A Method and System for Generating Anti-Money Laundering Risk Rules Based on Multimodal Large Language Models

Abstract

This application discloses a method and system for generating anti-money laundering risk rules based on a multimodal large language model. The method includes the following steps: obtaining an initial training set composed of transaction data; attaching money laundering labels based on a rule model; then extracting and expanding transaction features using a multimodal large language feature extraction model; forming a modeling training set from the initial training set and the expanded dataset; fitting a multimodal large language learning prediction model to the modeling training set to obtain the money laundering risk value of the transaction data, and calculating the actual contribution value of structured features; inputting a validation set into the multimodal large language learning prediction model to verify the money laundering label redetermined result, updating it to the initial training set; and then generating an updated rule model based on the validation result and the contribution value of structured features to the money laundering risk value. This application can integrate multimodal transaction data to form a money laundering transaction risk prediction model, and adjust and update it in real time, accurately identifying money laundering transaction behaviors.

Description

Technical Field

[0001] This application relates to the field of anti-money laundering risk identification, specifically to an anti-money laundering risk rule generation method and system based on a multimodal large language model. Background Technology

[0002] Currently, most anti-money laundering detection systems for transaction business rely on three methods: 1. Manual identification by business experts; 2. Judgment of anti-money laundering transaction behavior based on existing rule models; 3. Classification of anti-money laundering behavior based on existing machine learning and deep learning prediction models.

[0003] First, regarding manual identification by business experts, due to limited resources and high costs, they can generally only supplement the second or third detection methods when faced with modern large-scale transaction data, reviewing their outputs. Second, for anti-money laundering transaction behavior judgment based on existing rule models, these models are highly static and cannot be dynamically adjusted with changes in economic conditions over time, making it difficult to cope with new and evolving money laundering methods. Third, while machine learning and deep learning prediction models are more adaptable than rule-based methods, their outputs lack interpretability, failing to provide specific reasons for the identification of money laundering cases, and offering no way to adjust inaccurate results. Furthermore, limited by the amount of transaction sample information acquired by machine learning, its accuracy in identifying money laundering cases is also lower than that of manual review. Meanwhile, as criminals adopt increasingly sophisticated and covert money laundering methods, anti-money laundering mechanisms also need to be constantly updated and adjusted to cope with these changes. For the second or third detection method, if it is necessary to dynamically adapt to new money laundering methods, it is necessary to study, adjust, update and maintain the knowledge graph and judgment rules. This process requires continuous investment of expert resources and time costs, which is costly and has low timeliness. Summary of the Invention

[0004] The purpose of this application is to overcome the shortcomings of the prior art and provide a method and system for generating anti-money laundering risk rules based on a multimodal large language model. This method can integrate structured and unstructured transaction data to form a risk prediction model for money laundering transactions and adjust and update the rule model in real time, thereby accurately identifying money laundering transactions.

[0005] Firstly, this application provides a method for generating anti-money laundering risk rules based on a multimodal large language model, which adopts the following technical solution:

[0006] An initial training set consisting of transaction data is obtained. Based on the initial rule model, money laundering labels are added to the transaction data in the initial training set. Then, the transaction features of the transaction data in the initial training set are extracted by the multimodal large language feature extraction model, and the transaction features are expanded to form an expanded dataset. The initial training set and the expanded dataset form a modeling training set for modeling.

[0007] By modeling the training set, a multimodal large language learning prediction model is fitted. The transaction data of the modeling training set is input, and the money laundering label corresponding to the transaction data of the modeling training set is output to obtain the money laundering risk value of the transaction data. The Shapley value of the transaction feature in the transaction data is calculated. The Shapley value is the contribution value of a specific transaction feature to the money laundering risk value.

[0008] The validation set is used as input to the multimodal large language learning prediction model to obtain a new validation set for money laundering label determination. This validation set is then updated to the initial training set. Finally, an updated rule model is generated based on the validation results and the contribution of structured features to the money laundering risk value, thus updating the initial rule model.

[0009] The above technical solution first uses an initial rule model to predict money laundering risks on the initial training set. Then, it extracts multimodal features from transaction data using multimodal large-scale language feature extraction and expands the data. This data is then input into a multimodal large-scale language learning prediction model to obtain the correlation between multimodal features and money laundering risks. Finally, a validation set is used to supplement the training set and correct the rule model. This technical solution obtains multimodal data from transaction data through a multimodal large-scale language learning prediction model and maps it to money laundering risks for risk prediction. Compared to risk assessment using a single rule model, this approach has stronger identification capabilities and accuracy, and can handle complex money laundering methods. By integrating multimodal features and expanding transaction data, the sample size is increased, reducing the impact of sample imbalance and improving the model's recognition ability. The validation set supplements the initial training set and corrects the rule model, continuously optimizing the model and enhancing the system's real-time performance and adaptability.

