Anti-money laundering risk rule generation method and system based on multi-modal large language model

Abstract

The invention discloses an anti-money laundering risk rule generation method and system based on a multi-modal large language model, and the method comprises the following steps: obtaining an initial training set composed of transaction data, adding a money laundering label based on a rule model, extracting transaction features through a multi-modal large language feature extraction model, and carrying out the expansion, forming a modeling training set by the initial training set and the expanded data set; fitting the multi-modal large language learning prediction model through the modeling training set to obtain a money laundering risk value of the transaction data, and calculating an actual contribution value of the structured features; and inputting the verification set into the multi-modal large language learning prediction model, verifying to obtain a money laundering label re-determination result, updating the money laundering label re-determination result to the initial training set, and then generating an updated rule model according to the verification result and the contribution value of the structured feature to the money laundering risk value. According to the invention, multi-mode transaction data can be integrated, a risk prediction model of money laundering transaction is formed, real-time adjustment and updating are carried out, and money laundering transaction behaviors can be accurately identified.

Description

Technical Field

[0001] The present application relates to the field of anti-money laundering risk identification, and specifically to a method and system for generating anti-money laundering risk rules based on a multimodal large language model. Background Art

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

[0003] First, manual identification by business experts is limited in resources and expensive. When faced with large-scale transaction data, these experts can generally only supplement the second or third forms of detection by reviewing their output. Secondly, for anti-money laundering transaction behavior assessments based on existing rule-based models, these models are highly static and cannot dynamically adjust to changes in economic conditions over time, making them difficult to adapt to new and evolving money laundering techniques. Finally, while machine learning and deep learning predictive models are more adaptable than rule-based approaches, their outputs lack interpretability, making it difficult to identify the specific reasons for identifying anti-money laundering cases and making adjustments impossible when the output is inaccurate. Furthermore, due to the limited amount of transaction sample information captured by machine learning, their accuracy in identifying money laundering cases is also lower than that of manual review. At the same time, as criminals adopt increasingly complex and covert money laundering methods, in order to cope with these changes, the anti-money laundering mechanism also needs to be constantly updated and adjusted. For the second or third detection methods, if it is necessary to dynamically adapt to new money laundering methods, the knowledge graph and judgment rules need to be studied, adjusted, updated and maintained. This process requires continuous investment of expert resources and time costs, which is costly and inefficient. Summary of the Invention

[0004] The purpose of this application is to overcome the shortcomings of the existing technology and provide an anti-money laundering risk rule generation method and system based on a multimodal large language model, which 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 to accurately identify money laundering transaction behaviors.

[0005] In the first aspect, the present application provides a method for generating anti-money laundering risk rules based on a multimodal large language model, which adopts the following technical solutions: 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, transaction features are extracted from the transaction data in the initial training set using a multimodal large language feature extraction model. These transaction features are expanded to form an expanded dataset. The initial training set and the expanded dataset form a modeling training set for modeling. By modeling the training set, fitting a multimodal large language learning prediction model, inputting the transaction data of the modeling training set, outputting the money laundering labels corresponding to the transaction data of the modeling training set, obtaining the money laundering risk value of the transaction data, and calculating the Shapley value of the transaction features in the transaction data. The Shapley value is the contribution of a specific transaction feature to the money laundering risk value; The validation set is used to input the multimodal large language learning prediction model, and the validation set with the money laundering label re-determined is verified and updated to the initial training set. Then, an updated rule model is generated based on the validation results and the contribution value of the structured features to the money laundering risk value, and the initial rule model is updated.

[0006] Through the above technical solution, the initial training set is first predicted for money laundering risk through the initial rule model, and then the multimodal features of the transaction data are extracted through multimodal large language features, and data expansion is performed; then the multimodal large language learning prediction model is input to obtain the correlation between the multimodal features and the money laundering risk; finally, the validation set is used to supplement the modeling training set and correct the rule model. The above technical solution obtains the multimodal data of the transaction data through the multimodal large language learning prediction model, and maps it with the money laundering risk to predict the risk. Compared with risk judgment through a single rule model, it has stronger recognition ability and accuracy and can cope with complex money laundering methods; by integrating multimodal features and expanding transaction data, the number of samples is expanded, the impact of sample imbalance is reduced, and the model recognition ability can be improved; by supplementing the initial training and correcting the rule model through the validation set, the model can be continuously optimized and the real-time and adaptability of the system can be enhanced.

