AI-based automatic enterprise financial and economic analysis method, apparatus and device, and medium

Abstract

The invention relates to an AI-based automatic enterprise financial and economic analysis method, device and equipment and a medium, and the method comprises the steps: carrying out the conceptualized mapping of multi-source financial and economic data, and generating a conceptualized data set with financial semantic tags; reasoning prediction is carried out on the set based on a graph neural network, and an index prediction value and an interlayer dynamic attention matrix representing an internal reasoning process are generated; constructing a logic evidence chain by using the matrix, and generating an enhanced logic evidence chain conforming to financial logic through backtracking and screening; and finally, generating a visual analysis report supporting interactive traceability based on the evidence chain and the specific numerical value. According to the method, the information flow in the prediction model is converted into a traceable and verifiable professional logic chain, so that the problems that the prediction result of an existing AI financial analysis method is poor in interpretability and non-transparent in logic are effectively solved, and the credibility and auditing performance of the analysis result in professional decision making are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to an AI-based automated enterprise financial analysis method, apparatus, equipment, and medium. Background Technology

[0002] With the deep integration of artificial intelligence (AI) technology into the fintech field, AI-based automated corporate financial analysis methods have become important tools for enterprises to conduct financial forecasting, risk insight, and decision support. These methods typically utilize machine learning models, especially deep learning technology, to automatically process massive amounts of structured financial data (such as financial statements) and unstructured text information (such as annual reports and news), aiming to replace or assist manual work and achieve efficient and accurate quantitative analysis and forecasting of key indicators such as corporate financial condition, cash flow, and credit risk.

[0003] However, existing AI financial analysis methods generally focus on improving the accuracy and computational efficiency of predictive models, and their core paradigm can be viewed as a "black box" or "grey box" system. While the prediction results may possess high statistical reliability, the process by which the model derives the final conclusion from complex, multi-source data lacks transparency. Specifically, existing methods struggle to generate explanations with rigorous logical chains that align with the thinking habits of human financial experts. For instance, when a system predicts a company's future cash flow will be strained, analysts often cannot ascertain whether this judgment stems primarily from a deterioration in accounts receivable turnover, an abnormal backlog of inventory, or a sharp increase in short-term liabilities, let alone trace the causal transmission path between these factors. This lack of explanatory power makes it difficult to fully trust and adopt prediction results in serious scenarios requiring high reliability and auditability, such as financial risk control, investment decisions, and compliance supervision. It also makes it exceptionally difficult to identify the root cause and optimize the model when predictions deviate. Summary of the Invention

[0004] Based on this, the purpose of the present invention is to provide an AI-based automated enterprise financial analysis method, apparatus, equipment, and medium that can automatically generate chain-like explanations that conform to professional logic.

[0005] The objective of this invention is achieved through the following solution:

[0006] In a first aspect, the present invention provides an AI-based automated enterprise financial analysis method, comprising the following steps:

[0007] S1: Conceptualize the acquired multi-source heterogeneous financial data of enterprises, and uniformly map unstructured text descriptions and structured financial data to the semantic space of a pre-constructed financial concept graph to generate a conceptual data set with financial semantic labels.

[0008] S2: Based on the graph neural network architecture, perform reasoning and prediction on conceptual data sets, simulate the logical relationships between financial concepts for information propagation and aggregation, and generate the predicted values ​​of target financial indicators and inter-layer dynamic attention matrices based on the graph attention mechanism.

[0009] S3: Construct a logical evidence chain for the inter-layer dynamic attention matrix, backtrack the weight information recorded in the matrix based on the attention-guided path backtracking mechanism, calculate the contribution of each potential information flow path to the final prediction, select the key concept reasoning path with the greatest contribution, and combine it with the preset financial knowledge graph to complete and verify the logical coherence of the selected path, thereby generating a strengthened logical evidence chain.

[0010] S4: Based on the strengthening of the logical evidence chain, each abstract financial concept is matched and fused with the specific numerical values ​​anchored in the conceptual data set to generate quantitative explanatory statements. Combined with the indicator prediction values, the prediction results and multiple explanatory statements are assembled into a structured narrative to generate a visual analysis report that supports interactive traceability.

[0011] In one embodiment, S1 of the AI-based automated enterprise financial analysis method provided by the present invention specifically includes the following steps:

[0012] S11: Atomize the acquired accounting subjects, financial ratios and analytical indicators for concept extraction and relationship definition. Based on accounting standards and financial analysis framework, construct an initial financial knowledge graph containing concept nodes and logical edges to generate a financial concept graph.

[0013] S12: Preprocess and semantically analyze the acquired multi-source heterogeneous financial data of enterprises, convert structured financial statement data into time-series feature vectors, and perform entity recognition and sentiment analysis on unstructured financial texts to generate standard data units.

[0014] S13: Based on the financial concept graph, perform concept association processing on the standard data units, map each data unit to the corresponding node in the preset financial concept graph, and calculate the association strength between it and the relevant concept nodes according to accounting standards to generate weighted concept node data.

[0015] S14: The concept node data undergoes semantic space embedding processing. The domain pre-trained model is called to encode each concept node and its associated weights into a unified low-dimensional dense vector, generating a conceptual data set with financial semantic labels.

[0016] In one embodiment, S2 of the AI-based automated enterprise financial analysis method provided by the present invention specifically includes the following steps:

[0017] S21: Perform graph structure modeling on the conceptual dataset, construct an initial graph with financial concepts as nodes and the interrelationships between concepts as edges, and initialize the conceptual vectors as node features to generate a financial concept relationship graph;

[0018] S22: Perform multi-layer attention message propagation processing on the financial concept relationship graph. In each layer, each node aggregates the feature information of its neighboring nodes according to the attention weight, and updates its own features through non-linear transformation to generate a node state matrix containing high-order association information.

[0019] S23: Based on the node state matrix, perform prediction decoding and attention recording processing. Extract the features of the nodes corresponding to the target financial indicators from the final layer node state and perform linear mapping to obtain the indicator prediction value. At the same time, organize the attention weights between all nodes calculated in each layer by layer to generate an inter-layer dynamic attention matrix.

[0020] In one embodiment, S3 of the AI-based automated enterprise financial analysis method provided by the present invention specifically includes the following steps:

[0021] S31: Perform importance propagation path backtracking on the interlayer dynamic attention matrix. Starting from the predicted target node, find the upstream node and connection edge that contributes the most to it based on the weight values ​​of each layer in the matrix, and generate multiple candidate concept reasoning paths.

[0022] S32: Perform contribution quantification scoring on candidate concept reasoning paths, comprehensively calculate the cumulative intensity of attention weights, path length, and path smoothness on each path, and generate a contribution score for each path.

