Off-site derivative cross-border supervision intelligent adaptation system based on heterogeneous data fusion

Abstract

The invention discloses an off-site derivative cross-border supervision intelligent adaptation system based on heterogeneous data fusion, and relates to the technical field of big data processing. A data acquisition module is set to obtain off-site derivative transaction data, and an original data set is generated; setting a data structured processing module to perform semantic analysis on the original data set by adopting a financial text-oriented text model to generate a transaction information table, setting a transaction type classification module to judge whether cross-border transactions are involved or not according to transaction types in the transaction information table, and if the cross-border transactions are involved, setting a cross-border compliance evaluation module to perform cross-border compliance evaluation on the cross-border transactions. Performing dynamic rule matching on the transaction information table to generate a compliance report; otherwise, a risk prediction module is set, risk prediction is conducted on the transaction information table through a pre-trained transaction graph neural network model, a risk early warning signal is generated, and corresponding early warning measures are executed. The compliance cost is reduced, and the risk management and control capability is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of big data processing, and specifically to an intelligent adaptation system for cross-border supervision of over-the-counter derivatives based on heterogeneous data fusion. Background Art

[0002] As an important part of the global financial system, the over-the-counter (OTC) derivatives market has a large trading scale and high complexity. Different from exchange-traded standardized products, OTC derivatives are highly customized with flexible contract terms, which poses great challenges to supervision. Especially in cross-border trading scenarios, due to the involvement of laws, regulations, regulatory requirements, and compliance standards in different countries and regions, the supervision complexity increases exponentially.

[0003] In OTC derivatives trading, cross-border transactions need to meet the regulatory requirements of multiple jurisdictions simultaneously, involving multiple compliance considerations such as foreign exchange control, capital flow restrictions, and anti-money laundering regulations. At the same time, cross-border transactions may also trigger regulatory arbitrage, tax risks, and cross-border transmission of systemic risks, posing potential threats to financial stability.

[0004] Existing OTC derivatives supervision technologies have many deficiencies. First, the problem of data silos is serious, and it is difficult to effectively integrate data among exchanges, banks, and regulatory agencies, resulting in a one-sided supervision perspective. Second, traditional rule matching methods lack flexibility and are difficult to cope with the highly customized characteristics of OTC derivative contracts.

[0005] Therefore, the present invention proposes an intelligent adaptation system for cross-border supervision of OTC derivatives based on heterogeneous data fusion. Summary of the Invention

[0006] The present invention aims to solve at least one of the technical problems existing in the prior art. For this purpose, the present invention proposes an intelligent adaptation system for cross-border supervision of OTC derivatives based on heterogeneous data fusion, which reduces the compliance cost and improves the risk control ability.

[0007] To achieve the above object, an intelligent adaptation system for cross-border supervision of OTC derivatives based on heterogeneous data fusion is proposed, including a data collection module, a data structuring processing module, a transaction type classification module, a cross-border compliance assessment module, a risk prediction module, and a supervision decision module; wherein, each module is electrically connected; The data collection module is used to obtain OTC derivatives trading data, including structured data from exchanges, unstructured data from banks, and regulatory review information from regulatory agencies, generate an original data set, and send the original data set to the data structuring processing module; A data structuring and processing module, which is used to perform semantic parsing on the original data set by using a text model for financial texts, generate a transaction information table, and send the transaction information table to the transaction type classification module; A transaction type classification module, which determines whether a cross-border transaction is involved according to the transaction type in the transaction information table. If a cross-border transaction is involved, the transaction information table is sent to the cross-border compliance assessment module; otherwise, the transaction information table is sent to the risk prediction module; A cross-border compliance assessment module, which performs dynamic rule matching on the transaction information table based on a pre-constructed rule knowledge graph to generate a compliance report; A risk prediction module, which performs risk prediction on the transaction information table by using a pre-trained transaction graph neural network model, generates a risk warning signal, and executes corresponding warning measures.

[0008] The generation of the original data set includes the following steps: Step 11: Collect exchange structured data through multi-source heterogeneous interfaces, including transaction records, product information, and market quotation data; The collection of the exchange structured data includes the following sub-steps: Step 111: Establish an API connection channel with the world's major over-the-counter derivatives exchanges; Step 112: Extract transaction record data from the exchange system through the corresponding API interface according to a predefined data template; Step 113: Obtain the product information of over-the-counter derivatives from the exchange system through the corresponding API interface; Step 114: Collect market quotation data from the exchange system in real time through the corresponding API interface; Step 115: The transaction record data, the product information of over-the-counter derivatives, and the market quotation data form the exchange structured data; Step 12: Obtain bank unstructured data by using an adaptive web crawler technology, including transaction announcements, product manuals, and risk disclosure documents; The obtaining of the bank unstructured data includes the following steps: Step 121: Configure an intelligent web crawler task to directionally collect relevant documents of over-the-counter derivatives on the official websites of various financial institutions; Step 122: Identify and extract the product announcement information of the transaction announcements issued by banks through the intelligent web crawler task; Step 123: Obtain and parse the product manuals issued by banks through the intelligent web crawler task, and extract product description information; Step 124: Collect the risk disclosure documents issued by banks through the intelligent web crawler task, and extract product risk information; Step 125: The product announcement information, the product description information, and the product risk information constitute the bank's unstructured data; Step 13: Connect to the regulatory agency's protocol database based on a secure communication protocol to obtain regulatory filing and compliance review information; Obtaining regulatory filing and compliance review information includes the following steps: Step 131: Establish a secure data exchange channel with financial regulatory agencies in various countries; Step 132: Through the secure data exchange channel, regularly obtain regulatory filing data for over-the-counter derivatives transactions according to the data exchange protocol preset by regulatory requirements; Step 133: Through the secure data exchange channel, obtain the compliance review information issued by the regulatory agency; Step 134: The regulatory review information includes the regulatory filing data and the compliance review information; Step 14: Classify the structured data of the exchange, the unstructured data of the bank, and the regulatory review information of the regulatory agency according to the product name or product code, and save them as the original dataset; The steps for generating a transaction information table by performing semantic parsing on the original dataset using a text model for financial texts include the following: Step 21: Construct a pre-trained language model dedicated to the financial field; the pre-trained language model dedicated to the financial field uses a bidirectional encoder representation variant based on the Transformer architecture; Step 22: Apply natural language processing techniques to the bank's unstructured data to extract key transaction elements and clause information; Step 23: Perform structured conversion on the regulatory review information to extract compliance requirements and regulatory decision-making bases; Step 24: Based on the pre-trained language model, construct a multi-modal fusion model for financial texts to integrate the semantic information of structured data, key transaction elements and clause information, and compliance requirements and regulatory decision-making bases; the multi-modal fusion model uses a cross-modal attention architecture based on Transformer; Step 25: Perform in-depth information extraction on the fused semantic information to identify and extract the core elements and associated attributes of the transaction, and convert them into structured graph units; Step 26: Semantically integrate the graph units to construct a transaction knowledge graph and form an association network between transaction entities; Step 27: Based on the transaction knowledge graph, perform multi-dimensional feature extraction to generate a structured representation of the transaction; Step 28: Based on the structured representation of the transaction, generate a standardized transaction information table, which includes keyword fields such as basic transaction information, participant information, product information, risk indicators, and compliance status; the generation of the transaction information table includes two stages: field design and data filling.