[0007] As a preferred method, the specific method of establishing the modeling training set is: Establishing an initial training set: For a specific time period, obtain real transaction data from that period as the initial real transaction data set. Real transaction data includes both structured and unstructured data. Then, obtain historical suspicious transaction samples as the initial past typical training set. Integrate the real transaction data set and past typical training set to form the initial training set. Initialized label judgment: In the initial training set, the transaction data in the historical suspicious transaction samples are labeled as real money laundering. For the real transaction data set, the transaction data in the real transaction data is judged based on the initial rule model, and the risky transactions are labeled as money laundering. Data integration: Input the structured and unstructured data in the initial training set into the multimodal large language feature extraction model to generate multimodal transaction features; Data augmentation: Based on the multimodal transaction features generated in data integration, data augmentation is performed on transaction data with and without money laundering labels to form an augmented dataset. Forming a modeling training set: The initial training set and the expanded dataset are mixed to form a modeling training set for modeling.

[0008] Through the above technical solution, real transaction data and historical suspicious transaction samples are combined as the initial training set. After the money laundering label is added to the initial training set, data integration and data expansion are performed. On the premise of clarifying whether the target data entry has a money laundering label, it 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.

[0009] Preferably, in the data integration step, the integrated multimodal transaction features are divided into: basic transaction information features, transaction type features, unstructured extraction features and large language model summary features; wherein, the basic transaction information features are transaction information and data extracted from structured data, the transaction type features are preset transaction classifications extracted from structured data, the unstructured data extraction features are features extracted from unstructured data through a multimodal large language feature extraction model and described in text form, and the large language model summary features are features extracted from structured data and unstructured data through a multimodal large language feature extraction model and summarized to form a description of the features.

[0010] Preferably, when integrating and generating the summarized features of the large language model, preset prompt words are used to guide the multimodal large language feature extraction model, summarize the multimodal transaction features within the preset dimensions, and form a description of the features.

[0011] Through the above technical solution, the basic transaction information features and transaction type features are features extracted by the traditional feature extraction model. On this basis, the unstructured extraction features and the inductive features of the large language model are extracted through the multimodal large language feature extraction model, which can realize the feature extraction of multimodal transaction data, make full use of the multi-source heterogeneous data generated in actual business, and realize the collaborative analysis of multi-dimensional features.

[0012] Preferably, in the data expansion step, the objects of expansion are the basic transaction information features, transaction type features, unstructured extraction features and large language model summary features generated according to the data integration step to form an expanded data set. Based on the results of the initialization label judgment step, the transaction data with money laundering labels are used as positive examples, and the transaction data without money laundering labels are used as negative examples. The expansion result of the positive example is still a positive example, and the expansion result of the negative example is still a negative example.

[0013] Through the above technical solution, it is possible to achieve bidirectional expansion of structured features and unstructured features, enhance the expressiveness of the dataset through cross-modal association, and supplement scarce samples to support the modeling needs of detecting complex money laundering scenarios.

[0014] As a preferred method, the specific method of using the validation set for validation is: First, the validation set is judged based on the initial rule model, and risky transactions are labeled as money laundering to form a control set. The transaction data of the validation set is then fed into the multimodal large language learning prediction model 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 as a validation experiment set, and transaction data with a money laundering risk value exceeding a second threshold are treated as a high-risk transaction data set, where the second threshold is higher than the first threshold. The verification experiment set is compared with the control set. For transaction data with inconsistent money laundering labels between the verification experiment set and the control set, auditors 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.

[0015] Preferably, the method for generating an updated rule model specifically includes: 1) For transaction data with inconsistent money laundering labels between the validation set and the control set, if the auditor's review results are inconsistent with the initial rule model's judgment, the initial rule model will be adjusted to improve its adaptability to the current period's data; 2) Continuously calculate the Shapley value of each structured transaction feature and rank them. Adjust the weight or eliminate from the initial rule model any structured features whose Shapley value rankings do not reach a certain threshold within the specified evaluation period. 3) For transaction features whose Shapley rankings reach a certain threshold within the specified evaluation period, if they are not included in the initial rule model, new rules are established for the transaction feature rules and added to the initial rule model.

