A financial product sensitive information identification method and system based on AIGC fusion knowledge graph

Abstract

The present application belongs to the technical field of digital finance, and specifically relates to a financial product sensitive information identification method and system based on AIGC fusion knowledge graph. By constructing a financial knowledge graph containing multi-dimensional structures such as entities, time, labels, etc., and combining multi-channel text comparison and reverse node reasoning, potential sensitive information in financial product text is identified. The system introduces a credibility grading mechanism to dynamically evaluate candidate sensitive information and iteratively update based on frequency of occurrence and node relevance, enabling automatic labeling of implicit risk content and graph expansion, suitable for financial content compliance screening and early warning in complex expression contexts.

Description

Technical Field

[0001] This invention belongs to the technical field of digital finance, specifically relating to a method and system for identifying sensitive information of financial products based on AIGC-fused knowledge graphs. Background Technology

[0002] Against the backdrop of rapid development in fintech, numerous financial institutions are offering diversified financial products and services through digital means, resulting in an explosive growth in related documents, announcements, brochures, and public opinion texts. These documents often contain a large amount of sensitive information, such as product pricing strategies, risk level descriptions, customer profiles, and internal decision-making basis, which are crucial for ensuring financial security, maintaining regulatory compliance, and controlling reputational risks. Therefore, efficiently identifying sensitive information in financial documents has become an important direction for financial institutions to improve their information security capabilities and risk management levels.

[0003] Currently, mainstream methods for identifying sensitive information mainly fall into three categories: keyword matching, rule engines, and deep learning-based classification models. Keyword matching methods are easy to implement but suffer from insufficient recognition accuracy and sensitivity to semantic variations. Rule engines allow for custom feature extraction rules to adapt to specific scenarios, but their scalability and generalization capabilities are weak, and maintenance costs are high. In recent years, with the advancement of natural language processing technology, pre-trained language models such as BERT and ERNIE have been widely used in text classification and named entity recognition tasks, and have also found applications in the financial field. However, financial texts are characterized by complex structures, ambiguous semantics, and highly time-sensitive concepts, making it difficult to obtain accurate contextual semantic understanding by relying solely on language models. Furthermore, existing methods generally lack effective modeling of the financial knowledge background and cannot explicitly represent the attribute relationships and reasoning chains between financial products, leading to the risk of misjudgment or omission in the identification of sensitive content, thus limiting the applicability and reliability of the models in real-world financial scenarios.

[0004] Therefore, there is an urgent need for a new identification method that integrates domain knowledge and language generation capabilities to improve the accuracy, generalization and interpretability of identifying sensitive financial information, and to provide more intelligent and robust technical support for information security governance in the financial industry. Summary of the Invention

[0005] To address the aforementioned problems, the present invention aims to propose a method for identifying sensitive information in financial products based on AIGC-fused knowledge graphs, comprising the following steps:

[0006] S1. Construct a multi-dimensional structured knowledge graph: Construct a knowledge graph based on sensitive information of historical financial products and their upstream and downstream information entities, mark the associated financial behavior nodes and risk labels, and form a multi-dimensional structured knowledge graph containing time, location, entity and label.

[0007] S2. Extract candidate sensitive information nodes: Receive the text to be identified, automatically identify and extract node information that is the same as or similar to the sensitive node description in the knowledge graph, as candidate sensitive information content. The sensitive information includes at least one of risk level, access permission level, and transaction restriction level.

[0008] S3. Perform multi-channel text comparison and filtering: Automatically retrieve known texts of the same financial product on multiple publishing channels, mark all texts containing sensitive information, exclude records whose content structure, wording, publication time, etc., exceed the first threshold range, and identify them as highly credible sensitive information.

[0009] S4. Perform forward and reverse graph reasoning and calculate credibility: Perform forward and reverse reasoning on highly credible sensitive information, trace the graph node association information upward from the sensitive information and trace the subsequent influencing nodes downward. Combine the comprehensive weight evaluation of sensitive information and other nodes to calculate the credibility level of each sensitive information.

[0010] S5. Output review instructions and update graph nodes: If the number of times sensitive information appears in the set of highly credible sensitive information or the cumulative dwell time exceeds the second threshold, an automatic marking instruction is output to the background for reviewers to manually mark and confirm. After the marking is approved, the sensitive information and related upstream and downstream nodes are included in the knowledge graph. If the two review conclusions are inconsistent, a prompt message is output to the background and the output of sensitive information for the financial product text is suspended.