[0009] The method for determining whether a cross-border transaction is involved according to the transaction type in the transaction information table is as follows: Read the field value of the transaction type of each product in the transaction information table. If the field value indicates a cross-border transaction, it is determined that a cross-border transaction is involved; otherwise, it is determined that no cross-border transaction is involved.

[0010] The steps for dynamically matching rules for the transaction information table based on a pre-constructed rule knowledge graph to generate a compliance report are as follows: Step 31: Pre-collect the regulatory rule texts of each financial regulatory agency, and based on the regulatory rule texts, construct the rule knowledge graph; the construction of the rule knowledge graph includes three stages: rule acquisition, rule parsing, and rule modeling; Step 32: Extract rule nodes and their association relationships from the rule knowledge graph to construct a rule matching engine; the rule matching engine includes three core components: a rule parser, a rule executor, and a result collector, which are used for dynamically matching rules for the data in the transaction information table; Step 33: Generate a structured compliance report through the rule matching engine.

[0011] The steps for predicting risks for the transaction information table using a pre-trained transaction graph neural network model to generate risk warning signals and execute corresponding warning measures are as follows: Step 41: Convert the transaction information table into graph-structured data as a transaction graph network; Step 42: Pre-construct the transaction graph neural network model; the transaction graph neural network model includes three components: a graph convolutional encoder, a temporal attention layer, and a risk decoder; The transaction graph neural network model adopts an architecture that combines a multi-layer graph convolutional network and a temporal attention mechanism. The transaction graph neural network model includes at least three components: a graph convolutional encoder, a temporal attention layer, and a risk decoder.

[0012] The graph convolutional encoder consists of three layers of graph convolutional networks, with 64 convolutional kernels in each layer; the first layer of graph convolution aggregates information from the first-order neighborhood of nodes to capture direct transaction relationships; the second layer of graph convolution extends to the second-order neighborhood to capture indirect association patterns; the third layer of graph convolution further expands the receptive field to identify complex network structure features; after each layer of graph convolution, batch normalization and ReLU activation functions are sequentially passed through; The time-series attention layer processes the time-series features of transactions and captures the time-evolution features of abnormal transaction patterns. The time-series attention layer adopts a self-attention mechanism to calculate the correlation weights of transaction features at different time points. The risk decoder adopts a multi-task learning framework to simultaneously predict the risk scores in four dimensions: market risk, credit risk, operational risk, and compliance risk. First, the risk decoder maps the node-level risk representation to a hidden feature space through two fully connected networks. The dimension of the hidden layer is 128, and the LeakyReLU activation function and Dropout regularization are used. Then, for each risk type, a corresponding prediction head is designed, which includes a fully connected layer and a Softmax activation function to output the five-level scoring probability distribution of this risk dimension.

[0013] Step 43: Train the transaction graph neural network model with the pre-collected historical transaction data. Step 44: Input the transaction graph network into the pre-trained transaction graph neural network model to generate multi-dimensional risk scores. Step 45: Integrate the multi-dimensional risk scores according to the preset risk weights to generate a comprehensive risk level. The generation of the comprehensive risk level adopts a weighted fusion method, and the weights are dynamically adjusted according to the importance and relevance of different risk dimensions.

[0014] Step 46: Generate a risk warning signal according to the comprehensive risk level and the preset warning threshold. Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention obtains over-the-counter derivative transaction data through a data acquisition module to generate an original data set, sets a data structuring and processing module to perform semantic parsing on the original data set using a text model for financial texts to generate a transaction information table, sets a transaction type classification module to determine whether a cross-border transaction is involved according to the transaction type in the transaction information table. If a cross-border transaction is involved, a cross-border compliance assessment module is set to perform dynamic rule matching on the transaction information table to generate a compliance report; otherwise, a risk prediction module is set to perform risk prediction on the transaction information table using a pre-trained transaction graph neural network model to generate a risk warning signal and execute corresponding warning measures. By integrating heterogeneous data sources, combining knowledge graphs and graph neural network technologies, the effective integration of data from exchanges, banks, and regulatory agencies is achieved, providing a comprehensive regulatory perspective. Secondly, through the dynamic matching mechanism of the rule knowledge graph, the problem of the lack of flexibility of traditional rule matching engines is solved, and it can accurately adapt to different types of over-the-counter derivative contracts. Moreover, the risk prediction model based on the graph neural network successfully captures the complex correlations in the transaction network, realizes the effective identification and warning of systemic risks, greatly reduces the compliance cost, and improves the risk control ability. Description of the Drawings

[0015] Figure 1 This is the module connection diagram of the off-exchange derivatives cross-border supervision intelligent adaptation system based on heterogeneous data fusion in Embodiment 1 of the present invention. Detailed implementation manners

[0016] Next, the technical solutions of the present invention will be described clearly and completely in conjunction with the embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0017] As Figure 1 shown, the off-exchange derivatives cross-border supervision intelligent adaptation system based on heterogeneous data fusion includes a data collection module, a data structuring processing module, a transaction type classification module, a cross-border compliance assessment module, a risk prediction module, and a supervision decision-making module; among them, each module is connected electrically; The data collection module is used to obtain off-exchange derivatives trading data, including structured data from exchanges, unstructured data from banks, and regulatory review information from regulatory agencies, generate an original data set, and send the original data set to the data structuring processing module; The data structuring processing module is used to perform semantic parsing on the original data set using a text model for financial texts, generate a transaction information table, and send the transaction information table to the transaction type classification module; The transaction type classification module determines whether a cross-border transaction is involved according to the transaction type in the transaction information table. If a cross-border transaction is involved, the transaction information table is sent to the cross-border compliance assessment module; otherwise, the transaction information table is sent to the risk prediction module; The cross-border compliance assessment module performs dynamic rule matching on the transaction information table based on a pre-constructed rule knowledge graph to generate a compliance report; The risk prediction module uses a pre-trained transaction graph neural network model to perform risk prediction on the transaction information table, generate a risk warning signal, and execute corresponding warning measures.

[0018] In the embodiment of the present invention, the generation of the original data set includes the following steps: Step 11: Collect structured data from exchanges through a multi-source heterogeneous interface, including transaction records, product information, and market quotation data; Specifically, the collection of structured data from exchanges includes the following sub-steps: Step 111: Establish an API connection channel with the world's major off-exchange derivatives exchanges; Specifically, in the preferred embodiment of the present invention, the API connection channel adopts a hierarchical architecture design, including a communication layer, an authentication layer, a data exchange layer, and a monitoring layer. The communication layer realizes secure data transmission based on the HTTPS protocol, supports the TLS1.3 encryption standard, and ensures the confidentiality and integrity during the data transmission process. The authentication layer implements an authentication mechanism based on OAuth 2.0, supports digital certificates and two-factor authentication, and prevents unauthorized access. The data exchange layer adopts a dual-mode design of REST and WebSocket. The REST interface is used to obtain batch historical data, and the WebSocket interface is used to push real-time trading data.