[0023] S33: Perform critical path screening on the candidate paths after scoring, sort them based on contribution scores and select the Top-K paths that exceed the preset threshold to generate a set of key concept reasoning paths;

[0024] S34: Perform logical completeness checks and supplementation on the set of key concept reasoning paths, compare each path with the preset financial knowledge graph, insert intermediate concept nodes in missing logical links to form a complete causal chain, and generate a reinforced logical evidence chain that conforms to professional logic.

[0025] In one embodiment, the contribution score calculation formula of the AI-based automated enterprise financial analysis method provided by the present invention is as follows:

[0026]

[0027] in, Let P be the contribution score for path P. The first one obtained from the inter-layer dynamic attention matrix From the source concept node in the layer To the target concept node Attention weights Let P be the number of edges, i.e., the path length. This is the length penalty coefficient. This is the adjustment coefficient for the knowledge graph. Concept nodes obtained from a pre-defined financial knowledge graph and The strength of the logical connection between them.

[0028] In one embodiment, S4 of the AI-based automated enterprise financial analysis method provided by the present invention specifically includes the following steps:

[0029] S41: Perform concept-numerical anchoring processing on the enhanced logical evidence chain, traverse each concept node in the evidence chain, find all original values ​​and their weights associated with the conceptual data from the conceptual data set, and generate quantitative evidence units.

[0030] S42: Based on the evidence unit, perform natural language sentence generation processing, call the financial analysis logic template to transform the numerical changes and conceptual relationships in the evidence unit into a coherent text description, and generate professional explanatory statements;

[0031] S43: Perform report structure assembly processing on the explanatory statements, sort and group multiple explanatory statements according to the importance of the evidence chain, and integrate and format them with indicator prediction values ​​and key charts to generate a visual analysis report that supports interactive tracing.

[0032] Secondly, the present invention provides an AI-based automated enterprise financial analysis device, which is equipped with the following modules:

[0033] The financial data conceptual mapping module is used to perform conceptual mapping on the acquired multi-source heterogeneous financial data of enterprises, and to uniformly map unstructured text descriptions and structured financial data to the semantic space of a pre-built financial concept graph, generating a set of conceptual data with financial semantic labels.

[0034] The graph neural network reasoning and prediction module is used to reason and predict conceptual data sets based on graph neural network architecture, simulate the logical relationship between financial concepts for information propagation and aggregation, and generate the indicator prediction value of the target financial indicator and the inter-layer dynamic attention matrix based on the graph attention mechanism.

[0035] The logical evidence chain strengthening module is used to construct a logical evidence chain for the inter-layer dynamic attention matrix. Based on the attention-guided path backtracking mechanism, it backtracks the weight information recorded in the matrix. By calculating the contribution of each potential information flow path to the final prediction, it selects the key concept reasoning path with the greatest contribution and combines it with the preset financial knowledge graph to complete and verify the logical coherence of the selected path, thereby generating a strengthened logical evidence chain.

[0036] The visualization analysis report generation module is used to match and integrate each abstract financial concept in the reinforced logical evidence chain with the specific values ​​anchored in the conceptualized data set, generate quantitative explanatory statements, and combine the prediction results with multiple explanatory statements to assemble a structured narrative, generating a visualization analysis report that supports interactive traceability.

[0037] Thirdly, this application provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement any of the above-mentioned AI-based automated enterprise financial analysis methods.

[0038] Fourthly, this application provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements any of the aforementioned AI-based automated enterprise financial analysis methods.

[0039] In summary, the AI-based automated enterprise financial analysis method provided in this application transforms complex raw information into standardized semantic representations that are machine-understandable and conform to professional contexts through a unified conceptual mapping of multi-source heterogeneous data. The graph attention neural network-based inference and prediction mechanism not only generates accurate indicator predictions but, more importantly, simultaneously captures and solidifies the complete attention trajectory of information aggregation and decision-making within the model, thus transforming the "black box" prediction process into traceable "white box" data for the first time. Through path backtracking and logical completion based on the attention matrix, it automatically constructs chain evidence that conforms to accounting reconciliation and financial theory, ensuring that the final prediction conclusion has a clear, coherent, and professional causal explanation, greatly improving the credibility and acceptability of the analysis results. By generating interactive analysis reports that integrate quantitative values ​​and qualitative logic, it provides decision-makers with intuitive and credible insights and supports tracing back from the conclusion to the original data and intermediate inference stages, achieving full-process transparency that meets stringent audit requirements, fundamentally enhancing the practicality and reliability of AI systems in high-risk financial decision-making scenarios.

[0040] To better understand and implement this invention, the following detailed description is provided in conjunction with the accompanying drawings. Attached Figure Description

[0041] Figure 1A flowchart illustrating an AI-based automated enterprise financial analysis method provided in this application embodiment;

[0042] Figure 2 A flowchart illustrating the process of generating a strengthened logical evidence chain that conforms to professional logic, provided for embodiments of this application;

[0043] Figure 3 This is a schematic diagram of the structure of an AI-based automated enterprise financial analysis device, provided as another embodiment of this application. Detailed Implementation

[0044] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Preferred embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the invention.

[0045] 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 in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0046] In one embodiment, such as Figure 1 As shown, an AI-based automated enterprise financial analysis method is provided. This embodiment illustrates the method's application to a terminal, but it is understood that the method can also be applied to a server, or to a device including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0047] S1: Conceptualize the acquired multi-source heterogeneous financial data of enterprises, and uniformly map unstructured text descriptions and structured financial data to the semantic space of a pre-constructed financial concept graph to generate a conceptual data set with financial semantic labels.

[0048] Specifically, the multi-source heterogeneous financial data encompasses both structured financial data and unstructured text data. Structured financial data originates from quantitative indicators in companies' quarterly and annual balance sheets, profit and loss statements, and cash flow statements. Unstructured text data comes from relevant sections of companies' annual reports, industry research reports, responses to regulatory inquiries, and press releases from financial media. The system builds a unified semantic framework based on a pre-constructed financial concept graph, which is jointly constructed through ontology and a rule engine. Preferably, the system first defines a set of core financial concepts based on relevant accounting standards and industry practices, divides financial concepts into different levels, and assigns unique semantic identifiers. Then, it uses a specific language to define various relationships between concepts and corresponding logical constraints. Finally, a specific algorithm maps concepts and relationships to a low-dimensional semantic space, ensuring that the cosine similarity of semantically similar concept vectors remains consistent.

[0049] Furthermore, the system cleans the acquired structured data, fills in missing values, and standardizes indicator names. For unstructured text, it uses a language model finely tuned for the financial domain for word segmentation, stop word removal, entity recognition, and relation extraction. During semantic mapping, the system directly maps the standardized structured indicators to corresponding financial concepts using a pre-defined mapping table, generating structured data semantic identifiers containing semantic tags. Financial entities extracted from the text are input into a semantic matching model to calculate the semantic similarity between the entities and graph concepts. Concepts with matching scores are selected as mapping targets, generating unstructured data semantic identifiers containing semantic tags. The system integrates these two types of semantic identifier data to form a unified format conceptual data set. Each data item in this set includes a semantic tag identifier, a financial concept identifier, original data reference, semantic similarity score, and data timestamp, stored in a distributed database to support subsequent rapid querying and association.

