Report generation method, apparatus, device, medium, and program product

Abstract

The application provides a report generation method and device, equipment, storage medium and program product, which can be applied to the field of artificial intelligence technology. The method comprises the following steps: obtaining financial management related data of a target customer, filtering the financial management related data to obtain effective data, wherein the financial management related data at least comprises transaction data, customer portrait data and product data; inputting the effective data into an inference analysis model to output an inference analysis result, wherein the inference analysis model adopts a reinforcement learning algorithm and a thinking chain inference mechanism; and inputting the inference analysis result into a report generation model to output a structured financial management report, wherein the report generation model adopts a hybrid expert architecture and a multi-token parallel prediction mechanism.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and more specifically, to a report generation method, apparatus, device, medium, and program product. Background Technology

[0002] Currently, banking institutions place great emphasis on maintaining and developing customer resources. In the current economic environment, many customers have generally reduced their willingness to learn about and purchase wealth management products. Banks are facing challenges in accurately targeting their wealth management product promotions, and at the same time, personalized wealth management reports tailored to individual customer needs remain insufficient, failing to effectively address customer pain points. Summary of the Invention

[0003] In view of the above problems, embodiments of this application provide a report generation method, apparatus, device, medium, and program product.

[0004] According to a first aspect of this application, a report generation method is provided, comprising: acquiring financial management-related data of a target customer; filtering the financial management-related data to obtain valid data, wherein the financial management-related data includes at least transaction data, customer profile data, and product data; inputting the valid data into an inference analysis model and outputting an inference analysis result, wherein the inference analysis model employs a reinforcement learning algorithm and a thought chain inference mechanism; and inputting the inference analysis result into a report generation model and outputting a structured financial management report, wherein the report generation model employs a hybrid expert architecture and a multi-token parallel prediction mechanism.

[0005] According to an embodiment of this application, filtering the financial management-related data to obtain valid data includes: converting the structured data in the financial management-related data into a text sequence; calculating the correlation score between the text sequence and the target customer characteristics using a correlation evaluation algorithm, wherein the target customer characteristics are derived from the customer profile data; calculating the information entropy of the text sequence; and determining the valid data based on the correlation score and the information entropy, wherein the correlation score corresponding to the valid data is greater than a first preset threshold, and the information entropy corresponding to the valid data is greater than a second preset threshold.

[0006] According to an embodiment of this application, the training of the inference analysis model includes: fine-tuning the inference analysis model to be trained based on a preset amount of scenario long-chain thinking data, so that the inference analysis model generates intermediate inference steps according to a preset logical link during the inference analysis stage, wherein the preset logical link follows a logical order from transaction motivation to risk exposure to optimization direction, and wherein the inference analysis model performs consistency verification on the preset logical link based on a rule engine and semantic similarity during the inference analysis stage.

[0007] According to an embodiment of this application, the reinforcement learning algorithm is a group relative policy optimization algorithm, and the training of the inference analysis model includes: for each input sample, generating N candidate inference results through the inference analysis model to be trained, the N candidate inference results constituting an inference group, wherein N is an integer greater than 1; calculating the total reward value of each candidate inference result; and updating the model parameters of the inference analysis model to be trained based on the total reward value of each candidate inference result in the inference group.

[0008] According to an embodiment of this application, the reasoning analysis model is provided with a dual-drive reward function, which includes an accuracy reward item and a language consistency reward item. The total reward value is the weighted sum of the accuracy reward item and the language consistency reward item. The accuracy reward item is used to evaluate the consistency between the candidate reasoning result and the preset logical rules, and the language consistency reward item is used to evaluate the logical coherence and terminology standardization of the candidate reasoning result.

[0009] According to an embodiment of this application, the inference analysis model is a model that has undergone distillation processing, and the context window of the inference analysis model is extended to a preset length through rotational position encoding extension technology, wherein the preset length supports long-cycle transaction data.

[0010] According to an embodiment of this application, the hybrid expert architecture includes: constructing at least a portion of the converter layers in the report generation model as M expert sub-networks, where M is an integer greater than 1; for each input token, calculating the activation weight of each expert sub-network of the current converter layer through a routing network, selecting a predetermined number of expert sub-networks for activation based on the activation weights; and weighting and summing the outputs of the activated expert sub-networks based on their activation weights as the output of the current converter layer.

[0011] According to an embodiment of this application, the multi-token parallel prediction mechanism includes: decomposing the structured financial report to be generated into K report modules, where K is an integer greater than 1; and generating a token sequence for the K report modules in parallel.

[0012] According to an embodiment of this application, the method further includes: performing a compliance verification on the structured financial report, the compliance verification being based on a preset compliance rule base and a classification model; and outputting the structured financial report that has passed the compliance verification in a preset format.

[0013] According to an embodiment of this application, the compliance verification of the structured financial report includes: matching the structured financial report against each rule based on the preset compliance rule base; intercepting the structured financial report in response to a matched violation; and performing a compliance judgment using the classification model in response to no matched violation, and outputting a classification result.

[0014] According to a second aspect of this application, a report generation apparatus is provided, comprising: a data extraction module for acquiring financial management-related data of a target customer, filtering the financial management-related data to obtain valid data, wherein the financial management-related data includes at least transaction data, customer profile data, and product data; an inference analysis module for inputting the valid data into an inference analysis model and outputting inference analysis results, wherein the inference analysis model employs a reinforcement learning algorithm and a thought chain inference mechanism; and a report generation module for inputting the inference analysis results into a report generation model and outputting a structured financial management report, wherein the report generation model employs a hybrid expert architecture and a multi-token parallel prediction mechanism.

[0015] According to a third aspect of this application, an electronic device is provided, comprising: one or more processors; and a memory for storing one or more computer programs, wherein the one or more processors execute the one or more computer programs to implement the steps of the method described above.

[0016] According to a fourth aspect of this application, a computer-readable storage medium is also provided, on which a computer program or instructions are stored, wherein the computer program or instructions, when executed by a processor, implement the steps of the above-described method.

[0017] According to a fifth aspect of this application, a computer program product is also provided, comprising a computer program or instructions that, when executed by a processor, implement the steps of the above-described method. Attached Figure Description

[0018] The above-mentioned contents, other objects, features and advantages of this application will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0019] Figure 1 The illustrations depict application scenarios of report generation methods, apparatus, devices, media, and program products according to embodiments of this application.

[0020] Figure 2 A flowchart illustrating a report generation method according to an embodiment of this application is shown schematically;

[0021] Figure 3 This illustration shows a flowchart of filtering financial management-related data to obtain valid data according to an embodiment of this application;

[0022] Figure 4 This illustration shows a flowchart of compliance verification of a structured financial report according to an embodiment of this application;

[0023] Figure 5 A flowchart illustrating the generation of a compliant financial report according to an embodiment of this application is shown schematically;

[0024] Figure 6 This schematic diagram illustrates a structural block diagram of a report generation apparatus according to an embodiment of the present application;

[0025] Figure 7 A block diagram schematically illustrates an electronic device suitable for implementing a report generation method according to an embodiment of this application. Detailed Implementation

[0026] The embodiments of this application will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of this application. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of this application for ease of explanation. However, it will be apparent that one or more embodiments may be implemented without these specific details. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concepts of this application.

[0027] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0028] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0029] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).

[0030] Currently, banks face the following problems when providing wealth management transaction reports: the interpretation of wealth management products is superficial, the efficiency of accurately pushing wealth management products is low, it is difficult to meet customers' personalized needs, the deployment cost is high, the resource requirements are large, the operation and maintenance are complex, and the hardware requirements are stringent.

[0031] An embodiment of this application provides a report generation method, comprising: acquiring financial management-related data of a target customer; filtering the financial management-related data to obtain valid data, wherein the financial management-related data includes at least transaction data, customer profile data, and product data; inputting the valid data into an inference analysis model and outputting inference analysis results, wherein the inference analysis model employs a reinforcement learning algorithm and a thought chain inference mechanism; and inputting the inference analysis results into a report generation model and outputting a structured financial management report, wherein the report generation model employs a hybrid expert architecture and a multi-token parallel prediction mechanism.

[0032] The report generation method provided in this embodiment collaborates with the inference analysis model and the report generation model, integrates new technologies such as reinforcement learning and hybrid expert architecture, optimizes model parameters to adapt to financial scenarios, and achieves end-to-end intelligent generation and efficient deployment from customer data to professional reports. Ultimately, it improves the efficiency and customer experience of customers purchasing bank wealth management products and increases the coverage of wealth management products.

[0033] It should be noted that the report generation method and apparatus provided in this application can be used in the field of artificial intelligence technology, or in any field other than artificial intelligence technology, such as fintech technology. Here, the application field of the report generation method and apparatus provided in this application is not limited.

[0034] In the technical solution of this application, the user information (including but not limited to user personal information, user image information, user device information, such as location information) and data (including but not limited to data used for analysis, stored data, and displayed data) involved are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of related data all comply with relevant laws, regulations, and standards, take necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation entry points for users to choose to authorize or refuse.