[0010] As a preferred method, the specific approach for establishing the modeling training set is as follows:

[0011] Establish the initial training set: For a specific period, obtain the real transaction data of that period as the initial real transaction dataset. The real transaction data includes structured data and unstructured data. Then, obtain historical suspicious transaction samples as the initial past typical training set. Integrate the real transaction dataset and the past typical training set to form the initial training set.

[0012] Initial label judgment: In the initial training set, the transaction data in the historical suspicious transaction samples are inherently labeled as real money laundering; for the real transaction dataset, the transaction data in the real transaction data is judged based on the initial rule model, and money laundering labels are attached to risky transactions.

[0013] Data integration: Structured and unstructured data from the initial training set are input into the multimodal large language feature extraction model to generate multimodal transaction features;

[0014] Data augmentation: Based on the multimodal transaction features generated during data integration, data augmentation is performed on transaction data with and without money laundering labels to form an augmented dataset;

[0015] Forming the modeling training set: The initial training set and the expanded dataset are combined to form the modeling training set used for modeling.

[0016] By using the above technical solution, combining real transaction data and historical suspicious transaction samples as the initial training set, and then adding money laundering labels to the initial training set, data integration and expansion can be carried out. Under the premise of clearly knowing whether the target data item has a money laundering label, the data can be expanded based on its structured and unstructured transaction characteristics. The expanded data is highly targeted, closer to human cognition, and can achieve effective complementarity of information dimensions.

[0017] As a preferred approach, the multimodal transaction features generated during the data integration step are categorized into: basic transaction information features, transaction type features, unstructured data extraction features, and large language model inductive features. Specifically, basic transaction information features are transaction information and data extracted from structured data; transaction type features are predefined transaction categories extracted from structured data; unstructured data extraction features are features extracted from unstructured data and described in textual form using a multimodal large language feature extraction model; and large language model inductive features are features extracted from both structured and unstructured data using a multimodal large language feature extraction model and then summarized to form a feature description.

[0018] As a preferred approach, when integrating and generating large language models to summarize features, preset prompt words are used to guide the multimodal large language feature extraction model, summarizing multimodal transaction features within preset dimensions to form feature descriptions.

[0019] Through the above technical solution, the basic information features and transaction type features of the transaction are the features extracted by the traditional feature extraction model. On this basis, the unstructured features are extracted by the multimodal big language feature extraction model and the inductive features of the big language model are extracted. This enables feature extraction of multimodal transaction data, making full use of the multi-source heterogeneous data generated in actual business and realizing the collaborative analysis of multi-dimensional features.

[0020] As a preferred approach, in the data augmentation step, the augmented data consists of transaction basic information features, transaction type features, unstructured extraction features, and large language model inductive features generated based on the data integration step, forming an augmented dataset. Based on the results of the initialization label judgment step, transaction data with money laundering labels are taken as positive examples, and transaction data without money laundering labels are taken as negative examples. The augmentation results of positive examples remain positive examples, and the augmentation results of negative examples remain negative examples.

[0021] The above technical solutions enable bidirectional expansion of structured and unstructured features, enhance the expressive power of the dataset through cross-modal association, and supplement scarce samples to support the modeling needs of detecting complex money laundering scenarios.

[0022] As a preferred method, the specific approach for verification using a validation set is as follows:

[0023] First, the validation set is judged based on the initial rule model, and money laundering labels are attached to risky transactions to form a control set;

[0024] Then, the transaction data of the validation set is input into the multimodal big language learning prediction model to obtain the money laundering risk value of the transaction data of the validation set. For transaction data with a money laundering risk value exceeding the first threshold, a money laundering label is added as a validation experimental set. Transaction data with a money laundering risk value exceeding the second threshold is regarded as a high-risk transaction dataset. The second threshold is higher than the first threshold.

[0025] The validation set is compared with the control set. For transaction data whose money laundering labels are inconsistent between the validation set and the control set, the reviewers will manually confirm or reject whether the transaction data is money laundering data. The transaction data that is finally determined to have money laundering labels and high-risk money laundering transaction data will be updated to the initial training set.

[0026] As a preferred method, the specific methods for generating updated rule models include:

[0027] 1) For transaction data where the money laundering labels of the experimental set and the control set are inconsistent, and the audit results of the auditors are inconsistent with the initial rule model, the initial rule model shall be adjusted to improve the adaptability of the initial rule model to the data of the current period.