[0019] Step 112: Extract transaction record data from the exchange system through the corresponding API interface according to the predefined data template, including key fields such as information of both trading parties, trading time, product type, notional principal, currency, term, and settlement method. Specifically, the predefined data template is defined using the JSON Schema specification and includes two parts: core trading elements and extended attributes. The core trading elements include: a unique transaction identifier, a transaction execution timestamp, the legal entity identification codes of both trading parties; the trading product type; the notional principal amount and currency, the start date and maturity date of the transaction, the settlement method, including options such as physical delivery, cash settlement, or net settlement. The extended attributes set specific fields according to different derivative types. For example, interest rate derivatives include benchmark interest rate, fixed interest rate, floating interest rate, etc.; foreign exchange derivatives include spot exchange rate, forward exchange rate, delivery method, etc.; commodity derivatives include underlying commodity, measurement unit, delivery location, etc.

[0020] Step 113: Obtain the product information of over-the-counter derivatives from the exchange system through the corresponding API interface, including product code, contract specification, pricing model, risk parameters, and market liquidity indicators. Specifically, there are two modes for obtaining product information: active query and change subscription. In the active query mode, a full-volume product information synchronization is performed daily after the trading session ends to obtain the latest information of all active products. In the change subscription mode, product information change notifications are received in real time through a WebSocket connection, and updates are obtained immediately when a new product is launched, existing product parameters are adjusted, or a product is delisted.

[0021] The obtained product information includes the following key elements: a product unique identifier, using the ISIN or a custom coding system; product classification information, including main categories such as interest rate, credit, foreign exchange, commodities, etc. and sub-categories such as interest rate swaps, credit default swaps, foreign exchange options, etc.; contract specifications, including contract unit, minimum price change unit, trading time, and settlement rules; pricing model description, including model types such as Black-Scholes, Hull-White, etc. and key parameters; risk parameters, including Greek letter values such as historical volatility, Delta, Gamma, Vega, Theta, etc.; market liquidity indicators, including average daily trading volume, bid-ask spread, and market depth, etc.

[0022] Step 114: Real-time collect market quotation data from the exchange system through the corresponding API interface, including quotations, transaction prices, bid-ask depth, implied volatility, and yield curve, etc. Specifically, the market quotation data collection adopts a multi-level cache and stream processing architecture to ensure the real-time and integrity of the data. Establish a low-latency connection with the exchange quotation release system for key derivative varieties, such as foreign exchange options of major currency pairs, interest rate swaps of benchmark interest rates, etc.

[0023] In the specific implementation process of the present invention, the quotation data collection is divided into three frequency levels: the first level is the real-time quotation of core products, adopting the push mode, and immediately pushed when the data at the exchange end changes, and the update frequency can reach 10 times per second; the second level is the quotation of regular products, adopting the polling mode, and updated once every 5 seconds; the third level is the quotation of low-liquidity products, updated once every minute. Dynamically adjust the collection frequency according to the trading activity of the product, and automatically increase the collection frequency when the market volatility intensifies.

[0024] The collected market quotation data includes: the latest transaction price and volume; the best bid-ask quotations and order quantities; the bid-ask depth data, recording the cumulative order quantities at different price levels; the implied volatility surface, covering combinations of different strike prices and expiration dates; the yield curve data, including the spot rate and forward rate at key maturity points; the basis data, reflecting the price difference relationship between the spot and derivatives; market sentiment indicators, such as the panic index, etc.

[0025] Step 115: The trading record data, product information of over-the-counter derivatives, and market quotation data form the exchange structured data. Step 12: Adopt an adaptive crawler technology to obtain bank unstructured data, including transaction announcements, product manuals, and risk disclosure documents. Specifically, the obtaining of bank unstructured data includes the following steps: Step 121: Configure intelligent crawler tasks to directionally collect relevant documents of over-the-counter derivatives on the official websites of various financial institutions. In the specific implementation process of the present invention, the intelligent crawler adopts a distributed architecture design, and at least includes five core components: a task scheduler, a URL manager, a page downloader, a content extractor, and a data storage. The task scheduler is responsible for the creation, allocation, and monitoring of crawler tasks, and supports priority-based task scheduling and an automatic retry mechanism for failed tasks. The URL manager maintains the queue of URLs to be crawled and the set of crawled URLs, and realizes URL deduplication and priority sorting. The page downloader is responsible for obtaining web page content, and realizes customized request headers, proxy IP rotation, and access frequency control. The content extractor is responsible for extracting structured information from web pages, and supports various extraction methods such as XPath, CSS selectors, and regular expressions. The data storage is responsible for saving the extracted data to a distributed file system and a database; Step 122: Through the intelligent crawler task, identify and extract the product announcement information of the transaction announcements issued by the bank, including new product releases, transaction term changes, and major market event notifications; In the specific implementation process of the present invention, the extraction of transaction announcements adopts a hybrid method combining rules and machine learning. The rule extraction part uses predefined document structure templates and keyword matching rules to quickly locate the page areas that may contain transaction announcements. The machine learning extraction part uses the BERT-BiLSTM-CRF sequence labeling model to perform refined entity recognition and relationship extraction on the text.

[0026] The input of the BERT-BiLSTM-CRF model is the web page text paragraph. First, the pre-trained BERT model in the financial field is used to generate context-related word vector representations, then the bidirectional LSTM network is used to capture sequence features, and finally the conditional random field layer is used to realize the prediction of the optimal labeling sequence.

[0027] During the extraction of transaction announcements, the main transaction announcements identified are new product release announcements, including product names, types, target markets, and launch times; transaction term change notifications, including the comparison of terms before and after the change and the effective time; market interruption event announcements, including the reasons for the interruption, the scope of influence, and the recovery expectation; regulatory compliance updates, including the implementation of new rules and changes in compliance requirements; risk warning announcements, including risk warnings for specific markets or products; Step 123: Through the intelligent crawler task, obtain and parse the product manuals issued by the bank, and extract product description information such as product structure, pricing mechanism, risk-return characteristics, and applicable investor types; Specifically, the extraction of the product description information includes three stages: layout analysis, content recognition, and information extraction; In the layout analysis stage, a document layout analysis model based on Mask R-CNN is adopted to segment the product manual document into content areas such as titles, texts, tables, charts, and notes.

[0028] In the content recognition stage, specialized processing strategies are applied to different types of content areas. For example, for text areas, a pre-trained BERT model in the financial field is used for semantic understanding; for table areas, a table structure recognition algorithm is used to extract row-column relationships and cell contents; for chart areas, image recognition technology is used to extract key data points and trend information.