[0035] In scenarios where personal information is used for automated decision-making, the methods, devices, and systems provided in this application all provide users with corresponding operation entry points for users to choose to agree to or reject the automated decision results; if the user chooses to reject, the process enters the expert decision-making process.

[0036] As used in this paper, the term "model" refers to a model that learns the relationship between inputs and outputs from training data, enabling it to generate corresponding outputs for a given input after training. Model generation can be based on machine learning techniques. Deep learning is a machine learning algorithm that processes inputs and provides corresponding outputs using multiple layers of processing units. A neural network model is an example of a deep learning-based model. In this paper, "model" may also be referred to as a "machine learning model," "learning model," "machine learning network," or "learning network," and these terms are used interchangeably.

[0037] It's important to note that the term "neural network" can refer to a machine learning network based on deep learning. A neural network processes input and provides corresponding output, typically consisting of an input layer, an output layer, and one or more hidden layers between them. Neural networks used in deep learning applications often include many hidden layers, increasing the network's depth. The layers of a neural network are connected sequentially, so that the output of the previous layer serves as the input to the next layer. The input layer receives the input to the neural network, while the output layer's output becomes the final output. Each layer of a neural network includes one or more nodes (also called processing nodes or neurons), each processing the input from the layer above.

[0038] It should be understood that machine learning generally includes three phases: training, testing, and application (also known as inference). In the training phase, a given model is trained using a large amount of training data, iteratively updating parameter values ​​until the model can consistently generate inferences that meet the expected goals from the training data. Through training, the model can be considered to have learned the relationship between inputs and outputs (also known as the input-output mapping) from the training data. The parameter values ​​of the trained model are determined. In the testing phase, test inputs are applied to the trained model to test whether it can provide the correct output, thus determining the model's performance. In the application phase, the model can be used to process actual inputs based on the trained parameter values ​​to determine the corresponding output.

[0039] In one or more embodiments described herein, the term "large model" can refer to a deep learning model with a large number of model parameters, which can include hundreds of millions, tens of billions, hundreds of billions, trillions, or even tens of trillions of model parameters. Large models can also be called foundational models or basic models. They are pre-trained using large-scale unlabeled corpora to produce pre-trained models with hundreds of millions of parameters. Such models can adapt to a wide range of downstream tasks and have good generalization ability, such as large language models and multimodal pre-trained models. It should be understood that in practical applications, large models only require a small number of samples to fine-tune the pre-trained model before being applied to different tasks. Large models can be widely used in natural language processing, computer vision, and other fields. Specifically, they can be applied to computer vision tasks such as visual question answering, image captioning, and image generation, as well as natural language processing tasks such as text-based sentiment classification, text summarization, and machine translation. Major application scenarios for large models can include digital assistants, intelligent robots, search, online education, office software, e-commerce, and intelligent design.

[0040] Figure 1 The illustrations depict application scenarios of report generation methods, apparatus, devices, media, and program products according to embodiments of this application. For example... Figure 1 As shown, application scenario 100 according to an embodiment of this application may include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, and a server 105. The network 104 serves as a medium for providing a communication link between the first terminal device 101, the second terminal device 102, the third terminal device 103, and the server 105. The network 104 may include various connection types, such as wired or wireless communication links or fiber optic cables. For example, a user can use the first terminal device 101, the second terminal device 102, and the third terminal device 103 to interact with the server 105 through the network 104 to receive or send information, etc.

[0041] The first terminal device 101, the second terminal device 102, and the third terminal device 103 can be electronic devices such as smartphones, wearable devices, personal computers, intelligent voice interaction devices, smart home appliances, intelligent vehicles, in-vehicle terminals, aircraft, unmanned vending terminals, and extended reality devices. Extended reality devices can include virtual reality devices, augmented reality devices, and mixed reality devices. A client application for the target application can be installed and run on the terminal devices. This target application can include, but is not limited to, financial transaction applications, payment applications, shopping applications, web browser applications, search applications, instant messaging tools, email clients, and social media platform software (these are just examples). Furthermore, this application embodiment does not limit the form of the target application, and it can include, but is not limited to, applications, mini-programs, etc., installed on the terminal devices, and can also be in the form of web pages.

[0042] Server 105 can be a server providing various services, such as a backend management server supporting websites browsed by users using the first terminal device 101, the second terminal device 102, and the third terminal device 103 (this is just an example). The backend management server can analyze and process received user requests and other data, and feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal devices. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services such as cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks, and basic cloud computing services such as big data. The server can be the backend server of the aforementioned target application, used to provide backend services to the clients of the target application.

[0043] It should be noted that the report generation method provided in this application embodiment can generally be executed by server 105 and / or terminal devices 101-103. Accordingly, the report generation device provided in this application embodiment can generally be set in server 105 and / or terminal devices 101-103.

[0044] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0045] Figure 2 A flowchart illustrating a report generation method according to an embodiment of this application is shown schematically. Figure 2 As shown, the report generation method 200 according to the embodiments of this application may include steps S210 to S230.

[0046] It should be noted that, although the steps of the method are described in a specific order in the embodiments of this application, it should be understood that the execution order of these steps is not fixed and can be adjusted according to actual needs. Specifically, some steps can be executed in reverse order, or they can be executed in parallel to improve efficiency without affecting the overall method effect. In addition, some steps may be optional and can be selectively executed according to specific application scenarios or requirements.

[0047] In step S210, the financial management-related data of the target customer is obtained, the financial management-related data is filtered, and valid data is obtained. The financial management-related data includes at least transaction data, customer profile data, and product data.

[0048] In embodiments of this application, user consent or authorization can be obtained before acquiring user information. For example, a request to acquire user information can be sent to the user before step S210. If the user consents or authorizes the acquisition of user information, step S210 is executed.

[0049] In step S220, valid data is input into the inference analysis model, and the inference analysis results are output. The inference analysis model uses reinforcement learning algorithms and thought chain inference mechanisms.

[0050] In the embodiments of this application, a corresponding operation entry point can be provided to the user, allowing the user to choose to agree to or reject the automated decision-making result. That is, before processing / making a decision by inputting valid data into the inference analysis model and outputting the inference analysis result, the user can be provided with an instruction to agree to or reject the processing / decision through the corresponding operation entry point. If the user agrees to the processing / decision, the user information is processed to input valid data into the inference analysis model and output the inference analysis result, i.e., step S220 is executed. If the user rejects the processing / decision, the expert decision-making process is initiated.

[0051] In step S230, the reasoning analysis results are input into the report generation model, and a structured financial report is output. The report generation model adopts a hybrid expert architecture and a multi-token parallel prediction mechanism.

[0052] According to the embodiments of this application, by collaborating the reasoning analysis model and the report generation model, the problem that a single model is difficult to handle in terms of both complex financial reasoning and high-quality text generation is solved; fully automatic generation from raw customer data to structured financial reports is achieved, improving the degree of personalization and generation efficiency.

[0053] For example, acquiring financial management-related data from target customers may include: collecting core customer data through a data input layer. Core customer data may include financial transaction data, customer profile data, and financial product data. Financial transaction data may include transaction time, amount, product type, and holding returns; customer profile data may include risk level, investment period, and return target; and financial product data may include the rating of the financial subsidiary and historical return fluctuations.

[0054] For example, large language models are natively text-processing oriented. Structured transaction data from bank wealth management (transaction time, amount, holding returns, risk level, etc.) is embedded through structured data serialization to achieve input adaptation. The following will combine... Figure 3 This paper describes how to convert structured data in financial management-related data into a text sequence.

[0055] Figure 3The illustration shows a flowchart of filtering financial management-related data to obtain valid data according to an embodiment of this application.

[0056] According to the embodiments of this application, filtering financial management-related data to obtain valid data may include steps S310 to S340.

[0057] In step S310, the structured data in the financial management-related data is converted into a text sequence.

[0058] In step S320, a correlation evaluation algorithm is used to calculate the correlation score between the text sequence and the target customer features, where the target customer features are derived from customer profile data.

[0059] In step S330, the information entropy of the text sequence is calculated.

[0060] In step S340, valid data is determined based on the correlation score and information entropy, wherein the correlation score of the valid data is greater than a first preset threshold and the information entropy of the valid data is greater than a second preset threshold.

[0061] For example, financial transaction data, customer profile data, and product data in a bank's core system are typically stored in the form of structured key-value pairs (e.g., database tables, JSON format, etc.). To enable subsequent text analysis algorithms (e.g., relevance assessment algorithms) to process financial data, each record (or each combination of fields) needs to be converted into a natural language text sequence. Converting structured data in financial data into a text sequence can include: converting structured key-value pair data into a natural language text sequence; extracting field names and corresponding values, and concatenating them into complete sentences according to a preset template. Each record can generate a text sequence. For example, the natural language text sequence after converting structured key-value pair data could be {"Transaction Time":"20260101","Holding Amount":50000,"Risk Level":"R2","Recent 3-Month Return":2.1%}.