[0011] As a preferred technical solution, in step S2, when identifying sensitive information nodes in the text to be identified, a method of comparison with knowledge graph nodes is adopted. Specifically, this includes: identifying all directly matching node information, calculating the ratio of the information content of unmatched nodes to the information content of matched nodes, and if the ratio is greater than or equal to a preset confidence threshold, then the node is determined to be a sensitive information node and included in the comparison set.

[0012] As a preferred technical solution, in step S3, the specific steps for determining highly credible sensitive information based on multi-channel comparison are as follows: receiving financial product texts containing sensitive information, automatically connecting to internet finance platforms, social platforms, news information platforms, securities trading platforms, and third-party information publishing platforms, extracting text information and marking all texts containing sensitive fields, recording their content, publication time, and text structure information; calculating the overlap of sensitive fields, publication time interval, and text structure similarity between different texts; using preset thresholds for field overlap, publication time interval, and structure similarity as the basis for judgment, excluding sensitive information that does not simultaneously meet the three threshold conditions; if the publication time difference of a certain sensitive information in multiple texts is less than the allowable range, the credibility of the sensitive information is increased by one level; if any difference is greater than the allowable range, the credibility is decreased by one level.

[0013] As a preferred technical solution, the specific method for determining the credibility rating level of sensitive information in S4 is as follows: starting from a high-credibility node in the knowledge graph, a forward traversal is performed sequentially along the directed edges between nodes towards the outer layer nodes, while a reverse traversal is performed simultaneously; during the traversal, a weighted score is calculated based on the strength of the association information of each node, the node's dwell time, and the node's credibility score; if there are upstream and downstream nodes that are both sensitive nodes, the score of that node is increased by one level; finally, based on the comprehensive score, the sensitive information is marked as high credibility, medium credibility, or low credibility and output.

[0014] As a preferred technical solution, in step S5, if text information uploaded by a third-party user is received after the prompt information is output, the content similarity of the text information is calculated and manually reviewed; if the manual review is passed, the text information is included in the sensitive node set; if the review is not passed, the sensitive nodes that failed the review and their time information are recorded as a reference for updating the knowledge graph.

[0015] This invention also provides a financial product sensitive information identification system based on AIGC fused knowledge graph, used to implement the method, including:

[0016] The knowledge graph construction module is used to capture text data of financial products, tag fields and user query operations, mark sensitive information nodes, build a financial product knowledge graph model based on the relationship between nodes, automatically integrate related fields, tag fields, node fields, non-sensitive information nodes and common expression sets, generate a unified dataset and verify the format, build the edge weight relationship between nodes based on the relationship matrix, and output it after regularization.

[0017] The information extraction module is used to receive the text content to be identified through an interactive dialogue page, automatically aggregate and construct a node set, automatically generate candidate content of sensitive information based on the node information that is the same as or similar to the sensitive node description in the knowledge graph, and extract and record its running context and structural weight.

[0018] The multi-channel comparison module connects internet finance platforms, social platforms, news information platforms, securities trading platforms, and third-party information publishing platforms to extract text information and mark all texts containing sensitive fields, recording their content, publication time, and structural information; it also calculates the overlap of sensitive fields, publication time difference, and structural similarity between adjacent texts, compares the results with a set threshold, and outputs high-confidence sensitive information to the graph construction module.

[0019] The credibility calculation module is used to simultaneously perform forward and backward reasoning on highly credible sensitive information, tracing back the graph nodes from the sensitive information upwards and tracking the influencing nodes downwards, and calculating the credibility level of each sensitive information by combining the weight of the nodes; it is also used to receive risk level thresholds, group sensitive information according to credibility level for statistical analysis, and if the number of occurrences or cumulative dwell time of a certain sensitive information exceeds the second threshold, an audit instruction is output to the backend.

[0020] As a preferred technical solution, the information extraction module includes a sensitive node extraction submodule and a multi-channel extraction submodule. The sensitive node extraction submodule is used to identify sensitive information nodes that are consistent with the sensitive node description and calculate the proportion of matching node names. If the proportion is greater than or equal to a set threshold, the next step of judgment is performed; if it is lower than the threshold, the node is removed. The multi-channel extraction submodule is used to connect online and offline multi-channel text information, perform deduplication, cleaning and filtering operations, and retain sensitive information whose content, time and channel are consistent for use by the credibility calculation module.