[0050] S2: Based on the graph neural network architecture, it performs inference and prediction on conceptual data sets, simulates the logical relationships between financial concepts to carry out information propagation and aggregation, and generates the indicator prediction values ​​of target financial indicators and inter-layer dynamic attention matrices based on the graph attention mechanism.

[0051] Specifically, the system then uses a graph neural network architecture to perform inference and prediction on the conceptualized dataset. This transforms the conceptualized dataset into a graph structure input, where nodes correspond to financial concepts. Node feature vectors are formed by concatenating concept embedding vectors and statistical feature vectors from the financial concept graph. Edges correspond to the relationships between concepts defined in the financial concept graph, and edge weights are initialized to the confidence levels of these relationships. Preferably, the graph neural network can employ a three-layer stacked graph attention architecture. The input layer receives node feature vectors, the hidden layer sets multiple attention heads, calculates the raw attention scores between nodes using a specific formula, obtains the final attention weights after normalization, and aggregates neighboring node information based on these weights to update the current node's feature vector. The output layer uses a single attention head, outputs the predicted value of the target financial indicator through a linear activation function, and simultaneously outputs the attention matrix of each layer, i.e., the inter-layer dynamic attention matrix.

[0052] During model training, the system selects a large number of historical conceptual data sets of listed companies and corresponding actual indicator data, and divides them into training set, validation set and test set according to proportion. It adopts the mean squared error loss function combined with regularization mechanism to avoid overfitting, selects a specific optimizer to optimize model parameters, and adjusts the learning rate through the learning rate scheduler. When the training reaches the preset epoch or the validation set loss meets the preset conditions, the training stops and the optimal model parameters and the logic for generating the inter-layer dynamic attention matrix are saved.

[0053] S3: Construct a logical evidence chain for the inter-layer dynamic attention matrix. Based on the attention-guided path backtracking mechanism, backtrack the weight information recorded in the matrix. By calculating the contribution of each potential information flow path to the final prediction, select the key concept reasoning path with the greatest contribution and combine it with the preset financial knowledge graph to complete and verify the logical coherence of the selected path, thereby generating a strengthened logical evidence chain.

[0054] Specifically, the system constructs a logical evidence chain for the inter-layer dynamic attention matrix. Starting from the output layer node corresponding to the target financial indicator, it traverses the attention matrix of each layer in reverse, tracking the predecessor nodes in each layer that meet preset conditions, forming multiple potential information flow paths. To determine the critical path, the system uses a specific fair allocation algorithm to calculate the contribution of each potential path. This algorithm calculates the contribution of each path to the final prediction result by the prediction value deviation corresponding to the path subset. The system selects the potential paths with the highest contribution ranking as the key concept reasoning paths, and then calls the rule engine of the pre-built financial knowledge graph to verify the completeness of the relationships between concepts in the critical path.

[0055] Preferably, the rule engine, based on the causal chains in the knowledge graph, identifies missing intermediate concept nodes in the path and automatically inserts them into their corresponding positions to complete the path. In the logic verification phase, the system employs a two-way logic verification mechanism. On one hand, it verifies the rationality of path relationships through the ontology constraints of the financial knowledge graph, ensuring that the relationships conform to preset direction and dimension requirements. On the other hand, it verifies whether there are logical contradictions in the path through the domain rule base, excluding paths that violate common financial sense. If a critical path is found to have a logical contradiction, the system triggers a path reselection mechanism, selecting the path with the second highest contribution as a replacement, and repeats the completion and verification process. Each completed and verified critical path forms a reinforced logical evidence chain, represented by a triple sequence of concept node-relationship type-weight contribution, clearly presenting the logical connections and influence strength between concepts.

[0056] S4: Based on the strengthening of the logical evidence chain, each abstract financial concept is matched and fused with the specific numerical values ​​anchored in the conceptual data set to generate quantitative explanatory statements. Combined with the indicator prediction values, the prediction results and multiple explanatory statements are assembled into a structured narrative to generate a visual analysis report that supports interactive traceability.

[0057] Specifically, the system ultimately generates a visual analysis report supporting interactive traceability based on a strengthened logical evidence chain. It traverses the concept nodes in each strengthened logical evidence chain, anchoring the corresponding specific data through the mapping relationship between financial concept identifiers in the conceptualized data set and the original data references. Numerical data is directly associated with structured indicator values ​​and related statistical information, while textual data is associated with specific descriptive content in unstructured text. Preferably, the system can employ a weighted fusion strategy to integrate the anchored multi-source data, assigning data weights according to preset priorities and confidence levels to ensure that the fused data has quantitative support. Based on a preset financial domain natural language generation template library, the system automatically generates explanatory statements in conjunction with the fused data. The template library is categorized by relationship type, and the generated explanatory statements clearly present the relationships between concepts and the data support.

[0058] Furthermore, the system employs a pre-defined four-layer narrative framework to assemble a structured narrative. The first layer clearly defines the predicted values ​​and trend judgments of the target indicators; the second layer summarizes and strengthens the core logic of the logical evidence chain; the third layer presents the quantitative explanation statements corresponding to each strengthening logical evidence chain and marks the corresponding conceptual path and contribution; and the fourth layer lists the details of all anchored raw data. The system uses a web-based visualization framework to construct the report architecture, realizing three core functions: an overview of prediction results, visualization of logical evidence chains, and data tracing.

[0059] The prediction results overview function displays the predicted values ​​of target indicators, trend comparisons, and risk level assessment results in a specific format. The logical evidence chain visualization function displays the strengthened logical evidence chain in a graph format, reflecting the contribution through node size, and supports user interaction in viewing relevant data and weight information. The data traceability function provides a four-level traceability path, allowing users to filter and view according to different conditions, and jump to the corresponding original data and text fragments. The report backend is built using a specific framework, and data storage adopts a combination of multiple databases and file systems. The frontend uses specific technologies to achieve interactive operation and supports report export to meet compliance audit requirements. Through this process, the system assembles the prediction results and multiple explanatory statements into a structured narrative, generating a visual analysis report that is both readable and auditable.

[0060] In summary, the AI-based automated enterprise financial analysis method provided in this application transforms complex raw information into standardized semantic representations that are machine-understandable and conform to professional contexts through a unified conceptual mapping of multi-source heterogeneous data. The graph attention neural network-based inference and prediction mechanism not only generates accurate indicator predictions but, more importantly, simultaneously captures and solidifies the complete attention trajectory of information aggregation and decision-making within the model, thus transforming the "black box" prediction process into traceable "white box" data for the first time. Through path backtracking and logical completion based on the attention matrix, it automatically constructs chain evidence that conforms to accounting reconciliation and financial theory, ensuring that the final prediction conclusion has a clear, coherent, and professional causal explanation, greatly improving the credibility and acceptability of the analysis results. By generating interactive analysis reports that integrate quantitative values ​​and qualitative logic, it provides decision-makers with intuitive and credible insights and supports tracing back from the conclusion to the original data and intermediate inference stages, achieving full-process transparency that meets stringent audit requirements, fundamentally enhancing the practicality and reliability of AI systems in high-risk financial decision-making scenarios.