[0062] For example, structured transaction data, customer profile data, product data, etc., obtained from databases / interfaces can be serialized into standard JSON (JavaScript Object Notation) format text. Field names, values, and tags retain their complete semantic information, completing the conversion from "structured data" to "semantic text sequence," adapting to the text input paradigm of large language models (i.e., large models in this embodiment). For instance, the original data such as transaction time, holding amount, and risk level of bank wealth management products are structured tabular data stored in a database, belonging to key-value pair format, and cannot be directly input into large language models. Through serialization conversion, it is organized into a standard JSON string, essentially transforming structured data into a plain text format with complete semantics, preserving field names, values, and tags, enabling it to be directly read, parsed, and inferred by large language models.

[0063] For example, target customer characteristics are derived from customer profile data and may include risk level, investment period, return target, etc. Target customer characteristics are used to assess the relevance of each transaction sample to the customer; the more closely the characteristics match, the more representative the sample is, and the more it should be retained as valid data.

[0064] For example, a relevance assessment algorithm is used to measure the degree of matching between two pieces of text (or text and query). One or more relevance assessment algorithms can be used to measure the degree of matching between two pieces of text (or text and query). For example, at least one of the term frequency-inverse document frequency (TFM) and probabilistic retrieval models (e.g., BM25 (Best Matching 25) algorithm, query likelihood model, etc.) can be used to calculate the relevance score between a text sequence and the target customer features. TFM is based on term frequency statistics and calculates the cosine similarity between each text sequence and the customer feature query to obtain a score. The probabilistic retrieval model calculates the relevance score between the text sequence and the query based on the normalization of term frequency and document length. It should be noted that using at least one of the term frequency-inverse document frequency and probabilistic retrieval models to calculate the relevance score between a text sequence and the target customer features is only an example and is not a limitation of the embodiments of this application. In other embodiments, at least one of the following can also be used to calculate the relevance score between a text sequence and the target customer features: sublinear term frequency scaling, random divergence model, statistical language model, dual-tower model, etc.

[0065] The method in this embodiment can employ various relevance evaluation algorithms to calculate the relevance scores between the text sequence and the target customer features. For the multiple relevance scores obtained through calculation, a weighted summation can be used to calculate a comprehensive relevance score. It should be noted that using a weighted summation to calculate the comprehensive relevance score is merely an example and not a limitation of the embodiments of this application. In other embodiments, the comprehensive relevance score can also be calculated by taking the maximum value or by calculating scores that are all greater than a threshold (e.g., by weighted summation of relevance scores that satisfy the condition "relevance score greater than a threshold").

[0066] For example, before calculating the relevance score between the text sequence and the target customer characteristics using a relevance assessment algorithm, the target customer characteristics can be extracted: key features can be extracted from customer profile data, such as risk level (e.g., C2), investment period (e.g., 1-3 years), return target (e.g., conservative), etc. These features are then transformed into query keywords or query vectors. A relevance assessment algorithm is then used to calculate the relevance score between the text sequence and the target customer characteristics; for example, two algorithms can be used to calculate the relevance score: term frequency-inverse document frequency and probabilistic retrieval models.

[0067] In some embodiments, the first preset threshold is the lower limit of the relevance score. When the relevance score between the text sequence and the target customer's characteristics is greater than the threshold, the sample is determined to be highly relevant to the customer and is retained; otherwise, it is considered noise filtering. For example, if the overall relevance score of the data is greater than the first preset threshold, the data is retained. The first preset threshold can be dynamically adjusted according to the business scenario (e.g., conservative customers require higher relevance) or data distribution. For example, the first preset threshold can be set to 0.5~0.6 to balance sample quality and quantity.

[0068] For example, information entropy is used to measure the amount of information contained in a text sequence. A higher entropy value indicates a richer vocabulary and greater information density in the text. The steps for calculating the information entropy of a text sequence are as follows: Each text sequence is segmented into words to obtain a word sequence; the probability of each word appearing in the corpus composed of all text sequences of the client is calculated; the entropy value is then calculated using the information entropy formula. A higher entropy value indicates a more diverse vocabulary and a greater amount of information; a lower entropy value indicates excessive repetition or too many common words. Here, the information entropy can be calculated using commonly used information entropy formulas, such as Shannon information entropy, generalized entropy, and left entropy.

[0069] Taking Shannon information entropy as an example, based on the Shannon information entropy formula, the sample information entropy is calculated for core characteristics such as customer transaction frequency, product type, and profit fluctuation. After ablation testing with 100,000 real samples from banks, an information entropy threshold of 5.2 was determined to be the optimal threshold. Values ​​below this threshold introduce excessive noise data, while values ​​above it result in insufficient effective samples. 5.2 achieves a balance between noise filtering and customer coverage. The information entropy threshold is the second preset threshold in this embodiment. Setting the second preset threshold to 5.2 means that samples below this threshold are considered to have low information content (e.g., including a large number of repetitive and generic terms) and are discarded, while high-entropy samples are retained, filtering out low-quality noise data.

[0070] Valid data is determined based on relevance scores and information entropy. This process may include dual screening criteria: a relevance score greater than a first preset threshold and an information entropy greater than a second preset threshold. Text sequences that simultaneously meet both screening criteria are retained as valid data. Data that does not meet the dual screening criteria are considered noise or low-quality samples and are filtered out. The final output is a set of valid data for subsequent inference and analysis.

[0071] According to the embodiments of this application, valid data is determined based on relevance scores and information entropy, eliminating transaction records that are irrelevant to customer profiles, reducing noise, and reducing the burden of subsequent model processing; it can avoid a large number of repetitive and meaningless general descriptions occupying the model input window, retain high-information samples, improve the quality of input data, and improve inference efficiency and accuracy; it also reduces unnecessary data transmission and computational overhead.

[0072] For example, during the data preparation phase, approximately 1,000 original records of nearly 12 months' worth of financial data for Customer A (risk level C2, investment period 1-3 years) were extracted from the bank's core system. Each record generated a text sequence. The word frequency-inverse document frequency and probability retrieval model scores were calculated between the text sequence and the target customer's characteristics. For example, a record "redeeming 500,000 yuan of equity products" had a low correlation with "risk level C2 (conservative)" with a score of 0.3; while "increasing fixed income+ products by 100,000 yuan" had a high correlation with "risk level C2 (conservative)" with a score of 0.85. The information entropy of the text sequences was calculated, and high-quality samples with an information entropy of not less than 5.2 were retained. For example, a record "Customer A bought a product" contained a single term and had a low entropy value (e.g., information entropy of 2.1); another record "Customer A redeemed 500,000 yuan of equity products on XXXX year XX month XX day due to market volatility hedging" contained multiple terms and had a high entropy value (e.g., information entropy of 6.3). Samples with a correlation score > 0.5 and an entropy value > 5.2 were retained, resulting in over 500 valid data entries (including 3 large transactions and 8 held products). The retained valid data was then used for inference analysis using a pre-trained inference model, outputting the inference analysis results.

[0073] According to an embodiment of this application, the training of the inference analysis model includes: fine-tuning the inference analysis model to be trained based on a preset amount of scenario long-chain thinking data, so that the inference analysis model generates intermediate inference steps according to a preset logical link during the inference analysis stage, wherein the preset logical link follows a logical order from transaction motivation to risk exposure to optimization direction, and wherein the inference analysis model performs consistency verification on the preset logical link based on rule engine and semantic similarity during the inference analysis stage.

[0074] For example, a scenario-based long-chain thinking dataset is constructed, containing, for instance, 600,000 "question-reasoning chain-answer" triples from various financial management scenarios. The reasoning chain for each data point is written strictly according to the logical order of "trading motivation → risk exposure → optimization direction," and is annotated by financial experts or generated using a semi-automatic method.

[0075] Taking the financial long-chain thinking data as an example, the long-chain thinking reasoning logic is as follows. Input: Customer profile: C3 balanced type, with a 45% holding in the new energy industry in the past 3 months, a trading frequency of 22 times / month, and an average holding period of 8 days. The long-chain reasoning process may include: Identifying trading motives: Trading frequency > 20 times / month and holding period < 10 days, judged as short-term speculative motives; Quantifying risk exposure: The 45% holding in the new energy industry exceeds the industry concentration threshold of 30%, indicating a high concentration risk exposure; Assessing risk matching degree: The customer's risk level is C3, and the high concentration exposure does not match the customer's risk tolerance, with a matching degree score of 0.4; Generating optimization suggestions: Based on the suitability requirements for C3 customers, it is recommended to reduce the proportion of new energy holdings to 20%-30%, while allocating broad-based index funds to diversify industry risk. Output: Customer investment behavior insight: Significant short-term speculative behavior and high industry concentration risk; Assessing the health of holdings: The industry concentration risk exposure is high, with a low matching degree with the customer's risk tolerance; Optimization suggestions: Reduce the proportion of new energy holdings and allocate broad-based index funds to diversify risk.