[0016] Preferably, the updated rule model is tested and verified against the existing rule model before going online. If the performance is consistent with the existing rule model within a preset test period and the accuracy exceeds that of the existing rule model, the updated rule model is adopted.

[0017] Through the above technical solution, the accuracy of money laundering risk prediction of the multimodal large 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 remains to be verified. By introducing the verification set, the money laundering labels determined by the rule model and the money laundering risk prediction output by the multimodal large language learning prediction model are cross-validated, and through manual judgment by auditors, cross-validation of the rule model and the multimodal large 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 timeliness data of the latest cases.

[0018] Preferably, for transaction data with money laundering tags attached, a report on the identification of the transaction data is provided based on a multimodal large language feature extraction model. The report content includes the basis for identification, transaction features with higher contribution to the money laundering risk value, and warning information.

[0019] Through the above technical solution, the predicted risk warning transaction projects can output interpretable anti-money laundering reports and warning information based on their multimodal characteristics, which improves the transparency and credibility of the predictions and provides reference and convenience for subsequent operation and maintenance personnel to identify and conduct secondary reviews.

[0020] In a second aspect, the present 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; The data collection module collects real transaction data and historical suspicious samples to form an initial training set; The rule model identification module uses the rule model to evaluate the money laundering risk of transaction data in the initial training set, identifies money laundering in the input transaction data, and adds money laundering labels to high-risk transaction data. The multimodal feature extraction module extracts multimodal transaction features of transaction data; The data expansion module performs data expansion based on multimodal transaction features and merges the expanded dataset with the initial training set into the modeling training set; The prediction model module fits a multimodal large language learning prediction model based on the modeling training set and performs a secondary assessment of the money laundering risk of the input transaction data; The verification module verifies the multimodal large language learning prediction model through the verification set, adds transaction data and high-risk transaction data that have been manually reviewed by auditors in the verification set into the data collection module, and modifies the rule model in the rule model discrimination module.

[0021] In summary, this application includes at least one of the following beneficial technical effects: 1. This application can integrate multi-source heterogeneous data in transaction data, introduce semantic analysis and feature generation mechanisms, and map the structured and unstructured features of transaction data to money laundering risk labels, enabling the model to more accurately identify potential money laundering behaviors, reduce false positives, and improve accuracy.

[0022] 2. This application uses a multimodal large-scale language feature extraction model to perform data generative amplification on transaction data based on structured and unstructured features, effectively reducing the impact of sample imbalance, improving the model's ability to identify minority samples (such as suspicious transactions), and reducing the risk misjudgment rate; compared to simply amplifying structured features, this improves the model's generalization and robustness.

[0023] 3. Through supplementary verification of the validation set and human-machine collaborative feedback, this application can continuously optimize the rules and parameters of the model, making it adaptable to new money laundering methods and enhancing the real-time and adaptability of the system.

[0024] 4. This application can output explanatory suspicious reports and risk warnings, making the decision-making process visual, improving the transparency and credibility of the forecast, ensuring that the decision meets the requirements of industry regulatory standards, and improving the efficiency of professionals in reviewing and modifying forecast results. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 Schematic diagram of a method flow for generating anti-money laundering risk rules based on a multimodal large language model in an embodiment of the present application; Figure 2 This is a flow chart of S100 of the method for generating anti-money laundering risk rules based on a multimodal large language model in an embodiment of the present application; Figure 3 This is a flow chart of S300 of the method for generating anti-money laundering risk rules based on a multimodal large language model in an embodiment of the present application; Figure 4 A schematic diagram of the risk identification process for transaction data sets; 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 the present application. DETAILED DESCRIPTION

[0026] This specific embodiment is merely an explanation of the present application and is not a limitation of the present application. After reading this specification, those skilled in the art may make non-creative modifications to the present embodiment as needed, but as long as they are within the scope of the present application, they are protected by patent law.

