Training methods, rate table generation methods, and systems for rate table generation models

Abstract

This specification provides a training method, a rate table generation method, and a system for a rate table generation model. The training method includes: obtaining a first training dataset, which comprises a first data subset and a second data subset. The first data subset includes product clause training samples and corresponding rate table labels. The second data subset is a temporal inference dataset, which includes thought chain-style annotation information of time logic constraints associated with the first data subset. The first training dataset is then input into a base network model to train the base network model, thereby obtaining a rate table generation model. The rate table generation model is trained to automatically generate rate tables that satisfy time logic constraints based on the input product clauses. This enables the rate table generation model to not only learn to "generate correct tables" but also to "generate correct tables according to correct time logic."

Description

Technical Field

[0001] This specification relates to the field of artificial intelligence technology, and in particular to a training method, a rate table generation method, and a system for a rate table generation model. Background Technology

[0002] Rate tables are widely used in the insurance, finance, public services, and commercial sectors. A rate table is a standardized pricing tool used to clearly list corresponding fees or prices based on factors such as user attributes, usage volume, risk level, or service type.

[0003] Traditionally, rate tables were primarily compiled manually by actuaries or transaction (or product) requirements experts based on product terms, a time-consuming and error-prone process. In recent years, with the development of Large Language Models (LLMs), the industry has begun to experiment with using the text of product terms as input, training neural network models through end-to-end supervised learning to automatically generate corresponding rate tables. For example, by constructing a paired dataset of product terms and rate tables, a pre-trained model can be fine-tuned to learn the mapping relationship from natural language descriptions (i.e., product terms) to structured numerical tables (i.e., rate tables).

[0004] However, the aforementioned end-to-end training relies solely on input-output pairs and lacks explicit modeling of intermediate inference logic. This results in a lack of strength in the intermediate inference logic of the trained model, making it vulnerable when dealing with terms involving complex time dependencies.

[0005] It should be noted that the above-mentioned related technologies are only information known to the inventor personally, and do not mean that the above information had entered the public domain before the application date of this specification, nor do they mean that it can be considered prior art in this specification. Summary of the Invention

[0006] This specification provides a training method, a rate table generation method, and a system for a rate table generation model, to avoid at least one of the aforementioned technical problems.

[0007] Firstly, this specification provides a training method for a rate table generation model, including: Obtain a first training dataset, wherein the first training dataset includes a first data subset and a second data subset. The first data subset includes product terms training samples and corresponding rate table labels. The second data subset is a temporal inference dataset, and the temporal inference dataset includes thought chain-style annotation information of time logic constraints associated with the first data subset. The first training dataset is input into the base network model to train the base network model and obtain the rate table generation model, wherein the rate table generation model is trained to automatically generate a rate table that meets the time logic constraints based on the input product terms.

[0008] Secondly, this specification provides a method for generating a rate table, including: Obtain the terms and conditions for the products to be processed; The product terms to be processed are input into the rate table generation model, and the rate table corresponding to the product terms to be processed is output. The rate table generation model is trained based on the training method described in the first aspect.

[0009] Thirdly, this specification provides a training system for a rate table generation model, including: At least one storage medium storing at least one instruction set for training the rate table generation model; At least one processor is communicatively connected to the at least one storage medium, wherein when the at least one processor is running, it reads the at least one instruction set and executes the training method as described in the first aspect according to the instructions of the at least one instruction set.

[0010] Fourthly, this specification provides a rate table generation system, including: At least one storage medium storing at least one set of instructions for generating rate tables; At least one processor is communicatively connected to the at least one storage medium, wherein when the at least one processor is running, it reads the at least one instruction set and executes the generation method as described in the second aspect according to the instructions of the at least one instruction set.

[0011] Fifthly, this specification provides a computer-readable non-transitory storage medium, wherein the computer-readable non-transitory storage medium stores at least one instruction set, which is executed by at least one processor to implement the method as described in the first or second aspect.

[0012] As can be seen from the above technical solutions, the rate table generation model training method, rate table generation method, and system provided in this specification achieve this by: obtaining a first data subset (such as a dataset of input (product terms - output rate table pairs) and an intermediate inference logic dataset (such as time inference logic, etc.)); using these two datasets to supervise and fine-tune the basic network model, enabling the basic network model to learn both the ability to output rate tables based on the input product terms (i.e., the ability to learn the mapping relationship between product terms and rate tables) and the ability to output rate tables that conform to intermediate inference logic (such as time inference logic) based on product terms, thereby obtaining the rate table generation model. In other words, by adopting a dual-driven rate table generation model training method of "mapping relationship from product terms to rate tables + time inference chain," the key but implicit time logic constraints can be made explicit, enabling the rate table generation model to not only learn to "generate the correct table" but also to "generate the correct table according to the correct time logic."

[0013] The training methods for the rate table generation model, the rate table generation methods, and other functions of the system provided in this specification will be partially listed in the following description. The inventive aspects of the training methods for the rate table generation model, the rate table generation methods, and the system provided in this specification can be fully explained through practice or by using the methods, apparatus, and combinations described in the detailed examples below. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this specification, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 A schematic diagram illustrating an application scenario of the training method for the rate table generation model provided in the embodiments of this specification; Figure 2 A schematic diagram of the structure of the training system for the rate table generation model provided in the embodiments of this specification; Figure 3 A flowchart illustrating the training method for a rate table generation model provided in one embodiment of this specification; Figure 4 A schematic diagram illustrating the principle of a training method for a rate table generation model provided in one embodiment of this specification; Figure 5 A flowchart illustrating the training method for a rate table generation model provided in another embodiment of this specification; Figure 6 A schematic diagram illustrating the principle of a training method for a rate table generation model provided in another embodiment of this specification; Figure 7 A flowchart illustrating the rate table generation method provided in the embodiments of this specification; Figure 8 This is a schematic diagram illustrating the application of the rate table generation model provided in the embodiments of this specification to life insurance products. Detailed Implementation

[0016] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this specification. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this specification as detailed in the appended claims.

[0017] It should be understood that the terms “comprising” and “having” and any variations thereof in the embodiments of this specification are intended to cover but not exclude inclusion. For example, a product or device that includes a series of components is not necessarily limited to those components that are explicitly listed, but may include other components that are not explicitly listed or that are inherent to such product or device.

[0018] The term "and / or" in the embodiments of this specification describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship.

[0019] In the embodiments of this specification, the term "multiple" refers to two or more, and other quantifiers are similar.

[0020] The terms “first,” “second,” “third,” etc., used in this specification are used to distinguish similar or related objects or entities and do not necessarily imply a specific order or sequence, unless otherwise indicated. It should be understood that such terms can be used interchangeably where appropriate, for example, in situations where implementation can proceed in an order other than those given in the embodiments illustrated or described in this specification.

[0021] As used in this specification, the term "unit / module" means any known or subsequently developed hardware, software, firmware, artificial intelligence, fuzzy logic, or combination of hardware and / or software code capable of performing the functions associated with that element.

[0022] To avoid at least one of the technical problems mentioned in the background section above, this specification proposes a technical concept developed through inventive effort: obtaining multiple datasets (i.e., multi-source datasets), such as datasets including: datasets of input (product terms) - output (rate tables) pairs, datasets of intermediate inference logic (such as time inference logic), so as to perform supervised fine-tuning (SFT) on the basic network model (such as a pre-trained large language model) through multiple datasets, so that the basic network model can learn both the ability to output rate tables based on the input product terms (i.e., the ability to learn the mapping relationship between product terms and rate tables), and the ability to output according to intermediate inference logic (such as time inference logic) based on product terms, thereby obtaining a rate table generation model.

[0023] Correspondingly, the rate table generation model not only has the mapping capability to determine the output (rate table) based on the input (product terms), but also the intermediate reasoning capability to determine the output (rate table) based on the input (product terms). For example, the intermediate reasoning capability can be a time reasoning capability, that is, the rate table generation model is trained to automatically generate a rate table that meets time logic constraints based on the input product terms.

[0024] The technical solutions provided in this specification are based on the above-described technical concepts. As can be seen from the description of the above technical concepts, in the technical solutions provided in this specification, by adopting a dual-driven rate table generation model training method of "mapping relationship from product terms to rate table + time inference chain", the key but implicit time logic constraints can be made explicit, so that the rate table generation model not only learns to "generate the correct table", but also learns to "generate the correct table according to the correct time logic".