[0061] In one embodiment, S1 of the AI-based automated enterprise financial analysis method provided by the present invention specifically includes the following steps:

[0062] S11: Atomize the acquired accounting items, financial ratios, and analytical indicators by extracting concepts and defining relationships. Based on accounting standards and financial analysis frameworks, construct an initial financial knowledge graph containing concept nodes and logical edges to generate a financial concept graph.

[0063] Specifically, the system comprehensively collects accounting items from the current financial system, commonly used financial ratios in practice, and core indicators required for corporate financial analysis, using this data as the basic data source for concept extraction. Based on the current effective accounting standards and the core logic of mainstream financial analysis frameworks, the system decomposes all complex concepts. For accounting items, the system follows the six major classification criteria of assets, liabilities, owner's equity, revenue, expenses, and profit, breaking down multi-level nested complex items layer by layer until they reach the indivisible basic accounting unit. For financial ratios, the system analyzes the calculation logic of each ratio, breaking it down into the basic indicators corresponding to the numerator and denominator, clarifying the original concepts of each component. For analytical indicators, the system strips away their included accounting and evaluation dimensions, extracting core attributes to form a set of atomic concepts. After completing the extraction of atomized concepts, the system defines various relationships between concepts based on the accounting rules in accounting standards and the inherent logical connections in financial analysis.

[0064] Furthermore, the system clarifies the hierarchical relationships embodied in subordinate relationships, including logical relationships, numerical derivation logic corresponding to accounting relationships, causal transmission logic involved in influence relationships, and mutual mapping logic reflected in correspondence relationships. Through a rule engine, various relationships are precisely associated with atomic concepts, constructing the topological structure of the initial financial knowledge graph. Each concept node in the graph uniquely corresponds to an atomic concept, and each logical edge corresponds to a defined inter-concept relationship. All logical edges are accompanied by corresponding accounting standard clause numbers or financial analysis rule identifiers to ensure the compliance and traceability of relationships. The final output is a financial concept graph containing complete concept nodes and logical edges.

[0065] S12: Preprocess and semantically analyze the acquired multi-source heterogeneous financial data of enterprises, convert structured financial statement data into time-series feature vectors, and perform entity recognition and sentiment analysis on unstructured financial texts to generate standard data units.

[0066] Specifically, the system performs comprehensive preprocessing and semantic parsing operations on the acquired multi-source heterogeneous financial data of enterprises. For structured financial statement data, the system performs data consistency verification. Based on the format requirements and data logic in the financial statement preparation specifications, the system filters and removes data items that do not meet the specifications. For missing information caused by statistical caliber adjustments, data reporting delays, etc., the system completes the information supplementation by referring to the data processing methods of similar enterprises in the same industry and the supplementary provisions in the report preparation guidelines.

[0067] Furthermore, the system integrates financial statement data from consecutive periods according to the time dimension, forming a complete time-series data sequence. Based on the time attributes and statistical periods of the data, the system converts the time-series data sequence into a time-series feature vector. The vector dimension strictly corresponds to the time span and statistical dimension of the data, ensuring that the vector can fully preserve the time-series change characteristics and statistical attributes of the structured data. For unstructured financial text data, the system first performs text cleaning to remove meaningless characters, repetitive expressions, and information fragments irrelevant to corporate financial analysis. Then, it uses text processing algorithms adapted to the financial domain for word segmentation and syntactic analysis, breaking down the text structure and clarifying the logical connections between sentences. The system calls a pre-set financial entity recognition rule library to accurately extract entity information related to financial analysis from the processed text, such as company names, financial indicator names, transaction descriptions, and regulatory policy clauses. At the same time, the system combines a dedicated sentiment dictionary for the financial domain to perform sentiment analysis on the text content based on the wording and contextual logic of financial matters, determining the potential impact of the text information on the company's financial situation. The system integrates the processed structured time-series feature vectors with the entity information and sentiment analysis results of unstructured text, and encapsulates them into a standard data unit containing data type identifiers, core information fields, and processing status identifiers. This transforms raw data from different sources and in different formats into a data carrier with a unified format and standard information dimensions.

[0068] S13: Based on the financial concept graph, perform concept association processing on the standard data units, map each data unit to the corresponding node in the preset financial concept graph, and calculate the association strength between it and the relevant concept nodes according to accounting standards to generate weighted concept node data.

[0069] Specifically, the system uses a pre-constructed financial concept graph as a benchmark to perform concept association on standard data units, extracting core information fields for each standard data unit and establishing matching channels between these information fields and concept nodes in the financial concept graph. For standard data units corresponding to structured data, the system strictly adheres to the definitions of accounting subject scope and financial indicator calculation logic in accounting standards to directly locate the unique correspondence between data units and concept nodes, ensuring the accuracy of the matching. For standard data units corresponding to unstructured data, the system determines the corresponding concept node by calculating the semantic similarity between entity information and concept nodes, combined with the financial impact dimension indicated by sentiment analysis results.

[0070] After the matching operation is completed, the system calculates the association strength. Based on the account correspondence and financial indicator calculation logic specified in accounting standards, and taking into account factors such as the information completeness of standard data units and the semantic fit between core information fields and concept nodes, the system generates a quantified association weight through comprehensive calculation of multi-dimensional parameters. This weight directly reflects the closeness of the association between the standard data unit and the corresponding concept node. Subsequently, the system binds the core information of each standard data unit with the corresponding concept node identifier and association weight, forming weighted concept node data containing the concept node identifier, data unit reference address, and association weight.

[0071] S14: The concept node data undergoes semantic space embedding processing. The domain pre-trained model is called to encode each concept node and its associated weights into a unified low-dimensional dense vector, generating a conceptual data set with financial semantic labels.

[0072] Specifically, the system invokes a pre-trained model optimized for the financial domain. This model has undergone training and iterative optimization on massive amounts of financial text, accounting standard provisions, and historical financial data samples, possessing the ability to accurately represent financial semantics. The system uses the semantic information of each concept node and its corresponding association weights as joint inputs to the encoding layer of the pre-trained model. The encoding layer extracts the core semantic features and weight association features of the concept nodes using feature extraction algorithms, and then uses a dimensionality compression algorithm to convert the high-dimensional feature information into low-dimensional dense vectors, maintaining the integrity and semantic relevance of the features during the compression process. Through pre-defined semantic space constraint rules, the system ensures that concept node vectors within the same semantic category exhibit a clustered distribution in the semantic space, while concept node vectors from different semantic categories maintain clear distinguishability in the semantic space, thereby achieving consistency and rationality in the semantic space.

