analysis. A multi-dimensional financial knowledge graph is constructed through entity extraction and relation modeling. A computational engine integrating graph neural networks and deep reinforcement learning is employed to identify financial risks and predict trends, guided by user objectives and combined with budget execution data. This not only tracks complex financial relationship networks but also continuously optimizes the accuracy of risk predictions through reinforcement learning. Finally, a visualized financial analysis report containing risk identification results and trend predictions is automatically generated, achieving full automation from data collection, processing, and analysis to report generation.

[0106] Specifically, this invention addresses the pain points of traditional financial analysis, such as low data processing efficiency, delayed policy response, and high professional knowledge threshold. It reduces the matching time of various financial statements in accounting text data, enables timely integration with the latest policies and regulations for dynamic financial analysis, and automatically generates visual analysis reports, thereby improving the timeliness and convenience of financial analysis and providing more intelligent and automated financial decision support for corporate and individual users.

[0107] like Figure 9 As shown, this embodiment of the invention also provides an intelligent financial analysis device 900 using natural language processing, which can implement the above-described method. This device may include: The first module 910 is used to acquire historical financial data; the historical financial data includes account text data, user-described financial analysis requirements, and user-described account matching methods. The second module 920 is used to perform language parsing and intent recognition on the user's verbal financial analysis needs using the BERT model, and generate user goals. The third module 930 is used to perform language parsing on the user's verbal account matching method based on the T5 model and generate matching rules. The fourth module 940 is used to preprocess the account text data. Based on the user's goal, the preprocessed account text data is deconstructed and reconstructed, and irrelevant items to the user's goal are filtered out to obtain the filtered account text data. The fifth module 950 is used to concatenate and organize the filtered account text data based on matching rules through a pointer generator network to obtain the organized account text data. Module 6, 960, is used to obtain the latest policies and regulations for the current period. Based on the organized accounting text data, it uses a cross-attention dual-tower model enhanced with accounting knowledge to perform feature association on the policies and regulations and extract budget execution analysis data. Module 7, 970, is used to construct a multidimensional financial knowledge graph by taking the organized accounting text data as the main data and using entity extraction and relation modeling. Module 8, 980, is used to identify financial risks and predict financial risk trends based on a multidimensional financial knowledge graph, guided by user goals and with budget execution analysis data as correction items. It employs a graph analysis deep computing engine to generate visualized financial analysis reports.

[0108] It is understood that the content of the above method embodiments is applicable to the present device embodiments. The specific functions implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0109] This invention also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described above. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0110] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0111] like Figure 10 As shown, Figure 10 The hardware structure of an electronic device 1000 according to another embodiment is illustrated. The electronic device 1000 includes: The processor 1001 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (aSIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present invention. The memory 1002 can be implemented as a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RaM). The memory 1002 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1002 and is called and executed by the processor 1001. Input / output interface 1003 is used to implement information input and output; The communication interface 1004 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). Bus 1005 transmits information between various components of the device (e.g., processor 1001, memory 1002, input / output interface 1003, and communication interface 1004); The processor 1001, memory 1002, input / output interface 1003 and communication interface 1004 are connected to each other within the device via bus 1005.

[0112] The electronic device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0113] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0114] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0115] This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0116] It is understood that the content of the above method embodiments is applicable to the embodiments of this program product. The specific functions implemented by the embodiments of this program product are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0117] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0118] The intelligent financial analysis method, device, electronic device, storage medium, and program product using natural language processing provided in this invention first acquires complete historical financial data, including account text data, user-stated requirements, and matching methods. It employs the BERT model to perform deep semantic parsing and intent recognition on the user's stated financial analysis requirements, generating quantifiable and executable analysis objectives. Simultaneously, it utilizes the T5 model to parse the user's stated account matching logic, automatically generating executable matching rules, simplifying the manual operation process required for traditional financial statement matching, and reducing the time spent on comparison and communication arising from matching issues between different reports. Preprocessing the account text data ensures data quality consistency. Based on user objectives, the preprocessed data is deconstructed and reorganized, and irrelevant items are filtered out. A pointer generator network is used to intelligently stitch together the account data, improving data processing efficiency. By introducing the latest policies and regulations as dynamic correction factors, and using a cross-attention dual-tower model enhanced with accounting knowledge, the processed account data is deeply feature-associated with the latest policies and regulations to extract time-sensitive budget execution analysis data. This enables timely responses to regulatory changes and corresponding corrections, providing accurate data support for subsequent analysis. This invention constructs a multi-dimensional financial knowledge graph through entity extraction and relational modeling, and employs a computational engine that integrates graph neural networks and deep reinforcement learning. Guided by user objectives, it combines budget execution data to identify financial risks and predict trends. It not only tracks complex financial relationship networks but also continuously optimizes the accuracy of risk prediction through reinforcement learning. Finally, it automatically generates a visualized financial analysis report containing risk identification results and trend predictions, achieving full automation from data collection, processing, and analysis to report generation. This invention addresses the pain points of traditional financial analysis, such as low data processing efficiency, delayed policy response, and high professional knowledge barriers. It reduces the matching time of various financial statements in accounting text data, enables timely integration of the latest policies and regulations for dynamic financial analysis, and automatically generates visualized analysis reports, improving the timeliness and convenience of financial analysis and providing more intelligent and automated financial decision support for both institutional and individual users.

