Knowledge graph-based bank associated transaction identification and supervision method and system

Abstract

The invention discloses a bank associated transaction identification and supervision method and system based on a knowledge graph, and the method comprises the following steps: S1, multi-source data integration: integrating Internet data, and generating a unified data source through data cleaning, standardization and conflict resolution; s2, entity relationship joint extraction: converting the bank data into a text by adopting a multi-OCR fusion technology, identifying entities through a BERT-BiLSTM-CRF model, and extracting an equity relationship, a management relationship and a transaction relationship based on a graph neural network joint; s3, entity name standardization: adopting a mixed name standardization method, combining rule matching, character similarity and Sension-BERT semantic similarity calculation, and mapping a full name, an abbreviation and another name to a standard entity identifier; and S4, knowledge fusion and storage: combining multi-source entities based on an entity fusion algorithm of multi-dimensional similarity, and storing entities, relationships and attributes through a Neo4j graph database.

Description

Technical Field

[0001] The present invention relates to the field of knowledge graph construction methods. Background Art

[0002] With the rapid development of the financial sector, the banking industry plays a vital role in the national economy. As the core of the financial system, the sound operation of banks has a significant impact on the stability of the entire financial system and the national economy. However, the complexity of bank equity structures and management relationships poses significant challenges to financial supervision, particularly in the oversight of related-party transactions.

[0003] Related-party transactions refer to transactions between a bank and its related parties, including loans, guarantees, asset transfers, and other forms. While related-party transactions themselves are not illegal, they can easily lead to problems such as profit transfers and risk contagion due to the inherent ties of interest between related parties, necessitating strict oversight. However, traditional bank related-party transaction oversight relies primarily on manual review and fragmented data sources, which presents the following issues: Information fragmentation: Bank equity, management, and transaction information are scattered across multiple data sources such as industrial and commercial data, annual reports, and news, making it difficult to fully obtain and integrate.

[0004] Complex related parties: The relationship between a bank and its related parties may be formed through complex means such as multi-layered shareholdings and cross-management, making it difficult to fully identify through traditional methods.

[0005] Non-standard entity names: The same entity may have different names or abbreviations in different data sources, which increases the difficulty of entity identification and association analysis.

[0006] Low analysis efficiency: Traditional manual analysis methods are difficult to handle massive amounts of data and complex relationships, and are inefficient and prone to errors.

[0007] As a semantic network, knowledge graphs can represent and store knowledge in the form of entities and relationships, providing an effective means for representing and analyzing complex relationships. By constructing a knowledge graph of bank equity and management, dispersed bank equity, management, and transaction information can be integrated into a unified knowledge base, supporting the identification and analysis of complex relationships and providing technical support for the supervision of bank-related transactions. Summary of the Invention

[0008] The present invention aims to disclose a method for identifying and supervising bank-related transactions based on a knowledge graph, which is characterized by comprising the following steps: S1: Multi-source data integration: Integrate internet data and generate a unified data source through data cleaning, standardization, and conflict resolution. S2: Joint entity relationship extraction: Use multi-OCR fusion technology to convert bank data into text, identify entities through the BERT-BiLSTM-CRF model, and jointly extract equity, management, and transaction relationships based on a graph neural network. S3: Entity name normalization: Use a hybrid name normalization method, combining rule matching, character similarity, and Sentence-BERT semantic similarity calculation to map full names, abbreviations, and aliases to standard entity identifiers. S4: Knowledge fusion and storage: Merge multi-source entities using an entity fusion algorithm based on multi-dimensional similarity, and store entities, relationships, and attributes in a Neo4j graph database. The multi-source data integration specifically includes: using an IP proxy pool and User-Agent rotation strategy in acquiring internet data; obtaining industrial and commercial data through distributed scheduling via an API interface; and performing page segmentation, table structure recognition, and confidence-weighted fusion of multiple OCR models on bank annual report PDFs.