[0073] After vector encoding is completed, the system adds financial semantic labels to each low-dimensional dense vector. The label information is generated based on core features such as the attribute category, relationship type, and relationship weight level of the corresponding concept node, comprehensively covering the key semantic attributes of the concept. The system integrates all low-dimensional dense vectors with financial semantic labels and forms a conceptual data set according to a unified data format specification. Each data item in this set contains core fields such as a unique identifier for the concept node, low-dimensional dense vector data, financial semantic label, and reference address of the original standard data unit.

[0074] In one embodiment, S2 of the AI-based automated enterprise financial analysis method provided by the present invention specifically includes the following steps:

[0075] S21: Perform graph structure modeling on the conceptual dataset, construct an initial graph with financial concepts as nodes and the interrelationships between concepts as edges, and initialize the conceptual vectors as node features to generate a financial concept relationship graph.

[0076] Specifically, the system models the conceptual dataset using a graph structure, traverses the dataset, and extracts all independent nodes based on the unique identifier of each financial concept node, forming a node set. This ensures that each financial concept corresponds to a unique node in the graph structure, with no duplicates. The system also analyzes the interrelationships between concepts, using the accounting standards' requirements for account reconciliation, the computational dependencies between financial indicators, and the inherent logical connections in financial analysis as core criteria. Simultaneously, it references the defined types of inter-concept relationships in the financial concept graph to clarify the categories of interrelationships between nodes, and associates each edge with a corresponding interrelationship identifier.

[0077] The system constructs an initial graph topology using financial concepts as nodes and confirmed reconciliation relationships as edges. It then performs node feature initialization, using the following formula: ,in Let represent the initial feature vector of the i-th node. This represents the low-dimensional dense vector corresponding to the i-th financial concept in the conceptualized dataset. The system performs structural validation on the initial graph, checking the uniqueness of node identifiers, the completeness of the correlation between edges and their interrelationships, and the full coverage of node feature assignments. It eliminates isolated nodes without interrelationships and invalid edges without supporting identifiers, generating a financial concept relationship graph with a complete structure and clear features.

[0078] S22: Perform multi-layer attention message propagation processing on the financial concept relationship graph. In each layer, each node aggregates the feature information of its neighboring nodes according to the attention weight, and updates its own features through non-linear transformation to generate a node state matrix containing high-order association information.

[0079] Specifically, the system performs multi-layer attention message propagation on the financial concept relationship graph. After setting a preset number of propagation layers, it initiates an iterative propagation process. During each layer of propagation, the system first calculates the attention weight between each node and all its neighboring nodes. The calculation formula is as follows:

[0080]

[0081] in, This represents the original attention score of node i to node j. Represents the attention coefficient vector. Represents a linear transformation matrix. This represents the feature vector of node i in the current layer. This represents the feature vector of node j in the current layer. This indicates a vector concatenation operation.

[0082] Furthermore, the system uses the softmax function to... The final attention weights are obtained by normalization. Then, based on this weight, the features of all neighboring nodes of each node are weighted and aggregated. The aggregation formula is as follows:

[0083]

[0084] in, Let i represent the set of neighboring nodes. Represents a nonlinear transformation function. This represents the updated feature vector of node i. The system merges the aggregated features with the node's original features and then passes the result to a nonlinear transformation function to complete the feature transformation, updating the feature information of the current node. This logic is followed to process all nodes in this layer before proceeding to the next layer of propagation. The same operation logic is repeated, and after multiple layers of propagation, a node state matrix containing high-order association information is generated.

[0085] S23: Based on the node state matrix, perform prediction decoding and attention recording processing. Extract the features of the nodes corresponding to the target financial indicators from the final layer node state and perform linear mapping to obtain the indicator prediction value. At the same time, organize the attention weights between all nodes calculated in each layer by layer to generate an inter-layer dynamic attention matrix.

[0086] Specifically, the system performs prediction decoding and attention recording based on the node state matrix. It locates the node corresponding to the target financial indicator from the node state matrix, extracts the feature vector of that node in the final layer, and this vector integrates high-order association information from multiple propagation layers with its own semantic features. The predicted indicator value is obtained through linear mapping, and the prediction formula is as follows:

[0087]

[0088] in, This represents the predicted value of the target financial indicator. This represents the output layer linear mapping weight matrix. This represents the feature vector of the node corresponding to the target financial indicator at layer L, where L represents the total number of propagation layers. Simultaneously with generating the predicted value, the system initiates an attention recording process, collecting the attention weights between all nodes calculated during each layer of propagation. Each layer corresponds to an independent attention weight submatrix. ,in medium elements This represents the attention weight between node i and node j in layer k. The system standardizes the format of the attention weight submatrices for each layer to ensure that all submatrices have a unified data structure and dimensional specifications. Then, all submatrices are integrated in layer order, following the integration logic... ,in This represents the inter-layer dynamic attention matrix, generating a complete record of the changes in the association strength between nodes.

[0089] In one embodiment, such as Figure 2 As shown, S3 of the AI-based automated enterprise financial analysis method provided by this invention specifically includes the following steps:

[0090] S31: Perform importance propagation path backtracking on the interlayer dynamic attention matrix. Starting from the predicted target node, find the upstream node and connecting edge that contributes the most to it based on the weight values ​​of each layer in the matrix, and generate multiple candidate concept reasoning paths.

[0091] Specifically, the system performs importance propagation path backtracking on the inter-layer dynamic attention matrix, locating the node identifier corresponding to the prediction target in the inter-layer dynamic attention matrix, and using this node as the starting point for path backtracking, specifying that the backtracking direction is reverse progress from the final layer to the initial layer. The system first calculates the weight selection threshold for each layer, using the following formula:

[0092]

[0093] in, This represents the weighted filtering threshold, where N represents the total number of nodes in the current layer. This represents the maximum attention weight of the i-th node in the current layer. The system extracts the attention weight information recorded in the matrix layer by layer, which will exceed... The nodes corresponding to the weights are determined to be highly correlated nodes and given priority as upstream candidate nodes.

[0094] During each backtracking process, the system establishes a connection between the current node and upstream candidate nodes, associating corresponding edge identifiers to ensure a matching relationship between edges and nodes. The system repeats this process, tracing back layer by layer, continuously searching for upstream nodes and connecting edges that meet the weight selection criteria, until it backtracks to the initial layer node or there are no weight-supporting nodes that meet the threshold. Throughout this process, the system records each complete backtracking trajectory, which includes a node sequence, an edge sequence, and corresponding layer weight association information, forming multiple candidate concept reasoning paths.

[0095] S32: Perform contribution quantification scoring on candidate concept reasoning paths, comprehensively calculate the cumulative intensity of attention weights, path length, and path smoothness on each path, and generate a contribution score for each path.