[0029] In the information extraction stage, based on a predefined product manual knowledge model, product description information is extracted from the recognized content. The extracted product description information includes: basic product information, such as name, type, issuer, custodian; product structure, such as underlying assets, linking mechanisms, payment structures; pricing mechanisms, such as initial pricing, valuation methods, price adjustment mechanisms; risk-return characteristics, such as expected return rate, maximum loss, risk level; applicable investors, such as investor type, minimum investment amount, suitability requirements; legal terms, such as governing law, dispute resolution mechanism, early termination clause; Step 124: Through an intelligent crawler task, collect risk disclosure documents issued by banks, and extract product risk information such as risk types, risk levels, risk mitigation measures, and investor protection clauses; Specifically, natural language processing technology is used for the collection and analysis of risk disclosure documents to identify and quantify various risk factors. Its specific processing process includes four steps: risk paragraph identification, risk classification, and risk quantification.

[0030] In the risk paragraph identification stage, a sequence labeling model based on BiLSTM-CRF is adopted to identify the text paragraphs describing risks in the document.

[0031] In the risk classification stage, a hierarchical classification framework is adopted to classify the recognized risk content into six major categories: market risk, credit risk, liquidity risk, operational risk, legal risk, and compliance risk. Each major category has multiple sub-risk types. Classification uses a multi-label classification model based on XGBoost, which supports a risk description belonging to multiple risk categories simultaneously.

[0032] In the risk quantification stage, risk levels and quantification metrics are extracted through natural language processing techniques. For content with clearly marked risk levels (such as "high risk", "medium risk", etc.), they are directly extracted and standardized; for descriptive risk content, through sentiment analysis and keyword weight calculation, the severity of the risk is quantified into a numerical level from 1 to 5. At the same time, quantification metrics related to risks are extracted, such as the maximum loss ratio, value at risk, stress test results, etc. The risk category, as well as the risk level and quantification metrics, constitute the product risk information; Step 125: The product announcement information, the product description information, and the product risk information constitute the bank's unstructured data; Step 13: Based on the secure communication protocol, connect to the regulatory agency's protocol database to obtain regulatory filing and compliance review information; Specifically, obtaining regulatory filing and compliance review information includes the following steps: Step 131: Establish a secure data exchange channel with financial regulatory agencies in various countries; In the specific implementation process, the secure data exchange channel uses dedicated network connections and VPN technology to establish point-to-point secure channels with financial regulatory agencies in various countries to obtain regulatory filing and compliance review information from each financial regulatory agency in real time; Step 132: Through the secure data exchange channel, according to the preset data exchange protocol for regulatory requirements, regularly obtain regulatory filing data for over-the-counter derivatives transactions; Specifically, the obtained regulatory filing data includes: transaction report data, recording the details of over-the-counter derivatives transactions reported to the regulatory agency; product registration data, containing information on derivatives products registered with the regulatory agency; market participant qualification data, containing the regulatory qualifications and compliance status of trading entities; limit monitoring data, containing the position limits and usage of various derivatives; compliance inspection results, containing the compliance inspection findings of the regulatory agency on market participants; Step 133: Through the secure data exchange channel, obtain the compliance review information released by the regulatory agency, including compliance inspection results, violation penalty decisions, and regulatory policy interpretations; In the specific implementation process of the present invention, the compliance review information is obtained by regularly querying structured data related to compliance reviews. The query parameters include inspection objects, such as institution names, registration numbers, inspection times, inspection scopes, problems found, penalty decisions, such as penalty types, penalty amounts, rectification requirements, legal bases, and regulatory interpretations, etc.

[0033] Step 134: The regulatory review information includes the regulatory filing data and the compliance review information; Step 14: Classify the structured data of the exchange, the unstructured data of the bank, and the regulatory review information of the regulatory agency according to the name or code of the product, and save it as the original data set; It can be understood that this classification can be achieved by using the name or code of the product as the primary key of the database table. Through the secure data exchange channel, other information obtained from the compliance review information issued by the regulatory agency is used as the non-primary attributes of the database table; Furthermore, the semantic parsing of the original data set using a text model for financial texts to generate a transaction information table includes the following steps: Step 21: Construct a pre-trained language model dedicated to the financial field; Specifically, the pre-trained language model dedicated to the financial field adopts a bidirectional encoder representation variant based on the Transformer architecture and conducts domain adaptation pre-training for financial texts. This pre-trained language model consists of three main components: an encoder layer, an attention mechanism layer, and an output layer.

[0034] In the specific implementation process of the present invention, the encoder layer is stacked by 12 Transformer blocks. Each Transformer block contains a multi-head self-attention sub-layer and a feed-forward neural network sub-layer. The multi-head self-attention mechanism uses 8 attention heads, and the dimension of each attention head is 64, enabling the model to simultaneously focus on information in different positions. The feed-forward neural network adopts a two-layer structure, with a hidden layer dimension of 3072 and an activation function of GELU.

[0035] In the specific implementation process of the present invention, the pre-training of the pre-trained language model adopts a two-stage strategy: the first stage is to conduct pre-training on a general financial corpus, which includes financial news, research reports, and regulatory documents, etc.; the second stage is to conduct continued pre-training on over-the-counter derivative professional documents, including transaction documents, product manuals, and regulatory rules, etc. The pre-training tasks include two objectives: masked language model and next sentence prediction (NSP).

[0036] Step 22: Apply natural language processing technology to the bank's unstructured data to extract key transaction elements and clause information; Specifically, the extraction of key transaction elements and clause information adopts a method that combines named entity recognition and relation extraction.

[0037] Named entity recognition uses the BiLSTM-CRF model to identify financial entities in the text, including transaction parties, product names, transaction amounts, transaction dates, maturity dates, interest rate types, settlement methods, etc. The BiLSTM-CRF model consists of three layers: the embedding layer converts the input text into a vector representation; the BiLSTM layer captures context information and outputs a feature sequence; the CRF layer considers the dependencies between labels and outputs the optimal label sequence.

[0038] Relation extraction uses a graph convolutional network based on the attention mechanism to identify the relationships between entities, such as "transaction party - transaction product", "product - amount", "product - term", etc. The AGCN model first constructs a document-level entity relationship graph, then learns the representations of entity nodes through multi-layer graph convolutional operations, and finally aggregates node features through the attention mechanism to predict the relationship types between entities.

[0039] For more complex clause information, such as trigger conditions, early termination clauses, etc., a method combining rules and machine learning is used for extraction. The rule part utilizes the structural features and keyword patterns of the clauses; the machine learning part uses a sequence-to-sequence model to convert the clause text into a structured representation.

[0040] Step 23: Structurally transform the regulatory review information to extract compliance requirements and the basis for regulatory decisions; Specifically, the structural transformation of regulatory review information includes three steps: regulatory text classification, element extraction, and correlation analysis.

[0041] Regulatory text classification adopts a hierarchical classification framework, classifying regulatory texts into five categories: regulatory rules, compliance requirements, inspection findings, penalty decisions, and rectification suggestions. The classification model uses a text classifier based on XLNet, and through the permutation language modeling pre-training method, enhances the ability to understand long texts and complex contexts.