[0025] To facilitate readers' understanding of this manual, the application scenarios of this manual are introduced below.

[0026] The technical solutions provided in this manual are applicable to scenarios that require the automatic generation of rate tables corresponding to product terms.

[0027] For example, products can be insurance products, banking and financial services products, telecommunications and utility products, logistics and transportation products, public service products, etc.

[0028] Therefore, rate tables can be those corresponding to insurance products (such as annuity rate tables, life insurance rate tables, health insurance rate tables, accident insurance rate tables, and supplementary insurance rate tables); they can also be those corresponding to banking and financial service products (such as loan interest rate tables, deposit interest rate tables, and handling fee rate tables); they can also be those corresponding to telecommunications and public utility products (such as communication tariff packages, water, electricity, and gas tiered rate tables); they can also be those corresponding to logistics and transportation products (such as express delivery / freight rate tables, taxi / ride-hailing fare tables); and they can also be those corresponding to public service products (such as tax rate tables, social security / housing provident fund contribution ratio tables).

[0029] It is worth noting that the above examples are merely illustrative of the application scenarios to which the technical solutions in this specification can be applied, and should not be construed as limiting the application scenarios. When examples are given later in this specification, they will primarily be illustrated using insurance products and insurance premium rate tables as examples.

[0030] Figure 1 This diagram illustrates an application scenario of the training method (hereinafter referred to as the training method) for the rate table generation model in this specification. The training method can be applied to, for example... Figure 1 Scenario 100 is shown. (e.g.) Figure 1 As shown, scenario 100 includes target user 101, client 102, server 103, and network 104.

[0031] Target user 101 can be the user who triggers the training of the base network model to obtain the rate table generation model. For example, target user 101 can perform a target operation on client 102 to trigger the training of the base network model.

[0032] Client 102 may be an electronic device that provides interactive functionality to target user 101. For example, client 102 may provide an interactive interface to target user 101, where target user 101 can perform target operations.

[0033] In some embodiments, client 102 may include a mobile device, tablet, laptop, built-in device in a motor vehicle, or similar content, or any combination thereof. In some embodiments, the mobile device may include a smart home device, a smart mobile device, a virtual reality device, an augmented reality device, or similar device, or any combination thereof. In some embodiments, smart home devices may include a smart TV, a desktop computer, etc., or any combination thereof. In some embodiments, smart mobile devices may include a smartphone, a personal digital assistant, a gaming device, a navigation device, etc., or any combination thereof. In some embodiments, built-in devices in a motor vehicle may include an in-vehicle computer, an in-vehicle television, etc.

[0034] In some embodiments, client 102 may have one or more applications (APPs) installed. APPs provide target user 101 with the ability and interface to interact with the outside world via network 104. APPs include, but are not limited to: web browser APPs, search APPs, chat APPs, shopping APPs, video APPs, financial management APPs, instant messaging tools, email clients, social media platform software, etc.

[0035] like Figure 1 As shown, client 102 can establish a communication connection with server 103. Server 103 can communicate with one client 102 or multiple clients 102. In some embodiments, client 102 can interact with server 103 via network 104 to receive or send messages, etc.

[0036] Server 103 can be a server that provides various services. For example, server 103 can be a cloud server or a local server. Server 103 can communicate with one client 102 and receive data sent by that client 102, or it can communicate with multiple clients 102 and receive data sent by each client 102.

[0037] In some embodiments, the training method described herein can be executed on server 103. In this case, server 103 may store data or instructions for executing the training method described herein, and may execute or be used to execute the data or instructions. Server 103 may include hardware devices with data processing capabilities and the necessary programs required to drive the hardware devices.

[0038] For example, such as Figure 1 As shown, server 103 can execute the training method described in this specification to train the basic network model and obtain the rate table generation model. And as... Figure 1 As shown, the input to the rate table generation model is the product terms and conditions, such as the terms and conditions text of an insurance product. The output of the rate table is a rate table, such as an annuity rate table corresponding to an insurance product.

[0039] Network 104 is a medium used to provide a communication connection between client 102 and server 103. Network 104 can facilitate the exchange of information or data. Figure 1 As shown, client 102 and server 103 can each connect to network 104 and transmit information or data to each other through network 104.

[0040] In some embodiments, network 104 can be any type of wired or wireless network, or a combination thereof. For example, network 104 may include a cable network, a wired network, a fiber optic network, a telecommunications network, an intranet, the Internet, a local area network (LAN), a wide area network (WAN), a wireless local area network (WLAN), a metropolitan area network (MAN), a public switched telephone network (PSTN), a Bluetooth network™, a ZigBee™ short-range wireless network, a near field communication (NFC) network, or a similar network.

[0041] In some embodiments, network 104 may include one or more network access points. For example, network 104 may include wired or wireless network access points, such as base stations or internet switching points, through which one or more components of client 102 and server 103 can connect to network 104 to exchange data or information.

[0042] It is worth noting that, Figure 1 The number of clients 102, servers 103, and networks 104 shown is merely illustrative. Depending on implementation needs, there can be any number of clients 102, servers 103, and networks 104.

[0043] In other words, Figure 1 and targeting Figure 1 The above description is only used to illustrate the possible application scenarios for the training method in this specification, and should not be construed as limiting the application scenarios.

[0044] Figure 2 A hardware structure diagram of a training system 200 provided according to an embodiment of this specification is shown. The training system 200 can execute the training methods described in this specification. The training methods are described in other parts of this specification. In conjunction with the above example, when the training methods are executed on server 103, the training system 200 can be server 103.

[0045] like Figure 2 As shown, the training system 200 may include at least one storage medium 203 and at least one processor 202. In some embodiments, the training system 200 may also include a communication port 204 and an internal communication bus 201. The training system 200 may also include I / O components 205.

[0046] The internal communication bus 201 can connect to different system components. For example, the internal communication bus 201 can connect to storage medium 203, processor 202, communication port 204, and I / O component 205.

[0047] I / O component 205 supports input / output between training system 200 and other components.

[0048] Communication port 204 is used for data communication between the training system 200 and the outside world. For example, communication port 204 can be used for data communication between the training system 200 and network 104. Communication port 204 can be a wired communication port or a wireless communication port.

[0049] Storage medium 203 may include a data storage device. The data storage device may be a non-transitory storage medium or a temporary storage medium. For example, the data storage device may include one or more of a disk 2031, a read-only storage medium (ROM) 2032, or a random access storage medium (RAM) 2033. Storage medium 203 also includes at least one instruction set stored in the data storage device. The instruction set includes computer program code, which may include programs, routines, objects, components, data structures, procedures, modules, etc., that execute the training methods provided in this specification.

[0050] At least one processor 202 may be communicatively connected to at least one storage medium 203. The at least one processor 202 is used to execute the at least one instruction set described above. When the training system 200 is running, the at least one processor 202 reads the at least one instruction set and executes the training method provided in this specification according to the instructions of the at least one instruction set. The processor 202 may execute all steps included in the training method. The processor 202 may be in the form of one or more processors. In some embodiments, the processor 202 may include one or more hardware processors, such as a microcontroller, microprocessor, reduced instruction set computer (RISC), application-specific integrated circuit (ASIC), application-specific instruction set processor (ASIP), central processing unit (CPU), graphics processing unit (GPU), physical processing unit (PPU), microcontroller unit, digital signal processor (DSP), field-programmable gate array (FPGA), advanced RISC machine (ARM), programmable logic device (PLD), any circuit or processor capable of performing one or more functions, or any combination thereof.

[0051] For illustrative purposes only, only one processor 202 is shown in the accompanying drawings of the training system 200. However, it should be noted that the training system 200 may also include multiple processors. Therefore, the operations and / or method steps disclosed herein may be executed by one processor or by multiple processors in combination. For example, if the processor 202 of the training system 200 described in this specification executes steps A and B, it should be understood that steps A and B may also be executed jointly or separately by two different processors 202 (e.g., the first processor executes step A, the second processor executes step B, or the first and second processors jointly execute steps A and B).

[0052] Please see Figure 3 , Figure 3 This is a flowchart illustrating the training method for a rate table generation model provided in one embodiment of this specification. Figure 3 The training method shown can be implemented by a training system. For a description of the training system, please refer to the example above; it will not be repeated here.