[0009] The entity relationship joint extraction adopts an end-to-end framework: Entity recognition layer: Based on the Chinese financial BERT model with domain-adaptive fine-tuning, with a BiLSTM feature enhancement layer and a CRF label constraint layer; Relationship extraction layer: Builds a text dependency graph, aggregates contextual features through a graph attention network, and outputs equity / management / transaction relationships. Entity name normalization includes preprocessing and multi-level matching. The entity fusion algorithm specifically generates entity vectors based on a BERT-Siamese network, integrating character features, structural features, and contextual features; and determines entity alignment through multi-threshold decision making.

[0010] This application also discloses a bank equity and management knowledge graph application system, which is characterized by including: Knowledge graph construction module: executes the method described in any one of claims 1 to 5 to output a knowledge graph of bank equity and management; Related-party transaction judgment module: Based on equity penetration rules, management association rules, and kinship rules, it identifies direct / indirect related parties and generates related-party transaction risk scores; Visualization analysis module: realizes equity network force-oriented layout, related path coloring display and risk heat map, and supports interactive time series analysis. The related transaction judgment module includes: Association path analysis algorithm: Based on the Cypher language of the graph database Neo4j, multi-hop paths are queried and the association strength is calculated as 1 / path length × shareholding ratio; Violation risk assessment model: input transaction size, pricing deviation and disclosure completeness characteristics, and output violation probability.

[0011] The bank equity and management knowledge graph application system according to claim 7 is characterized by further including an equity penetration analysis module: this module traces the equity control chain to the ultimate beneficiary, identifies circular shareholdings and hidden equity concentration, calculates the Herfindahl Index (HHI), and triggers an equity structure warning when the HHI exceeds 0.25. The knowledge graph construction module supports incremental updates: it identifies newly added entities / relationships based on change logs and matches the knowledge base in real time using a BERT-Siamese network; it also implements a version snapshot mechanism to support data backtracking and consistency verification. The visualization analysis module provides overlay functionality for management association network diagrams and penalty graphs. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 This is the overall architecture diagram of the bank-related transaction identification system based on knowledge graph provided by the present invention; Figure 2 is a flow chart of the multi-source data acquisition and integration method provided by the present invention; Figure 3 This is a model structure diagram of the bank entity recognition method based on BERT-BiLSTM-CRF provided by the present invention; Figure 4 It is a model structure diagram of the rule-based relationship extraction method provided by the present invention; Figure 5 This is a model structure diagram of the entity name alignment method based on Sentence-BERT provided by the present invention; Figure 6 This is a flow chart of the method for identifying related transactions based on knowledge graph provided by the present invention. DETAILED DESCRIPTION

[0013] This application constructs three main datasets for bank entity recognition, entity relationship extraction, and entity name normalization tasks: 1. Bank Entity Recognition Dataset (BankNER): This dataset collects text from publicly available bank annual reports, regulatory documents, and financial news, and annotates it by financial experts. The annotated entity types include bank names, company names, personal names, place names, time periods, and amounts. The dataset contains 10,000 text entries and approximately 50,000 annotated entities.

[0014] 2. Bank Relationship Extraction Dataset (BankRE): Based on the BankNER dataset, this dataset further annotates relationships between entities, including investment relationships, controlling relationships, management relationships, and transaction relationships. The dataset contains 8,000 texts and approximately 30,000 annotated relationships.

[0015] 3. Bank Entity Name Normalization Dataset (BankNorm): This dataset collects various names (full names, abbreviations, and nicknames) of banks and their affiliated companies, constructs entity pairs, and annotates whether they refer to the same entity. The dataset contains 50,000 entity name pairs, of which 25,000 are positive examples (referring to the same entity) and 25,000 are negative examples (referring to different entities).

[0016] The data preprocessing process includes the following steps: 1. Text cleaning: remove HTML tags, special characters, and redundant spaces, and unify full-width and half-width characters and simplified and traditional Chinese characters.

[0017] 2. Word segmentation and annotation conversion: Use the jieba word segmentation tool to perform Chinese word segmentation and convert entity annotations into BIO format (Beginning, Inside, Outside).