[0042] Element extraction uses a sequence labeling model based on BERT-CRF to extract key elements from regulatory texts, including regulatory clauses, violations, penalty types, penalty amounts, rectification deadlines, etc. The model input is a regulatory text paragraph, and the output is a sequence with BIO annotations, identifying the boundaries and types of various regulatory elements.

[0043] Correlation analysis establishes the logical relationships between regulatory elements, forming a complete regulatory chain of "violation - regulatory clause - penalty decision - rectification requirement". Correlation analysis uses a relationship reasoning model based on the graph attention network (GAT), taking the extracted regulatory elements as graph nodes, learning the association strength between nodes through multi-layer graph attention mechanisms, and outputting the relationship network of regulatory elements.

[0044] Step 24: Based on the pre-trained language model, construct a multi-modal fusion model for financial texts, integrating the semantic information of structured data, key transaction elements and clause information, as well as compliance requirements and regulatory decision-making bases; Specifically, the multi-modal fusion model adopts a cross-modal attention architecture based on Transformer to achieve semantic alignment and information fusion of different types of data.

[0045] The core components of the multi-modal fusion model include modality-specific encoders, cross-modal attention layers, and fusion representation layers. The modality-specific encoders design specialized encoding methods for different types of data: structured data generates an initial representation through field embedding and positional encoding; text data generates context-related representations through the pre-trained language model; tabular data generates representations considering row-column relationships through tabular Transformer.

[0046] The cross-modal attention layer adopts a bidirectional attention mechanism to calculate the correlation between different modality data. For example, calculate the attention weights between the product code in the transaction record and the product description text to achieve complementary enhancement of information. In the specific implementation process, for modality A and modality B, first calculate the attention scores of each element in modality A to all elements in modality B, and then weight-aggregate the information of modality B based on these scores as the enhanced representation of modality A.

[0047] The fusion representation layer adopts a gating mechanism to adaptively control the fusion ratio of different modality information. The gating unit receives the representations of each modality and the cross-modal enhanced representation as inputs, and outputs a weight vector between 0 and 1 to control the fusion ratio of the original representation and the enhanced representation. The final fusion representation is further optimized through residual connection and layer normalization to generate a unified representation containing multi-modal information.

[0048] Step 25: Perform in-depth information extraction on the fused semantic information, identify and extract the core elements and associated attributes of the transaction, and convert them into structured graph units; Specifically, the in-depth information extraction includes three sub-modules: entity recognition, relation extraction, and attribute filling. Entity recognition is based on the fusion representation, and adopts a method combining conditional random field (CRF) and deep neural network to identify key entities such as transaction parties, products, amounts, dates, etc. in the text. Relation extraction uses a relation classifier based on graph convolutional network to identify the semantic relations between entities, such as "transaction party - product", "product - amount", etc. Attribute filling supplements attribute information for the identified entities through a context-aware feature extractor, such as the registered location of the transaction party, the risk level of the product, etc. Organize the identified entities, relations, and attributes into a semantic graph structure. It can be understood that each group of entities, relations, and attributes constitutes a group of graph units; Step 26: Semantically integrate the graph units to construct a transaction knowledge graph, forming an association network among transaction entities; Specifically, the semantic integration uses a graph fusion algorithm to connect discrete graph units into a complete transaction knowledge graph. The graph fusion process includes two stages: entity alignment and relationship mapping.

[0049] In the entity alignment stage, a multi-feature fusion entity matching algorithm is used to identify nodes that refer to the same entity in different data sources. The matching algorithm comprehensively considers the character similarity of entity names, the distribution similarity of attribute values, and the topological similarity of structural contexts, and calculates the comprehensive similarity between entities. For entity pairs with a similarity exceeding a preset threshold, they are marked as different representations of the same entity and merged in subsequent processing.

[0050] In the relationship mapping stage, based on a predefined relationship ontology model, relationship types from different sources are mapped to a unified semantic space. The relationship ontology model adopts a three-layer architecture: the top layer contains core relationship types, such as "participate in transactions", "include products", "apply rules", etc.; the middle layer contains domain-specific relationships, such as "serve as a trading counterparty", "provide guarantee", etc.; the bottom layer contains specific business relationships, which directly correspond to the relationship expressions in the original data. The mapping process uses a relationship classifier based on a graph neural network to map the original relationships to the standard relationship types in the ontology model.

[0051] Step 27: Based on the transaction knowledge graph, perform multi-dimensional feature extraction to generate a structured representation of the transaction; Specifically, the multi-dimensional feature extraction includes three parts: entity feature extraction, relationship feature extraction, and path feature extraction.

[0052] In the entity feature extraction part, the transaction entity nodes in the knowledge graph are characterized. For transaction subject entities, the extracted features include subject types, such as banks, investment companies, hedge funds, etc., registered regions, asset sizes, credit ratings, historical transaction frequencies, and regulatory qualifications, etc.; for product entities, the extracted features include product categories, risk levels, liquidity indicators, price volatilities, and market shares, etc.; for transaction event entities, the extracted features include transaction amounts, transaction times, settlement methods, transaction purposes, and associated markets, etc. Feature extraction uses a representation learning method based on a graph attention network to fuse the attribute information and structural information of entities into a fixed-dimensional entity vector representation.

[0053] In the relational feature extraction part, the system analyzes the direct relational features between entities. The relational features include relation type, relation strength, relation duration, and relation frequency, etc. The relation strength is calculated based on the interaction frequency and transaction scale between associated entities; the relation duration is determined based on the time span of historical transaction records; and the relation frequency reflects the number of interactions within a specific time window. The system adopts a time-series aware relational embedding model to encode the relational features into relational vector representations.

[0054] In the path feature extraction part, the system analyzes the indirect association paths between entities. The system adopts a meta-path analysis method to pre-define a series of semantically clear path patterns, such as "trading party - product - trading party" which reflects the products traded in common, and "trading party - regulatory agency - trading party" which reflects being regulated by the same regulatory agency, etc. For each meta-path, the system calculates features such as the number of path instances, path diversity, and path importance. The path feature extraction adopts a method combining random walk and attention mechanism to generate path feature representations considering path semantics.

[0055] The entity vectors, relational vectors, and path feature representations serve as the structured representations of transactions; Step 28: Based on the structured representations of transactions, generate a standardized transaction information table, which includes key fields such as basic transaction information, participant information, product information, risk indicators, and compliance status, etc. Specifically, the generation of the transaction information table includes two stages: field design and data filling.

[0056] In the field design stage, based on the internationally common over-the-counter derivatives transaction reporting standards and local regulatory requirements, design the table structure of the transaction information table. In the specific implementation process of the present invention, the transaction information table includes five major categories of fields: basic transaction information fields, including the transaction unique identifier, transaction type, such as whether it is a cross-border transaction, transaction execution time, transaction effective date, transaction expiration date, and transaction status, etc.; participant information fields, including the legal entity identification codes, registered regions, institution types, and qualification statuses of both parties to the transaction; product information fields, including product type, underlying asset, notional amount, currency, and settlement method, etc.; risk indicator fields, including market risk value, credit risk score, liquidity risk level, and operational risk indicators, etc.; compliance status fields, including regulatory reporting status, compliance inspection results, and risk warning signs, etc.