[0076] In some embodiments, the fine-tuning process can employ supervised learning: customer financial data is used as input, and the corresponding reasoning chain is used as output to update the parameters of a pre-trained reasoning analysis model (large language model). Through training on a large number of long-chain thinking dataset samples, the model learns a structured paradigm for financial reasoning: instead of directly outputting conclusions, it first generates intermediate reasoning steps and gradually derives the final recommendation. This fine-tuning enables the model to automatically generate text according to a preset logical chain during the reasoning stage. For example: the transaction motivation is: "The customer has reduced its holdings of equity products in the past 6 months to avoid market volatility"; the risk exposure is: "The proportion of bond holdings is 82%, and the liquidity coverage ratio is only 18%, which is lower than the safety threshold of 25%"; the optimization direction is: "It is recommended to supplement with short-term liquidity products worth 50,000 to 100,000 yuan".

[0077] When deploying the fine-tuned model for inference, the generated inference chain still needs to undergo internal consistency verification to ensure logical coherence and consistency. Verification methods can include rule engines and semantic similarity calculations.

[0078] The rules engine detects logical conflicts and handles deterministic and quantifiable compliance rules, such as the completeness of risk warnings, verification of prohibited statements, and compliance of data definitions. The rules engine predefines a base of financial logic rules. For example, if a previous step determined that "the client has a conservative risk appetite," then subsequent rules such as "suggest allocating high-volatility equity products" should not appear; similarly, if the risk exposure is assessed as "insufficient liquidity," then the optimization direction must include "supplementing with short-term products." The rules engine matches each rule in the reasoning chain. Once a conflict is detected (e.g., a contradiction between trading motivation and risk exposure, or a mismatch between risk exposure and optimization direction), it triggers correction or regeneration.

[0079] Semantic similarity is used to assess coherence and handles unstructured, semantic-based compliance risks, such as implicit exaggeration and inappropriate statements. The three links in the inference chain (transaction motive, risk exposure, and optimization direction) are extracted as text fragments. These fragments are then converted into feature vectors using a pre-trained semantic embedding model, and the cosine similarity between each pair is calculated: the cosine similarity between transaction motive and risk exposure reflects the logical coherence between the motive and risk assessment; the cosine similarity between risk exposure and optimization direction reflects the consistency between the assessment results and recommendations. If the similarity is below a preset threshold (e.g., 0.7, which can be dynamically adjusted based on business scenarios or data distribution), it indicates a jump in the inference chain or a deviation from the theme, and is considered inconsistent, requiring regeneration or adjustment.

[0080] For example, the core logic of determining the internal consistency of the inference chain through semantic similarity scoring can be as follows: taking the logical chain of transaction motivation → risk exposure → optimization direction as the verification object, firstly, the rule engine is used to check for logical conflicts, and then the similarity between the text and semantics of each link in the chain is calculated. The higher the semantic matching degree and the fewer the logical contradictions, the higher the corresponding score. This quantifies whether the internal inference chain is coherent and conflict-free. Finally, together with the verification result of the rule engine, it constitutes the scoring basis for the internal consistency of the inference chain.

[0081] The rule engine provides deterministic logical constraints, while semantic similarity provides soft semantic constraints. The combination of rule engine verification and language similarity verification can comprehensively verify the internal consistency of the inference chain. During the training phase, the verification results can be used as part of the language consistency reward for reinforcement learning training; during the inference phase, it serves as a real-time quality control module, ensuring that the output inference analysis results are both consistent with financial logic and linguistically coherent.

[0082] According to embodiments of this application, fine-tuning the inference analysis model to be trained using long-chain thinking data can enhance the interpretability and logical coherence of the model's reasoning, avoid reasoning jumps or contradictions, and make the final recommendations more consistent with financial logic. In the inference analysis stage, consistency verification of the preset logical links based on rule engines and semantic similarity can further improve the reliability of the inference conclusions.

[0083] The above describes the introduction of a thought chain reasoning mechanism when training the reasoning analysis model. This involves fine-tuning the model based on a predetermined amount of long-chain thought data from various scenarios, enabling the model to generate intermediate reasoning steps according to a pre-defined logical chain during the reasoning analysis phase. Training the reasoning analysis model can also include using reinforcement learning algorithms.

[0084] According to an embodiment of this application, the reinforcement learning algorithm is a group relative policy optimization algorithm. The training of the inference analysis model includes: for each input sample, generating N candidate inference results through the inference analysis model to be trained, the N candidate inference results forming an inference group, where N is an integer greater than 1; calculating the total reward value of each candidate inference result; and updating the model parameters of the inference analysis model to be trained based on the total reward value of each candidate inference result in the inference group.

[0085] According to an embodiment of this application, the reasoning analysis model is equipped with a dual-drive reward function, which includes an accuracy reward item and a language consistency reward item. The total reward value is the weighted sum of the accuracy reward item and the language consistency reward item. The accuracy reward item is used to evaluate the consistency between the candidate reasoning result and the preset logical rules, and the language consistency reward item is used to evaluate the logical coherence and terminology standardization of the candidate reasoning result.

[0086] For example, the group relative policy optimization algorithm is a reinforcement learning algorithm applied to large language models. It removes the value model, generates multiple candidate outputs for the same input to form a group, calculates the mean and standard deviation of the reward within each group, and updates the policy based on relative advantage (the degree of reward deviation), causing the model to tend towards higher reward outputs, thereby reducing training overhead and resource consumption. The standard deviation and mean, combined, eliminate reward scale bias between different groups by standardizing the reward value, and also enable adaptive gradient step size adjustment. It should be noted that when training the inference analysis model, fine-tuning the model using long-chain reasoning data is usually performed before training with the group relative policy optimization algorithm, thus providing the model with initial inference chain generation capabilities and avoiding completely invalid outputs during the cold start of reinforcement learning.

[0087] For example, generating candidate inference results (constructing inference groups) may include: for a sample of customer financial data input (e.g., customer A's transaction records and profile data), the inference analysis model to be trained will generate N different candidate inference results in parallel, where N is typically 4 to 8. These candidate results constitute an inference group. Each candidate inference result is organized in a thought chain manner, containing a complete inference chain from "trading motivation" to "risk exposure" to "optimization direction".

[0088] The core of the group-relative policy optimization algorithm lies in intra-group comparison, rather than relying on an external value network. For example, specific steps may include: calculating the mean of the total reward for all candidate outcomes within the group; comparing the total reward value of each candidate outcome with the mean: those with a total reward value higher than the mean are considered "good outcomes," and those lower than the mean are considered "bad outcomes"; when updating model parameters, increasing the probability of generating "good outcomes" and decreasing the probability of generating "bad outcomes"; through repeated iterations, the model gradually learns to generate inference outcomes with higher rewards, which are both accurate and logically coherent.

[0089] According to the embodiments of this application, the group relative strategy optimization algorithm is used to train the inference analysis model, which can simplify the training architecture, reduce memory usage, accelerate the convergence speed, and is not sensitive to hyperparameters, making it more suitable for the standardized deployment environment of bank branches.

[0090] For example, calculating the total reward value for each candidate outcome may include: using a dual-drive reward function to calculate the total reward value, as shown in the following formula.

[0091]

[0092] in, For accuracy-related bonus items, As a language consistency reward item, The total reward value, b is a dynamically adjustable weighting coefficient that can be adjusted based on business scenarios or data distribution. For example, The value of b is 0.5.

[0093] Accuracy rewards automatically assess the correctness, consistency with financial facts, and adherence to pre-defined rules based on a pre-defined financial logic rule base (such as risk matching and liquidity safety thresholds). Accuracy rewards can also automatically verify the reasonableness and effectiveness of the model's output numerical recommendations based on a quantitative recommendation result verification module. For example, the module can call a backend calculation engine to simulate the execution of the model's quantitative recommendations based on the client's current holdings and financial data, verifying whether the results match the model's claims. If the verification passes, a positive reward is given. Semantic rewards assess the coherence of the reasoning chain (e.g., whether "trading motivation" and "optimization direction" contradict each other) and the standardization of terminology usage through a rule engine and semantic similarity calculations. Each candidate result receives a comprehensive total reward value; a higher score indicates better reasoning quality.

[0094] Based on financial business labeled data, the accuracy reward is calculated by weighting two dimensions: node hit rate and conclusion matching rate. Node hit rate: The F1 score of key nodes in the inference chain, such as transaction motives, risk exposure, and matching degree, compared to the expert-annotated results. The F1 score is a core indicator in machine learning used to measure the degree of matching between the model's prediction results and the human standard answer. It ranges from 0 to 1, with values ​​closer to 1 indicating a more accurate match. The F1 score comprehensively measures whether the model has missed any key nodes or misjudged any nodes, accurately reflecting the model's node identification ability. Conclusion matching rate: The cosine similarity between the model's final conclusion and the business judgment benchmark. The business judgment benchmark is a predefined standard for customer risk and transaction behavior, including: regulatory / industry risk control rules (e.g., large and frequent transfers in a single day = high risk); unified expert conclusions (the final judgment of business experts on the customer's risk level and optimization direction); and compliance standards (official judgments on anti-money laundering and credit risk control). The weighted average of node hit rate and conclusion matching rate constitutes the accuracy reward score, directly reflecting the accuracy of the conclusion.