[0018] 3. Data augmentation: Apply data augmentation techniques to the training set, including synonym replacement, random deletion, random insertion, etc., to expand the diversity of training samples.

[0019] 4. Negative sampling: For relation extraction and entity name normalization tasks, a hard example mining strategy is used to generate high-quality negative samples to improve the model's discriminative ability.

[0020] To ensure data quality, a rigorous quality control process was implemented: two annotators independently annotated the data, followed by a third expert review and resolution of any inconsistent annotations. The final annotation consistency (Cohen's Kappa coefficient) reached 0.92, indicating high annotation quality.

[0021] 1. Evaluation Metrics and Baseline Methods Evaluation Metrics To comprehensively evaluate the model performance, a variety of evaluation indicators are used, including precision, recall, F1 value and accuracy. The definitions of these indicators are as follows: - Precision = TP / (TP + FP) - Recall = TP / (TP + FN) - F1 value = 2 * precision * recall / (precision + recall) - Accuracy = (TP + TN) / (TP + TN + FP + FN) Among them, TP (True Positives) represents the number of correctly predicted positive examples, FP (False Positives) represents the number of incorrectly predicted positive examples, FN (False Negatives) represents the number of incorrectly predicted negative examples, and TN (True Negatives) represents the number of correctly predicted negative examples.

[0022] For different tasks, specific evaluation strategies are adopted: 1. Entity recognition task: A strict matching strategy is adopted. An entity is considered correctly identified only when its predicted entity type and boundary are exactly the same as the annotation.

[0023] 2. Relation extraction task: Two evaluation modes are adopted: (a) relation classification mode, which assumes that the entity boundaries are known and only evaluates the accuracy of relation classification; (b) end-to-end mode, which simultaneously evaluates the performance of entity recognition and relation classification.

[0024] 3. Entity Name Normalization Task: Using a pairwise evaluation strategy, the computational model’s ability to correctly determine whether two entity names refer to the same entity is tested.

[0025] In addition, micro-average and macro-average are used to calculate the overall performance of multi-category tasks to comprehensively evaluate the performance of the model in each category.

[0026] Baseline Method In order to fully evaluate the advantages of the method proposed in this study, a variety of classic and latest baseline methods were selected for comparative experiments. These baseline methods include: 1. Entity Recognition Baseline Method: - CRF: A classic conditional random field model that uses handcrafted features for sequence labeling.

[0027] - BiLSTM-CRF: A deep learning model that combines bidirectional LSTM and CRF, a classic method for sequence labeling.

[0028] - BERT-CRF: A model that combines BERT with CRF, leveraging the powerful representation capabilities of pre-trained language models.

[0029] - SpanBERT: A span-based named entity recognition method that directly predicts the start and end positions of an entity.

[0030] - FLAT: Flattened attention-enhanced sequence tagging model designed for Chinese entity recognition.

[0031] 2. Relation Extraction Baseline Method: - CNN: Use convolutional neural networks to extract relationship features between entity pairs.

[0032] - BiLSTM-Attention: A relation extraction model that combines bidirectional LSTM and attention mechanism.

[0033] - BERT-RE: Use BERT to encode entity pairs and context for relation classification.

[0034] - SpERT: a span-based model for joint entity and relation extraction.

[0035] - CasRel: A relation extraction model based on cascade decoding to handle overlapping relations.

[0036] 3. Entity Name Normalization Baseline Method: - Edit distance: A character similarity method based on Levenshtein distance.

[0037] - Jaccard similarity: a method based on the similarity of character sets.

[0038] - TF-IDF + Cosine Similarity: Use TF-IDF vectors to represent entity names and calculate cosine similarity.

[0039] - Word2Vec + Cosine Similarity: Use Word2Vec word vectors to represent entity names and calculate cosine similarity.

[0040] - Standard Sentence-BERT: Uses the original Sentence-BERT model to calculate the semantic similarity of entity names.

[0041] All baseline methods were trained and evaluated on the same dataset and experimental environment to ensure fair comparison. For methods that require pre-trained models, the same pre-trained model (such as Chinese BERT-base) was used as initialization to eliminate the impact of the pre-trained model selection on the experimental results.