[0021] As a preferred technical solution, the credibility calculation module includes a forward and reverse reasoning submodule, which is used to sort highly credible sensitive information by weight, obtain related node information from a single sensitive node, record the weighted score between each related node and the sensitive information, and calculate the score according to the weighted formula by combining structural weight, frequency weight, and dwell time weight, and output the credibility level.

[0022] As a preferred technical solution, after receiving the text information of financial products uploaded by the back-end management terminal, the text information is updated in real time, and all sensitive information is integrated in units of a set time period. The sensitive information level is determined by weighting the data on a node-by-node basis.

[0023] As a preferred technical solution, the knowledge graph construction module outputs knowledge graph construction based on multi-turn dialogue, including three basic tasks: first, to update the labeling of sensitive nodes based on features such as time, location, nodes, and tags; second, to update a set of sensitive nodes by mutual verification between sensitive nodes and non-sensitive nodes; and third, to label four sets of sensitive nodes from high to low according to the frequency of occurrence and risk level of sensitive information. Beneficial effects

[0024] This invention deeply integrates AIGC (Generative Artificial Intelligence) capabilities with structured knowledge graphs to identify sensitive information in financial products. Through the synergy of semantic matching, structural alignment, and behavioral trajectory backtracking, it achieves implicit identification and reasoning of non-explicit sensitive information, breaking through the limitations of traditional methods in literal keyword recognition and significantly improving the ability to perceive implicit risk expressions. It is particularly suitable for analyzing complex expressions, evasive rhetoric, and new financial product release texts.

[0025] This invention introduces a "multi-channel credibility consensus mechanism," which involves horizontally comparing the content of the same financial product published on multiple channels such as internet finance platforms, social platforms, and third-party information. It sets three thresholds for field overlap, publication time difference, and structural similarity to filter out information silos and abnormally propagated content, retaining only sensitive nodes with high consensus. This effectively reduces false alarm redundancy and provides a more unified and robust information foundation for subsequent decision-making, adapting to the complex dissemination environment of fragmented and frequently updated financial information.

[0026] This invention incorporates a "forward and reverse graph reasoning module" at the system level. By combining structural paths, node connection strength, and weighted feedback of upstream and downstream sensitive nodes, a dynamic credibility evolution path is constructed. This mechanism not only traces the causes of sensitive information but also predicts the node regions it may affect in advance, enabling causal chain identification and early warning of sensitive information. This enhances the interpretability and foresight of risk information management, making it particularly suitable for regulatory-oriented review processes and scenarios requiring manual assistance. It also possesses scalable compliance risk adaptation capabilities. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0028] To enhance understanding of the present invention, the present invention will be further described in detail below with reference to embodiments. These embodiments are only used to explain the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0029] Example 1: Implementation of a Method for Identifying Sensitive Information in Financial Products Based on AIGC-Fused Knowledge Graph

[0030] like Figure 1 As shown, this embodiment provides a method for identifying sensitive information in financial products based on AIGC-integrated knowledge graphs. It is applicable to fintech platforms, intelligent risk control systems, or compliance information monitoring platforms. By integrating generative artificial intelligence language capabilities with structured knowledge graph reasoning mechanisms, it achieves accurate identification and dynamic management of potentially sensitive information in financial product texts. The method includes the following steps:

[0031] S1. Construct a multi-dimensional structured knowledge graph:

[0032] First, a large amount of historical financial product data was collected, including prospectuses, product descriptions, marketing materials, rating reports, investor Q&A records, and other textual documents. Key information entities were extracted from the text using Named Entity Recognition (NER) technology, including product name, issuing institution, risk level, return method, target audience, regulatory classification, and suitability requirements.

[0033] Based on this, entities are structured according to time, location, subject type, and tags to establish a multi-dimensional information entity set. For example, "the leveraged ETF launched in market A in July 2023 for C-type investors" is decomposed into five dimensions: time node "2023-07", location "A", product type "ETF", target audience "C-type investors", and tag "leveraged".