[0028] 2) Continuously calculate the Shapley value of each structured transaction feature, rank them, and adjust or eliminate the weights of structured features whose ranking of Shapley value does not reach a certain threshold within the specified evaluation period in the initial rule model.

[0029] 3) For transaction features whose rankings in the Shapley ranking system reach a certain threshold within the specified evaluation period, if they are not in the initial rule model, a new rule will be created for that transaction feature and added to the initial rule model.

[0030] As a preferred option, before an updated rule model is launched, it is tested and verified against the existing rule model. If the updated rule model performs consistently and has a higher accuracy rate than the existing rule model within a preset testing period, then the updated rule model is adopted.

[0031] Through the above technical solution, the accuracy of the money laundering risk prediction of the multimodal big data language learning prediction model fitted to the training set formed by the money laundering labels determined by the rule model and the extracted structured and unstructured features needs to be verified. By introducing a validation set, the money laundering labels determined by the rule model are cross-validated with the money laundering risk prediction output by the multimodal big data language learning prediction model. Through manual review by auditors, the cross-validation of the rule model and the multimodal big data language learning prediction model is achieved, and the two are supplemented and corrected, thereby improving the accuracy of money laundering risk prediction and enabling the prediction model to be updated and adjusted according to the latest case data.

[0032] As a preferred option, for transaction data with attached money laundering labels, a report is provided based on a multimodal large language feature extraction model to identify the transaction data. The report includes the identification criteria, transaction characteristics that contribute significantly to the money laundering risk value, and warning information.

[0033] The above technical solutions enable the predicted risk warning transactions to output interpretable anti-money laundering reports and warning information based on their multimodal characteristics, thereby improving the transparency and credibility of the predictions and providing reference and convenience for subsequent operation and maintenance personnel to identify and conduct secondary review.

[0034] Secondly, this application provides an anti-money laundering risk rule generation system based on a multimodal large language model, comprising: a data collection module, a rule model discrimination module, a multimodal feature extraction module, a data expansion module, a prediction model module, and a verification module;

[0035] The data collection module gathers real transaction data and historical suspicious samples to form an initial training set;

[0036] The rule model discrimination module uses a rule model to assess the money laundering risk of transaction data on the initial training set, performs money laundering discrimination on the input transaction data, and adds money laundering labels to high-risk transaction data;

[0037] The multimodal feature extraction module extracts multimodal transaction features from the transaction data;

[0038] The data augmentation module augments data based on multimodal transaction features and combines the augmented dataset with the initial training set to form a modeling training set.

[0039] The prediction model module is based on the modeling training set, fits a multimodal large language learning prediction model, and performs a secondary assessment of money laundering risk on the input transaction data.

[0040] The validation module uses this validation set to validate the multimodal large language learning prediction model. Transaction data and high-risk transaction data that have been manually reviewed by auditors are added to the data collection module, and the rule model in the rule model discrimination module is corrected.

[0041] In summary, this application includes at least one of the following beneficial technical effects:

[0042] 1. This application can integrate multi-source heterogeneous data in transaction data, introduce semantic analysis and feature generation mechanisms, so that the structured features and unstructured features of transaction data are mapped with money laundering risk labels, enabling the model to more accurately identify potential money laundering behavior, reduce false alarm rate and improve accuracy.

[0043] 2. This application uses a multimodal large language feature extraction model to perform data generative augmentation on transaction data based on structured and unstructured features, which effectively reduces the impact of sample imbalance, improves the model's ability to identify minority samples (such as suspicious transactions), and reduces the risk misjudgment rate; compared with simply augmenting structured features, it improves the model's generalization ability and robustness.

[0044] 3. This application can continuously optimize the rules and parameters of the model through supplementary verification of the verification set and human-machine collaborative feedback, so that it can adapt to new money laundering methods and enhance the real-time performance and adaptability of the system.

[0045] 4. This application can output explanatory doubt reports and risk warnings, making the decision-making process more visible, improving the transparency and credibility of predictions, ensuring that decisions meet the requirements of industry regulatory standards, and improving the efficiency of professionals in reviewing and modifying prediction results. Attached Figure Description

[0046] Figure 1 This is a flowchart illustrating the method for generating anti-money laundering risk rules based on a multimodal large language model in an embodiment of this application.

[0047] Figure 2 This is a flowchart illustrating step S100 of the anti-money laundering risk rule generation method based on a multimodal large language model in an embodiment of this application.

[0048] Figure 3This is a flowchart illustrating step S300 of the anti-money laundering risk rule generation method based on a multimodal large language model in an embodiment of this application.