[0042] 2. Experimental Results of Bank Entity Recognition Overall performance comparison The BERT-BiLSTM-CRF model proposed in this study was compared with various baseline methods on the BankNER dataset.

[0043] The experimental results show that the proposed BERT-BiLSTM-CRF model significantly outperforms other baseline methods. In particular, after incorporating domain adaptation fine-tuning, the F1 score reaches 92.29%, 2.84 percentage points higher than the strongest baseline method, FLAT. This demonstrates the effectiveness of the BiLSTM feature enhancement layer and domain adaptation fine-tuning.

[0044] Notably, all pre-trained approaches (BERT-CRF, SpanBERT, FLAT, and

[15] ) significantly outperformed traditional CRF and BiLSTM-CRF approaches, demonstrating the clear advantages of pre-trained language models for bank entity recognition tasks. Furthermore,

[15] 's model achieved balanced improvements in both precision and recall, which is crucial for comprehensive performance in real-world applications.

[0045] Performance analysis of different entity types In order to deeply analyze the performance of the model on different types of entities, the recognition performance of each major entity type was carefully evaluated The results show that the model performs differently for different entity types. The model performs best for highly structured entities with relatively fixed expressions (such as time, amount, and percentage), with F1 scores exceeding 96%. This is primarily because these entities have distinct formal features and regular patterns, making them easy to identify.

[0046] The model also achieved good performance for bank names and person names, with F1 scores reaching 94.21% and 93.21% respectively, indicating that the model can effectively capture the semantic and contextual features of these entities.

[0047] In contrast, recognition performance for company names was relatively low, with an F1 score of 90.84%. This is primarily due to the diverse variations and lengths of company names, as well as their frequent inclusion of complex information such as industry and region, which increases recognition difficulty. Furthermore, confusion between company names and bank names in some cases (e.g., "China Construction Bank" versus "China Jianyin Investment Co., Ltd.") also contributed to the poor performance.

[0048] 5.3.3.3 Ablation Experiment In order to verify the contribution of each component, an ablation experiment was conducted. Ablation experiments show that each component contributes significantly to model performance, with the CRF and BiLSTM layers having the greatest impact. Removing any of these components leads to a significant performance drop. This confirms the design philosophy: BERT provides powerful contextual representations, BiLSTM enhances sequence modeling capabilities, and CRF ensures consistent label transfer.

[0049] Domain adaptation fine-tuning also resulted in an F1 score improvement of approximately 1.58 percentage points, demonstrating the importance of domain knowledge for bank entity recognition. Optimizations such as multi-layer BiLSTM, attention mechanisms, and residual connections each contributed 0.9-1.8 percentage points to performance improvements, demonstrating their importance in enhancing the model's representation capabilities.

[0050] Although the improvement of adversarial training is relatively small (0.7 percentage points), it significantly improves the robustness of the model, especially when dealing with noisy data and unseen entities, showing more stable performance.

[0051] III. Experimental Results of Bank Entity Name Normalization Overall Performance Comparison A comparative experiment was conducted on the BankNorm dataset between the BERT-Siamese network entity fusion model proposed in this study and various baseline methods.

[0052] From the experimental results, it can be seen that the proposed BERT-Siamese network entity fusion model is significantly superior to other baseline methods. Especially after adding domain optimization, the F1 value reaches 94.49%, which is 4.76 percentage points higher than the standard Sentence-BERT. This proves the effectiveness of the proposed model and optimization strategy.

[0053] Traditional character similarity-based methods (edit distance and Jaccard similarity) have relatively poor performance, mainly because they cannot capture the semantic information of bank entity names and only rely on the similarity of surface forms. Methods based on distributed representations (Word2Vec and Sentence-BERT) have significantly improved performance, indicating that semantic modeling is crucial for entity name normalization.

[0054] It is worth noting that the proposed model has achieved a balanced improvement in both precision and recall, which is crucial for the comprehensive performance in practical applications. Especially the significant improvement in precision (reaching 95.23%) indicates that the model can effectively reduce incorrect matches, which is particularly important in banking business scenarios.