[0034] When constructing nodes, key sensitive nodes that frequently appear in compliance disputes or risk warning documents are highlighted, such as "off-exchange margin trading," "transfer of beneficial rights," "principal protection guarantee," and "structured issuance." The association paths between sensitive nodes and upstream and downstream entities are also marked, such as the "product type" → "investment asset" → "risk warning" link. A graph modeling tool (such as Neo4j) is used to graphically model the above entities and their relationships, generating a complete knowledge graph containing entity nodes, attribute nodes, and relationship edges. The system regularly updates the graph structure based on newly released financial products, achieving dynamic evolution and risk adaptability of the graph.

[0035] S2. Extract candidate sensitive information nodes:

[0036] After receiving the text to be recognized, the system first completes sentence segmentation, word segmentation, part-of-speech tagging, and syntactic dependency structure construction through the natural language preprocessing module. Then, it uses an AIGC generative dialogue model (such as a finely tuned GPT-4) to extract semantic features from the entire text and identify sentence fragments that may contain sensitive expressions.

[0037] The processed text is then compared with the constructed knowledge graph. The comparison is divided into three categories:

[0038] Complete matching comparison: The entities in the text are completely consistent with the content of the graph nodes.

[0039] Semantic similarity comparison: Sentence vector matching is used to identify semantically similar expressions (e.g., "potential principal loss" ≈ "no principal protection").

[0040] Structural co-occurrence comparison: Based on dependency parsing, identify content whose expression structure is consistent with knowledge graph nodes.

[0041] The system extracts nodes that meet any matching pattern as candidate sensitive nodes and determines their corresponding sensitive information dimension (such as access permission level, risk level, and transaction restriction level). In the optimal solution, the system further introduces a self-supervised learning model to enhance training on historical identification results and review feedback, thereby improving identification accuracy.

[0042] S3. Perform multi-channel text comparison and filtering:

[0043] To enhance the stability and reliability of the recognition results, the system cross-validates the text of the same financial product across multiple publicly available channels. Data sources include, but are not limited to: securities company websites, bank wealth management platforms, third-party wealth management platforms, social media (such as Weibo and Zhihu), news and information (such as Caixin and Bloomberg), and investor forums.

[0044] The system aggregates different versions of the same product text by using the product's unique identifier (such as product code, product name + issuing institution + release date), extracts the parts containing sensitive content fields, and records their release time, release channel, and text structure characteristics (such as whether they contain risk warning paragraphs, compliance disclaimers, etc.).

[0045] Filtering is based on three key parameters:

[0046] Sensitive field overlap: measures the textual similarity rate of sensitive expressions between different texts.

[0047] Release time interval: If the release time difference between multiple versions is less than a preset threshold (such as 48 hours), it can be considered as high correlation.

[0048] Structural similarity: refers to the degree of similarity in paragraph structure, sentence layout, etc.

[0049] Only sensitive information mentioned in texts that simultaneously meet the thresholds of the above three indicators will be retained and marked as high-credibility sensitive nodes. If a sensitive statement appears repeatedly in multiple channels, and the publication time is close and the semantics are consistent, the system will automatically increase its credibility score for use in subsequent reasoning stages.

[0050] S4. Perform forward and reverse graph inference and calculate the confidence level:

[0051] After obtaining highly credible candidate sensitive information, the system starts the graph reasoning module to perform upstream and downstream path backtracking and weight analysis for each sensitive node.

[0052] Forward reasoning starts with the current sensitive node and traverses its downstream nodes in the graph, analyzing the potential subsequent risk associations that the sensitive statement may trigger. For example, "expected return of over 8%" → "non-principal guaranteed" → "structured products" → "high suitability rating". Reverse reasoning, on the other hand, ends with the sensitive node and traces back to the possible sources of risk. For example, "highly volatile assets" → "trigger-based structural design" → "trigger-based return clauses".

[0053] During each traversal, the system calculates a comprehensive score for each node visited. The scoring model considers the following three factors:

[0054] Node dwell time (i.e., the proportion of sensitive statements in the graph path);

[0055] Node credibility (e.g., whether it is a sensitive expression that is a key focus of regulatory oversight);

[0056] Node relationship strength (obtained from co-occurrence frequency and training with historical corpus).

[0057] If both the upstream and downstream nodes of a sensitive node are high-risk nodes, the system will automatically raise its score level and output a corresponding credibility label (such as high / medium / low). At the same time, the credibility level will be passed to the front end for subsequent review and graph updates.