[0096] Specifically, the system performs contribution metric scoring on candidate concept reasoning paths. The system first extracts the attention weight data corresponding to each candidate path from the inter-layer dynamic attention matrix, determining the attention weights for each layer along each path. , Represents the source concept node in the l-th layer. To the target concept node Attention weights are assigned. The system counts the number of edges in each path as the path length. The length penalty coefficient is determined based on preset rules. With knowledge graph adjustment coefficient , Used to balance the impact of path length on contribution. The weighting of adjacent concept nodes is used to adjust the strength of logical connections in the financial knowledge graph. The system extracts adjacent concept nodes from the pre-defined financial knowledge graph. and logical correlation strength This strength is determined based on the types of relationships between concepts in the knowledge graph and the rules governing them. Preferably, the system can use the formula:

[0097]

[0098] in, Let P be the contribution score for path P. The first one obtained from the inter-layer dynamic attention matrix From the source concept node in the layer To the target concept node Attention weights Let P be the number of edges, i.e., the path length. This is the length penalty coefficient. This is the adjustment coefficient for the knowledge graph. Concept nodes obtained from a pre-defined financial knowledge graph and The strength of the logical connection between them.

[0099] S33: Perform critical path screening on the candidate paths after scoring, sort them based on contribution scores and select the Top-K paths that exceed the preset threshold to generate a set of key concept reasoning paths.

[0100] Specifically, the system performs critical path filtering on the scored candidate paths. The system first calculates the statistical characteristics of the contribution scores of all candidate paths and then determines a preset threshold using a formula:

[0101]

[0102] in, Indicates the filtering threshold. This represents the mean of the contribution scores of all candidate paths. This represents the adjustment coefficient. This represents the standard deviation of the contribution score. The system sorts all candidate concept reasoning paths in descending order of contribution score, and then compares the sorted candidate paths with a threshold. Compare the scores and select those exceeding the limit. The system selects the Top-K paths based on their scores, with the K value determined according to the actual analysis scenario's requirements for the number of critical paths.

[0103] Furthermore, the system validates the selected paths based on the completeness of the node sequence and the consistency of edge-weight associations, excluding paths with missing nodes or abnormal weights. During validation, the system uses a path structure validation algorithm to check the continuity of node identifiers and the rationality of edge associations, ensuring that the logical transmission of each path is unbroken. After validation, the system integrates the qualified paths to generate a set of key concept reasoning paths, with each path in the set accompanied by contribution scores and weight association information.

[0104] S34: Perform logical completeness checks and supplementation on the set of key concept reasoning paths, compare each path with the preset financial knowledge graph, insert intermediate concept nodes in missing logical links to form a complete causal chain, and generate a reinforced logical evidence chain that conforms to professional logic.

[0105] Specifically, the system performs logical completeness checks and fills in the set of reasoning paths for key concepts, retrieves a pre-set financial knowledge graph, and calculates the logical completeness score for each path using a formula:

[0106]

[0107] in, This represents the logical completeness score of path P. This represents the number of existing logical connections in the path. This represents the total number of logical connections that a corresponding path should have in the financial knowledge graph. The system compares each path with the logical connection system of the financial knowledge graph node by node and edge by edge, based on the score. The system determines whether the logical links are continuous and identifies missing intermediate concept nodes or logical connections. When the score is lower than the preset standard, the system extracts intermediate concept nodes from the financial knowledge graph that conform to the logical transmission direction of the path. These nodes can connect the concept nodes on both sides of the gap.

[0108] Furthermore, the system inserts intermediate concept nodes into the corresponding positions on the path, supplements the associated edge identifiers and logical bases, and updates the node and edge sequences of the path. After completion, the system recalculates. The system verifies logical consistency to ensure that the paths conform to accounting standards and professional financial analysis logic. Following the above process, the system checks and completes all critical paths, generating a complete logical chain for each path and integrating them to form a reinforced logical evidence chain that conforms to professional logic.

[0109] In one embodiment, S4 of the AI-based automated enterprise financial analysis method provided by the present invention specifically includes the following steps:

[0110] S41: Perform concept-numerical anchoring processing on the enhanced logical evidence chain, traverse each concept node in the evidence chain, find all original values ​​and their weights associated with the conceptual data from the conceptual data set, and generate quantitative evidence units.

[0111] Specifically, the system performs concept-numerical anchoring on the strengthening logical evidence chain, traversing each concept node in the chain, extracting the unique identifier and semantic attributes of each node, and establishing a connection channel between the concept node and the conceptualized data set. The system retrieves all original data records from the conceptualized data set that match the semantic attributes of the current concept node, and calculates the degree of data matching using the anchoring matching degree formula.

[0112]

[0113] in, This represents the matching degree between the i-th concept node and the j-th original data. Indicates the semantic similarity between the two. This represents the association weight of the j-th original data item. This represents the maximum semantic similarity benchmark value corresponding to the concept node. The system selects raw data that meets the preset matching requirements, extracts the raw values ​​and association weights, and binds the concept node identifier, raw values, association weights, and data source information.

[0114] The system performs consistency checks on the bound information to ensure that the semantic association between numerical values ​​and concept nodes complies with accounting standards and data recording specifications, and excludes data combinations with mismatched values ​​and concepts or abnormal weights. After verification, the system encapsulates each set of bound information into an independent data unit, generating a quantitative evidence unit containing numerical information corresponding to all concept nodes.

[0115] S42: Based on the evidence unit, perform natural language sentence generation processing, call the financial analysis logic template to transform the numerical changes and conceptual relationships in the evidence unit into a coherent text description, and generate professional explanatory statements.

[0116] Specifically, the system generates natural language sentences based on quantitative evidence units and loads a pre-defined financial analysis logic template library. The templates in the library are categorized and organized according to conceptual relationship types, numerical change trends, and financial analysis scenarios. Each template contains a fixed logical framework and variable data placeholders. The system extracts numerical change information, relationship types between conceptual nodes, and association weights from each quantitative evidence unit, and calculates the template fit degree using a template fitting formula.

[0117]

[0118] in, This represents the fit between the k-th template and the i-th evidence unit. This represents the feature items in the template. This indicates the degree of matching between the feature terms and the evidence unit. This represents the total number of feature items for the k-th template. The system selects the template with the highest fit, fills the template placeholders with the numerical change data, conceptual relationship descriptions, and association weight information from the evidence unit, and optimizes the syntax and logical coherence of the statements using a natural language generation algorithm to ensure that the text descriptions conform to the professional expression standards of financial analysis. The system performs semantic verification on the generated statements, checking the accuracy of the correspondence between numerical values ​​and concepts, and the clarity of the expression of logical relationships to ensure that the statements are unambiguous and the data support is clear. Finally, it generates a professional explanation statement corresponding to each evidence unit.

[0119] S43: Perform report structure assembly processing on the explanatory statements, sort and group multiple explanatory statements according to the importance of the evidence chain, and integrate and format them with indicator prediction values ​​and key charts to generate a visual analysis report that supports interactive tracing.