[0053] like Figure 3 As shown, the method includes the following steps S301 to S302: S301: Obtain the first training dataset, wherein the first training dataset includes a first data subset and a second data subset. The first data subset includes product terms training samples and corresponding rate table labels. The second data subset is a time-series inference dataset, and the time-series inference dataset includes thought chain-style annotation information of time logic constraints associated with the first data subset.

[0054] The first training dataset is a collection of training data used for supervised fine-tuning of the pre-trained base network model. The first training dataset includes at least two subsets: a first subset of labeled term-rate pairs and a second subset of labeled data with temporal logical reasoning processes.

[0055] In other words, the first training dataset is a multi-source dataset, and the multi-source dataset consists of at least two datasets, such as a first data subset and a second data subset.

[0056] The first subset of data can be traditional supervised learning data. For example, the first subset of data includes multiple samples, each of which includes product terms and corresponding standard rate tables.

[0057] Continuing with the example of insurance products, the product terms and conditions include the insurance product text, such as the insurance contract definition, details of insurance liability, payment rules, and underwriting parameters (such as insured age, payment method, and coverage period). The corresponding standard rate table can be a structured JSON format rate table. Specifically, for example, if the product terms and conditions are the text of a whole life insurance policy, the corresponding rate table is a JSON format table showing the annual premiums for ages 0-100.

[0058] The second subset of data is a temporal reasoning dataset, which includes thought chain-style annotation information. For example, the second subset of data includes intermediate reasoning steps on how to derive time-related logic from product terms.

[0059] For example, mind chain-style annotation information is used to annotate the step-by-step reasoning process of temporal logic. For instance, if an insurance product's terms and conditions stipulate "insured age 18-65 years old," then the corresponding premium rate table's age range should be 18-65 years old. If the insurance product's terms and conditions stipulate "the policy year starts from 00:00 on the day following the effective date," then the age calculation uses "policy year - year of birth." This type of annotation can guide the basic network model to learn interpretable temporal reasoning capabilities.

[0060] Time-related constraints can be understood as compliance rules involving the time dimension in the rate table, such as: "the insured's age at the time of application must be ≥ 18 years old", "the waiting period is 90 days, i.e., the earliest payment date = the effective date + 90 days", and "the payment period ≤ the coverage period". These constraints are reflected in the rate calculation or the rate table structure.

[0061] For example, taking life insurance products as an example, the training samples of product terms in the first data subset can be the text of the terms of life insurance products. For example, the content of the terms text includes "The insured age of this product is 18 to 65 years old (inclusive); the insurance period is for life; the payment method is annual payment, and the payment period can be 10 years, 20 years or until the age of 60; the waiting period is 90 days; the death benefit payment date shall not be earlier than the policy effective date plus 90 days." The first data subset also includes the life insurance rate table corresponding to the terms text.

[0062] The second data subset includes thought chain-style annotation information, such as: Step 1: The insured age range is 18–65 years old → The life insurance rate table is for ages 18 to 65; Step 2: Pay until age 60 → This payment method is not available for insured individuals aged 60 and above (marked "Not Supported"). Step 3: Waiting period of 90 days → The timestamps of all payment events must be ≥ the effective date + 90 days (although they do not appear directly in the life insurance rate table, they need to be modeled in the internal logic of the model). Step 4: Select the payment period as a discrete value (10 / 20 / to 60) → The life insurance premium rate table column headings are fixed to these three items.

[0063] Alternatively, mind chain-based annotation information can solve complex time logic calculation problems, such as accurately calculating the payout age for insurance.

[0064] For example, the second data subset could include a mind-chain annotation for the question: "The insured was born on March 15, 1990, the contract effective date is October 11, 2025, and the terms stipulate that annuities will be paid from the first policy anniversary after the insured reaches age 60. From which year will the annuity payments begin?" This mind-chain annotation could include the following time calculation process: 1. Calculate the date the insured reaches age 60: 1990 + 60 = March 15, 2050; 2. Find the first policy anniversary after March 15, 2050; 3. The policy anniversary is October 11th of each year; 4. October 11, 2050 is after March 15, 2050, therefore it is the first eligible policy anniversary; 5. Therefore, the annuity payments will begin in 2050.

[0065] S302: Input the first training dataset into the base network model to train the base network model and obtain the rate table generation model, wherein the rate table generation model is trained to automatically generate a rate table that meets the time logic constraints based on the input product terms.

[0066] The base network model is a pre-trained network model. This embodiment does not limit the type and architecture of the base network model. For example, the base network model can be a pre-trained large language model.

[0067] In this specification, the large language model may also be referred to simply as the large model. A large language model is a natural language processing model based on deep learning techniques, typically with billions to hundreds of billions or even more parameters, possessing powerful language understanding and generation capabilities. Large language models can employ the Transformer architecture or its variants (such as GPT, BERT, etc.), which utilizes an attention mechanism to globally model sequential data, efficiently handling long-distance dependencies and thus performing exceptionally well in natural language tasks. Large language models learn the statistical features and semantic relationships of language through pre-training on large-scale corpora, giving them outstanding generalization capabilities. The core capabilities of large language models include, but are not limited to: understanding contextual semantics, generating coherent and grammatically correct text, performing logical reasoning, and handling multi-task scenarios. Their usage typically includes two modes: direct inference and fine-tuning. In direct inference mode, the user guides the large language model to generate specific outputs by designing prompts. Prompts can be task descriptions or instructions in text form, used to stimulate the large language model's semantic understanding and generation capabilities. In fine-tuning mode, large language models are further trained on small-scale datasets within a specific domain to optimize their performance on specific tasks. The powerful generalization capabilities and flexibility of large language models make them an important tool in the field of artificial intelligence, providing efficient and accurate solutions for automated text generation and understanding.

[0068] In some embodiments, large language models can also understand and generate data from other modalities (such as visual and audio data). In this case, large language models can also be called multimodal large language models (MLLMs). MLLMs provide a richer and more natural interactive experience by integrating multiple types of input and output, such as text, images, and sound. The core advantage of MLLMs lies in their ability to process and understand information from different modalities and fuse this information to complete complex tasks. For example, MLLMs can analyze an image and generate descriptive text, or generate a corresponding image based on a text description. This cross-modal understanding and generation capability makes MLLMs widely applicable across multiple fields.

[0069] It should be noted that the key technologies of large language models can be found in the detailed description in the paper "A Survey of Large Language Models" (paper number: arXiv:2303.18223v16, published on March 11, 2025, public link: https: / / doi.org / 10.48550 / arXiv.2303.18223), and will not be repeated here.

[0070] like Figure 4 As shown, after obtaining the first training dataset (first data subset and second data subset), the training system can train the basic network model (such as a pre-trained large language model) based on the first data subset and the second data subset to obtain the rate table generation model.

[0071] Based on the above analysis, the first data subset can provide end-to-end correct output supervision for the basic network model; the second data subset can provide explicit guidance for the time-sensitive inference process of the basic network model. Therefore, by jointly training the basic network model using the first and second data subsets, the rate table generation model can both "generate" and "know why it generates in this way," making it more robust, especially when dealing with complex time dependencies.

[0072] In other words, if only the first subset of data is used to train the basic network model, the trained model may not be able to guarantee temporal logic consistency. However, by combining the second subset of data to train the basic network model, the implicit temporal rules can be made explicit and structured, so that the rate table generation model has strong actuarial logic capabilities.

[0073] Continuing with the example of insurance products, insurance premium rate tables are highly dependent on time variables such as date of birth, effective date, policy year, waiting period, and payment period. If the insurance premium rate table generation model ignores or miscalculates time logic, it will lead to errors in the inference of the above time-related information.

[0074] For example, the payout date of insurance products, such as annuity products, is usually affected by multiple factors, including the effective date, the insured's age, and the premium payment period, involving complex time-series reasoning such as leap years, age thresholds, and liability transitions. Existing schemes often have errors in calculating key time-series factors such as payout dates spanning multiple years, birthdays, and the start and end points of liability, leading to inaccurate rate table data.

[0075] For example, the age range in the insurance premium rate table may be incorrect (e.g., including infants aged 0 years that are prohibited by the product), the payment date may be earlier than the waiting period, or the payment period may exceed the maximum age limit. In this embodiment, the training system can intervene to address these issues by combining the time-series inference dataset.