[0027] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application. It should be noted that in the optional embodiments of the present application, when the embodiments in the present application are applied to specific products or technologies, the object information and other related data involved need to obtain the object's permission or consent, and the collection, use and processing of the relevant data need to comply with the relevant laws, regulations and standards of the relevant countries and regions. In other words, if the embodiments of the present application involve data related to the object, it needs to be obtained through 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 needs to be obtained. The embodiments also need to be implemented with the authorization and consent of the object.

[0028] The embodiments of the present application are described in further detail below with reference to the accompanying drawings.

[0029] The anti-money laundering (AML) field has begun leveraging 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 abnormal results. These systems rely on a pre-set rule base and static thresholds, comparing real-time transaction data with established rules to assess account risk, monitor suspicious transactions, and, to a certain extent, provide automated early warnings. However, this technical architecture has significant limitations: First, the data used by the predictive model for learning is mostly structured data, ignoring 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 device fingerprint anomalies, irregular transaction timing, and dispersed capital flows, making it difficult to identify carefully designed, complex money laundering operations. Furthermore, the model cannot understand the semantic connections between transaction behaviors, such as logical relationships such as "short-term logins to multiple accounts from the same device." It relies solely on numerical feature matching and lacks semantic analysis and dynamic feature generation capabilities. Second, the predictive model primarily relies on statistical patterns and rule bases from historical data, unable to independently infer the logical connections between characteristic elements in new money laundering methods. When criminal methods innovate, the system fails due to its lack of generalization. Furthermore, rule base updates lag significantly behind the evolution of criminal methods. After regulatory authorities notify new cases, banks must manually analyze, extract features, and adjust rules. This process often takes weeks, resulting in the system's complete loss of early warning capabilities in the early stages of criminal method evolution. These issues are the goals of this application to improve and enhance.

[0030] In one embodiment of the present application, see Figure 1 , a method for generating anti-money laundering risk rules based on a multimodal large language model, comprising the following steps: S100, obtaining an initial training set consisting of transaction data, adding money laundering labels to the transaction data in the initial training set based on an initial rule model, and then extracting transaction features of the transaction data in the initial training set using a multimodal large language feature extraction model, wherein the transaction features include structured features and unstructured features. The transaction features are then expanded to form an expanded data set, and a modeling training set for modeling is formed from the initial training set and the expanded data set.

[0031] S200, through modeling training sets, takes structured features and unstructured features as independent variables, and takes transaction data money laundering labels as dependent variables, constructs a multimodal large language learning prediction model to obtain the money laundering risk value of transaction data.

[0032] In this embodiment, a pre-trained multimodal language feature extraction model is first used to construct a corresponding multimodal data feature vector. This feature vector is then merged with the data's original structured feature vector to form a multimodal structured feature vector, which is then mapped to whether the transaction data carries a money laundering label. The multimodal language learning prediction model then outputs the probability of whether the transaction data ultimately carries a money laundering label based on its transaction features, using the probability level classification 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 the probability of determining whether the transaction data carries a money laundering label, i.e., its contribution to the money laundering risk value. In this step, the multimodal language learning prediction model is fitted to integrate the multimodal data into the transaction data money laundering label recognition model, and to provide feedback and secondary confirmation to the rule model in S100.

[0033] S300: The validation set is input into the multimodal large language learning prediction model to verify the validation set with a newly determined money laundering label. This validation set is then updated to the initial training set. An updated rule model is then generated based on the validation results and the contribution of the structured features to the money laundering risk value, thereby updating the initial rule model. In addition to cross-validating the accuracy of the multimodal large language learning prediction model and the rule model using the validation set, transaction data from the validation set is also used to supplement the transaction data in the modeling dataset, correcting 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 actual transaction data from the initial training set, or separately prepared validation data with clear money laundering labels.

[0034] Specifically, in one embodiment, see Figure 2 , the specific steps of S100 are: S110, establishing an initial training set: for a specific period, obtaining real transaction data for that period as an initial real transaction data set, where the real transaction data includes structured data and unstructured data; then obtaining historical suspicious transaction samples as an initial past typical training set; integrating the real transaction data set and the past typical training set to form the initial training set.