[0055] Performance Analysis of Different Entity Types To deeply analyze the performance of the model on different types of bank entity names, a detailed evaluation was carried out on the normalization performance of each main type.

[0056] From the results, it can be seen that the performance of the model on different types of bank entity name pairs varies. For the matching of bank full names and abbreviations (such as "China Construction Bank Corporation Limited" and "CCB"), the model performs the best, with an F1 value reaching 97.11%. This is mainly because bank abbreviations are usually direct abbreviations of full names, with strong regularity and consistency. <00,00161>

[0057] The matching of bank full names and aliases (such as "China Construction Bank Corporation Limited" and "Construction Bank") and the matching of bank abbreviations and aliases (such as "CCB" and "Construction Bank") also achieved good performance, with F1 values of 93.73% and 96.10% respectively.

[0058] In contrast, matching tasks involving branches is more challenging, particularly matching bank nicknames with branches (e.g., "Construction Bank" with "Construction Bank Beijing Branch"), with an F1 score of 90.33%. This is primarily because branch names contain additional geographical and hierarchical information, which increases matching complexity.

[0059] IV. Experimental Results on Identifying Bank-Related Transactions Overall performance comparison The knowledge graph-based bank related transaction identification method proposed in this study was compared with multiple baseline methods.

[0060] Experimental results show that the proposed knowledge graph-based approach to identifying bank-related transactions significantly outperforms other baseline methods. In particular, after inference enhancement, the F1 score reaches 92.60%, 5.45 percentage points higher than the strongest baseline method (a relationship-based approach). This demonstrates the powerful ability of knowledge graphs to represent and analyze complex banking relationship networks.

[0061] Traditional rule-based and statistical methods have relatively poor performance, primarily because they struggle to capture the complex patterns and hidden relationships in bank-related transactions. Machine learning and deep learning methods have seen some improvement, but remain limited by feature engineering and model expressiveness. Entity- and relationship-based methods have further improved performance, but they typically only consider direct connections and struggle to detect multi-hop relationships and hidden paths.

[0062] In contrast, the knowledge graph approach comprehensively represents bank entities, relationships, and attributes, and discovers complex association paths through graph algorithms and inference rules, resulting in superior performance in identifying hidden related-party transactions. In particular, the reasoning enhancement strategy resulted in an F1 score improvement of approximately 2.22 percentage points, demonstrating the significant value of knowledge reasoning in identifying related-party transactions.

[0063] 5.3.5.2 Performance Analysis of Different Association Types In order to deeply analyze the performance of the model on different types of related transactions, the recognition performance of each major type of relationship was carefully evaluated.

[0064] The results show that the model performs differently for different types of related-party transactions. The model performs best for direct and obvious relationships, such as equity and management relationships, with F1 scores reaching 95.61% and 93.43%, respectively. This is primarily because these relationships are typically well-documented and legally defined, resulting in higher-quality data.

[0065] The recognition performance for business connections and family connections was slightly lower, with F1 scores of 91.44% and 89.82%, respectively. This is because business connections take many forms and have varying criteria for identification, while family connections often lack public records and rely on indirect evidence.

[0066] For indirect relationships, identification performance gradually decreases as the length of the relationship path increases. The F1 value for a one-hop indirect relationship (e.g., if A controls B, and B controls C, then A and C are indirectly related) is 92.33%, while the F1 value for indirect relationships of three or more hops drops to 85.08%. This indicates that as the relationship chain lengthens, the strength of the relationship weakens, uncertainty increases, and the difficulty of identification increases accordingly.

[0067] Application Case of Bank Equity Management Knowledge Graph I. Application of Related-Party Transaction Supervision Related-party transaction identification and early warning Based on the knowledge graph of bank equity management, we developed an application for related-party transaction identification and early warning. Its main functions include: 1. Automatic identification of related parties: Based on the knowledge graph, the bank's related parties are automatically identified, including equity related parties, management related parties, family related parties, and business related parties.