[0057] In the data filling stage, first, obtain the field values in the structured representation of the transaction, including core elements such as transaction number, transaction time, transaction party information, transaction amount, transaction type, transaction status, etc. Then, according to the preset data filling rules, map the above field values to the corresponding positions in the transaction information table. For example, for the transaction party information, parse the detailed information of the transaction initiator and recipient, including name, address, contact information, account information, etc., and fill this information into the corresponding data fields. For the transaction amount, convert it to the standard format, and calculate the equivalent local currency amount according to the exchange rate information at the time of the transaction, and fill it into the amount-related fields. For the transaction type, according to the predefined type mapping table, convert the original transaction type to the standard type code used internally by the system, and fill it into the type field. For the transaction status, through the status mapping table, convert the original status to the system standard status code, and fill it into the status field. In addition, according to the time information of the transaction, the settlement date, clearing date, and other time points of the transaction can be automatically calculated and filled into the corresponding time fields. For the goods or service information included in the transaction, parse it into detailed items, each detailed item includes information such as product name, quantity, unit price, total price, etc., and fill it into the transaction details-related fields. Finally, according to the filled data, automatically calculate the values of some derived fields, such as total transaction amount, taxes, net amount, etc., and fill these calculation results into the corresponding fields; Further, the method for determining whether a cross-border transaction is involved according to the transaction type in the transaction information table is as follows: Read the field value of the transaction type of each product in the transaction information table. If the field value indicates a cross-border transaction, it is determined that a cross-border transaction is involved; otherwise, it is determined that no cross-border transaction is involved. Further, the steps for dynamically matching rules for the transaction information table based on a pre-constructed rule knowledge graph to generate a compliance report are as follows: Step 31: Pre-collect the regulatory rule texts of each financial regulatory agency, and build the rule knowledge graph based on the regulatory rule texts; Specifically, the building of the rule knowledge graph includes three stages: rule acquisition, rule parsing, and rule modeling.

[0058] In the rule acquisition stage, the system collects regulatory rule texts from multiple sources, including international agreements such as the Basel Accord, the ISDA Master Agreement, national regulations such as the Securities Law and the Regulations on the Administration of Futures Trading, regulatory agency regulations such as the regulations issued by the central bank, the securities regulatory commission, and the foreign exchange administration, and industry self-regulatory rules. Rule acquisition adopts multiple methods: direct docking with the structured rule library; regular crawling of the official websites of regulatory agencies; API calls to regulatory databases; OCR recognition and conversion of paper documents.

[0059] In the rule parsing stage, natural language processing is performed on the obtained rule text to extract the core elements of the rules. The parsing process adopts a method combining deep learning and rule templates. First, named entity recognition technology is applied to identify the subject entity, action entity, condition entity, and result entity in the rule text; then, relation extraction technology is applied to identify the semantic relations between entities, such as "apply to", "prohibit", "require", etc.; finally, semantic role labeling technology is applied to identify the condition part, action part, and result part of the rule. The parsing result forms a structured rule representation, including rule ID, rule source, release time, effective time, scope of application, condition expression, and result expression.

[0060] In the rule modeling stage, the structured rule representation is converted into a knowledge graph model. The rule knowledge graph adopts a multi-level ontology model: the top-level ontology defines core concept categories, such as regulatory subjects, regulatory objects, regulatory actions, and regulatory results; the domain ontology defines concepts in specific regulatory fields, such as cross-border transactions, foreign exchange management, anti-illegal fund transfer, etc.; the application ontology defines concepts in specific business scenarios, such as over-the-counter derivatives, structured products, etc. Rules are represented in the graph as complex relationship patterns connecting multiple entities. Each rule contains condition nodes, action nodes, and result nodes, as well as the logical relationships between them. The relationships between rules are also explicitly modeled, including inclusion relationships, inheritance relationships, conflict relationships, and supplementary relationships. This knowledge graph model is the rule knowledge graph; Step 32: Extract rule nodes and their associated relationships from the rule knowledge graph to construct a rule matching engine; Specifically, the rule matching engine includes three core components: a rule parser, a rule executor, and a result collector, which are used to perform dynamic rule matching on the data in the transaction information table.

[0061] Among them, the rule parser reads the rule nodes in the rule knowledge graph and converts each rule into an executable logical expression. For each rule, parse its trigger condition, scope of application, execution action, and priority to form a structured rule object.

[0062] Traverse the transaction information table. For each transaction record, call the rule executor to perform rule matching. The rule executor filters out the set of rules relevant to the current transaction record according to the scope of application of the rule. The rule executor evaluates the trigger conditions of each rule in order of the rule's priority. When a transaction record meets the trigger condition of a rule, the system marks the rule as "triggered" and records the specific reason and relevant data items for the trigger.

[0063] For each triggered rule, the system executes the actions defined by the rule, including but not limited to: compliance marking, risk level assessment, abnormal behavior identification, and regulatory requirement verification. The execution results are recorded in the result collector.

[0064] The result collector summarizes the results of all rule executions and generates preliminary compliance analysis data. The system aggregates the analysis data to identify compliance issues, risk points, and abnormal patterns in the transaction records.

[0065] Step 33: Generate a structured compliance report through the rule matching engine.

[0066] The compliance report at least includes the following main parts: compliance overview, risk assessment, details of abnormal transactions, compliance with regulatory requirements, and recommended measures. The compliance overview part presents the overall compliance status of transaction information through statistical analysis, including statistical data such as compliance rate, risk distribution, and main problem types. The risk assessment part details all identified risk points, including risk types, risk levels, risk descriptions, and indexes of relevant transaction records. The system displays the risks in a hierarchical manner according to the severity and impact scope of the risks; the details of abnormal transactions part lists all transaction records marked as abnormal, including transaction identifiers, abnormal types, triggered rule numbers, and abnormal descriptions. The system provides detailed information on abnormal transactions for further investigation and handling; the compliance with regulatory requirements part evaluates the compliance degree of transaction information with various regulatory requirements, including lists of compliance items, non-compliance items, and items to be confirmed. According to the regulatory rule nodes in the rule knowledge graph, the compliance rate of each regulatory requirement is calculated; the recommended measures part provides targeted improvement suggestions and compliance measures based on the identified compliance issues and risk points. Relevant solution nodes are extracted from the rule knowledge graph to generate an actionable list of recommendations.