[0035] In S110, structured data for real transaction data 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 image records of transfer receipts or remittance receipts, unstructured data extracted from the SWIFT messaging system (such as transaction descriptions and notes), transaction IP addresses and device fingerprints, offline branch location data, transaction notes, transaction descriptions, and customer behavior data. Structured and unstructured data are associated with customer IDs. Historical suspicious transaction samples are transaction samples identified as money laundering cases, obtained from user-stored databases or public information repositories (such as the central bank's anti-money laundering case database). Depending on the source of these transaction samples, some structured or unstructured data may be missing.

[0036] S120, initialization label judgment: In the initial training set, the transaction data in the historical suspicious transaction samples are inherently labeled with real money laundering labels; for the real transaction data set, the transaction data in the real transaction data is judged based on the initial rule model, and the risky transactions are labeled with money laundering labels.

[0037] In S120, since the transaction data in the historical suspicious transaction sample has already been identified as money laundering cases, there's no need to directly assign a money laundering label. For the real transaction dataset, however, a preliminary money laundering risk assessment is performed using a pre-set initial rule model. If a money laundering risk is identified, a money laundering label is assigned. This initial rule model is an existing discrimination model based on the structured features of transaction data. For example, if the transaction frequency within a single account unit period exceeds a threshold, the transaction is identified as a money laundering risk and a money laundering label is assigned. This initial rule model is subsequently updated or replaced by newly added or revised rule models through continuous iterative verification, generating an updated rule model.

[0038] S130, data integration: input the structured data and unstructured data in the initial training set into the multimodal large language feature extraction model to generate structured transaction features and unstructured transaction features.

[0039] In S130, the integrated structured transaction features and unstructured transaction features are divided into: basic transaction information features, transaction type features, unstructured extraction features, and large language model summary features. Among them, basic transaction information features are transaction information and data extracted from structured data; transaction type features are preset transaction categories extracted from structured data, such as payment settlement, salary payment, tax payment, fund collection, transfer settlement, and investment transfer; unstructured data extraction features are features extracted from unstructured data using a multimodal large language feature extraction model and described in text form. For example, text information semantic features extracted from transaction notes and transaction descriptions, company main business features, transaction IP address features, device fingerprint features, and offline branch location data features extracted from business license images, etc.; large language model summary features are features extracted from structured and unstructured data using a multimodal large language feature extraction model and summarized to form feature descriptions.

[0040] More specifically, in the above embodiment, when integrating and generating the summarized features of the large language model, a built-in prompt (preset prompt word) is used to guide the multimodal large language feature extraction model. This guidance is a question-based guidance based on the anti-money laundering strategy and the transaction behavior and transaction pattern, so that the multimodal large language feature extraction model can summarize the transaction features within the preset dimensions and form a feature description. For example, the multimodal large language feature extraction model is asked whether a single account receives a large number of irregular small transfers and then transfers them to a fixed account, or a large amount of funds is transferred in at one time and then split into multiple small transfers to different accounts; whether funds are transferred out quickly after receiving them, with extremely short stays in the account and extremely low balances; whether there are frequent fund transfers through different banks and payment platforms; whether there are frequent transfers from corporate accounts to personal accounts, especially involving transactions unrelated to the main business listed on the business license; whether there are frequent transactions in a long-unused account in a short period of time; whether multiple accounts use the same IP address or use overseas proxy servers to hide their true location, etc. Through the multimodal large language feature extraction model, based on structured data and unstructured data, and guided by Prompt, the extraction and induction of multimodal features are achieved to form the inductive features of the large language model.

[0041] S140, data expansion: Based on the structured transaction features and unstructured transaction features generated in the data integration, data enhancement is performed on the transaction data with money laundering labels and the transaction data without money laundering labels to form an expanded data set.

[0042] In S140, the expanded data sets are the basic transaction information features, transaction type features, unstructured extracted features, and summarized features from the large language model generated in the data integration step, forming an expanded dataset. Based on the results of the initial label determination step, transaction data with money laundering labels is considered positive examples, while transaction data without money laundering labels is considered negative examples. The expanded results for positive examples remain positive, and the expanded results for negative examples remain negative. Due to the scarcity of positive examples in anti-money laundering scenarios, data augmentation primarily focuses on expanding positive examples to increase the proportion of positive examples in the total dataset. Using a multimodal large language feature extraction model, it is possible to expand both structured and unstructured features. For example, the large language model can identify the multi-layered money laundering chain "gambling and fraud funds - jewelry purchases - overseas investment" as an unstructured feature. Then, "jewelry purchases" is expanded to "luxury goods purchases" to increase the number of relevant positive examples.