[0120] Specifically, the system performs structured report assembly processing on the interpretation statements. First, the system extracts the contribution score of the strengthening logical evidence chain corresponding to each interpretation statement. Then, combining this with the logical relationships between statements, it calculates the ranking priority of each statement using a ranking priority formula.

[0121]

[0122] in, This indicates the priority of the i-th interpretation statement. This indicates the contribution score of the evidence chain corresponding to the statement. Indicates the logical relationship between a statement and other statements. and These represent the adjustment coefficients for contribution and correlation, respectively. The system sorts all explanatory statements in descending order of priority and groups them according to the financial analysis dimensions corresponding to the evidence chain, presenting explanatory statements of the same dimension together. The system integrates the sorted and grouped explanatory statements with the indicator prediction values, embedding them into a preset report framework, and simultaneously retrieves key charts generated during the analysis process, including financial concept relationship diagrams, logical evidence chain visualization maps, and numerical change trend charts.

[0123] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0124] Based on the same inventive concept, this application also provides an AI-based automated enterprise financial analysis device for implementing the AI-based automated enterprise financial analysis method described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations of one or more AI-based automated enterprise financial analysis device embodiments provided below can be found in the limitations of the AI-based automated enterprise financial analysis method described above, and will not be repeated here.

[0125] Preferably, such as Figure 3 As shown, the present invention provides an AI-based automated enterprise financial analysis device 500, which is configured with the following modules:

[0126] The financial data conceptual mapping module 510 is used to perform conceptual mapping on the acquired multi-source heterogeneous financial data of enterprises, and to uniformly map unstructured text descriptions and structured financial data to the semantic space of a pre-built financial concept graph, generating a set of conceptual data with financial semantic labels.

[0127] The graph neural network reasoning and prediction module 520 is used to reason and predict conceptual data sets based on graph neural network architecture, simulate the logical relationship between financial concepts for information propagation and aggregation, and generate the indicator prediction value of the target financial indicator and the inter-layer dynamic attention matrix based on the graph attention mechanism.

[0128] The logical evidence chain strengthening module 530 is used to construct a logical evidence chain for the inter-layer dynamic attention matrix. Based on the attention-guided path backtracking mechanism, it backtracks the weight information recorded in the matrix. By calculating the contribution of each potential information flow path to the final prediction, it selects the key concept reasoning path with the greatest contribution and combines it with the preset financial knowledge graph to complete and verify the logical coherence of the selected path, thereby generating a strengthened logical evidence chain.

[0129] The visualization analysis report generation module 540 is used to match and integrate each abstract financial concept in the reinforced logical evidence chain with the specific values ​​anchored in the conceptual data set, generate quantitative explanatory statements, and combine the prediction results with multiple explanatory statements to form a structured narrative, generating a visualization analysis report that supports interactive traceability.

[0130] Preferably, the financial data conceptual mapping module 510 provided in this application is configured with the following units:

[0131] The financial concept graph construction unit is used to perform atomic concept extraction and relationship definition processing on the acquired accounting items, financial ratios and analysis indicators, and to construct an initial financial knowledge graph containing concept nodes and logical edges based on accounting standards and financial analysis framework, thereby generating a financial concept graph.

[0132] The multi-source financial data preprocessing and parsing unit is used to preprocess and semantically parse the acquired multi-source heterogeneous financial data of enterprises, convert structured financial statement data into time-series feature vectors, and perform entity recognition and sentiment analysis on unstructured financial text to generate standard data units.

[0133] The data unit concept association mapping unit is used to perform concept association processing on standard data units based on the financial concept graph. It maps each data unit to the corresponding node in the preset financial concept graph and calculates the association strength between it and the relevant concept nodes according to accounting standards, generating weighted concept node data.

[0134] The semantic embedding unit for concept nodes is used to perform semantic space embedding processing on concept node data. It calls the domain pre-trained model to encode each concept node and its associated weights into a unified low-dimensional dense vector, generating a conceptual data set with financial semantic labels.

[0135] Preferably, the graph neural network reasoning and prediction module 520 provided in this application is configured with the following units:

[0136] The Financial Concept Graph Structure Modeling Unit is used to perform graph structure modeling on the conceptual dataset. It constructs an initial graph with financial concepts as nodes and the interrelationships between concepts as edges, and initializes the conceptual vectors as node features to generate a financial concept relationship graph.

[0137] The multi-layer attention message propagation unit is used to perform multi-layer attention message propagation processing on the financial concept relationship graph. In each layer, each node aggregates the feature information of its neighboring nodes according to the attention weight, and updates its own features through nonlinear transformation to generate a node state matrix containing high-order association information.

[0138] The prediction decoding and attention recording unit is used to perform prediction decoding and attention recording based on the node state matrix. It extracts the features of the nodes corresponding to the target financial indicators from the final layer node states and performs linear mapping to obtain the indicator prediction values. At the same time, it organizes the attention weights between all nodes calculated in each layer by layer to generate an inter-layer dynamic attention matrix.

[0139] Preferably, the logical evidence chain strengthening module 530 provided in this application is configured with the following units:

[0140] The reasoning path backtracking unit is used to perform importance propagation path backtracking processing on the inter-layer dynamic attention matrix. Starting from the predicted target node, it reverses the search for the upstream node and connecting edge that contributes the most to it based on the weight values ​​of each layer in the matrix, and generates multiple candidate concept reasoning paths.

[0141] The Path Contribution Measurement Unit is used to perform contribution measurement scoring on candidate concept reasoning paths. It comprehensively calculates the cumulative intensity of attention weights, path length, and path smoothness on each path to generate a contribution score for each path.

[0142] The critical path filtering unit is used to filter the candidate paths after scoring, sort them based on contribution scores and select the Top-K paths that exceed a preset threshold to generate a set of critical concept reasoning paths.

[0143] The logical evidence chain completion and reinforcement unit is used to check and complete the logical completeness of the set of reasoning paths for key concepts. It compares each path with the preset financial knowledge graph, inserts intermediate concept nodes in the missing logical links to form a complete causal chain, and generates a reinforced logical evidence chain that conforms to professional logic.

[0144] Preferably, the visualization analysis report generation module 540 provided in this application is configured with the following units:

[0145] Concept-numerical anchoring unit is used to perform concept-numerical anchoring processing on the enhanced logic evidence chain. It traverses each concept node in the evidence chain, finds all original values ​​and their weights associated with the conceptual data from the conceptual data set, and generates a quantitative evidence unit.

[0146] The explanatory statement generation unit is used to generate natural language sentences based on the evidence unit. It calls the financial analysis logic template to transform the numerical changes and conceptual relationships in the evidence unit into a coherent text description and generate professional explanatory statements.