[0049] Figure 4 This is a schematic diagram illustrating the risk identification process for transaction datasets;

[0050] Figure 5 This is a schematic diagram of the architecture of the anti-money laundering risk rule generation system based on a multimodal large language model in an embodiment of this application. Detailed Implementation

[0051] This specific embodiment is merely an explanation of this application and is not intended to limit it. After reading this specification, those skilled in the art can make modifications to this embodiment without contributing any inventive step, but such modifications are protected by patent law as long as they are within the scope of this application.

[0052] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application. It should be noted that in the optional embodiments of this application, the object information and other related data involved require the permission or consent of the object when the embodiments of this application are applied to specific products or technologies, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. That is to say, if the embodiments of this application involve data related to the object, it needs to be obtained with the authorization and consent of the object, the authorization and consent of the relevant departments, and in compliance with the relevant laws, regulations, and standards of the country and region. If personal information is involved in the embodiments, the acquisition of all personal information requires the consent of the individual. If sensitive information is involved, the separate consent of the information subject is required, and the embodiments also need to be implemented with the authorization and consent of the object.

[0053] The embodiments of this application will now be described in further detail with reference to the accompanying drawings.

[0054] The anti-money laundering technology field has begun to utilize technologies such as artificial intelligence, machine learning, and knowledge graphs to develop efficient detection, risk assessment, and early warning systems. These systems generate predictive models by fitting feature data with anomaly results. During operation, they rely on pre-set rule bases and static thresholds, comparing real-time transaction data with patterns in established rules to assess account risk, monitor suspicious transactions, and achieve a degree of automated early warning. However, this type of technological architecture has significant limitations:

[0055] First, the data used for predictive model learning is mostly structured data, neglecting the rich information contained in unstructured data (such as transaction descriptions and notes). The model's ability to collaboratively analyze multi-dimensional features is weak; for example, it cannot effectively correlate cross-domain features such as abnormal device fingerprints, irregular transaction times, and dispersed fund flows, making it difficult to identify carefully designed complex money laundering behaviors. Furthermore, the model cannot understand the semantic relationships of transaction behaviors, such as logical relationships like "logging into multiple accounts on the same device in a short period," relying solely on numerical feature matching and lacking semantic analysis and dynamic feature generation capabilities. Second, the predictive model mainly relies on the statistical patterns and rule bases of historical data, unable to autonomously reason about the logical relationships between feature elements in new money laundering methods. When criminal methods innovate, the system fails due to a lack of generalization ability. Simultaneously, rule base updates lag significantly behind the evolution of criminal methods. After regulatory authorities report new cases, banks need to manually analyze and extract features and adjust rules, a process that often takes weeks, causing the system to completely lose its early warning capabilities in the early stages of evolving criminal methods. These issues are the goals of this application to improve and enhance the system.

[0056] In one embodiment of this application, please refer to Figure 1 A method for generating anti-money laundering risk rules based on a multimodal large language model includes the following steps:

[0057] S100: Obtain an initial training set consisting of transaction data. Based on the initial rule model, attach money laundering labels to the transaction data in the initial training set. Then, use a multimodal large language feature extraction model to extract transaction features from the transaction data in the initial training set. The transaction features include structured features and unstructured features. Then, expand the transaction features to form an expanded dataset. The initial training set and the expanded dataset form a modeling training set for modeling.

[0058] S200, through modeling the training set, uses structured and unstructured features as independent variables and money laundering labels on transaction data as dependent variables to construct a multimodal large language learning prediction model to obtain the money laundering risk value of transaction data.

[0059] In this embodiment, a pre-trained multimodal big data feature extraction model is first used to construct a corresponding multimodal data feature vector. This feature vector is then merged with the original structured feature vector of the data to form a multimodal structured feature vector, which is then mapped to whether the transaction data carries a money laundering label. The multimodal big data learning prediction model outputs the probability that the transaction data ultimately carries a money laundering label based on its transaction features, using the probability level as the money laundering risk value. Simultaneously, the Shapley value of the structured features in the transaction data is calculated. This Shapley value represents the contribution of a specific structured feature vector to determining the probability of whether the transaction data carries a money laundering label, i.e., its contribution to the money laundering risk value. In this step, the multimodal big data learning prediction model is fitted to integrate the multimodal data into the transaction data money laundering label identification model, and to provide feedback and secondary confirmation to the rule model in S100.