[0076] Alternatively, continuing with the description of time-related logical constraints in the above example, when training the basic network model using only the first subset, the trained model might incorrectly output the rate "pay until age 60" for a 65-year-old user (because 65 > 60). However, by training the basic network model using the second subset of data, the model can learn to first determine the relationship between age and the payment cutoff age before deciding whether to fill in the value, thus avoiding time-related logical conflicts.

[0077] Furthermore, this embodiment does not limit the method by which the training system trains the basic network model based on the first and second data subsets. For example, the training system may use a multi-task learning framework to train the basic network model, or it may use multi-dimensional data feature concatenation to train the basic network model based on a single-task learning framework.

[0078] Taking the training of a basic network model using a single-task learning framework as an example: The base network model can be constructed by concatenating the first and second data subsets as a long sequence input. The training system can train the base network model using cross-entropy loss.

[0079] For example, for each element (token) in a long sequence, the base network model can predict what the next element will be and calculate the error (i.e., loss) based on the actual next element. This loss is used to measure how well the model predicts and aims to optimize the model parameters to make the predictions as accurate as possible.

[0080] In some embodiments, for structural symbols (such as parentheses, commas, etc.) in long sequences, the training system may employ a tokenization method to process these structural symbols in order to reduce their impact on the overall training process.

[0081] In addition, to help the basic network model better learn to handle data of various types (including complex and long tabular data), the training system can randomly shuffle the order of the input data during training and select samples according to certain rules (such as data complexity), so that the trained rate table generation model learns more general rules rather than just memorizing certain specific patterns.

[0082] Furthermore, to ensure a smooth and effective training process, the training system can set several important parameters. These include the learning rate (which controls the speed at which the base network model is updated), the batch size (the number of samples used in each update), and gradient clipping (to prevent excessively large gradient values ​​from causing training instability). This allows the base network model to learn effectively without issues such as overfitting or gradient explosion.

[0083] Furthermore, to accelerate training speed and improve efficiency, the training system can employ multi-GPU parallel training methods, such as determining different training tasks based on different data, thereby greatly shortening the training time and ensuring training efficiency and stability.

[0084] Taking the training of a base network model using a multi-task learning framework as an example: The basic network model extracts first features from the first subset of data, such as product attributes (e.g., age range, payment options), and maps them to a structured numerical table. The basic network model extracts second features from the second subset of data, such as causal chains based on temporal logic.

[0085] The multi-task approach comprises two tasks: Task 1 is the rate table generation task (the main task), and Task 2 is the temporal logic reasoning task (the auxiliary task). The goal of Task 1 is to accurately output a structured rate table; the goal of Task 2 is to generate the correct temporal reasoning thought process steps. Relatively speaking, the implementation of Task 1 depends on a first subset of data and a second subset of data; the implementation of Task 2 depends on the second subset of data.

[0086] The training system can construct loss functions for Task 1 and Task 2 respectively, such as constructing cross-entropy loss functions separately, and add the two loss functions to obtain the total loss function. The parameters of the basic network model are adjusted with the goal of minimizing the total loss function to obtain the rate table generation model.

[0087] Based on the above analysis of S301 and S302, it can be seen that in this embodiment, the training system adopts a dual-driven rate table generation model training method of "mapping relationship from insurance terms to rate table (first data subset) + time inference chain (second data subset)", which can make the key but implicit time logic constraints explicit, so that the rate table generation model not only learns to "generate the correct table", but also learns to "generate the correct table according to the correct time logic".

[0088] To help readers gain a deeper understanding of the technical principles behind the training methods provided in this manual, the following explanation is provided in conjunction with... Figure 5 The training methods provided in this manual will be described in more detail. For example... Figure 5 As shown, the method includes: S501: Obtain the first training dataset, wherein the first training dataset includes a first data subset and a second data subset. The first data subset includes product terms training samples and corresponding rate table labels. The second data subset is a time-series inference dataset, and the time-series inference dataset includes thought chain annotation information of time logic constraints associated with the first data subset.

[0089] S502: Input the first training dataset into the base network model to train the base network model and obtain the rate table generation model, wherein the rate table generation model is trained to automatically generate a rate table that meets the time logic constraints based on the input product terms.

[0090] It is understood that, in order to avoid tedious descriptions, this embodiment will not repeat the same or similar content as the examples above. For example, for the understanding of S501 and S502, please refer to the description of S301 and S302 in the examples above.

[0091] Additionally, in some embodiments, the base network model may include a pre-trained base model and a LoRA adapter. S302 above may include: freezing the parameters of the base model and updating the low-rank parameters of the LoRA adapter based on a first training dataset to train the base network model and obtain a rate table generation model.

[0092] The base model can be understood as a large language model pre-trained on a large corpus. Relatively speaking, the base model has general language understanding and generation capabilities, but it needs fine-tuning for tasks such as rate table generation.

[0093] LoRA adapter is a parameter-efficient fine-tuning technique that introduces a low-rank matrix decomposition structure by bypassing the original parameters (weights) of the base model.

[0094] For example, the base network model includes a base model and a LoRA adapter, and the LoRA adapter can be inserted into key layers of the base model (such as the projection matrix of the attention module of the base model). When training the base network model based on the first training dataset, the parameters of the base model are not updated, but the number of trainable parameters (i.e., low-rank parameters) introduced by the LoRA adapter are updated. This allows for the injection of rate table generation expertise (such as expertise in the actuarial field) into the base network model with a small number of parameters, building upon the linguistic commonalities (such as syntax, logical secondary connections, numerical representations, etc.) of the base model. Furthermore, supervised fine-tuning of the base network model by freezing the base model can avoid catastrophic forgetting and overfitting to small-scale domain data.

[0095] In other words, the first data subset can provide output target supervision for the LoRA adapter, and the second data subset can provide intermediate temporal logic reasoning guidance for the LoRA adapter, so that the LoRA adapter can learn how to activate the potential capabilities related to temporal logic and numerical tables in the base model when the two are combined.

[0096] Continuing with the example above, let's take the generation of a life insurance rate table as an example: The base model can be a pre-trained Qwen3-32B, and the LoRA adapter can insert a low-rank LoRA matrix into the (Q) and (V) projection matrices of the Transformer layer in Qwen3-32B.

[0097] During training, the pedestal model processes product-related text from the first training dataset to generate hidden states; the LoRA adapter dynamically adjusts the (Q, V) vector during attention computation to enhance attention to time-related keywords in the first training data. The low-rank parameters in the LoRA low-rank matrix are updated via backpropagation by calculating the total loss, while the parameters of the pedestal model remain unchanged.

[0098] In this embodiment, a base network model including a pedestal model and a LoRA adapter is used. The pedestal model is frozen, and the low-rank parameters of the LoRA adapter are updated to train the base network model. This reduces the number of training parameters, thereby improving training efficiency. Furthermore, it allows for scalable domain adaptation of the training of the base network model.

[0099] Based on the above analysis and Figure 4 As can be seen, the first training dataset mainly includes a first data subset and a second data subset. The rate table generation model is trained based on the first data subset and the second training data subset: it can automatically generate the corresponding rate table based on the input product terms, and the generated rate table can conform to the time logic constraints of the product terms.

[0100] In other embodiments, such as Figure 6 As shown, the first training dataset may also include datasets from additional sources, such as a third and / or a fourth subset. The third subset includes domain knowledge associated with the first subset. The fourth subset includes a negative sample dataset associated with the first subset.

[0101] The third data subset, also known as the domain knowledge subset, can be understood as structured or semi-structured domain knowledge associated with product terms and rates. It can be used to enhance the model's understanding of transaction requirement rules, actuarial principles, or detection logic. The third data subset includes, but is not limited to, domain knowledge in the form of actuarial formulas, detection clause summaries, product design specifications, and terminology definition tables.

[0102] For example, taking insurance products as an example, the third data subset can be an insurance transaction requirements knowledge base dataset, which can provide rules, definitions, background knowledge, etc. in the insurance field.