[0043] S150, forming a modeling training set: mixing the initial training set and the expanded data set to form a modeling training set for modeling.

[0044] Specifically, in one embodiment, in S200, the specific construction process of the multimodal large language learning prediction model may be: S210 , using a pre-trained multimodal large language feature extraction model, such as the text vectorization model of iFlytek Spark in this embodiment, to take the structured input and unstructured data portion of a data item as input, generates a structured numerical feature vector. This structured numerical feature vector is then combined with the original structured feature vector of the data to form a multimodal structured feature vector.

[0045] S220 uses a multimodal structured feature vector as an independent variable and a money laundering label as a dependent variable, and adopts a neural network training model, such as a logistic regression model and an xgboost model, to obtain a multimodal large language learning prediction model that can predict the money laundering risk value.

[0046] Specifically, in one embodiment, see Figure 3 , S300, the specific method of using the verification set for verification is: S310 , the verification set is judged based on the initial rule model, and money laundering labels are added to risky transactions to form a control set.

[0047] At step S320, the transaction data from the validation set is input into the multimodal large language learning prediction model from step S200 to obtain a money laundering risk value for the validation set. Transaction data with a money laundering risk value exceeding a first threshold is assigned a money laundering label as part of the validation experiment set. This process effectively involves using the multimodal large language learning prediction model to determine the money laundering label based on the multimodal features of the transaction data. Transaction data with a money laundering risk value exceeding a second threshold, where the second threshold is higher than the first threshold, is designated as a high-risk transaction dataset.

[0048] At step S330, the validation set is compared with the control set. For transaction data with inconsistent money laundering labels between the validation set and the control set, human reviewers manually confirm or reject the data as money laundering data. This data represents conflicting judgments between the rule-based model and the multimodal large-scale language learning prediction model. This manual review and final determination of the conflicting judgments has significant instructive implications for both the rule-based model and the multimodal large-scale language learning prediction model, and therefore warrants inclusion in the historical typical training set. Furthermore, the high-risk money laundering transaction dataset, which contains transaction data with significantly higher money laundering risk, has a high degree of fit between its transaction characteristics and the money laundering labels, resulting in a high degree of judgment accuracy. Therefore, this dataset is also suitable for inclusion in the historical typical training set.

[0049] More specifically, in another embodiment, the past typical training set is divided into several periodic training sets distinguished by fixed periods. The past typical training set output by the regularly updated validation set enters the corresponding periodic training set. After entering a new period, the weight of each periodic training set for the initial training set is fine-tuned to ensure that the model adapts to recent new types of money laundering behaviors.

[0050] More specifically, in another embodiment, in S300, the method for generating an updated rule model specifically includes: 1) For transaction data with inconsistent money laundering labels between the validation set and the control set, if the auditor's review results are inconsistent with the initial rule model's judgment, the initial rule model will be adjusted to improve its adaptability to the current period's data; 2) Continuously calculate the Shapley value of each structured transaction feature and rank them. For structured features whose Shapley value rankings do not reach a certain threshold within a specified evaluation period, the corresponding rules in the initial rule model will be weighted or eliminated; 3) For transaction features whose Shapley rankings 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 established for the transaction feature and added to the initial rule model.

[0051] Before going online, updated rule models must undergo A / B testing and verification, comparing them to existing rule models. If the updated rule model performs consistently across different data sets within a pre-set testing period and exceeds the accuracy of the existing rule model, the updated rule model will be adopted. Otherwise, the rule model will not be updated. In another embodiment, if there are cases where existing rules cannot consistently identify anti-money laundering cases for new data, new money laundering transaction identification rules will be directly constructed based on the existing rules and the new samples that cannot be identified.

[0052] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean 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.