[0095] Language consistency rewards are quantified through two-way verification: Internal consistency within the inference chain: This checks for conflicts in the logical chain of "trading motivation → risk exposure → optimization direction," using a rule engine for verification and semantic similarity scoring; and consistency between reports and customer profiles: This verifies whether the language used is appropriate for the customer's risk level and profile tags. For example, using low-risk language for conservative customers and avoiding overload with technical jargon. The combination of internal consistency within the inference chain and consistency between reports and customer profiles constitutes the language consistency reward, ensuring logical coherence and scenario adaptation.

[0096] Verifying whether language expression matches customer risk levels and profile tags involves text semantic matching, tag rule verification, and large model alignment verification. In the financial field, text semantic matching involves converting customer profile tags (conservative / aggressive, high-risk / low-risk) and risk levels into vectors, and simultaneously converting the model-generated report text into a vector. The similarity between the two is calculated to verify whether the expression style matches. Keyword + rule engine verification involves building a language rule base: for conservative customers, use terms like "stable," "low volatility," and "low risk," and prohibit the use of terms like "aggressive," "leveraged," and "high-yield"; for high-risk customers, use expressions like "risk hedging" and "compliance management." The text keywords are directly verified for compliance. Classification model verification involves training a lightweight text classification model, inputting the model's output report, and determining whether the text tone / style matches the customer risk profile tags.

[0097] According to embodiments of this application, the dual-drive reward function-guided model simultaneously optimizes the correctness of reasoning and the professionalism of expression in reinforcement learning, making the generated financial advice both factually accurate and compliant with bank compliance requirements, and improving the accuracy of understanding professional terminology in the report.

[0098] The report generation method provided in this example embeds a group relative strategy optimization algorithm, employing a dual-drive mechanism of accuracy rewards and language consistency rewards to ensure coherent reasoning and accurate conclusions. Combined with 600,000 data points from long-chain financial scenarios, it enables hierarchical reasoning of customer behavior, such as transaction motivation → risk matching degree → optimization direction. Outputs can include: insights into customer investment behavior, portfolio health assessments (including quantitative values ​​of risk exposure), and potential optimization suggestions.

[0099] According to the embodiments of this application, the group relative strategy optimization algorithm is used to optimize the coherence, professionalism and compliance of long chain reasoning, and the exclusive long chain thinking data provides the group relative strategy optimization algorithm with scenario-based decision-making basis. The two are adapted to each other and optimized in a coordinated manner to form an exclusive optimization scheme for reasoning analysis.

[0100] For example, the inference analysis model can embed a multi-head latent attention mechanism. After inputting data, the multi-head latent attention mechanism quickly locates key trading nodes. The group relative strategy optimization algorithm performs inference through a double reward mechanism and outputs the following conclusion: "Client A has reduced its holdings of equity products (to avoid market volatility) and increased its holdings of fixed-income+ products (fitting for C2 level) in the past 6 months. The proportion of bonds in its holdings is 82%, and the liquidity coverage ratio is only 18% (below the safety threshold of 25%). It is recommended to supplement with short-term liquidity products of RMB 50,000 to RMB 100,000." This inference process takes only 2.3 seconds.

[0101] The inference analysis model provided in the embodiments can integrate multiple algorithms related to large models to further improve model performance, speed up response, and enhance the quality of inference analysis.

[0102] According to an embodiment of this application, the inference analysis model is a model that has undergone distillation processing, and the context window of the inference analysis model is extended to a preset length through rotational position encoding extension technology, wherein the preset length supports long-term transaction data.

[0103] For example, the inference analysis model provided in the embodiment can adopt, for example, a distilled large language model with 32B (32 billion) parameters, which retains the core inference capabilities and enables lightweight deployment, thereby balancing the performance and cost of the inference analysis model.

[0104] By extending the context window to 128K (1K=1024 bytes) through rotational positional encoding extension technology, it is possible to support long-term transaction data inference. For example, by adjusting the positional encoding method of the model, a model that could originally only process a few thousand words (e.g., 4K) can now process tens of thousands of words (e.g., 128K) of long text after a few minor adjustments.

[0105] As mentioned above, the inference analysis model can also activate a multi-head latent attention mechanism to reduce the memory overhead of the key-value cache; and enable 8-bit floating-point quantization technology to further reduce the GPU memory requirements. For example, by using a multi-head attention mechanism, the dimensionality of the key-value cache is reduced through low-rank compression, and combined with 8-bit floating-point quantization, the GPU memory pressure is further alleviated.

[0106] According to embodiments of this application, knowledge distillation reduces the number of parameters in the inference analysis model, saves system resources, and improves inference analysis efficiency. The context window of the inference analysis model is expanded to handle transaction records spanning several years, enhancing the model's processing capabilities. By integrating distillation techniques, 8-bit floating-point quantization, and multi-head attention mechanisms, a balance between model performance and deployment costs is achieved, adapting to the lightweight requirements of bank branches.

[0107] The above describes the execution steps for inputting valid data into the inference analysis model and outputting the inference analysis results. After obtaining the inference analysis results, the report generation method described in this embodiment uses the report generation model to output a structured financial report. By collaborating with the inference analysis model and the report generation model, end-to-end intelligent generation and efficient deployment from customer data to professional reports are achieved, improving the efficiency of customers purchasing bank financial products.

[0108] For example, the output of the inference analysis model can be input into the report generation model in the form of structured JSON data, including core fields such as customer investment behavior insights, portfolio health assessment (including quantitative values ​​of risk exposure), and optimization suggestions. At the same time, structured tags for key nodes in the inference chain can be attached to ensure that the report generation model can directly call business data and logical basis when generating reports, achieving a precise mapping from analysis conclusions to report text.

[0109] For example, the report generation model can adopt a basic large language model with 70B (70 billion) parameters, and the number of transformer layers can be set to 61 layers. The first 3 layers are dense layers to ensure the semantic foundation, and the 4th to 61st layers enable a hybrid expert architecture (containing 256 expert sub-networks, with each token activating 8 experts) to improve computational efficiency.

[0110] According to an embodiment of this application, the hybrid expert architecture includes: constructing at least a portion of the converter layers in the report generation model as M expert sub-networks, where M is an integer greater than 1; for each input token, calculating the activation weight of each expert sub-network of the current converter layer through a routing network, selecting a predetermined number of expert sub-networks for activation based on the activation weights; and weighting the outputs of the activated expert sub-networks based on their activation weights as the output of the current converter layer.

[0111] For example, at least a portion of the transformer layers in the report generation model can be a preset number of transformer layers, such as 10 layers, 42 layers, 45 layers, etc. For instance, layers 4 to 61 of the transformer can be set as hybrid expert architecture layers, with each layer from layer 4 to 61 constructed as 256 (M=256) expert subnetworks. Each expert gradually specializes in different financial subtasks during training, such as transaction behavior analysis, risk exposure quantification, compliance script generation, and structured text writing.

[0112] In some embodiments, the 256 expert subnetworks can adopt a division of labor model based on random initialization, business-driven automatic differentiation, and functionally oriented constraints. Initially, the parameters of all expert subnetworks can be randomly initialized without pre-defined functional divisions. During fine-tuning training on the financial report dataset, expert automatic clustering and differentiation are achieved based on domain task-driven approaches. Simultaneously, routing constraint loss is introduced to guide the expert subnetworks to specialize in differentiated financial sub-tasks such as transaction behavior analysis, risk exposure quantification, compliance script generation, and structured text writing, thus achieving specialized division of labor at the functional level.

[0113] For each input token, the activation weights of each expert subnetwork in the current converter layer are calculated using a routing network. For example, a gated routing network can be used to calculate the activation weights of each expert subnetwork in the current converter layer. The gated routing network is a learnable linear layer that takes the representation vector of the current token as input and outputs a 256-dimensional weight vector. This weight is normalized so that the sum of all weights is 1. The magnitude of the weight reflects the degree of matching between the token and each expert subnetwork. The routing network selects the top 8 experts by weight as activation experts, and the remaining 248 experts do not participate in the calculation of the current token. This design ensures sparse activation, using only 3.125% of the experts for each token, significantly reducing the computational cost. It should be noted that using a gated routing network to calculate the activation weights of each expert subnetwork in the current converter layer is only an example and not a limitation of the embodiments of this application. In other embodiments, other methods can also be used to calculate the activation weights of each expert subnetwork in the current converter layer, such as linear routing, attention routing, reinforcement learning routing, etc.