[0119] The embodiments described in this invention are for the purpose of more clearly illustrating the technical solutions of the embodiments of this invention, and do not constitute a limitation on the technical solutions provided by the embodiments of this invention. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this invention are also applicable to similar technical problems.

[0120] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of the present invention, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0121] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0122] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0123] The preferred embodiments of the present invention have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present invention. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and spirit of the present invention should be within the scope of the claims of the present invention.

Claims

1. A natural language processing-based intelligent financial analysis method, characterized in that, The method includes the following steps: Acquire historical financial data; wherein, the historical financial data includes account text data, user-reported financial analysis requirements, and user-reported account matching methods; The BERT model is used to perform language parsing and intent recognition on the user's verbal financial analysis needs to generate user goals; Based on the T5 model, language parsing is performed on the user's verbal account matching method to generate matching rules; The account text data is preprocessed, and the preprocessed account text data is deconstructed and reconstructed based on the user target, and irrelevant items to the user target are filtered out to obtain filtered account text data. Based on the matching rules, the filtered account text data is spliced ​​and organized through a pointer generator network to obtain the organized account text data. The latest policies and regulations for the current period are obtained. Based on the organized accounting text data, the cross-attention dual-tower model enhanced with accounting knowledge is used to perform feature association on the policies and regulations, and extract budget execution analysis data. Using the organized accounting text data as the main data, a multidimensional financial knowledge graph is constructed through entity extraction and relation modeling. Based on the multidimensional financial knowledge graph, with the user's goal as the objective guide and the budget execution analysis data as the correction item, a graph analysis deep computing engine is used to identify financial risks and predict financial risk trends, generating a visualized financial analysis report.

2. The method according to claim 1, characterized in that, The process of obtaining historical financial data includes the following steps: Data is extracted from a multi-source system consisting of personal financial statements, corporate financial systems, and payment platforms to obtain the account text data; wherein, the account text data is multi-source heterogeneous text data; Collect and transcribe the first audio data segment of the user's spoken financial analysis needs, and classify and merge them to obtain the user's spoken financial analysis needs; The second voice data segment, which collects and transcribes the interrelationships between different data reports and the matching content of key items in the user's dictated account text data, is classified and merged to obtain the user's dictated account matching method.

3. The method according to claim 1, characterized in that, The process of using the BERT model to parse and identify the user's verbal financial analysis needs and generate user goals includes the following steps: The BERT model was used to perform word segmentation, part-of-speech tagging, and dependency parsing on the user's verbal financial analysis requirements to obtain the initial user requirement analysis results. Based on the initial user needs analysis results, keywords are identified to pinpoint the user's core intent; these keywords include at least one of cost analysis, budget execution rate, and financial risk analysis. The semantic components of the keywords representing the user's core intent are structured to generate a quantified user goal.

4. The method according to claim 1, characterized in that, The process of parsing the user's verbal account matching method based on the T5 model to generate matching rules includes the following steps: The user-described account matching method is converted into a standardized account matching text sequence, and then the T5 model encoder is used to parse the grammatical structure to identify the matching conditions and logical relationships between different accounts. Based on the matching conditions and logical relationships between different accounts, the T5 model decoder converts the colloquial description in the user's verbal account matching method into a logical expression of attribute-operator-value, resulting in a structured rule sequence. The T5 model is used to dynamically adjust the rule generation strategy of the structured rule sequence through a self-attention mechanism to obtain the matching rules.