[0058] S5. Output the audit instructions and update the graph nodes:

[0059] When the cumulative number of times a certain sensitive information node appears or the duration of its stay exceeds the second threshold set by the system in nearly N financial texts (such as appearing more than 3 times or staying for more than 72 hours in total), the system will trigger the background output mechanism.

[0060] Specifically, it includes:

[0061] Output sensitive information marking instructions to the content moderation system;

[0062] Includes information on its upstream and downstream related nodes and inference links;

[0063] This message indicates the current credibility level and source of the sensitive information.

[0064] Human reviewers can confirm based on system-recommended content. If the review is successful, the system will incorporate the sensitive node and its reasoning chain into the knowledge graph as a structural basis for subsequent judgments. If the conclusions of the two rounds of review are inconsistent, the system will temporarily classify the node as "pending confirmation" and suspend its weight contribution in subsequent identification.

[0065] Further, "label feedback learning" can be performed on sensitive nodes based on human review feedback. By fine-tuning the model parameters, the accuracy of identifying similar text in the next iteration can be improved.

[0066] This embodiment, based on the fusion strategy of AIGC and knowledge graphs, combines multi-channel consensus mechanisms, upstream and downstream causal path reasoning, and credibility level modeling to construct a highly interpretable, adaptable, and dynamically evolving financial sensitive information identification mechanism. Compared with traditional methods based on rules, keywords, or single semantic models, this method has the following advantages:

[0067] It can identify hidden sensitive information such as non-explicit expressions, semantic avoidance, and structural translation;

[0068] Supports cross-platform, cross-language, and cross-time period multi-source content fusion and comparison;

[0069] To achieve reverse tracing of the generation mechanism and dissemination path of sensitive content;

[0070] It possesses the ability to dynamically update knowledge graphs, adapting to new types of risk representations;

[0071] It can be coupled with a manual review mechanism to improve controllability and compliance.

[0072] The complete implementation of this embodiment can be widely applied to areas such as financial compliance, regulatory technology (RegTech), investor protection, and risk warning, which helps to improve the transparency of financial information disclosure and systemic risk governance capabilities.

[0073] Example 2: A Sensitive Information Identification System for Financial Products Based on AIGC Fusion Knowledge Graph

[0074] This embodiment provides a sensitive information identification system for financial products based on AIGC-fused knowledge graphs. It implements the steps of the aforementioned methods, effectively identifying sensitive information that may be contained in financial product texts, and improving the automation and intelligence of financial risk monitoring. The system integrates multiple functional modules such as knowledge graph construction, multimodal information extraction, multi-source data comparison, and credibility calculation. It possesses advantages such as a reasonable structure, clear logic, and strong adaptability, and can be widely applied to sensitive information identification tasks in different scenarios such as banking, securities, insurance, and internet finance.

[0075] The system mainly includes the following structural modules:

[0076] 1. Knowledge Graph Construction Module:

[0077] The knowledge graph construction module serves as the foundational support for the entire system. Its core task is to construct a graph representation of entities related to financial products, including various heterogeneous information such as sensitive information nodes, non-sensitive nodes, tag fields, and user query behavior, ultimately forming a knowledge graph model with associated weights.

[0078] This module includes the following specific functions:

[0079] Data scraping and field parsing: Through the set data source interface, the text information of financial products is scraped periodically or in real time, including financial product prospectuses, user comments, platform announcements, etc., and preprocessing operations such as word segmentation, syntactic analysis, and part-of-speech tagging are performed on the raw text to extract tag fields (such as "annual yield", "fund investment direction", "redemption period") and node fields (such as "trust company", "asset package", "credit rating").

[0080] Node tagging and classification: Based on algorithms such as keyword matching and word vector similarity analysis, nodes that may contain sensitive information are automatically tagged, such as "returns exceeding 15%", "guaranteed principal", "insider information", etc. The tagging results are divided into three levels: high sensitivity, medium sensitivity and low sensitivity. At the same time, a manual review interface is introduced for manual review of some important nodes.

[0081] Edge weight calculation and regularization: An initial relation matrix is ​​established based on node co-occurrence frequency, semantic similarity, and structural adjacency, and edge weights are calculated on this basis. To prevent local extrema from affecting the overall judgment, normalization and regularization operations are performed on all edge weights, and a standardized knowledge graph structure is output for subsequent modules to use.