[0114] The eight selected experts each perform independent forward computation on the input token, outputting eight vectors. These output vectors are then multiplied by the corresponding weights given by the routing network, and summed to obtain the output of the current transformer layer, which is then passed to the next layer for further processing.

[0115] According to embodiments of this application, sparse activation improves computational efficiency, enabling the model to run in real-time on branch-level graphics processing units. The model can dynamically combine the most suitable expert outputs for the current token to generate high-quality structured financial reports.

[0116] For example, to ensure training stability and balanced expert utilization, the method in this embodiment can also introduce three constraints: Load balancing loss function constraint: Monitoring the activation frequency of each expert, if some experts are activated too infrequently or too frequently, a penalty term is added to encourage uniform activation of all experts, avoiding idleness or overload. Orthogonalization training constraint: Adding a regularization term to the loss function to ensure that the output vectors of different experts are orthogonal to each other, thereby guiding each expert to learn differentiated financial sub-tasks and achieving functional complementarity. Learnable routing network constraint: The weight calculation method described above itself guarantees that duplicate experts are not activated, because the top 8 experts are all different.

[0117] For example, to further improve the report generation speed, the method of the embodiment introduces a multi-token parallel prediction mechanism to generate multi-module content of the report in parallel.

[0118] According to an embodiment of this application, the multi-token parallel prediction mechanism includes: decomposing the structured financial report to be generated into K report modules, where K is an integer greater than 1; and generating a token sequence for the K report modules in parallel.

[0119] For example, a structured financial report may include four loosely coupled and relatively independent modules (K=4): a transaction summary module (structured organization of client holdings, transaction flow, and return data); a data interpretation module (quantitative analysis such as return attribution, risk exposure, and volatility characteristics); an optimization suggestion module (asset allocation schemes such as rebalancing direction, allocation ratio, and product recommendations); and a risk warning module (generation of warnings regarding compliance risk, market risk, and liquidity risk). These four modules are independent in content and have no sequential dependency, providing feasibility for parallel generation. Each module relies only on the structured analysis results output by the inference analysis model, with no strong temporal dependency, and can be generated simultaneously by the model, and then finally merged and assembled according to the report format. It should be noted that the inclusion of four relatively independent modules in a structured financial report is merely an example and not a limitation of the embodiments of this application. In other embodiments, the structured financial report may also include, for example, a risk handling module.

[0120] Taking the decomposition of the structured financial report to be generated into four report modules as an example, the multi-token parallel prediction mechanism achieves parallelism in the following ways: Setting independent prediction heads: The model sets an independent prediction head (i.e., an independent linear classification layer) for each module in the output layer. Each prediction head is responsible for generating the token sequence for the corresponding module. Parallel decoding: After the input (inference analysis result) is processed by the shared hybrid expert architecture layer to obtain the hidden state, the four prediction heads can start working simultaneously. Each head independently performs autoregressive generation: first, it generates the first token of the module, and then continues to generate the next token based on the generated intra-module prefix. The generation processes of each module do not wait for or interfere with each other. Synchronous output: When the generation of all modules is completed, the texts of the four modules are concatenated into a complete structured report in a predetermined order.

[0121] For example, the outputs of the four modules could be: the transaction summary module outputs the number of transactions in the past 12 months and details of large transactions; the data interpretation module outputs the proportion of bonds held and the liquidity coverage ratio assessment; the optimization suggestion module outputs "Recommend XX Bank 7-day notice deposit wealth management product of RMB 80,000"; and the risk warning module outputs standard warning messages.

[0122] For example, the report generation model in this embodiment can also be combined with low-rank adaptation technology, using, for example, 12,000 financial reporting instructions for fine-tuning to improve the accuracy of understanding domain terminology. Examples of selected instructions are as follows: Based on the client's transaction history, holdings, and profit data over the past six months, analyze trading motives, quantify portfolio risk exposure, and generate a structured financial analysis report suitable for conservative clients; combine the client's asset allocation structure to identify potential portfolio risks, output compliant asset optimization suggestions and risk warnings, and control the density of professional terminology; conduct logical analysis of high-frequency trading behavior for aggressive clients, quantify portfolio volatility risk, and output actionable portfolio adjustment optimization directions; verify the degree of matching between client trading behavior and risk level, and output client behavior insights, risk quantification results, and compliant financial improvement suggestions; integrate multi-dimensional financial data from clients to generate a standardized financial report covering transaction summaries, data interpretation, optimization suggestions, and risk warnings. The core parameter configuration is fine-tuned using low-rank adaptation technology. Only the query and value matrices of the attention layer are unfrozen, while the backbone parameters of the model are frozen. The core hyperparameters are configured as follows: rank dimension r=32, initial learning rate is set to 3x10^-4, and a cosine annealing learning rate scheduling strategy is adopted; batch size is 8, training epochs=4, and weight decay coefficient is 0.01; gradient checkpointing and 8-bit floating-point quantization are enabled during training, which effectively reduces memory overhead while ensuring the fine-tuning effect of the 70B large model.

[0123] For example, generating a structured financial report based on a report generation model can include: dynamically activating eight expert sub-networks such as "report structure organization" and "financial terminology adaptation" through a hybrid expert architecture; generating four modules of content in parallel through multi-token parallel prediction technology; and accurately outputting the following through a model fine-tuned by low-rank adaptation: "Recommended allocation of 'XX Bank 7-day notice deposit wealth management' (expected return 2.5%, redemption on the next trading day), allocation amount of 80,000 yuan, to increase liquidity coverage ratio to 28%, and control overall return fluctuation within ±3%."

[0124] The output of the report generation model can be a structured report, which may include transaction summaries, data interpretations, optimization suggestions, risk warnings, etc., and supports multiple output formats such as portable document format, hypertext markup language, and voice.

[0125] According to the embodiments of this application, a multi-token parallel prediction mechanism is used to generate multiple modules of report content in parallel, which improves the report generation speed and meets the high throughput requirement of 100,000 reports per day; at the same time, it ensures that the content between modules is independent, which facilitates subsequent compliance verification and local correction.

[0126] According to embodiments of this application, a structured financial report is generated through dual-model collaboration, embedding group relative strategy optimization algorithms, long-chain thinking data reasoning, and a hybrid expert architecture. This solves the problems of "shallow interpretation, low efficiency, and weak personalization" in bank financial transaction reports, improving reasoning efficiency and report generation efficiency, and enhancing the accuracy of risk warnings. The generated structured financial report can meet customers' personalized needs, allowing them to clearly obtain financial advice tailored to their individual circumstances, thus improving the degree of personalization.

[0127] The report generation method provided in this embodiment supports lightweight deployment: both private and cloud deployments are supported. With cache hits, it significantly reduces the cost of application programming interface calls, resolving the issue of "heavy deployment" in bank wealth management transaction reports.

[0128] According to embodiments of this application, the contradiction between long-cycle financial data inference and low-cost deployment is resolved by customizing the parameters of the inference analysis model / report generation model (e.g., long context window, quantitative techniques) to suit the characteristics of bank wealth management data.

[0129] For example, to avoid compliance risks in the generated structured financial reports, the method in this embodiment can also perform compliance checks on the generated structured financial reports. The following will combine... Figure 4 Please provide a detailed explanation.

[0130] Figure 4 The flowchart illustrating compliance verification of a structured financial report according to an embodiment of this application is shown in the illustration.

[0131] According to an embodiment of this application, compliance verification of a structured financial report may include steps S410 to S420.

[0132] In step S410, the structured financial report is subjected to compliance verification, which is based on a preset compliance rule base and classification model.

[0133] In step S420, the structured financial report that has passed the compliance verification is output in a preset format.

[0134] A pre-defined compliance rule base can be a rule library built through compliance analysis of rules across a large number of domains. It can be constructed using large models or through expert annotation. The compliance rule base can include a pre-defined number of compliance rules, for example, over 200 rules covering regulatory requirements and internal bank risk control standards. For example, compliance rules might prohibit exaggerated returns such as "principal guaranteed" or "guaranteed profit"; require the inclusion of standard risk warnings; and prohibit recommending wealth management products with equity components exceeding 30% for conservative clients (e.g., C2 level and below).

[0135] The classification model can be a lightweight text classifier fine-tuned with a small number of samples (e.g., a small model based on a bidirectional encoder representation using a transformer). Its training data can consist of a small number of positive and negative samples (approximately 1000-2000) labeled by bank compliance experts. The input to the classification model can be the full text of a structured financial report, and the output can be a binary classification result of "compliant" or "non-compliant". The classification model can be used to identify compliance risks in reports that are "semantic or ambiguous," such as exaggerated returns or missing risk warnings.

[0136] For example, if the compliance verification of the structured financial report is performed based on the compliance rule base and no violation rule is matched (i.e., the structured financial report has passed the check of all explicit rules), then the second level of judgment is entered, at which point the classification model is called to perform compliance verification of the structured financial report.