[0103] In other words, in this embodiment, the third data subset can provide external knowledge for training the basic network model. This external knowledge is domain-related, allowing the basic network model to learn prior domain knowledge during training, rather than simply fitting data. Especially for insurance products, where the corresponding knowledge pertains to the insurance domain, insurance clauses are relatively more cautious and logically more complex. Therefore, by introducing the third data subset for supervised fine-tuning of the basic network model, the "illusion" phenomenon (such as misunderstanding, omission, confusion, or missing fields) in the basic network model's understanding of insurance liabilities and other knowledge in the insurance clauses can be avoided, thereby improving the effectiveness and reliability of supervised fine-tuning of the basic network model.

[0104] The fourth data subset, also known as the negative sample dataset, can be understood as a set of erroneous samples, such as the constructed erroneous terms-rate table pairings, whose rate tables violate explicit or implicit constraints in the product terms (such as age limits, time logic conflicts, unreasonable values, etc.). It can be used to train the basic network model to identify and avoid generating such invalid outputs.

[0105] For example, taking insurance products as an example, a negative sample dataset can be a collection of negative samples that focuses on rule compliance errors (such as rate and age segmentation), date calculation errors, output step errors, timing logic errors, etc.

[0106] In this embodiment, the negative samples in the fourth data subset (i.e., the negative sample dataset) can be used to explicitly demonstrate "what cannot be done," thus compensating for the insufficient coverage of the first data subset. Therefore, the base network model can learn to generate a more accurate rate table by incorporating the fourth data subset.

[0107] S503: Obtain the second training dataset, which includes several preference sample pairs. Each preference sample pair includes input text and corresponding annotation output. The input text is text associated with product terms. The annotation output includes positive sample rate table output and negative sample rate table output corresponding to the input text.

[0108] The second training dataset is used to train the rate table generation model for preference alignment, and its core unit is the preference sample pair.

[0109] Preference sample pairs are two different outputs (rate tables) generated from the same input text (such as product terms), labeled as better (positive sample) or worse (negative sample). For example, a preference sample pair might take the form of (input text, positive sample output, negative sample output).

[0110] In contrast, positive sample outputs are those that better align with transactional needs and preferences. For example, a positive sample rate table output could be one that, while meeting product terms, better reflects actuarial preferences, market practices, or user experience (e.g., smooth numerical values, clear structure). A negative sample rate table output, while potentially logically compliant, might have flaws in terms of transactional needs (e.g., sudden premium changes, redundant columns, inconsistencies with pricing strategies), serving as a comparative signal for the positive sample rate table output.

[0111] Among these, the coverage of preference sample pairs can be relatively broad. Continuing with the example of insurance products, the coverage of preference sample pairs can include multiple liability types, multiple payment methods, multiple claim periods, boundary conditions, complex temporal reasoning, and abnormal transaction requirements, ensuring a comprehensive representation of transaction needs.

[0112] This embodiment does not limit the method of obtaining the second training dataset. For example, the training system can first construct several triples (input terms, positive sample rate table output, negative sample rate table output), and then perform preference labeling on the triples through automatic labeling or manual labeling to obtain several preference sample pairs in the second training dataset.

[0113] For example, for the same input text First, the training system can generate two possible rate table outputs either through a model or manually. and Then, the training system can be based on labeled tasks (such as...). and (Which one better matches the corresponding preference) to obtain the corresponding preference sample pair.

[0114] In addition, the training system can perform annotation based on preset quality control strategies. These strategies may include: multiple tables with fewer cuts, review of highly controversial samples, and random checks.

[0115] The "multiple labeling, fewer judgments" approach can be understood as having each pair of samples labeled with preferences by two or more actuarial / product experts. For highly controversial samples, a review process can be conducted, such as submitting highly controversial samples to senior experts for review, in order to apply preference labeling to these samples. "Random checks," such as periodic random checks, can be understood as periodically selecting labeled samples for review to ensure long-term labeling quality.

[0116] In contrast, preference labeling differs from "tagging." Preference labeling is the explicit process of value judgments regarding transaction needs, transforming implicit expert experience into learnable signals.

[0117] S504: Input the second training dataset into the rate table generation model to perform preference alignment training on the rate table generation model using the Direct Preference Optimization (DPO) method, and obtain the optimized rate table generation model.

[0118] Direct preference optimization is a preference alignment algorithm that does not require an explicit reward model. It generates model parameters by directly fine-tuning the rate table on preference pairs by maximizing the implicit reward advantage of positive samples relative to negative samples. It belongs to supervised preference learning.

[0119] Preference alignment training refers to using preference samples to guide the output of the rate table generation model to be closer to the target preferences of human experts or business needs, thereby achieving a leap in the ability to generate rate tables from "correct" to "high-quality".

[0120] Continuing with the examples above and Figure 6 After the rate table generation model is trained by S501 and S502, the rate table generation model can be optimized by using the scheme described in S503 and S504, combined with the second training dataset (several preference sample pairs), so that the optimized rate table generation model can generate a correct and relatively better rate table corresponding to the product terms.

[0121] Based on the above analysis of S503 and S504, it can be seen that in this embodiment, the training system introduces direct preference optimization based on preference sample pairs. This injects preferences into the rate table generation model in an efficient, stable, and low-cost manner, further improving the effectiveness and reliability of training the basic network model. In other words, by introducing preference sample pairs, this embodiment optimizes the performance of the rate table generation model obtained through supervised fine-tuning in terms of the completeness and structural consistency of the rate table output, strengthening its ability to generate complete and comprehensive rate tables. This results in the optimized rate table generation model having higher accuracy and reliability in generating rate tables, and producing higher-quality rate tables that better meet user expectations and experience.

[0122] In some embodiments, "training the rate table generation model with preference alignment using the direct preference optimization method" in S504 includes the following steps 11 and 12: Step 11: Construct a preference contrast loss function based on the positive and negative sample rate table outputs. The preference contrast loss function includes a regularization term, which is determined based on the KL (Kullback-Leibler) divergence between the rate table generation model and the base network model.

[0123] The preference contrast loss function is the objective function used for direct preference optimization. By comparing the relative generation probabilities of positive and negative samples, it guides the optimized rate table generation model to output a better preference.

[0124] The preference contrast loss function includes a regularization term based on KL divergence to constrain the update magnitude of the rate table generation model. This regularization term can be understood as a penalty introduced into the loss function to prevent the rate table generation model from deviating excessively from the initial strategy, thereby improving training stability and generalization ability. KL divergence, also known as relative entropy, is an asymmetric measure of the difference between two probability distributions.

[0125] For example, the regularization term can be explicitly defined as the rate table generation model and the base network model, while KL divergence can be used to quantify the difference in the distribution of the outputs generated by the rate table generation model and the base network model under a given input, as a regularization constraint.

[0126] Step 12: Optimize the parameters of the rate table generation model by minimizing the preference contrast loss function.

[0127] Correspondingly, after constructing the preference contrast loss function, the training system can optimize the parameters of the rate table generation model with the goal of minimizing the preference contrast loss function, so that the rate table generation model can learn to generate relatively better rate tables.

[0128] Based on the analysis of steps 11 and 12 above, it can be seen that in this embodiment, the training system achieves synergistic optimization of preference alignment and rate table generation model stability by explicitly introducing a regularization term based on KL divergence into the preference contrast loss function. That is, the optimized rate table generation model achieves "high-quality (e.g., preference-compliant) rate table output" without losing the ability to generate rate tables that satisfy time logic.

[0129] S505: Obtain the third training dataset, which includes multiple product terms reinforcement samples, and perform reinforcement learning training on the optimized rate table generation model based on the third training dataset.

[0130] Combining the above examples and Figure 6 It can be seen that after the training system optimizes the rate table generation model and obtains the optimized rate table generation model, it can further perform reinforcement learning training on the optimized rate table generation model.

[0131] Accordingly, the dataset used for reinforcement learning training of the optimized rate table generation model is the third training dataset. The basic unit of the third training dataset is the product terms reinforcement sample, that is, the product terms text used for reinforcement learning of the optimized rate table generation model.

[0132] This embodiment does not limit the method used for reinforcement learning training. For example, the training system can use reinforcement learning methods such as policy gradient and proximal policy optimization (PPO) to train the optimized rate table generation model.

[0133] In contrast, reinforcement learning methods based on policy gradients aim to directly adjust the "decision strategy" of the optimized rate table model (i.e., the performance of the optimized rate table generation model in generating rate tables based on product terms). Reinforcement learning methods based on proximal policy optimization aim to prevent the optimized rate table generation model from "going too far" in its updates, thus avoiding training failure, while ensuring learning efficiency.