5. The method according to claim 1, characterized in that, The preprocessing of the account text data, which involves deconstructing and reconstructing the preprocessed account text data based on the user objective, and filtering out items irrelevant to the user objective to obtain filtered account text data, includes the following steps: The accounting text data is preprocessed to obtain preprocessed accounting text data; wherein, the preprocessing includes filling in the missing values ​​of financial amount related items and unifying the financial data attribute values; Based on the user objective, the preprocessed account text data is deconstructed, redundant data items and noisy records are actively filtered out as items irrelevant to the user objective, and key data fields and entities related to the user objective are extracted. The key data fields and entities related to the user's target are used as the filtered account text data.

6. The method according to claim 1, characterized in that, The process of concatenating and organizing the filtered account text data using a pointer generator network based on the matching rules to obtain the organized account text data includes the following steps: The filtered account text is deeply encoded by the encoder of the pointer generator network, and the semantic features and contextual information of each account record are extracted as the encoded account features. An attention mechanism is used to associate the encoded account features with the matching rules to identify the encoded features; wherein, the encoded features represent the relevant account items to be concatenated and the strength of the association. Using the correlation strength as the attention weight, and based on the relevant account entries, the decoder of the pointer generator network is used to decode and generate an account text sequence as the concatenation result; The sequence of account text is logically validated and formatted. The splicing result is then evaluated using preset validation rules. The splicing result that passes the evaluation is output as the organized account text data.

7. The method according to claim 1, characterized in that, The process of obtaining the latest policies and regulations for a given period, based on the compiled accounting text data, utilizes a cross-attention dual-tower model enhanced with accounting knowledge to perform feature association on the policies and regulations, and extracts budget execution analysis data, including the following steps: The latest policy and regulation data is crawled and cleaned using web crawling technology, and the structured text of the policy and regulation data is output; wherein, the latest time period is a preset time period starting from the current time. The structured text and the organized account text data are input into the cross-attention dual-tower model for enhanced accounting knowledge, and the account context vector is processed as the output to form cross-attention key-value pairs. Based on the cross-attention key-value pairs, using the cross-attention layer of the cross-attention dual-tower model, and with the account context vector as the query, the vector of the policy and regulations is weighted to obtain the financial and accounting integration features; The financial and accounting fusion features are input into a multilayer perceptron for decoding, and an investment quantitative benefit score is output. The budget execution analysis data is obtained by tracing the source of the quantitative benefit score of the investment.

8. The method according to claim 1, characterized in that, The process of using the organized accounting text data as the main data and constructing a multidimensional financial knowledge graph through entity extraction and relation modeling includes the following steps: Financial entities are identified and extracted from the organized accounting text data; wherein, the financial entities include project name, expense account, transaction item and unit information; Based on syntactic analysis and semantic role labeling, financial relationships are established between each of the financial entities; the financial relationships include the execution status of financial accounting and budgetary accounting, transaction exchanges, nature of funds, and project linkage. By mapping the financial entities and financial relationships to a unified graph model, a multidimensional financial knowledge graph is constructed.

9. The method according to claim 1, characterized in that, The graph analysis deep computing engine is a fusion model of graph neural networks and deep reinforcement learning networks. Based on the multidimensional financial knowledge graph, with the user's goal as the objective guide and the budget execution analysis data as the correction term, the graph analysis deep computing engine is used to identify financial risks and predict financial risk trends, generating a visualized financial analysis report, including the following steps: Based on the multidimensional financial knowledge graph, and taking the user's goal as the target guide, the graph neural network is used to extract the structure of the related subgraph, and the financial risk transmission path and financial risk trend characteristics are tracked through node embedding and relationship propagation. The budget execution data is used as an environmental correction item, and the deep reinforcement learning network is used to simulate regulatory dynamics to construct a risk identification model. The financial risk transmission path and the financial risk trend characteristics are input into the risk identification model to analyze the results of the graph analysis and generate a visual financial analysis report.

10. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method according to any one of claims 1 to 9.

Images