[0060] S300 uses a validation set as input to the multimodal large language learning prediction model to validate and obtain a new validation set for money laundering label determination. This validation set is then updated to the initial training set. Next, based on the validation results and the contribution of structured features to the money laundering risk value, an updated rule model is generated, updating the initial rule model. The validation set, in addition to cross-validating the accuracy of the multimodal large language learning prediction model and the rule model, also uses transaction data from the validation set to supplement the transaction data in the modeling dataset, revising and supplementing the rules in the rule model, further adjusting the modeling dataset and the multimodal large language learning prediction model. Specifically, the data in the validation set can be real transaction data from the initial training set, or separately prepared validation data with clearly defined money laundering labels.

[0061] Specifically, in one embodiment, please refer to Figure 2 The specific steps for S100 are as follows:

[0062] S110, Establish the initial training set: For a specific period, obtain the real transaction data of that period as the initial real transaction dataset. The real transaction data includes structured data and unstructured data. Then, obtain historical suspicious transaction samples as the initial past typical training set. Integrate the real transaction dataset and the past typical training set to form the initial training set.

[0063] In S110, structured data for real transactions includes basic account information, transaction records, asset and liability information, and payment information. Unstructured data includes contracts, signed documents, scanned copies or photos of ID cards or business licenses, electronic records of transfer receipts or remittance slips, unstructured data extracted from the SWIFT messaging system (such as transaction descriptions and remarks), transaction IP addresses and device fingerprints, offline branch location data, transaction notes, transaction descriptions, and customer behavior data. Structured and unstructured data are linked to the customer ID. Historical suspicious transaction samples are obtained from the user's own database or publicly available information repositories (such as the central bank's anti-money laundering case database) and are considered money laundering cases. Depending on the source of the data, these samples may contain missing structured or unstructured data.

[0064] S120, Initialize label judgment: In the initial training set, the transaction data in the historical suspicious transaction samples are inherently labeled as real money laundering; for the real transaction dataset, the transaction data in the real transaction data is judged based on the initial rule model, and money laundering labels are attached to risky transactions.

[0065] In S120, since the transaction data in the historical suspicious transaction samples has already been pre-identified as money laundering cases, no further judgment is needed before directly attaching a money laundering label. For the real transaction dataset, a preliminary money laundering risk assessment is first performed using a pre-defined initial rule model. If a money laundering risk is identified, a money laundering label is attached. This initial rule model is an existing discrimination model based on the structured features of the transaction data. For example, transactions with a single account exceeding a threshold frequency within a unit period are identified as having money laundering risk and are labeled with a money laundering label. In subsequent updates, this initial rule model will be modified or replaced by newly added or revised rule models through continuous iterative verification, generating an updated rule model.

[0066] S130, Data Integration: Input the structured and unstructured data from the initial training set into the multimodal large language feature extraction model to generate structured and unstructured transaction features.

[0067] In S130, the integrated structured and unstructured transaction features are categorized into: basic transaction information features, transaction type features, unstructured extracted features, and large language model inductive features. Basic transaction information features are transaction information and data extracted from structured data; transaction type features are predefined transaction categories extracted from structured data, such as payment settlement, salary disbursement, tax payment, fund collection, transfer settlement, and investment conversion; unstructured data extracted features are features extracted from unstructured data and described in textual form using a multimodal large language feature extraction model, such as semantic features extracted from transaction notes and descriptions, and features related to the company's main business, transaction IP address, device fingerprint, and offline branch location data extracted from business license images; large language model inductive features are features extracted from both structured and unstructured data using a multimodal large language feature extraction model and then summarized to form a feature description.

[0068] More specifically, in the above embodiments, when integrating and generating features from a large language model, a built-in Prompt (preset prompt word) is used to guide the multimodal large language feature extraction model. This guidance is based on anti-money laundering strategies and uses question-based guidance on transaction behavior and patterns, enabling the multimodal large language feature extraction model to summarize transaction features within preset dimensions and form feature descriptions. For example, the multimodal large language feature extraction model may be asked whether a single account receives a large number of irregular small transfers and then transfers them out to a fixed account, or whether a large amount of funds is transferred in at once and then split into multiple small transfers to different accounts; whether funds are quickly transferred out after arriving in the account, with a very short duration and a very low account balance; whether there are frequent fund transfers through different banks and payment platforms; whether corporate accounts frequently transfer funds to personal accounts, especially involving transactions unrelated to the main business listed on the business license; whether long-unused accounts are frequently used for transactions in a short period of time; whether multiple accounts use the same IP address, or whether the real location is hidden through overseas proxy servers, etc. By using a multimodal large language feature extraction model, based on structured and unstructured data and guided by a Prompt, multimodal features are extracted and summarized to form large language model inductive features.