[0134] For each product term reinforcement sample in the third training dataset, the "reinforcement learning training" in S505 includes the following steps 21 to 23: Step 21: Generate an enhanced sample rate table based on the optimized rate table generation model, corresponding to the enhanced sample of the product terms.

[0135] A reinforced sample rate table refers to a rate table output generated autonomously by an optimized rate table generation model based on reinforced samples of product terms, serving as an "action" in reinforcement learning.

[0136] For example, for a certain product terms enhancement sample, the product terms enhancement sample can be used as input to the optimized rate table generation model. The output of the optimized rate table generation model is the rate table corresponding to the product terms enhancement sample. For easy distinction, this rate table can be called the enhancement sample rate table.

[0137] Step 22: Evaluate the quality of the enhanced sample rate table and generate reward signals.

[0138] The reward signal is a numerical feedback generated after scoring the enhanced sample rate table from multiple dimensions. It is used to guide the parameter update of the optimized rate table generation model and reflects the overall quality of the output in terms of compliance, smoothness, and completeness.

[0139] In other words, the reward signal can be understood as a comprehensive quality score of the enhanced sample rate table generated by the optimized rate table generation model.

[0140] In some embodiments, step 22 may include: evaluating the quality of the enhanced sample rate table based on preset automatic verification rules and generating a reward signal, wherein the automatic verification rules include at least one of gradient relationship, cross-dimensional logical consistency, and coverage.

[0141] Automatic validation rules can be represented by scripts, which can score the quality of the generated enhanced sample rate table to achieve quality assessment of the enhanced sample rate table and obtain corresponding reward signals.

[0142] Automatic validation rules can include one of gradient relationships, cross-dimensional logical consistency, and coverage, or multiple of these (such as two or three).

[0143] The gradient relationship can be understood as enhancing the smoothness and monotonicity of the changes in values ​​in the sample rate table with key dimensions (such as age and coverage amount).

[0144] Continuing with the example of insurance products, the gradient relationship can include age-increasing factors. For instance, life insurance premiums should increase rather than decrease with the insured's age, and the variation between adjacent ages should be within a reasonable range (e.g., an annual growth rate ≤ 5%). Violation of this rule is considered an abnormal gradient. Specifically, it's important to examine whether an age range in the enhanced sample rate table strictly increases. If so, it can be given bonus points; if not (e.g., reverse order or skipping steps), it should be given deduction points.

[0145] Cross-dimensional logical consistency can be understood as strengthening the logical constraints of transaction requirements that the sample rate table must meet across multiple attribute dimensions (such as payment period, gender, and coverage period).

[0146] Continuing with the example of insurance products, cross-dimensional logical consistency can include compliance with payment timing and rates. For instance, comparing the enhanced sample rate table with the insurance terms and conditions, based on the payment start time, payout period, and rate changes stipulated in the insurance policy. If the enhanced sample rate table complies with the insurance policy, points are awarded; otherwise, points are deducted.

[0147] Coverage can be understood as ensuring that the sample rate table fully includes all product combinations that should be supported.

[0148] Continuing with the example of insurance products, the enhanced sample rate table is examined to see if it omits age groups, does not cover coverage areas, or has special boundary situations, compared to the insurance terms. Comprehensive coverage earns points; conversely, incomplete coverage results in points being deducted. Specifically, if the insurance terms allow "insurance for ages 18–65 + payment period of 10 / 20 years / to age 60," then the enhanced sample rate table must provide valid values ​​for each age-payment period combination, without omissions.

[0149] Based on the above analysis, in this embodiment, gradient relationships can ensure numerical rationality (avoiding jumps, inversions, etc.), cross-dimensional logical consistency can ensure the logical correctness of transaction requirements (avoiding internal contradictions), and coverage can ensure product integrity (avoiding functional deficiencies). Therefore, by using one or more of gradient relationships, cross-dimensional logical consistency, and coverage as automatic verification rules, i.e., by performing quality assessment based on automatic verification rules, the accuracy and effectiveness of quality assessment can be achieved.

[0150] Step 23: Update the parameters of the optimized rate table generation model with the goal of maximizing the reward signal.

[0151] Continuing with the example above, this step can be understood as follows: The training system adjusts the parameters (such as the low-rank parameter in the example above) of the optimized rate table generation model based on the comprehensive quality score (i.e., reward signal) of the enhanced sample rate table generated by the optimized rate table generation model, so that the enhanced rate table generation model can generate and output rate tables with relatively higher quality scores.

[0152] Based on the analysis of steps 21 to 23 above, it can be seen that in this embodiment, by using the third training dataset to train the optimized rate table generation model through reinforcement learning, and specifically adjusting the parameters of the optimized rate table generation model with the goal of maximizing the reward signal obtained from the evaluation, the optimized rate table generation model can be optimized in terms of multi-objective reward aspects such as inference compliance, inference accuracy, and rate table structural integrity. This can further improve the enhanced rate table generation model's ability to understand and execute relatively more complex product terms (such as rules in insurance terms). In other words, the enhanced rate table generation model can generate and output rate tables with relatively higher quality scores, thus improving the quality and reliability of the rate tables generated by the enhanced rate table generation model.

[0153] In some embodiments, reinforcement learning training further includes the following steps 31 and 32: Step 31: During the reinforcement learning training process, determine the mean value of the reward signal within N steps every N steps, where N is an integer greater than or equal to 1.

[0154] In reinforcement learning training, a step refers to the number of iterations for optimizing model parameters. In this embodiment, a step refers to the number of iterations performed to optimize the parameters of the optimized rate generation model.

[0155] Correspondingly, every N steps can be understood as performing a detection operation once every fixed number of parameter updates (such as every 500 gradient updates) during reinforcement learning training to determine the mean of the reward signal within N steps.

[0156] Therefore, the mean reward signal can be understood as the arithmetic mean of the automatic reward scores obtained by all enhanced sample rate tables generated by the optimized rate table generation model within the last N steps. The mean reward signal can be used to measure the overall output quality level of the optimized rate table generation model within the last N steps.

[0157] Step 32: If the average reward signal value is detected to be lower than the preset threshold within multiple consecutive N steps, data reinforcement is performed to adjust the parameters of the optimized rate table generation model based on the obtained sample data.

[0158] This embodiment does not limit the preset threshold, which can be determined by the training system based on requirements, historical records, experiments, etc.

[0159] Data augmentation can be understood as expanding or optimizing a third training dataset. For example, data augmentation may include: introducing more sample terms of similar products, generating adversarial edge cases, and injecting manually corrected samples to enhance the learning ability of the optimized rate table generation model in its weak areas.

[0160] For example, if the mean reward signal is lower than the preset mean over multiple consecutive N steps, such as a mean reward signal <0.7 (preset mean) over three consecutive 500-step cycles, it may be because the training data coverage of the third training dataset is insufficient. In this case, data augmentation can be performed to obtain more sample data.

[0161] For example, the next round of reinforcement learning training set can be determined by extracting more similar terms from the historical product database, generating boundary cases, or having low-scoring samples corrected by actuaries and then returned to the database as high-quality supervised samples. The next round of reinforcement learning training can then be conducted based on the corresponding training set.

[0162] Based on the analysis of steps 31 and 32 above, it can be seen that in this embodiment, the training system, through the mechanism of "periodic detection of the average reward + continuous low scores triggering data reinforcement," can transform reinforcement learning from a static training process into a dynamic repair process with a certain data reinforcement capability. This improves the effectiveness and reliability of reinforcement learning on the optimized rate table generation model, thereby enabling the reinforced rate table generation model to generate rate tables of relatively higher quality.

[0163] In some embodiments, step 32 may include the following sub-steps 321 and 322: Sub-step 321: If the mean value of the reward signal is detected to be lower than the preset threshold within multiple consecutive N-step periods, then determine the loss distribution information corresponding to the mean value of the reward signal being lower than the preset threshold within multiple consecutive N-step periods.

[0164] The information on the distribution of lost points can be understood as the statistical analysis of the types of quality defects and their frequencies corresponding to all low-reward samples in multiple consecutive N-step training cycles.

[0165] Among them, the information on the distribution of lost points can be determined by back-calculating the scores of automatic verification rules (such as gradient relationship, cross-dimensional logical consistency, coverage, etc.) to accurately locate the main weak links of the optimized rate table generation model in multiple consecutive N-step training cycles.