[0082] Unified dataset generation and format validation: Integrate related fields, non-sensitive nodes, and commonly used expression sets (such as "net value type", "floating income", "filing number", etc.) to build a dataset with a unified format. Ensure the consistency of data type, naming format, and structure of each field through the set validation rules, and finally store and update it in the form of a graph database.

[0083] 2. Information Extraction Module:

[0084] The information extraction module is used to receive text content to be identified from user terminal input or automatic push from the platform, and to mine sensitive information content that may be contained therein with reference to knowledge graph.

[0085] The specific implementation process is as follows:

[0086] User interactive input reception: Supports natural language dialogue interaction mode. Users can input financial product-related content through text boxes, speech-to-text, etc. The system automatically converts the text into a structured format and sends it into the processing flow.

[0087] Sensitive Node Extraction Submodule: Performs entity recognition and syntactic relation analysis on the input text, matches the extracted nodes with the list of sensitive nodes in the knowledge graph, and calculates the matching ratio of node names; when the ratio is higher than a set threshold (e.g., 70%), the node is determined to be potentially sensitive information, otherwise it is removed.

[0088] Context and structural weight analysis: For the retained candidate sensitive nodes, their context in the original text (including the two sentences before and after) is further extracted. Combining sentence structure and sentiment polarity scores, a context weight matrix is ​​constructed and recorded in the node attributes along with structural hierarchy information.

[0089] Multi-channel extraction submodule: This submodule connects multiple online and offline channels (such as financial information websites, WeChat official accounts, short video comments, customer service dialogue records, etc.), performs text deduplication, redundant information cleaning and sensitive node filtering operations, and retains a set of information that is consistent in terms of content, time and channel, providing a data source for subsequent comparison.

[0090] 3. Multi-channel comparison module:

[0091] The main responsibility of this module is to cross-validate and fuse sensitive information from different platforms or channels to improve the accuracy and credibility of the identification results.

[0092] Its functions include:

[0093] Platform data connectivity: Automatically acquire financial text content from information sources including but not limited to internet finance platforms, social media platforms (such as Weibo and Zhihu), news and information platforms (such as Caixin and Eastmoney), securities trading platforms (such as Shenzhen Stock Exchange announcements), and third-party forums through APIs or web crawlers.

[0094] Sensitive field marking and recording: The text is scanned using the built sensitive word library, paragraphs containing sensitive fields are marked, and their occurrence time, text structure characteristics (such as whether it is a title, abstract, or body text), and text publication channels are recorded.

[0095] Overlap and structural similarity calculation: The overlap of sensitive fields, the proximity of publication time, and the similarity of text structure (such as TF-IDF vector cosine similarity) are calculated for text data from different platforms to form a multidimensional similarity matrix. This matrix is ​​then compared with a set threshold to filter out sensitive information with higher credibility and output it to downstream modules.

[0096] 4. Credibility Calculation Module:

[0097] This module is the core intelligent analysis unit of this system. It is used to assess the credibility of identified sensitive information and calculate its potential risk level through the upstream and downstream propagation paths of the graph nodes.

[0098] It includes the following key functions:

[0099] Forward and backward reasoning submodule: Starting from a highly sensitive node, it traces upward along the knowledge graph to its cause (such as the corresponding asset structure and related platforms) and downward to the nodes that may be affected (such as risk level and user investment decisions), recording the structural weight, co-occurrence frequency, and duration of each node in the entire path.

[0100] Weighted scoring and grading: based on the weighted calculation formula:

[0101] Credibility score = α·Structure weight + β·Frequency weight + γ·Dwell time weight;

[0102] Among them, α, β, and γ are configurable parameters.

[0103] After scoring each piece of sensitive information, it is divided into three levels: high, medium, and low, according to the set threshold.

[0104] Instruction output mechanism: When the number of times a certain sensitive information node appears or the cumulative time spent in multiple channels exceeds the second threshold (e.g., 5 times / 48 hours), the system automatically generates an audit instruction, which is fed back to the back-end management system to trigger a manual audit process or risk warning mechanism.