[0069] S140, Data Augmentation: Based on the structured and unstructured transaction features generated during data integration, data augmentation is performed on transaction data with and without money laundering labels to form an augmented dataset.

[0070] In S140, the expanded dataset consists of transaction basic information features, transaction type features, unstructured extraction features, and large language model inductive features generated based on the data integration step. Based on the results of the initial label judgment step, transaction data with a money laundering label is used as positive examples, and transaction data without a money laundering label is used as negative examples. The expanded results of positive examples remain positive, and the expanded results of negative examples remain negative. Due to the scarcity of positive examples in anti-money laundering scenarios, the actual data expansion focuses on expanding positive examples to increase the proportion of positive examples in the total dataset. Through a multimodal large language feature extraction model, structured and unstructured features can be expanded separately. For example, the large language model identifies the multi-layered money laundering chain of "gambling and fraud funds - jewelry purchases - overseas investment" as an unstructured feature, and then expands "jewelry purchases" to "luxury goods purchases," increasing the number of relevant positive examples.

[0071] S150, Forming the modeling training set: The initial training set and the expanded dataset are combined to form the modeling training set used for modeling.

[0072] Specifically, in one embodiment, the construction process of the multimodal large language learning prediction model in S200 can be as follows:

[0073] S210 utilizes a pre-trained multimodal large language feature extraction model, such as the iFlytek Xinghuo text vectorization model in this implementation, to generate a structured numerical feature vector from the structured and unstructured data portions of a single data point. This structured numerical feature vector is then merged with the original structured feature vector of the data to form a multimodal structured feature vector.

[0074] S220 uses multimodal structured feature vectors as independent variables and money laundering labels as dependent variables. It employs neural network training models, such as logistic regression and XGBoost models, to obtain a multimodal big language learning prediction model capable of predicting money laundering risk values.

[0075] Specifically, in one embodiment, please refer to Figure 3 S300, the specific method for verification using a validation set is as follows:

[0076] S310 involves judging the validation set based on the initial rule model, attaching money laundering labels to risky transactions, and forming a control set.

[0077] In step S320, the transaction data from the validation set is input into the multimodal large language learning prediction model in step S200 to obtain the money laundering risk value of the validation set transaction data. Transaction data with a money laundering risk value exceeding a first threshold are labeled with a money laundering tag and used as the validation experimental set. This process is essentially the process of using the multimodal large language learning prediction model to determine the money laundering tag based on the multimodal features of the transaction data. Transaction data with a money laundering risk value exceeding a second threshold are designated as high-risk transaction data; the second threshold is higher than the first threshold.

[0078] S330 compares the validation experimental set with the control set. For transaction data where the money laundering labels in the validation experimental set and the control set are inconsistent, manual reviewers confirm or reject whether the transaction data is money laundering data. This part of the data represents the conflicting judgments between the rule-based model and the multimodal large language learning prediction model. The final determination through manual review has significant educational value for both the rule-based model and the multimodal large language learning prediction model, and therefore needs to be added to the existing typical training set. Additionally, the high-risk money laundering transaction dataset contains transaction data with significantly higher money laundering risk values. Its transaction features have a high degree of fit with the money laundering labels, resulting in high accuracy. Therefore, it is also suitable for adding to the existing typical training set.

[0079] More specifically, in another embodiment, the previous typical training set is divided into several periodic training sets that are distinguished by a fixed period. The previous typical training set output by the periodically updated validation set is entered into the corresponding periodic training set. After entering a new period, the weights of each periodic training set relative to the initial training set are fine-tuned to ensure that the model adapts to recent new money laundering behaviors.

[0080] More specifically, in another embodiment, in S300, the method for generating the updated rule model specifically includes:

[0081] 1) For transaction data where the money laundering labels of the experimental set and the control set are inconsistent, and the audit results of the auditors are inconsistent with the initial rule model, the initial rule model shall be adjusted to improve the adaptability of the initial rule model to the data of the current period.

[0082] 2) Continuously calculate the Shapley value of each structured transaction feature, rank them, and adjust or eliminate the corresponding rules in the initial rule model for structured features whose Shapley value ranking does not reach a certain threshold within the specified evaluation period.

[0083] 3) For transaction features whose rankings in the Shapley ranking system reach a certain threshold within the specified evaluation period, if there is no corresponding rule in the initial rule model, a new rule will be created for that transaction feature and added to the initial rule model.