[0166] For example, if the average reward signal over three consecutive 500-step cycles is detected to be lower than a preset threshold, the training system can statistically analyze the quality defect types and frequencies of the samples corresponding to the average reward signal over those three consecutive 500-step cycles to obtain the corresponding score distribution information. This score distribution information might include: "70% of low-scoring samples are due to 'insufficient coverage,' 20% are due to 'payment period logic conflicts,'" and so on.

[0167] Sub-step 322: Perform data reinforcement based on the loss distribution information, and adjust the parameters of the optimized rate table generation model based on the obtained sample data.

[0168] Based on the above example, in this embodiment, if the average reward signal over three consecutive 500-step cycles is detected to be lower than a preset threshold, the training system does not blindly perform data reinforcement. Instead, it first determines the quality defect type and frequency statistics (i.e., the failure distribution information) of the samples corresponding to the average reward signal over those three consecutive 500-step cycles. Based on the failure distribution information, it determines on which type of samples the optimized rate table generation model repeatedly fails, then supplements the corresponding samples accordingly, and adjusts the parameters of the optimized rate table generation model based on the supplemented samples. For example, reinforcement learning training can continue on the current model (such as the optimized rate table generation model that has already undergone reinforcement learning training).

[0169] Based on the above analysis of sub-steps 321 and 322, it can be seen that in this embodiment, the training system first determines what kind of data should be reinforced based on the loss distribution information, and then performs data reinforcement on this basis. This can improve the pertinence, effectiveness, and reliability of data reinforcement, thereby improving the accuracy of the rate table generation model after reinforcement.

[0170] Based on the above analysis, the first training dataset can include a fourth data subset, which is a negative sample dataset. In some embodiments, samples whose mean reward signal value is lower than a preset threshold over multiple consecutive N steps can be added to the fourth data subset as negative sample data.

[0171] Furthermore, based on the above analysis, if the average reward signal value is lower than the preset threshold for multiple consecutive N steps, it may be due to the sample, and the training system can perform data reinforcement.

[0172] However, in other embodiments, if the average reward signal over multiple consecutive N steps is detected to be lower than a preset threshold, it may be due to an unreasonable configuration of the reward function used to determine the reward signal. In such cases, the training system can update the script used to determine the reward signal to avoid drawbacks caused by an unreasonable script.

[0173] In some embodiments, after obtaining the enhanced rate table generation model, the enhanced rate table generation model can be evaluated, and the evaluation entity can also be the training system.

[0174] For example, after S505, the training system can evaluate the enhanced rate table generation model based on the obtained test set and obtain the evaluation results.

[0175] The test set includes multiple test samples, and each test sample includes a product terms test sample and a corresponding rate table test label.

[0176] For each test sample in the test set, the training system can evaluate the quality of that test sample based on the enhanced rate table generation model. For example, the input to the enhanced rate table generation model is the test sample, and the enhanced rate table generation model is the rate table corresponding to the test sample (for ease of distinction, it can be called the test rate table). Accordingly, the training system can compare the test rate table with the rate table test labels to obtain the evaluation result.

[0177] The evaluation results include the quality score corresponding to the test sample, and the error type label corresponding to the test sample when the quality score is relatively low (which can be achieved by comparing with a preset threshold, such as when the quality score is less than the preset threshold).

[0178] Quality scoring can be achieved using the automatic validation rules shown in the example above. It can also be achieved by using or combining other factors. For example, the training system can analyze the reasonableness of the test rate table's values ​​(specifically, a higher score for more reasonable values) and verify its structural integrity (specifically, a higher score for a more complete structure) to determine the quality score of the test rate table.

[0179] Error type labels can be understood as identifiers that provide a structured classification of errors in the output of the enhanced rate table generation model (such as a life insurance rate table). They are not vague "bad," but rather precisely indicate "where the error occurred and what type of problem it belongs to."

[0180] Based on the above analysis, error type labels can be determined automatically based on gradient relationships, cross-dimensional logical consistency, coverage, and other validation rules, or they can be determined manually.

[0181] For example, error type labels can include: compliance error, logical consistency error, gradient / smoothness error, and so on.

[0182] A compliance error can be understood as the content of the generated test rate table violating the upper limit of the corresponding standard. For example, taking the test rate table corresponding to an insurance product as an example, the insured age in the rate table exceeds the coverage age in the insurance product terms and conditions.

[0183] A logical consistency error can be understood as the generated test rate table not conforming to the logic of the corresponding product terms. For example, taking the test rate table for an insurance product as an example, the annual premium for a 10-year payment plan is lower than that for a 20-year payment plan. However, in the insurance product terms, the 20-year payment plan is higher.

[0184] Gradient / smoothness errors can be understood as the generated test rate table not conforming to the increasing or decreasing logic of the corresponding product terms. For example, taking the test rate table corresponding to an insurance product as an example, the premium age limit in the rate table increases and then decreases, resulting in an "inverted" phenomenon.

[0185] Accordingly, the training system can further determine the error distribution report for the enhanced rate table generation model based on the evaluation results (including quality scores and error type labels), so as to characterize the performance shortcomings of the enhanced rate table generation model through error analysis reports. Based on this, targeted adjustments can be made to the enhanced rate table generation model to avoid its performance shortcomings.

[0186] In some embodiments, for high-risk test samples, the training system can interact with actuarial professionals to manually review and correct those test samples, forming a more standardized corrected sample.

[0187] The corrected samples can be fed back into the first training dataset, such as Figure 6 As shown, closed-loop data augmentation is performed on the first training dataset. The corrected samples can be used for subsequent retraining of the base network model, the optimized rate table generation model, and the augmented rate table generation model.

[0188] Based on the above analysis, in this embodiment, the training system can dynamically detect the performance of the enhanced rate table generation model by performing evaluations, and fine-tune the enhanced rate table generation model in a targeted manner based on the evaluation results, so as to improve the effectiveness and reliability of the final rate table generation model.

[0189] According to another aspect of this specification, this specification also provides a method for generating a rate table.

[0190] Similarly, this embodiment will not repeat the same or similar content as the examples above. For example, regarding the application scenarios of the rate table generation method, please refer to the examples above, and will not be repeated here.

[0191] Furthermore, the entity executing the rate table generation method can be a rate table generation system. This system can be the same as the training system or a different system. For the structure of the rate table generation system, please refer to the description of the training system in the example above; it will not be repeated here.

[0192] Please see Figure 7 , Figure 7 This is a flowchart illustrating the rate table generation method provided in the embodiments of this specification. Figure 7 As shown, the method includes: S701: Obtain terms for products pending processing.

[0193] S702: Input the product terms to be processed into the rate table generation model and output the rate table corresponding to the product terms to be processed. The rate table generation model is trained based on the training method described in the above embodiment.

[0194] Pending product terms refer to the terms and conditions text of products for which corresponding rate tables need to be generated.

[0195] Based on the above analysis, it can be seen that the rate table generation method provided in this embodiment can be applied to different scenarios. Therefore, the product terms to be processed can be the term texts of different products in different scenarios.

[0196] Taking the rate table generation method provided in this embodiment as an example applied to the insurance field, the product terms to be processed can be the text of insurance product terms, such as life insurance terms.

[0197] like Figure 8 As shown, the input to the rate table generation model is the product terms to be processed, specifically life insurance terms text. The rate table generation model is as follows: Figure 6 The model is generated based on the principle shown, specifically an enhanced rate table generation model. The output of the rate table generation model is a rate table, specifically as follows: Figure 8 The life insurance premium rate table shown.

[0198] Based on the above analysis, it can be seen that because the rate table generation model incorporates more dimensions of training data and has undergone preference optimization, reinforcement learning training, and evaluation fine-tuning, the rate table generated based on the rate table generation model has high accuracy and reliability.

[0199] It is worth noting that the above examples are merely illustrative of possible implementations of the training method described in this specification, and should not be construed as limiting the implementation of the training method described in this specification. For example, based on the above technical concept, some of the technical features described above can be combined to obtain new embodiments; new technical features can be added to the above examples to obtain new embodiments; some technical features can be removed from the above examples to obtain new embodiments; some technical features in the above examples can be replaced with other technical features; some technical features and their order in the above examples can be adjusted to obtain new embodiments, and so on, which will not be listed here.