[0067] Furthermore, the risk prediction of the transaction information table is performed using a pre-trained transaction graph neural network model to generate risk warning signals, and the corresponding warning measures include the following steps: Step 41: Convert the transaction information table into graph-structured data as the transaction graph network; The conversion of the transaction information table into graph-structured data can be in the form of backtracking to re-read the transaction knowledge graph corresponding to the transaction information table, and the graph structure of the transaction knowledge graph can be used as the transaction graph network; Step 42: Pre-build the transaction graph neural network model; the transaction graph neural network model adopts an architecture that combines a multi-layer graph convolutional network with a temporal attention mechanism. The transaction graph neural network model includes three components: a graph convolutional encoder, a temporal attention layer, and a risk decoder; In the specific implementation process of the present invention, the graph convolutional encoder consists of three layers of graph convolutional networks, with each layer containing 64 convolutional kernels. The first layer of graph convolution aggregates information from the first-order neighborhood of nodes to capture direct transaction relationships; the second layer of graph convolution extends to the second-order neighborhood to capture indirect association patterns; the third layer of graph convolution further expands the receptive field to identify complex network structure features. After each layer of graph convolution, batch normalization and the ReLU activation function are sequentially passed through to prevent overfitting and enhance the non-linear expression ability.

[0068] The temporal attention layer processes the time series features of transactions to capture the time evolution features of abnormal transaction patterns. The temporal attention layer adopts a self-attention mechanism to calculate the correlation weights of transaction features at different time points. In specific implementation, the transaction sequence in the past 12 months is used as the input, and query, key, and value matrices are generated through linear projection, and the attention scores are calculated and weighted aggregated to obtain the node representation considering time dependence.

[0069] The risk decoder adopts a multi-task learning framework to simultaneously predict the risk scores in four dimensions: market risk, credit risk, operational risk, and compliance risk. The risk decoder first maps the node-level risk representation to a hidden feature space through two layers of fully connected networks. The dimension of the hidden layer is 128, and the LeakyReLU activation function and Dropout regularization are adopted. Then, for each risk type, a dedicated prediction head is designed, which includes a fully connected layer and a Softmax activation function to output the five-level score probability distribution of this risk dimension. The market risk prediction head focuses on price volatility and liquidity indicators; the credit risk prediction head focuses on the default probability of the counterparty; the operational risk prediction head focuses on abnormal transaction processes and signs of system failures; the compliance risk prediction head focuses on violations of regulatory rules. The scores of each risk dimension adopt a grading system from 1 to 5, where level 1 indicates extremely low risk and level 5 indicates extremely high risk.

[0070] Step 43: Train the transaction graph neural network model through the pre-collected historical transaction data; Specifically, pre-collect the transaction history records in the past 12 months and use them as sample data for the transaction graph neural network model, and adopt a self-supervised learning method for pre-training during the training process of the transaction graph neural network model. The pre-training tasks include: node attribute prediction, randomly masking some attribute features of nodes, and training the model to predict these masked attributes; edge prediction, randomly removing some edges in the graph, and training the model to predict the existence of these edges; subgraph structure prediction, training the model to identify common risk-related subgraph patterns. Adam optimizer is used for pre-training, with an initial learning rate of 0.001 and a weight decay of 0.0001; Step 44: Input the transaction graph network into the pre-trained transaction graph neural network model to generate multi-dimensional risk scores; It is understandable that, based on the fact that the trading graph neural network model has been trained, the trading graph network is input into the trading graph neural network model, and the risk decoder of the trading graph neural network model can directly output risk scores in four different dimensions; Step 45: Integrate the multi-dimensional risk scores according to the preset risk weights to generate a comprehensive risk level; Specifically, the generation of the comprehensive risk level adopts a weighted fusion method, and the weights are dynamically adjusted according to the importance and relevance of different risk dimensions. First, calculate the weighted average of the four risk dimension scores, and the initial weights are set as 0.3 for market risk, 0.3 for credit risk, 0.2 for operational risk, and 0.2 for compliance risk.

[0071] In a further preferred embodiment, the synergistic effect of multi-dimensional risks can be further identified through a risk resonance detection algorithm. For example, when it is detected that two or more risk dimensions are simultaneously in a high-level state, the evaluation value of the comprehensive risk level can be increased. The risk resonance detection adopts a method combining rules and machine learning. The rule part defines the logical relationship between risk dimensions, and the machine learning part learns the risk cooperation pattern through historical cases. Finally, add the weighted average and the risk resonance adjustment value, and map the comprehensive risk level to different levels of risk levels. In the specific implementation process of the present invention, it can be set as extremely low risk (level 1), low risk (level 2), medium risk (level 3), high risk (level 4), and extremely high risk (level 5).

[0072] Step 46: Generate a risk warning signal according to the comprehensive risk level and the preset warning threshold; Specifically, the generation of the risk warning signal is based on a three-level warning mechanism. The system compares the comprehensive risk level with the preset warning threshold to generate a warning signal of the corresponding level. The warning levels include three types: reminder level, warning level, and emergency level. When the comprehensive risk level is level 3 (medium risk), the system generates a reminder-level warning, suggesting that the supervisor pay attention to the relevant transaction; when the comprehensive risk level is level 4 (high risk), the system generates a warning-level warning, suggesting that the supervisor conduct an in-depth review; when the comprehensive risk level is level 5 (extremely high risk), the system generates an emergency-level warning, suggesting that immediate regulatory measures be taken.

[0073] The above embodiments are only used to illustrate the technical method of the present invention and not to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical method of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical method of the present invention.

Claims

1. An intelligent adaptation system for cross-border supervision of over-the-counter derivatives based on heterogeneous data fusion, characterized by: It includes a data collection module, a data structured processing module, a transaction type classification module, a cross-border compliance assessment module, a risk prediction module, and a regulatory decision-making module; wherein each module is electrically connected; A data collection module is used to obtain OTC derivatives transaction data, including structured data from exchanges, unstructured data from banks, and regulatory review information from regulatory agencies, generate an original data set, and send the original data set to a data structuring processing module; A data structured processing module, which uses a text model for financial text to perform semantic analysis on the original data set, generates a transaction information table, and sends the transaction information table to a transaction type classification module; A transaction type classification module determines whether a cross-border transaction is involved according to the transaction type in the transaction information table. If a cross-border transaction is involved, the transaction information table is sent to the cross-border compliance assessment module; otherwise, the transaction information table is sent to the risk prediction module; The cross-border compliance assessment module dynamically matches the transaction information table based on the pre-built rule knowledge graph and generates a compliance report; The risk prediction module uses a pre-trained transaction graph neural network model to predict risks in the transaction information table, generate risk warning signals, and execute corresponding warning measures.

2. The intelligent adaptation system for cross-border supervision of over-the-counter derivatives based on heterogeneous data fusion according to claim 1 is characterized in that: Generating the original data set comprises the following steps: Step 11: Collect structured data from exchanges through multi-source heterogeneous interfaces, including transaction records, product information, and market data; Step 12: Use adaptive crawler technology to obtain unstructured bank data, including transaction announcements, product instructions, and risk disclosure documents; Step 13: Connect to the regulatory agency's protocol database based on the secure communication protocol to obtain regulatory filing and compliance review information; Step 14: Classify the structured data of the exchange, the unstructured data of the bank, and the regulatory review information of the regulatory agency according to the product name or product code, and save them as the original data set.