[0105] 5. Information updates and dynamic management of sensitivity levels:

[0106] The system supports the regular uploading of new financial product text information in the backend, including descriptions of newly issued funds, promotional materials for wealth management products, etc., and triggers the following process:

[0107] Perform formatting, word segmentation, and named entity recognition on the newly uploaded text;

[0108] All sensitive information that appears is integrated based on a set time window (e.g., 00:00-02:00 daily);

[0109] Sensitive content is weighted and its level is updated on a node-by-node basis;

[0110] For nodes whose level changes exceed the threshold, the nodes in the graph are remarked and the propagation path is updated to ensure that the system continuously reflects the latest financial risk dynamics.

[0111] This system can be deployed on a local server or a cloud platform. The front end can realize interactive input and result display through a web page or client application, while the back end is responsible for data crawling, graph updating, node comparison and credibility calculation.

[0112] To ensure operational efficiency and data security, the system adopts a microservice architecture, supporting horizontal scaling and hot-swappable modules. Graph data storage is based on graph databases (such as Neo4j), and text processing uses AIGC language models (such as ChatGLM and Wenxin Yiyan) to achieve intent recognition and semantic parsing, ensuring efficient processing of unstructured financial text.

[0113] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for identifying sensitive information in financial products based on AIGC-fused knowledge graphs, characterized in that, Includes the following steps: S1. Construct a multi-dimensional structured knowledge graph: Construct a knowledge graph based on sensitive information of historical financial products and their upstream and downstream information entities, mark the associated financial behavior nodes and risk labels, and form a multi-dimensional structured knowledge graph containing time, location, entity and label. S2. Extract candidate sensitive information nodes: Receive the text to be identified, automatically identify and extract node information that is the same as or similar to the sensitive node description in the knowledge graph, and use it as candidate sensitive information content. S3. Perform multi-channel text comparison and filtering: Automatically retrieve known texts of the same financial product on multiple publishing channels, mark all texts containing sensitive information, exclude records whose content structure, wording, and publication time differences exceed the first threshold range, and identify them as highly credible sensitive information; S4. Perform forward and reverse graph reasoning and calculate credibility: Perform forward and reverse reasoning on highly credible sensitive information, trace the graph node association information upward from the sensitive information and trace the subsequent influencing nodes downward. Combine the comprehensive weight evaluation of sensitive information and other nodes to calculate the credibility level of each sensitive information. S5. Output review instructions and update graph nodes: If the number of times sensitive information appears in the set of highly credible sensitive information or the cumulative dwell time exceeds the second threshold, an automatic marking instruction is output to the background for reviewers to manually mark and confirm. After the marking is approved, the sensitive information and related upstream and downstream nodes are included in the knowledge graph. If the two review conclusions are inconsistent, a prompt message is output to the background and the output of sensitive information for the financial product text is suspended.

2. The method for identifying sensitive information of financial products based on AIGC fused knowledge graph according to claim 1, characterized in that: In step S2, the sensitive information includes at least one of the following: risk level, access permission level, and transaction restriction level. When identifying sensitive information nodes in the text to be identified, a comparison with knowledge graph nodes is adopted. Specifically, this includes: identifying all directly matching node information, calculating the ratio of the information content of unmatched nodes to the information content of matched nodes, and if the ratio is greater than or equal to a preset confidence threshold, then the node is determined to be a sensitive information node and included in the comparison set.

3. The method for identifying sensitive information of financial products based on AIGC fused knowledge graph as described in claim 1, characterized in that: In step S3, the specific steps for determining highly credible sensitive information based on multi-channel comparison are as follows: Receive texts of financial products containing sensitive information, automatically connect to internet finance platforms, social platforms, news information platforms, securities trading platforms and third-party information publishing platforms, extract text information and mark all texts containing sensitive fields, and record their content, publication time and text structure information; Calculate the overlap of sensitive fields, publication time interval, and text structure similarity between different texts; The preset thresholds for field overlap, publication time interval, and structural similarity are used as the basis for judgment to exclude sensitive information that does not meet all three threshold conditions at the same time. If the time difference between the release of a certain sensitive information in multiple texts is less than the allowed range, the credibility of the sensitive information will be increased by one level; if any difference is greater than the allowed range, the credibility will be decreased by one level.