[0147] The structured report assembly unit is used to perform structured report assembly processing on explanatory statements. It sorts and groups multiple explanatory statements according to the importance of the evidence chain, and integrates and typesets them with indicator prediction values ​​and key charts to generate a visual analysis report that supports interactive traceability.

[0148] In one embodiment, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described AI-based automated enterprise financial analysis method.

[0149] In one embodiment, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned AI-based automated enterprise financial analysis method.

[0150] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0151] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0152] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in this application, and these should all be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An AI-based automated enterprise financial analysis method, characterized in that, Includes the following steps: S1: Conceptualize the acquired multi-source heterogeneous financial data of enterprises, and uniformly map unstructured text descriptions and structured financial data to the semantic space of a pre-constructed financial concept graph to generate a conceptual data set with financial semantic labels. S2: Based on the graph neural network architecture, reason and predict the conceptual data set, simulate the logical relationship between financial concepts to carry out information propagation and aggregation, and generate the indicator prediction value of the target financial indicator and the inter-layer dynamic attention matrix based on the graph attention mechanism. S3: Construct a logical evidence chain for the interlayer dynamic attention matrix, backtrack the weight information recorded in the matrix based on the attention-guided path backtracking mechanism, calculate the contribution of each potential information flow path to the final prediction, select the key concept reasoning path with the greatest contribution, and combine the preset financial knowledge graph to complete and verify the logical coherence of the selected path to generate a strengthened logical evidence chain. S4: Based on the strengthening logical evidence chain, each abstract financial concept is matched and fused with the specific numerical values ​​anchored in the conceptualized data set to generate quantitative explanatory statements. Combined with the indicator prediction values, the prediction results and multiple explanatory statements are assembled into a structured narrative to generate a visual analysis report that supports interactive traceability.

2. The method according to claim 1, characterized in that, S1 includes: S11: Atomize the acquired accounting subjects, financial ratios and analytical indicators for concept extraction and relationship definition. Based on accounting standards and financial analysis framework, construct an initial financial knowledge graph containing concept nodes and logical edges to generate a financial concept graph. S12: Preprocess and semantically analyze the acquired multi-source heterogeneous financial data of enterprises, convert structured financial statement data into time-series feature vectors, and perform entity recognition and sentiment analysis on unstructured financial texts to generate standard data units. S13: Based on the financial concept graph, perform concept association processing on the standard data units, map each data unit to the corresponding node in the preset financial concept graph, and calculate its association strength with related concept nodes according to accounting standards to generate weighted concept node data. S14: The concept node data is processed by semantic space embedding. The domain pre-trained model is called to encode each concept node and its associated weights into a unified low-dimensional dense vector, generating a conceptual data set with financial semantic labels.

3. The method according to claim 1, characterized in that, S2 includes: S21: Perform graph structure modeling on the conceptualized dataset, construct an initial graph with financial concepts as nodes and the interrelationships between concepts as edges, and initialize the conceptualized vectors as node features to generate a financial concept relationship graph; S22: Perform multi-layer attention message propagation processing on the financial concept relationship graph. In each layer, each node aggregates the feature information of its neighboring nodes according to the attention weight, and updates its own features through nonlinear transformation to generate a node state matrix containing high-order association information. S23: Based on the node state matrix, perform prediction decoding and attention recording processing, extract the features of the nodes corresponding to the target financial indicators from the final layer node states, perform linear mapping to obtain the indicator prediction value, and organize the attention weights between all nodes calculated in each layer by layer to generate an inter-layer dynamic attention matrix.

4. The method according to claim 1, characterized in that, S3 includes: S31: Perform importance propagation path backtracking on the interlayer dynamic attention matrix. Starting from the predicted target node, find the upstream node and connecting edge that contributes the most to it in reverse according to the weight value of each layer in the matrix, and generate multiple candidate concept reasoning paths. S32: Perform contribution quantification scoring on the candidate concept reasoning paths, comprehensively calculate the cumulative intensity of attention weights, path length, and path smoothness on each path, and generate a contribution score for each path. S33: Perform key path screening on the candidate paths after scoring, sort them based on the contribution scores and select the Top-K paths that exceed the preset threshold to generate a set of key concept reasoning paths; S34: Perform logical completeness checks and completion processing on the set of key concept reasoning paths, compare each path with the preset financial knowledge graph, insert intermediate concept nodes in the missing logical links to form a complete causal chain, and generate a reinforced logical evidence chain that conforms to professional logic.

5. The method according to claim 4, characterized in that, The formula for calculating the contribution score is as follows: in, Let P be the contribution score for path P. The first one obtained from the inter-layer dynamic attention matrix From the source concept node in the layer To the target concept node Attention weights Let P be the number of edges, i.e., the path length. This is the length penalty coefficient. This is the adjustment coefficient for the knowledge graph. Concept nodes obtained from a pre-defined financial knowledge graph and The strength of the logical connection between them.

6. The method according to any one of claims 1-5, characterized in that, S4 includes: S41: Perform concept-numerical anchoring processing on the enhanced logic evidence chain, traverse each concept node in the evidence chain, find all original values ​​and their weights associated with the conceptual data from the conceptual data set, and generate quantitative evidence units. S42: Based on the evidence unit, perform natural language sentence generation processing, call the financial analysis logic template to transform the numerical changes and conceptual relationships in the evidence unit into a coherent text description, and generate professional explanatory statements; S43: Perform report structure assembly processing on the explanatory statements, sort and group multiple explanatory statements according to the importance of the evidence chain, and integrate and format them with indicator prediction values ​​and key charts to generate a visual analysis report that supports interactive tracing.

7. An AI-based automated enterprise financial analysis device, characterized in that, The device includes: The financial data conceptual mapping module is used to perform conceptual mapping on the acquired multi-source heterogeneous financial data of enterprises, and to uniformly map unstructured text descriptions and structured financial data to the semantic space of a pre-built financial concept graph, generating a set of conceptual data with financial semantic labels. The graph neural network reasoning and prediction module is used to reason and predict the conceptualized data set based on the graph neural network architecture, simulate the logical relationship between financial concepts for information propagation and aggregation, and generate the indicator prediction value of the target financial indicator and the inter-layer dynamic attention matrix based on the graph attention mechanism. The logical evidence chain strengthening module is used to construct a logical evidence chain for the inter-layer dynamic attention matrix. Based on the attention-guided path backtracking mechanism, the weight information recorded in the matrix is ​​backtracked. By calculating the contribution of each potential information flow path to the final prediction, the key concept reasoning path with the greatest contribution is selected and combined with the preset financial knowledge graph to complete and verify the logical coherence of the selected path, thereby generating a strengthened logical evidence chain. The visualization analysis report generation module is used to match and fuse each abstract financial concept in the enhanced logical evidence chain with the specific numerical values ​​anchored in the conceptualized data set, generate quantitative explanatory statements, and combine the prediction results with multiple explanatory statements to form a structured narrative, thereby generating a visualization analysis report that supports interactive traceability.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 6.

Images