3. The intelligent adaptation system for cross-border supervision of over-the-counter derivatives based on heterogeneous data fusion according to claim 2 is characterized in that: The collecting of exchange structured data includes the following sub-steps: Step 111: Establish API connection channels with major global over-the-counter derivatives exchanges; Step 112: extracting transaction record data from the exchange system through the corresponding API interface according to the predefined data template; Step 113: Obtain product information of over-the-counter derivatives from the exchange system through the corresponding API interface; Step 114: Collect market data from the exchange system in real time through the corresponding API interface; Step 115: The transaction record data, product information of over-the-counter derivatives and market data constitute exchange structured data.

4. The intelligent adaptation system for cross-border supervision of over-the-counter derivatives based on heterogeneous data fusion according to claim 3 is characterized in that: The method of obtaining the bank's unstructured data includes the following steps: Step 121: Configure the intelligent crawler task to collect relevant documents of over-the-counter derivatives on the official websites of various financial institutions; Step 122: Identify and extract product announcement information of transaction announcements issued by the bank through intelligent crawler tasks; Step 123: Obtain and parse the product instructions issued by the bank through the intelligent crawler task, and extract product description information; Step 124: Collect risk disclosure documents issued by banks and extract product risk information through intelligent crawler tasks; Step 125: The product announcement information, the product description information and the product risk information constitute the bank's unstructured data.

5. The intelligent adaptation system for cross-border supervision of over-the-counter derivatives based on heterogeneous data fusion according to claim 4 is characterized in that: Obtaining regulatory filing and compliance review information involves the following steps: Step 131: Establish secure data exchange channels with financial regulators of various countries; Step 132: Regularly obtain regulatory filing data on OTC derivatives transactions through a secure data exchange channel in accordance with the data exchange protocol preset in accordance with regulatory requirements; Step 133: Obtain compliance review information issued by the regulatory authority through a secure data exchange channel; Step 134: The regulatory review information includes the regulatory filing data and the compliance review information.

6. The intelligent adaptation system for cross-border supervision of over-the-counter derivatives based on heterogeneous data fusion according to claim 5 is characterized in that: The semantic parsing of the original data set using a text model oriented to financial text to generate a transaction information table comprises the following steps: Step 21: construct a pre-trained language model dedicated to the financial field; the pre-trained language model dedicated to the financial field adopts a bidirectional encoder representation variant based on the Transformer architecture; Step 22: Apply natural language processing technology to the bank's unstructured data to extract key transaction elements and terms information; Step 23: Perform structured transformation on regulatory review information to extract compliance requirements and regulatory decision-making basis; Step 24: Based on the pre-trained language model, a multimodal fusion model of financial text is constructed to integrate structured data, key transaction elements and terms information, and semantic information of compliance requirements and regulatory decision-making basis; the multimodal fusion model adopts a Transformer-based cross-modal attention architecture; Step 25: Perform deep information extraction on the fused semantic information, identify and extract the core elements and associated attributes of the transaction, and convert them into structured graph units; Step 26: semantically integrate the graph units, construct a transaction knowledge graph, and form an association network between transaction entities; Step 27: Based on the transaction knowledge graph, perform multi-dimensional feature extraction to generate a structured representation of the transaction; Step 28: Based on the structured representation of the transaction, a standardized transaction information table is generated; the generation of the transaction information table includes two stages: field design and data filling.

7. The intelligent adaptation system for cross-border supervision of over-the-counter derivatives based on heterogeneous data fusion according to claim 6 is characterized in that: The method of judging whether a cross-border transaction is involved according to the transaction type in the transaction information table is as follows: The transaction type field value of each product in the transaction information table is read. If the field value indicates a cross-border transaction, it is determined to involve a cross-border transaction; otherwise, it is determined not to involve a cross-border transaction.

8. The intelligent adaptation system for cross-border supervision of over-the-counter derivatives based on heterogeneous data fusion according to claim 7 is characterized in that: The method of dynamically matching the transaction information table with rules based on the pre-built rule knowledge graph and generating a compliance report includes the following steps: Step 31: Collect regulatory rules texts of various financial regulatory agencies in advance, and construct the rule knowledge graph based on the regulatory rules texts; the construction of the rule knowledge graph includes three stages: rule acquisition, rule parsing, and rule modeling; Step 32: extracting rule nodes and their associations from the rule knowledge graph and constructing a rule matching engine; the rule matching engine includes three core components: a rule parser, a rule executor, and a result collector, and is used to dynamically match rules on the data in the transaction information table; Step 33: Generate a structured compliance report through the rule matching engine.

9. The intelligent adaptation system for cross-border supervision of over-the-counter derivatives based on heterogeneous data fusion according to claim 8 is characterized in that: The method of using a pre-trained transaction graph neural network model to predict risks on the transaction information table, generating risk warning signals, and executing corresponding warning measures includes the following steps: Step 41: Convert the transaction information table into graph structure data as a transaction graph network; Step 42: pre-build the transaction graph neural network model; the transaction graph neural network model includes three components: a graph convolution encoder, a temporal attention layer, and a risk decoder; Step 43: Train the transaction graph neural network model using the historical transaction data collected in advance; Step 44: Input the transaction graph network into the pre-trained transaction graph neural network model to generate a multi-dimensional risk score; Step 45: Integrate the multi-dimensional risk scores according to the preset risk weights to generate a comprehensive risk level; the comprehensive risk level is generated by using a weighted fusion method to dynamically adjust the weights according to the importance and relevance of different risk dimensions; Step 46: Generate a risk warning signal based on the comprehensive risk level and the preset warning threshold.

10. The intelligent adaptation system for cross-border supervision of over-the-counter derivatives based on heterogeneous data fusion according to claim 9 is characterized in that: The graph convolution encoder consists of a three-layer graph convolution network, each layer contains 64 convolution kernels; the first layer of graph convolution aggregates information on the first-order neighborhood of the node to capture direct transaction relationships; the second layer of graph convolution expands to the second-order neighborhood to capture indirect association patterns; the third layer of graph convolution further expands the receptive field and recognizes complex network structure features; each layer of graph convolution is followed by batch normalization and ReLU activation function; The temporal attention layer processes the time series characteristics of transactions and captures the time evolution characteristics of abnormal transaction patterns; the temporal attention layer uses a self-attention mechanism to calculate the correlation weights of transaction characteristics at different time points; The risk decoder adopts a multi-task learning framework to simultaneously predict risk scores in four dimensions: market risk, credit risk, operational risk, and compliance risk; The risk decoder first maps the node-level risk representation to the hidden feature space through a two-layer fully connected network with a hidden layer dimension of 128, using the LeakyReLU activation function and Dropout regularization; then, for each risk type, a corresponding prediction head is designed, which includes a fully connected layer and a Softmax activation function, and outputs the five-level score probability distribution of the risk dimension.

Images