[0068] 2. Related-party transaction detection: Analyze bank transaction data, automatically detect potential related-party transactions, and conduct risk assessments.

[0069] 3. Related-party transaction warning: Generate risk warning information based on the scale, frequency and pricing fairness of related-party transactions and push it to relevant regulators.

[0070] 4. Related-party transaction report: Automatically generate a related-party transaction report, including a list of related parties, related-party transaction details, and risk assessment results.

[0071] Case Study: A regulatory agency used this application to monitor related-party transactions within its jurisdiction. The system successfully identified a hidden related-party transaction. The transaction involved a large loan from a bank to a company. Knowledge graph analysis revealed that the company had multiple equity relationships with another company controlled by a relative of the bank's chairman, constituting a related-party transaction. However, the company had not followed the required approval and disclosure procedures. The system issued a timely warning, prompting the regulatory agency to investigate and avert potential risks.

[0072] Compliance analysis of related-party transactions The Related Party Transaction Compliance Analysis application conducts a comprehensive analysis of the compliance of banks' related party transactions based on regulatory requirements. Its main functions include: 1. Regulatory rules library: Establish a regulatory rules library for related-party transactions of banks, including the definition of related parties, related-party transaction approval procedures, information disclosure requirements, etc.

[0073] 2. Compliance Check: Conduct compliance checks on identified related-party transactions to determine whether they comply with regulatory requirements.

[0074] 3. Risk assessment of non-compliance: Conduct risk assessments on non-compliant related-party transactions and analyze the potential impact and risk level.

[0075] 4. Corrective action suggestions generation: Automatically generate corrective action suggestions and measures for non-compliance situations.

[0076] Case Study: A bank used the app to conduct a self-examination of internal related-party transactions. The system discovered that several transactions with related parties had not been included in related-party transaction management, posing a compliance risk. Based on the system's recommendations, the bank improved its related-party identification mechanism, revised its related-party transaction management system, and conducted additional review and disclosure of historical transactions, effectively mitigating regulatory risks.

[0077] 2. Application of Equity Structure Analysis Equity penetration analysis The equity penetration analysis application uses knowledge graph technology to achieve multi-level penetration analysis of bank equity structures. Its main functions include: 1. Equity chain tracing: Trace the multi-level control chain of bank equity and identify the ultimate beneficiary.

[0078] 2. Cross-shareholding analysis: Analyze cross-shareholding relationships among banks and identify potential circular shareholding risks.

[0079] 3. Equity Concentration Assessment: Evaluate the concentration of bank equity and identify the risk of excessive concentration of control.

[0080] 4. Equity change monitoring: monitor changes in bank equity and analyze trends and impacts.

[0081] Case Study: A regulatory agency used this application to conduct an equity penetration analysis of a city commercial bank. The agency discovered that the bank ostensibly had more than 20 shareholders, resulting in a relatively dispersed ownership structure. However, through multi-level penetration analysis, the agency identified a single conglomerate that, through multiple affiliated companies and individuals, effectively controlled over 45% of the bank's shares, posing a risk of hidden equity concentration. Consequently, the regulatory agency required the bank to adjust its equity structure to mitigate the risk of improper interference by a major shareholder.

[0082] Shareholder network analysis The shareholder network analysis application uses knowledge graph technology to analyze the complex relationships between bank shareholders. Its main functions include: 1. Shareholder Relationship Map: Build a relationship map between bank shareholders, including equity relationships, management relationships, business relationships, etc.

[0083] 2. Shareholder group identification: Based on the relationship between shareholders, identify the groups of shareholders with consistent interests.

[0084] 3. Association risk assessment: Evaluate risk points and transmission paths in the shareholder association network.

[0085] 4. Shareholder Behavior Analysis: Analyze shareholders’ voting behavior, trading behavior, and the exercise of their influence.

[0086] Case Study: A bank used this application to analyze its shareholder structure and discovered that several seemingly independent shareholders actually had close business and investment partnerships, constituting a de facto relationship of concerted action. Based on this information, the bank adjusted its shareholder voting rights management and related-party transaction approval processes to mitigate the risks of shareholder abuse of voting rights and improper related-party transactions.