[0084] For updated rule models, A / B testing is required before deployment. The updated rule model is compared with existing rule models. If it performs consistently across different datasets within a preset testing period and its accuracy exceeds that of existing rule models, the updated rule model is adopted; otherwise, the rule model is not updated. In another embodiment, if existing rules consistently fail to identify anti-money laundering cases with new data, new money laundering transaction identification rules are directly constructed based on the unidentifiable new samples, building upon the existing rules.

[0085] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0086] For practical applications of risk identification using transaction datasets, please refer to [link / reference needed]. Figure 4 First, the transaction dataset is preprocessed to obtain its multimodal features. Then, the latest rule-based model performs an initial money laundering risk identification on the transaction dataset. Next, the transaction dataset is input into a multimodal big data language learning prediction model for a secondary money laundering risk identification. Finally, a human-machine collaborative review of the differences between the two identifications determines the final output. For transaction data with added money laundering labels, a report is provided based on the multimodal big data language feature extraction model to classify the transaction data, including the basis, structured features that contribute significantly to the money laundering risk value, and warnings. This report provides interpretable anti-money laundering judgment criteria and warning information, improving the transparency and credibility of the prediction, and providing reference and convenience for subsequent identification and secondary review by maintenance personnel. At the same time, this part of the transaction dataset is fed back to the real transaction dataset to supplement the initial training set to adapt to the current period's data and also to adaptively adjust the rule-based model.

[0087] In another embodiment, please refer to Figure 5 This application discloses an anti-money laundering risk rule generation system based on a multimodal large language model, comprising: a data collection module 1, a rule model discrimination module 2, a multimodal feature extraction module 3, a data expansion module 4, a prediction model module 5, and a verification module 6.

[0088] Data collection module 1 collects real transaction data and historical suspicious samples to form an initial training set.

[0089] Rule model discrimination module 2 uses a rule model to assess the money laundering risk of transaction data in the initial training set and adds money laundering labels to high-risk transaction data.

[0090] The multimodal feature extraction module 3 extracts multimodal transaction features from the transaction data.

[0091] Data augmentation module 4 augments data based on multimodal transaction features and combines the augmented dataset with the initial training set to form a modeling training set.

[0092] Prediction model module 5 uses a modeling training set to fit a multimodal large language learning prediction model, and performs a secondary assessment of money laundering risk on the input transaction data.

[0093] The verification module 6 uses this verification set to verify the multimodal large language learning prediction model. The transaction data and high-risk transaction data that have been manually reviewed by the auditors in the supplementary verification set are entered into the data collection module, and the rule model in the rule model discrimination module is corrected.

[0094] In another embodiment, the anti-money laundering risk rule generation system based on a multimodal large language model also includes a report generation module 7, which outputs and provides an analysis report on the transaction data that has been manually reviewed by auditors in the verification module 6.

[0095] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the anti-money laundering risk rule generation system based on the multimodal large language model described above can be referred to the corresponding process in the aforementioned method embodiments, and will not be repeated here.

[0096] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0097] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for generating anti-money laundering risk rules based on a multi-modal large language model, characterized in that, The method comprises the following steps: S100, obtaining an initial training set composed of transaction data, and labeling the transaction data in the initial training set based on an initial rule model; The initial rule model is a model for distinguishing based on the structured features of the transaction data; Then, transaction features of the transaction data in the initial training set are extracted by a multi-modal large language feature extraction model, and the transaction features are expanded to form an expanded data set, wherein the transaction features include structured features and unstructured features; The modeling training set is formed by the initial training set and the expanded data set; S200, constructing a multi-modal data feature vector corresponding to the transaction data by the modeling training set, merging the multi-modal data feature vector with the original structured feature vector of the transaction data to form a multi-modal structured feature vector, and constructing a multi-modal large language learning prediction model with the multi-modal structured feature vector as the independent variable and the money laundering label of the transaction data as the dependent variable, and outputting the money laundering label corresponding to the transaction data of the modeling training set by the multi-modal large language learning prediction model to obtain a money laundering risk value; At the same time, the Shapley value of the structured feature in the transaction data is calculated, and the Shapley value is the contribution value of the structured feature to the money laundering risk value; S300, setting a validation set, discriminating the transaction data of the validation set based on the initial rule model, and labeling the transaction data to form a control set; Then, the transaction data of the validation set is input into the multi-modal large language learning prediction model to obtain the money laundering risk value of the transaction data of the validation set, and the transaction data with a money laundering risk value exceeding a first threshold is labeled as a validation experiment set; The validation experiment set and the control set are compared, and for the transaction data whose money laundering labels of the validation experiment set and the control set are inconsistent, whether the transaction data is money laundering data is confirmed or rejected by an auditor, and the transaction data finally determined to have a money laundering label is supplemented into the past typical training set of the initial training set of the rule model; Then, the Shapley values of each structured transaction feature are continuously calculated and ranked, the corresponding rules of the structured features that do not reach a ranking threshold in a specified evaluation period are adjusted or eliminated in the initial rule model, and if the structured features that do not reach the ranking threshold do not have corresponding rules in the initial rule model, new rules are established and supplemented to the initial rule model to generate an updated rule model and update the initial rule model.