[0053] When conducting risk identification on transaction datasets, please refer to Figure 4 First, the transaction dataset is preprocessed to obtain its multimodal features. The latest rule model then performs an initial identification of money laundering risks on the transaction dataset. The transaction dataset is then fed into a multimodal large language learning prediction model for a secondary identification of money laundering risks. A human-machine collaborative review of the differences between the two identifications determines the final output. For transaction data with money laundering labels, a report is provided based on the multimodal large language feature extraction model, including the basis for identifying the transaction data, structural features that contribute significantly to the money laundering risk value, and warnings. This report provides interpretable anti-money laundering judgments and warning information, enhancing the transparency and credibility of the predictions and providing a reference and convenience for subsequent identification and secondary review by operations and maintenance personnel. At the same time, this portion of the transaction dataset is fed back into the real transaction dataset to supplement the initial training set to accommodate the current cycle data and enable adaptive adjustment of the rule model.

[0054] In another embodiment, see Figure 5 The present 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.

[0055] The data collection module 1 collects real transaction data and historical suspicious samples to form an initial training set.

[0056] The rule model identification module 2 uses the rule model to evaluate the money laundering risk of transaction data in the initial training set and adds money laundering labels to high-risk transaction data.

[0057] The multimodal feature extraction module 3 extracts multimodal transaction features of the transaction data.

[0058] The data expansion module 4 performs data expansion based on the multimodal transaction features and merges the expanded data set with the initial training set into a modeling training set.

[0059] The prediction model module 5 performs a secondary assessment of the money laundering risk of the input transaction data by modeling the training set and fitting the multimodal large language learning prediction model.

[0060] The verification module 6 verifies the multimodal large language learning prediction model through the verification set, adds transaction data and high-risk transaction data in the verification set that have been manually reviewed by auditors into the data collection module, and modifies the rule model in the rule model discrimination module.

[0061] In another embodiment, the anti-money laundering risk rule generation system based on the multimodal large language model further includes a report generation module 7, which outputs the transaction data manually reviewed by the auditors in the verification module 6 and provides an analysis report.

[0062] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the anti-money laundering risk rule generation system based on the multimodal large language model can refer to the corresponding process in the aforementioned method embodiment and will not be repeated here.

[0063] Those skilled in the art will clearly understand that for the sake of convenience and brevity of description, only the division of the above-mentioned functional units and modules is used as an example. In actual applications, the above-mentioned functions can be distributed and completed by 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.

[0064] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention, and should all be included in the scope of protection of the present invention.

Claims

1. A method for generating anti-money laundering risk rules based on a multimodal large language model, characterized in that: The steps include: 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, transaction features are extracted from the transaction data in the initial training set using a multimodal large language feature extraction model. These transaction features are expanded to form an expanded dataset. The initial training set and the expanded dataset form a modeling training set for modeling. By modeling the training set, fitting a multimodal large language learning prediction model, inputting the transaction data of the modeling training set, outputting the money laundering labels corresponding to the transaction data of the modeling training set, obtaining the money laundering risk value of the transaction data, and calculating the Shapley value of the transaction features in the transaction data. The Shapley value is the contribution of a specific transaction feature to the money laundering risk value; The validation set is used to input the multimodal large language learning prediction model, and the validation set with the money laundering label re-determined is verified and updated to the initial training set. Then, an updated rule model is generated based on the validation results and the contribution value of the structured features to the money laundering risk value, and the initial rule model is updated.

2. The method for generating anti-money laundering risk rules based on a multimodal large language model according to claim 1, characterized in that: The specific method of establishing the modeling training set is: Establishing an initial training set: For a specific time period, obtain real transaction data from that period as the initial real transaction data set. Real transaction data includes both structured and unstructured data. Then, obtain historical suspicious transaction samples as the initial past typical training set. Integrate the real transaction data set and past typical training set to form the initial training set. Initialized label judgment: In the initial training set, the transaction data in the historical suspicious transaction samples are labeled as real money laundering. For the real transaction data set, the transaction data in the real transaction data is judged based on the initial rule model, and the risky transactions are labeled as money laundering. Data integration: The structured and unstructured data in the initial training set are input into the multimodal large language feature extraction model to generate structured and unstructured transaction features; 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. Forming a modeling training set: The initial training set and the expanded dataset are mixed to form a modeling training set for modeling.