[0087] 3. Network Analysis Application at the Management Level Management Relationship Analysis The management-level association analysis application uses knowledge graph technology to analyze the complex association network of bank management. Its main functions include: 1. Management Relationship Map: Build a relationship map of the bank's management, including career experience, educational background, social relationships, etc.

[0088] 2. Management cross-appointment analysis: Analyze the cross-appointment of management in multiple institutions to identify potential conflicts of interest.

[0089] 3. Management Change Analysis: Analyze management changes and trends and assess their impact on bank governance.

[0090] 4. Management background investigation: Conduct a comprehensive investigation of the management background to identify potential risk factors.

[0091] Case Study: A regulatory agency used this application to conduct a correlation analysis of bank management within its jurisdiction. It discovered that senior executives at multiple banks had close professional and social connections, forming a tight "circle." This network of relationships could lead to risks of unfair competition and profiteering within the industry. Consequently, the regulator strengthened its corporate governance oversight of the relevant banks, requiring them to improve their independent director systems and internal control mechanisms.

[0092] Management performance evaluation The Management Performance Assessment application uses knowledge graphs and data analysis technology to comprehensively evaluate the performance of bank management. Its main functions include: 1. Job Performance Analysis: Analyze management's job performance behaviors, such as meeting attendance, decision-making participation, and speaking engagements.

[0093] 2. Job Performance Assessment: Evaluate management's professional background, experience level, decision-making quality, and other aspects of their job performance.

[0094] 3. Identification of performance risks: Identifying potential risks and improper behaviors in the performance of management duties.

[0095] 4. Performance report generation: Automatically generate management performance evaluation reports to provide reference for supervision and internal governance.

[0096] Case Study: A bank used the app to assess the performance of its board members. They discovered that some independent directors were insufficiently involved in major decision-making and their professional opinions were insufficiently expressed, posing a risk of becoming "figurehead directors." Based on this, the bank optimized its board operations, strengthened support and evaluation of independent directors' performance, and enhanced the effectiveness of corporate governance.

[0097] 4. Risk early warning and prevention and control applications Risk transmission analysis The risk transmission analysis application uses knowledge graph technology to analyze the transmission path and impact range of bank risks. Its main functions include: 1. Identification of risk transmission paths: Identify the risk transmission paths within and outside the bank, including the equity chain, management chain, and business chain.

[0098] 2. Risk transmission simulation: simulate the transmission process and impact range of different risk events and assess potential losses.

[0099] 3. Risk Isolation Analysis: Analyze the effectiveness of risk isolation measures and identify weaknesses in risk firewalls.

[0100] 4. Systemic risk assessment: Assess the systemic risks that may arise from inter-bank risk transmission.

[0101] Case Study: A regulatory agency used this application to analyze the risk transmission of a large conglomerate that holds controlling stakes in multiple banks and non-bank financial institutions. The analysis revealed complex cross-guarantee and fund flows within the conglomerate. Risks within the group's core enterprises could quickly spread to the controlling financial institutions through multiple channels, creating a chain reaction. Consequently, the regulatory agency required the relevant financial institutions to strengthen their management of interrelated risks and establish risk isolation mechanisms to prevent contagion within the group.

[0102] Abnormal behavior monitoring The Abnormal Behavior Monitoring application uses knowledge graphs and machine learning technologies to monitor abnormal behavior of banks and their affiliates in real time. Its main functions include: 1. Abnormal transaction monitoring: Monitor abnormal transactions between banks and related parties, including abnormal size, frequency, and pricing.

[0103] 2. Abnormal relationship monitoring: Monitor abnormal relationship changes in the bank's related network, such as sudden changes in equity and management.

[0104] 3. Abnormal behavior warning: Generate warning information for detected abnormal behavior and push it to relevant personnel.

[0105] 4. Abnormal behavior analysis: Analyze the causes, impacts, and risk levels of abnormal behaviors and provide treatment recommendations.