2. The anti-money laundering risk rule generation method based on the multi-modal large language model according to claim 1, wherein the specific method for establishing the modeling training set is: Establishing an initial training set: within a predetermined time period, obtaining real transaction data as an initial real transaction data set, wherein the real transaction data includes structured data and unstructured data; then obtaining historical suspicious transaction samples as an initial past typical training set; and integrating the initial real transaction data set and the initial past typical training set to form the initial training set. ​ Initialization label judgment: in the initial training set, the transaction data in the historical suspicious transaction sample itself has a true money laundering label; for the real transaction data set, the transaction data in the real transaction data is judged based on the initial rule model, and the risk transaction is additionally labeled as money laundering; Data integration: input the structured data and unstructured data in the initial training set into the multi-modal large language feature extraction model to generate structured transaction features and unstructured transaction features; Data augmentation: based on the structured transaction features and unstructured transaction features generated in the data integration, the transaction data with and without money laundering labels are respectively data enhanced to form an augmented data set; Forming a modeling training set: mixing the initial training set and the augmented data set to form a modeling training set for modeling.

3. The anti-money laundering risk rule generation method based on the multi-modal large language model according to claim 2, characterized in that: In the data integration step, the generated structured transaction features and unstructured transaction features are divided into: transaction basic information features, transaction type features, unstructured extraction features, and large language model induction features; Among them, the transaction basic information features are the transaction information and data extracted from the structured data; The transaction type features are the preset transaction classification extracted from the structured data; The unstructured data extraction features are the features extracted from the unstructured data and described in words by the multi-modal large language feature extraction model; The large language model induction features are the features extracted from the extracted features in the structured data and unstructured data by the multi-modal large language feature extraction model, and the description of the features is formed.

4. The anti-money laundering risk rule generation method based on the multi-modal large language model according to claim 3, characterized in that: When integrating the generated large language model induction features, a preset prompt word is used to guide the multi-modal large language feature extraction model to induce the transaction features within a preset dimension to form the description of the features.

5. The anti-money laundering risk rule generation method based on the multi-modal large language model according to claim 3, characterized in that: In the data augmentation step, the augmented objects are the transaction basic information features, transaction type features, unstructured extraction features, and large language model induction features generated according to the data integration step, and the augmented data set is formed based on the augmented objects; Based on the results of the initialization label judgment step, the transaction data with the money laundering label is the positive example, and the transaction data without the money laundering label is the negative example. The augmented results of the positive examples are still positive examples, and the augmented results of the negative examples are still negative examples.

6. The anti-money laundering risk rule generation method based on the multi-modal large language model according to claim 1, characterized in that: The method of using the validation set for verification further comprises: For the transaction data with a money laundering risk value exceeding a first threshold, a money laundering label is added as a validation experiment set, and the transaction data with a money laundering risk value exceeding a second threshold is used as a high-risk transaction data set, the second threshold is higher than the first threshold; The high-risk transaction data set is supplemented into the initial training set of the rule model as the past typical training set.

7. The method of claim 6, wherein the method of generating the updated rule model further comprises: adjusting the initial rule model based on the audit results of the auditors to improve the adaptability of the initial rule model to the current period data, in the case that the audit results of the auditors are inconsistent with the initial rule model in the case that the transaction data of the validation experiment set and the control set are inconsistent in the money laundering label.

8. The method of claim 1, wherein the method of generating the updated rule model further comprises: comparing the updated rule model with the existing rule model before going online, and if the updated rule model performs consistently and has a higher accuracy than the existing rule model within a predetermined test period, then the updated rule model is adopted. For transaction data with additional money laundering labels, a report is provided based on the multi-modal large language feature extraction model to distinguish the transaction data, including outputting the top structured features and warnings that contribute to the money laundering risk value. Comprise: a data collection module, a rule model discrimination module, a multi-modal feature extraction module, a data expansion module, a prediction model module, and a verification module for implementing the method of claim 1-9.

9. The method of claim 1, wherein the method is characterized by, ​ 10. A money laundering risk rule generation system based on a multi-modal large language model, characterized by, ​ ​

Images