[0137] For example, the preset format can refer to the target file format output after the financial report is generated. This could be a portable document format, Hypertext Markup Language (HTML), or voice. The portable document format is used for formal printing or archiving, HTML is used for embedded display in mobile banking or online banking, and voice is used for visually impaired customers or in voice broadcast scenarios. For instance, taking the portable document format as an example, after successful verification, a portable document format report is generated and pushed to the customer's mobile banking app.

[0138] According to the embodiments of this application, compliance verification of structured financial reports reduces compliance risks and saves on manual verification costs. The compliance verification adopts a two-layer architecture of a pre-set compliance rule base and a classification model. The rule engine handles explicit and enumerable compliance rules, while the classification model handles ambiguous or complex situations that are difficult for the rules to cover. The two work together to achieve highly accurate automated verification. Structured financial reports that pass compliance verification are output in a pre-set format, supporting multiple format outputs and improving customer experience.

[0139] For example, compliance verification is performed on structured financial reports. The verification process can include complete risk warnings and ensure that there are no illegal statements. Once the verification is passed, a report in a portable document format is generated and pushed to the customer's mobile banking app within 100 milliseconds.

[0140] According to the embodiments of this application, compliance verification of structured financial reports includes: matching each rule of the structured financial report against a preset compliance rule base; blocking the structured financial report in response to a matched violation; and using a classification model to determine compliance in response to no matched violation, and outputting the classification result.

[0141] For example, a rule engine (a software module that performs automated judgment based on predefined logical rules) matches the structured financial report against a pre-defined compliance rule base one by one. If a match is found, the report is blocked to ensure compliance requirements. For example, the rule engine traverses every sentence in the structured financial report and performs pattern matching (keyword matching, regular expressions, semantic templates, etc.) against the pre-defined compliance rule base. If any violation is found, the report is immediately determined to be non-compliant, the blocking operation is performed, and the reason for the violation is recorded for subsequent manual review or model correction.

[0142] When the rule engine does not detect any violations, a classification model makes a compliance judgment. The role of the classification model is to capture compliance risks that the rule engine cannot quantify, such as: whether the tone is too aggressive (e.g., using words like "leverage" or "high returns" for conservative clients); whether the risk warning is sufficiently prominent (e.g., although there is a warning, its location is obscure or the font is too small); and whether there are suggestive violations in the context (e.g., "Although we cannot guarantee principal, we have never had a loss in the past"). The classification model performs semantic analysis on the entire text to determine whether there are risks that are difficult to quantify according to the rules, and outputs a binary classification result of "compliant / non-compliant".

[0143] For example, the classification model's judgment result can be used as the final output. If it is determined to be non-compliant, a report is blocked; if it is determined to be compliant, the next output step is performed.

[0144] Figure 5 The flowchart illustrating the generation of a compliant financial report according to an embodiment of this application is shown in the illustration.

[0145] like Figure 5 As shown, generating a compliant financial management report may include steps S510 to S540.

[0146] In step S510, data preparation is performed. This step may include filtering financial data to obtain valid data.

[0147] In step S520, the inference analysis model performs inference. This step may include inputting valid data into the inference analysis model and outputting inference analysis results.

[0148] In step S530, the report generation model generates a report. This step may include inputting the inference analysis results into the report generation model and outputting a structured financial report.

[0149] In step S540, the report undergoes compliance verification. This step may include verifying the structured financial report for compliance and outputting the structured financial report that passes the compliance verification in a preset format.

[0150] According to embodiments of this application, compliance rules are automatically verified based on a rule engine and a classification model fine-tuned with a small number of samples, significantly improving verification accuracy. The rule engine performs real-time matching, and the lightweight and fast classification model meets the throughput requirements of 100,000 reports per day. Combining the high accuracy of the rule engine and the generalization ability of the classification model for compliance verification improves the overall verification accuracy while ensuring the real-time generation of reports.

[0151] The method described in this embodiment provides a full-chain solution of "model parameter optimization + financial scenario adaptation + lightweight deployment", which saves labor costs and reduces deployment costs, and promotes the large-scale application of artificial intelligence in the field of financial reporting.

[0152] For example, the report generation method provided in the embodiment can be deployed on a branch's private image processor cluster, supporting the generation of 100,000 reports per day, and running smoothly without lag.

[0153] Based on the above report generation method, embodiments of this application also provide a report generation apparatus. The following will be combined with... Figure 6 The device is described in detail.

[0154] Figure 6 A schematic block diagram of a report generation apparatus according to an embodiment of this application is shown.

[0155] like Figure 6 As shown, the report generation device 600 of this embodiment includes a data extraction module 610, an inference analysis module 620, and a report generation module 630.

[0156] The data extraction module 610 is used to acquire financial management-related data of the target customer, filter the financial management-related data, and obtain valid data. The financial management-related data includes at least transaction data, customer profile data, and product data. In one embodiment, the data extraction module 610 can be used to perform step S210 described above, which will not be repeated here.

[0157] The inference analysis module 620 is used to input valid data into the inference analysis model and output inference analysis results. The inference analysis model employs reinforcement learning algorithms and a thought chain inference mechanism. In one embodiment, the inference analysis module 620 can be used to execute step S220 described above, which will not be repeated here.

[0158] The report generation module 630 is used to input the reasoning analysis results into the report generation model and output a structured financial report. The report generation model employs a hybrid expert architecture and a multi-token parallel prediction mechanism. In one embodiment, the report generation module 630 can be used to execute step S230 described above, which will not be repeated here.

[0159] According to an embodiment of this application, the data extraction module 610 includes: a data conversion unit 611, used to convert structured data in financial management-related data into a text sequence; a correlation calculation unit 612, used to calculate the correlation score between the text sequence and the target customer characteristics using a correlation evaluation algorithm, wherein the target customer characteristics are derived from customer profile data; an information entropy calculation unit 613, used to calculate the information entropy of the text sequence; and a data determination unit 614, used to determine valid data based on the correlation score and information entropy, wherein the correlation score corresponding to the valid data is greater than a first preset threshold, and the information entropy corresponding to the valid data is greater than a second preset threshold.

[0160] According to an embodiment of this application, the training of the inference analysis model includes: fine-tuning the inference analysis model to be trained based on a preset amount of scenario long-chain thinking data, so that the inference analysis model generates intermediate inference steps according to a preset logical link during the inference analysis stage, wherein the preset logical link follows a logical order from transaction motivation to risk exposure to optimization direction, and wherein the inference analysis model performs consistency verification on the preset logical link based on rule engine and semantic similarity during the inference analysis stage.

[0161] According to an embodiment of this application, the reinforcement learning algorithm is a group relative policy optimization algorithm. The training of the inference analysis model includes: for each input sample, generating N candidate inference results through the inference analysis model to be trained, the N candidate inference results forming an inference group, where N is an integer greater than 1; calculating the total reward value of each candidate inference result; and updating the model parameters of the inference analysis model to be trained based on the total reward value of each candidate inference result in the inference group.

[0162] According to an embodiment of this application, the reasoning analysis model is equipped with a dual-drive reward function, which includes an accuracy reward item and a language consistency reward item. The total reward value is the weighted sum of the accuracy reward item and the language consistency reward item. The accuracy reward item is used to evaluate the consistency between the candidate reasoning result and the preset logical rules, and the language consistency reward item is used to evaluate the logical coherence and terminology standardization of the candidate reasoning result.

[0163] According to an embodiment of this application, the inference analysis model is a model that has undergone distillation processing, and the context window of the inference analysis model is extended to a preset length through rotational position encoding extension technology, wherein the preset length supports long-term transaction data.

[0164] According to an embodiment of this application, the hybrid expert architecture includes: constructing at least a portion of the converter layers in the report generation model as M expert sub-networks, where M is an integer greater than 1; for each input token, calculating the activation weight of each expert sub-network of the current converter layer through a routing network, selecting a predetermined number of expert sub-networks for activation based on the activation weights; and weighting the outputs of the activated expert sub-networks based on their activation weights as the output of the current converter layer.

[0165] According to an embodiment of this application, the multi-token parallel prediction mechanism includes: decomposing the structured financial report to be generated into K report modules, where K is an integer greater than 1; and generating a token sequence for the K report modules in parallel.

[0166] According to an embodiment of this application, the report generation device 600 further includes a compliance verification module 640, which is used to: perform compliance verification on the structured financial report, the compliance verification being based on a preset compliance rule base and classification model; and output the structured financial report that has passed the compliance verification in a preset format.

[0167] According to an embodiment of this application, the compliance verification module 640 includes: a rule matching unit 641, used to perform rule matching on the structured financial report based on a preset compliance rule library; and a compliance processing unit 642, used to: intercept the structured financial report in response to a matched violation; and to perform compliance judgment using a classification model in response to no matched violation, and output the classification result.