[0106] Case Study: A bank used this application to monitor internal transactions. The system detected a series of unusual fund transfers and guarantee transactions involving multiple affiliated companies, creating a complex capital flow path. Further analysis revealed that these transactions were intended to conceal non-performing assets and circumvent regulatory requirements. The bank promptly identified and addressed this issue, avoiding further risk losses.

Claims

1. A method for identifying and supervising bank-related transactions based on knowledge graph, characterized in that: The following steps are involved: S1: Multi-source data integration: Integrate Internet data and generate a unified data source through data cleaning, standardization and conflict resolution; S2: Joint extraction of entity relationships: Multi-OCR fusion technology is used to convert bank data into text, entities are identified through the BERT-BiLSTM-CRF model, and equity relationships, management relationships, and transaction relationships are jointly extracted based on graph neural networks; S3: Entity name normalization: A hybrid name normalization method is used, combining rule matching, character similarity, and Sentence-BERT semantic similarity calculation to map full names, abbreviations, and aliases to standard entity identifiers; S4: Knowledge fusion and storage: An entity fusion algorithm based on multi-dimensional similarity is used to merge multi-source entities, and entities, relationships, and attributes are stored in the Neo4j graph database.

2. The method for identifying and supervising bank-related transactions based on knowledge graph according to claim 1 is characterized in that: The multi-source data integration specifically includes: using IP proxy pool and User-Agent rotation strategy in acquiring Internet data; acquiring industrial and commercial data through distributed scheduling of API interface; performing page segmentation, table structure recognition and multi-OCR model confidence weighted fusion on bank annual report PDF.

3. The method for identifying and supervising bank-related transactions based on knowledge graph according to claim 2 is characterized in that: The entity relationship joint extraction adopts an end-to-end framework: Entity recognition layer: Based on the Chinese financial BERT model with domain-adaptive fine-tuning, with a BiLSTM feature enhancement layer and a CRF label constraint layer; Relationship extraction layer: Builds a text dependency graph, aggregates contextual features through a graph attention network, and outputs equity / management / transaction relationships.

4. The method for identifying and supervising bank-related transactions based on knowledge graph according to claim 3 is characterized in that: The entity name normalization includes: preprocessing; multi-level matching.

5. The method for identifying and supervising bank-related transactions based on knowledge graph according to claim 4 is characterized in that: The entity fusion algorithm specifically comprises: generating entity vectors based on the BERT-Siamese network, integrating character features, structural features, and context features; Entity alignment is determined through multi-threshold decision making.

6. A bank equity and management knowledge graph application system, characterized by: include: Knowledge graph construction module: executes the method described in any one of claims 1 to 5 to output a knowledge graph of bank equity and management; Related-party transaction judgment module: Based on equity penetration rules, management association rules, and kinship rules, it identifies direct / indirect related parties and generates related-party transaction risk scores; Visualization Analysis Module: Implements force-oriented layout of equity networks, color-coded display of associated paths, and risk heat maps, and supports interactive time series analysis.

7. The bank equity and management knowledge graph application system according to claim 6 is characterized in that: The related-party transaction judgment module includes: Association path analysis algorithm: Based on the Cypher language of the graph database Neo4j, multi-hop paths are queried and the association strength is calculated as 1 / path length × shareholding ratio; Violation risk assessment model: input transaction size, pricing deviation and disclosure completeness characteristics, and output violation probability.

8. The bank equity and management knowledge graph application system according to claim 7 is characterized in that: It also includes an equity penetration analysis module: it traces the equity control chain to the ultimate beneficiary, identifies circular shareholdings and hidden equity concentration, calculates the Herfindahl Index (HHI) of equity concentration, and triggers an equity structure warning when HHI>0.

25.

9. The system according to claim 6, wherein: The knowledge graph construction module supports incremental updates: identifying new entities / relationships based on change logs, and matching the knowledge base in real time through the BERT-Siamese network; setting up a version snapshot mechanism to support data backtracking and consistency verification.

10. The system according to claim 6, wherein: The visual analysis module provides the function of overlaying the management-level association network diagram and the penalty map.

Images