[0200] Based on the above technical concept, this specification also provides a computer-readable non-transitory storage medium storing at least one instruction set, which, when executed by a processor, implements the steps of the training method described in this specification.

[0201] In some possible implementations, various aspects of this specification can also be implemented as a program product comprising program code. Taking the training method described above as an example, when the program product is run on the training system 200, the program code causes the training system 200 to execute the steps of the training method described in this specification. The program product used to implement the above method can employ a portable compact disc read-only memory (CD-ROM) containing program code and can run on the training system 200. However, the program product of this specification is not limited thereto. In this specification, a 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. The program product can employ any combination of one or more readable media. A readable medium can be a readable signal medium or a readable storage medium. A readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. The computer-readable storage medium may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable storage medium may also be any readable medium other than a readable storage medium that can send, propagate, or transmit programs for use by or in connection with an instruction execution system, apparatus, or device. Program code contained on a readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof. Program code for performing the operations described herein can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on training system 200, partially on training system 200, as a standalone software package, partially on training system 200 and partially on a remote training system, or entirely on remote training system 200.

[0202] It should be noted that the collection, storage, use, processing, transmission, provision, and disclosure of user-related information (such as user information required for insurance products) involved in the technical solutions of this specification all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0203] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0204] In summary, after reading this detailed disclosure, those skilled in the art will understand that the foregoing detailed disclosure is presented by way of example only and is not restrictive. Although not explicitly stated herein, those skilled in the art will understand that this specification requires various reasonable changes, improvements, and modifications to the embodiments. These changes, improvements, and modifications are intended to be made by this specification and are within the spirit and scope of the exemplary embodiments described herein.

[0205] Furthermore, certain terms in this specification have been used to describe embodiments of this specification. For example, "an embodiment," "an embodiment," and / or "some embodiments" mean that a particular feature, structure, or characteristic described in connection with that embodiment may be included in at least one embodiment of this specification. Therefore, it is to be emphasized and understood that two or more references to "an embodiment" or "an embodiment" or "alternative embodiment" in various parts of this specification do not necessarily refer to the same embodiment. Moreover, specific features, structures, or characteristics may be suitably combined in one or more embodiments of this specification.

[0206] It should be understood that in the foregoing description of the embodiments in this specification, various features are combined in a single embodiment, drawing, or description for the purpose of simplifying the description and to aid in understanding a feature. However, this does not mean that the combination of these features is necessary, and those skilled in the art, upon reading this specification, may readily identify some of the devices as separate embodiments. That is, the embodiments in this specification can also be understood as an integration of multiple secondary embodiments. And the content of each secondary embodiment is valid even if it contains fewer than all the features of a single foregoing disclosed embodiment.

[0207] Every patent, patent application, publication of a patent application, and other material cited herein, such as articles, books, specifications, publications, documents, and literature (excluding any related historical examination documents), is referenced for all purposes relevant to this document, including in the specification and claims herein. However, in the event of any inconsistency or conflict between the descriptions, definitions, and / or terms used in the foregoing and those used herein, the descriptions, definitions, and / or terms used herein shall prevail.

[0208] Finally, it should be understood that the embodiments disclosed herein are illustrative of the principles of the embodiments described in this specification. Other modified embodiments are also within the scope of this specification. Therefore, the embodiments disclosed in this specification are merely examples and not limitations. Those skilled in the art can implement the applications described in this specification using alternative configurations based on the embodiments in this specification. Therefore, the embodiments in this specification are not limited to the embodiments precisely described in the applications.

Claims

1. A training method for a rate table generation model, comprising: Obtain a first training dataset, wherein the first training dataset includes a first data subset and a second data subset. The first data subset includes product terms training samples and corresponding rate table labels. The second data subset is a temporal inference dataset, and the temporal inference dataset includes thought chain-style annotation information of time logic constraints associated with the first data subset. The first training dataset is input into the base network model to train the base network model and obtain the rate table generation model, wherein the rate table generation model is trained to automatically generate a rate table that meets the time logic constraints based on the input product terms.

2. The method according to claim 1, wherein, The base network model includes a pre-trained base model and a LoRA adapter; the first training dataset is input into the base network model to train the base network model, resulting in a rate table generation model, including: The parameters of the base model are frozen, and the low-rank parameters of the LoRA adapter are updated based on the first training dataset to train the base network model and obtain the rate table generation model.

3. The method according to claim 1, wherein, The first training dataset also includes a third data subset and / or a fourth data subset; wherein, The third data subset includes domain knowledge associated with the first data subset; The fourth data subset includes the negative sample dataset associated with the first data subset.

4. The method according to claim 1, wherein, The method further includes: Obtain a second training dataset, which includes several preference sample pairs. Each preference sample pair includes input text and labeled output corresponding to the input text. The input text is text associated with product terms. The labeled output includes positive sample rate table output and negative sample rate table output corresponding to the input text. The second training dataset is input into the rate table generation model, and the rate table generation model is trained with preference alignment using the direct preference optimization method to obtain the optimized rate table generation model.

5. The method according to claim 4, wherein, The preference alignment training of the rate table generation model using the direct preference optimization method includes: A preference contrast loss function is constructed based on the positive sample rate table output and the negative sample rate table output. This preference contrast loss function includes a regularization term, which is determined based on the KL divergence between the rate table generation model and the underlying network model. The rate table generation model is optimized by minimizing the preference contrast loss function.

6. The method according to claim 4, wherein, The method further includes: Obtain a third training dataset, which includes multiple product terms enhancement samples; and The optimized rate table generation model is trained using reinforcement learning based on the third training dataset, and for each product term reinforcement sample in the third training dataset, the reinforcement learning training includes: Based on the optimized rate table generation model, an enhanced sample rate table corresponding to the enhanced sample of the product terms is generated. The quality of the enhanced sample rate table is evaluated, and a reward signal is generated; and With the goal of maximizing the reward signal, the parameters of the optimized rate table generation model are updated.

7. The method according to claim 6, wherein, The reinforcement learning training also includes: During reinforcement learning training, the mean value of the reward signal within N steps is determined every N steps, where N is an integer greater than or equal to 1. If the average reward signal value is detected to be lower than a preset threshold within multiple consecutive N steps, data reinforcement is performed to adjust the parameters of the optimized rate table generation model based on the obtained sample data.

8. The method according to claim 7, wherein, If the average reward signal value is detected to be lower than a preset threshold over multiple consecutive N steps, data reinforcement is performed to adjust the parameters of the optimized rate table generation model based on the obtained sample data, including: If the average reward signal over multiple consecutive N steps is detected to be lower than a preset threshold, then the loss distribution information corresponding to the average reward signal over multiple consecutive N steps being lower than the preset threshold is determined; and Data reinforcement is performed on the aforementioned loss distribution information to adjust the parameters of the optimized rate table generation model based on the obtained sample data.

9. The method according to claim 6, wherein, The quality of the enhanced sample rate table is evaluated, and a reward signal is generated, including: The quality of the enhanced sample rate table is evaluated based on preset automatic verification rules to generate the reward signal. The automatic verification rules include at least one of gradient relationship, cross-dimensional logical consistency, and coverage.

10. The method according to claim 1, wherein, The product terms and conditions include insurance product terms and conditions, telecommunications tariff product terms and conditions, and utility pricing product terms and conditions. The rate tables accordingly include insurance rate tables, telecommunications tariff package tables, and tiered electricity pricing tables.

11. A method for generating a rate table, comprising: Obtain the terms and conditions for the products to be processed; The product terms to be processed are input into the rate table generation model, and the rate table corresponding to the product terms to be processed is output. The rate table generation model is trained based on the training method as described in any one of claims 1 to 10.

12. A training system for a rate table generation model, comprising: At least one storage medium storing at least one instruction set for training the rate table generation model; At least one processor is communicatively connected to the at least one storage medium, wherein when the at least one processor is running, it reads the at least one instruction set and executes the training method as described in any one of claims 1 to 10 according to the instructions of the at least one instruction set.

13. A rate table generation system, comprising: At least one storage medium storing at least one set of instructions for generating rate tables; At least one processor is communicatively connected to the at least one storage medium, wherein when the at least one processor is running, it reads the at least one instruction set and executes the generation method as described in claim 11 according to the instructions of the at least one instruction set.

Images