4. The method for identifying sensitive information of financial products based on AIGC fused knowledge graph as described in claim 1, characterized in that: In step S4, the specific method for determining the credibility rating of sensitive information is as follows: Starting from the high-confidence nodes in the knowledge graph, the system traverses forward along the directed edges between nodes to the outer nodes, while simultaneously performing a reverse traversal. During the traversal, a weighted score is calculated based on the strength of the association information of each node, the time spent at the node, and the node's credibility score. If both upstream and downstream nodes are sensitive nodes, the score of that node will be increased by one level. Finally, based on the overall score, the sensitive information is marked as high credibility, medium credibility, or low credibility for output.

5. The method for identifying sensitive information of financial products based on AIGC fused knowledge graph according to claim 1, characterized in that: In step S5, if after outputting the prompt message, a text message uploaded by a third-party user is received, the content similarity of the text message is calculated and manually reviewed. If the manual review is approved, the text information will be included in the sensitive node set; If the review fails, the sensitive points that failed the review and their time information will be recorded as a reference when updating the knowledge graph.

6. A system for identifying sensitive information in financial products based on AIGC-fused knowledge graphs, used to implement the method as described in any one of claims 1-5, characterized in that, include: The knowledge graph construction module is used to capture text data of financial products, tag fields and user query operations, mark sensitive information nodes, build a financial product knowledge graph model based on the relationship between nodes, automatically integrate related fields, tag fields, node fields, non-sensitive information nodes and common expression sets, generate a unified dataset and verify the format, build the edge weight relationship between nodes based on the relationship matrix, and output it after regularization. The information extraction module is used to receive the text content to be identified through an interactive dialogue page, automatically aggregate and construct a node set, automatically generate candidate content of sensitive information based on the node information that is the same as or similar to the sensitive node description in the knowledge graph, and extract and record its running context and structural weight. The multi-channel comparison module is used to connect internet finance platforms, social platforms, news information platforms, securities trading platforms and third-party information publishing platforms, extract text information and mark all text containing sensitive fields, and record its content, publication time and structural information; It also calculates the overlap of sensitive fields, the time difference of publication, and the structural similarity between adjacent texts, compares the results with the set threshold, and determines the high-confidence sensitive information to output to the graph construction module; The credibility calculation module is used to simultaneously perform forward and backward reasoning on highly credible sensitive information, trace the graph nodes upward from the sensitive information and track the influencing nodes downward, and calculate the credibility level of each piece of sensitive information by combining the weight of the nodes. It is also used to receive risk level thresholds, group sensitive information according to credibility level, and if the number of times a certain sensitive information appears or the cumulative time spent on it exceeds the second threshold, an audit instruction will be sent to the backend.

7. The financial product sensitive information identification system based on AIGC fusion knowledge graph according to claim 6, characterized in that: The information extraction module includes a sensitive node extraction submodule and a multi-channel extraction submodule. The sensitive node extraction submodule is used to identify sensitive information nodes that are consistent with the sensitive node description and to calculate the proportion of matching node names. If the ratio is greater than or equal to the set threshold, proceed to the next step of judgment; If the value is below the threshold, the node is removed. The multi-channel extraction submodule is used to connect online and offline multi-channel text information, perform deduplication, cleaning and filtering operations, and retain sensitive information that is consistent with the content, time and channel for use by the credibility calculation module.

8. The financial product sensitive information identification system based on AIGC fusion knowledge graph according to claim 6, characterized in that: The credibility calculation module includes a forward and reverse reasoning submodule, which sorts highly credible sensitive information by weight, retrieves related node information from a single sensitive node, records the weighted score between each related node and the sensitive information, and calculates the score according to the weighted formula by combining structural weight, frequency weight, and dwell time weight, and outputs the credibility level.

9. The financial product sensitive information identification system based on AIGC fusion knowledge graph according to claim 6, characterized in that: After receiving the text information of financial products uploaded by the backend management terminal, the text information is updated in real time, and all sensitive information is integrated in units of set time period, and weighted by node to determine the level of sensitive information.

10. The financial product sensitive information identification system based on AIGC fusion knowledge graph according to claim 6, characterized in that: The knowledge graph construction module outputs knowledge graph construction based on multi-turn dialogue, including three basic tasks: first, to update the labeling of sensitive nodes based on time, location, node, and tag features; second, to update a set of sensitive nodes by mutual verification between sensitive nodes and non-sensitive nodes; and third, to label four sets of sensitive nodes from high to low according to the frequency of occurrence and risk level of sensitive information.

Images