3. The method for generating anti-money laundering risk rules based on a multimodal large language model according to claim 2, characterized in that: In the data integration step, the structured transaction features and unstructured transaction features generated by integration are divided into: basic transaction information features, transaction type features, unstructured extraction features and large language model summary features; among them, basic transaction information features are transaction information and data extracted from structured data, transaction type features are preset transaction classifications extracted from structured data, unstructured data extraction features are features extracted from unstructured data through a multimodal large language feature extraction model and described in text form, and large language model summary features are features extracted from structured data and unstructured data through a multimodal large language feature extraction model, and summarized to form a description of the features.

4. The method for generating anti-money laundering risk rules based on a multimodal large language model according to claim 3, characterized in that: When integrating and generating the summarized features of the large language model, preset prompt words are used to guide the multimodal large language feature extraction model, summarize the transaction features within the preset dimensions, and form a description of the features.

5. The method for generating anti-money laundering risk rules based on a multimodal large language model according to claim 3, characterized in that: In the data expansion step, the objects of expansion are the basic transaction information features, transaction type features, unstructured extraction features and large language model summary features generated according to the data integration step to form an expanded data set. Based on the results of the initialization label judgment step, the transaction data with money laundering labels are used as positive examples, and the transaction data without money laundering labels are used as negative examples. The expansion result of the positive example is still a positive example, and the expansion result of the negative example is still a negative example.

6. The method for generating anti-money laundering risk rules based on a multimodal large language model according to claim 1, characterized in that: The specific method of using the validation set for verification is: First, the validation set is judged based on the initial rule model, and risky transactions are labeled as money laundering to form a control set. The transaction data of the validation set is then fed into the multimodal large language learning prediction model 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 as a validation experiment set, and transaction data with a money laundering risk value exceeding a second threshold are treated as a high-risk transaction data set, where the second threshold is higher than the first threshold. The verification experiment set is compared with the control set. For transaction data with inconsistent money laundering labels between the verification experiment set and the control set, auditors 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.

7. The method for generating anti-money laundering risk rules based on a multimodal large language model according to claim 6, characterized in that: The method for generating an updated rule model specifically includes: 1) For transaction data with inconsistent money laundering labels between the validation set and the control set, if the auditor's review results are inconsistent with the initial rule model's judgment, the initial rule model will be adjusted to improve its adaptability to the current period's data; 2) Continuously calculate the Shapley value of each structured transaction feature and rank them. For structured features whose Shapley value rankings do not reach a certain threshold within a specified evaluation period, the corresponding rules in the initial rule model will be weighted or eliminated; 3) For transaction features whose Shapley rankings reach a certain threshold within a specified evaluation period, if there is no corresponding rule in the initial rule model, a new rule is established for the transaction feature and added to the initial rule model.

8. The method for generating anti-money laundering risk rules based on a multimodal large language model according to claim 1, characterized in that: Before going online, the updated rule model is tested and verified against the existing rule model. If the performance is consistent within the preset test cycle and the accuracy exceeds that of the existing rule model, the updated rule model will be adopted.

9. The method for generating anti-money laundering risk rules based on a multimodal large language model according to claim 1, characterized in that: For transaction data with money laundering labels, a report on the identification of the transaction data is provided based on a multimodal large language feature extraction model, including the basis, structural features with higher contribution to the money laundering risk value, and warnings. 10.An anti-money laundering risk rule generation system based on a multimodal large language model, characterized in that: include: Data collection module, rule model discrimination module, multimodal feature extraction module, data expansion module, prediction model module and verification module; The data collection module collects real transaction data and historical suspicious samples to form an initial training set; The rule model discrimination module uses the 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; The multimodal feature extraction module extracts multimodal transaction features of transaction data; The data expansion module performs data expansion based on multimodal transaction features and merges the expanded dataset with the initial training set into the modeling training set; The prediction model module uses a model training set to fit a multimodal large language learning prediction model and conducts a secondary assessment of the money laundering risk of the input transaction data. The verification module verifies the multimodal large language learning prediction model through the verification set, adds transaction data and high-risk transaction data that have been manually reviewed by auditors in the verification set into the data collection module, and modifies the rule model in the rule model discrimination module.

Images