[0168] According to embodiments of this application, any plurality of modules among the data extraction module 610, inference analysis module 620, and report generation module 630 can be merged into one module, or any one of these modules can be split into multiple modules. Alternatively, at least part of the functionality of one or more of these modules can be combined with at least part of the functionality of other modules and implemented in one module. According to embodiments of this application, at least one of the data extraction module 610, inference analysis module 620, and report generation module 630 can be at least partially implemented as hardware circuitry, such as field-programmable gate arrays, programmable logic arrays, systems-on-a-chip, systems-on-a-substrate, systems-on-package, application-specific integrated circuits, or any other reasonable means of integrating or packaging circuitry, or implemented in software, hardware, or firmware, or in any appropriate combination of any of these three implementation methods. Alternatively, at least one of the data extraction module 610, inference analysis module 620, and report generation module 630 can be at least partially implemented as a computer program module, which, when run, can perform corresponding functions.

[0169] Figure 7 A block diagram schematically illustrates an electronic device suitable for implementing a report generation method according to an embodiment of this application.

[0170] like Figure 7 As shown, an electronic device 1200 according to an embodiment of this application includes a processor 1201, which can perform various appropriate actions and processes according to a program stored in a read-only memory 1202 or a program loaded from a storage portion 1208 into a random access memory 1203. The processor 1201 may include, for example, a general-purpose microprocessor, an instruction set processor and / or an associated chipset and / or a dedicated microprocessor. The processor 1201 may also include onboard memory for caching purposes. The processor 1201 may include a single processing unit or multiple processing units for executing different steps of the method flow according to an embodiment of this application.

[0171] Random access memory 1203 stores various programs and data required for the operation of electronic device 1200. Processor 1201, read-only memory 1202, and random access memory 1203 are interconnected via bus 1204. Processor 1201 executes various steps of the method flow according to embodiments of this application by executing programs in read-only memory 1202 and / or random access memory 1203. It should be noted that the programs may also be stored in one or more memories other than read-only memory 1202 and random access memory 1203. Processor 1201 may also execute various steps of the method flow according to embodiments of this application by executing programs stored in said one or more memories.

[0172] According to embodiments of this application, the electronic device 1200 may further include an input / output interface 1205, which is also connected to the bus 1204. The electronic device 1200 may also include one or more of the following components connected to the input / output interface 1205: an input section 1206 including a keyboard, mouse, etc.; an output section 1207 including a cathode ray tube, liquid crystal display, etc., and a speaker, etc.; a storage section 1208 including a hard disk, etc.; and a communication section 1209 including a network interface card, such as a local area network card, modem, etc. The communication section 1209 performs communication processing via a network such as the Internet. A drive 1210 is also connected to the input / output interface 1205 as needed. A removable medium 1211, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 1210 as needed so that computer programs read from it can be installed into the storage section 1208 as needed.

[0173] Embodiments of this application also provide a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiments of this application.

[0174] According to embodiments of this application, the computer-readable storage medium can be a non-volatile computer-readable storage medium, such as including but not limited to: portable computer disks, hard disks, random access memory, read-only memory, erasable programmable read-only memory, portable compact disk read-only memory, optical storage devices, magnetic storage devices, or any suitable combination thereof. In embodiments of this application, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, according to embodiments of this application, the computer-readable storage medium may include the read-only memory 1202, and / or random access memory 1203, and / or one or more memories other than read-only memory 1202 and random access memory 1203 described above.

[0175] Embodiments of this application also include a computer program product comprising a computer program containing program code for performing the methods shown in the flowchart. When the computer program product is run on a computer system, the program code is used to cause the computer system to implement the methods provided in the embodiments of this application.

[0176] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and may be downloaded and installed via the communication section 1209, and / or installed from the removable medium 1211. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0177] In embodiments of this application, the computer program can be downloaded and installed from a network via communication section 1209, and / or installed from removable medium 1211. When the computer program is executed by processor 1201, it performs the functions defined in the system of this application embodiment. According to embodiments of this application, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0178] According to embodiments of this application, program code for executing the computer programs provided in the embodiments of this application can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. The program code can be executed entirely on the user's computing device, partially on the user's device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0179] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0180] Those skilled in the art will understand that the features described in the various embodiments of this application can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, the features described in the various embodiments of this application can be combined and / or combined in various ways without departing from the spirit and teachings of this application. All such combinations and / or combinations fall within the scope of this application.

Claims

1. A report generation method, characterized in that, The method includes: Obtain financial management-related data of target customers, filter the financial management-related data to obtain valid data, and the financial management-related data includes at least transaction data, customer profile data and product data; The valid data is input into the inference analysis model, and the inference analysis results are output. The inference analysis model employs reinforcement learning algorithms and a thought chain inference mechanism. The reasoning and analysis results are input into the report generation model, which outputs a structured financial report. The report generation model adopts a hybrid expert architecture and a multi-token parallel prediction mechanism.

2. The method according to claim 1, characterized in that, The filtering of the financial management-related data to obtain valid data includes: Convert the structured data in the financial management-related data into a text sequence; A relevance assessment algorithm is used to calculate the relevance score between the text sequence and the target customer features, wherein the target customer features are derived from the customer profile data; Calculate the information entropy of the text sequence; and The effective data is determined based on the correlation score and the information entropy, wherein the correlation score corresponding to the effective data is greater than a first preset threshold, and the information entropy corresponding to the effective data is greater than a second preset threshold.

3. The method according to claim 1, characterized in that, The training of the inference analysis model includes: The inference analysis model to be trained is fine-tuned based on a preset amount of scenario-based long-chain thinking data. This allows the inference analysis model to generate intermediate inference steps according to a preset logical chain during the inference analysis phase. The preset logical chain follows a logical order from transaction motivation to risk exposure and then to optimization direction. In the inference analysis phase, the inference analysis model performs consistency verification on the preset logical links based on the rule engine and semantic similarity.

4. The method according to claim 1 or 3, characterized in that, The reinforcement learning algorithm is a group relative policy optimization algorithm, and the training of the inference analysis model includes: For each input sample, N candidate inference results are generated by the inference analysis model to be trained. The N candidate inference results constitute an inference group, where N is an integer greater than 1. Calculate the total reward value for each candidate inference result; and The model parameters of the inference analysis model to be trained are updated based on the total reward value of each candidate inference result in the inference group.

5. The method according to claim 4, characterized in that, The inference analysis model is equipped with a dual-drive reward function, which includes an accuracy reward and a language consistency reward. The total reward value is the weighted sum of the accuracy reward and the language consistency reward. The accuracy reward is used to evaluate the consistency between the candidate reasoning result and the preset logical rules, and the language consistency reward is used to evaluate the logical coherence and terminology standardization of the candidate reasoning result.

6. The method according to any one of claims 1-3 and 5, characterized in that, The inference analysis model is a model that has undergone distillation processing. The context window of the inference analysis model is extended to a preset length through rotation position encoding extension technology, wherein the preset length supports long-period transaction data.

7. The method according to any one of claims 1-3 and 5, characterized in that, The hybrid expert architecture includes: At least a portion of the transformer layers in the report generation model are constructed as M expert subnetworks, where M is an integer greater than 1; For each input token, the activation weight of each expert subnetwork in the current converter layer is calculated through the routing network, and a predetermined number of expert subnetworks are selected for activation based on the activation weights; and The outputs of the activated expert subnetworks are weighted and summed based on their activation weights, and this sum is used as the output of the current converter layer.

8. The method according to any one of claims 1-3 and 5, characterized in that, The multi-token parallel prediction mechanism includes: The structured financial report to be generated is decomposed into K report modules, where K is an integer greater than 1; and The token sequences of the K reporting modules are generated in parallel.

9. The method according to any one of claims 1-3 and 5, characterized in that, The method further includes: The structured financial report undergoes compliance verification, which is based on a pre-defined compliance rule base and classification model; and The structured financial report that passes the compliance verification will be output in a preset format.

10. The method according to claim 9, characterized in that, The compliance verification of the structured financial report includes: The structured financial report is matched against each rule based on the preset compliance rule base. In response to the detection of violations, the structured financial report is blocked. In response to no matching violation, the classification model is used to determine compliance and output the classification result.

11. A report generation device, characterized in that, The device includes: The data extraction module is used to acquire financial management-related data of target customers, filter the financial management-related data, and obtain valid data. The financial management-related data includes at least transaction data, customer profile data, and product data. The inference analysis module is used to input the valid data into the inference analysis model and output the inference analysis results, wherein the inference analysis model employs reinforcement learning algorithms and a thought chain inference mechanism; and The report generation module is used to input the reasoning and analysis results into the report generation model and output a structured financial report. The report generation model adopts a hybrid expert architecture and a multi-token parallel prediction mechanism.

12. An electronic device, comprising: One or more processors; Memory, used to store one or more computer programs. The characteristic feature is that the one or more processors execute the one or more computer programs to implement the steps of the method according to any one of claims 1 to 10.

13. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 10.

14. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 10.

Images