Method and device for generating simulation situation based on large language model and storage medium

Abstract

The invention provides a method and device for generating a simulation situation based on a large language model and a storage medium, and the method comprises the steps: carrying out the key element extraction of first input data based on a hierarchical architecture through the large language model, and generating a first key element corresponding to each layer in the hierarchical architecture; performing hierarchical diffusion deduction according to the extracted first key element and second input data through a large language model, and generating a diffusion deduction result corresponding to each layer in the hierarchical architecture; the diffusion deduction result is combined with an initiative link parameter to generate a situation result, and the initiative link parameter is generated by a scheduling model according to the second input data; wherein the hierarchical architecture is associated with the first input data, so that the fitting degree with customer requirements can be greatly improved, the computing power level is improved, and the consumption of time and resources is greatly reduced.

Description

Technical Field

[0001] The present application relates to the field of computer and artificial intelligence technology, and in particular to a method, device and storage medium for generating simulated scenarios based on a large language model. Background Art

[0002] Traditional implementations of complex situation simulations are mostly manual, traditional programming, or agent-based, including several typical methods:

[0003] 1. Rule system based on preset templates: using fixed templates + variable filling, relying on manually preset rule base, and simple parameter replacement generation. This implementation method has low efficiency of manual customization and usually takes 1-2 weeks to achieve customer approval.

[0004] 2. Programmatic generation technology: Use random algorithms to generate basic elements, ensure rationality through constraints, and combine algorithms to splice scenes. This implementation method can only meet the customization of specific industries such as the IT industry, which is a relatively structured industry, and cannot meet the scene reproduction of the entire industry.

[0005] 3. General AI-agent: Let the big language model define the workflow, define customer needs and scenarios, generate relevant materials, and use the big language model for quality inspection. This solution uses customized technology that consumes a lot of computing power, takes a long time, and has a low fit with customer scenarios.

[0006] The present invention aims to provide a method and system for generating simulation scenarios based on a large language model, so as to improve generation efficiency, reduce resource consumption, and expand the scope of application. Summary of the invention

[0007] The embodiments of the present application provide a method, device and storage medium for generating a simulated situation based on a large language model, so that the computing efficiency is greatly improved and the generated complex situation is highly consistent with the user's actual situation.

[0008] In order to solve the above technical problems, the embodiment of the present application provides a method for generating a simulation scenario based on a large language model, including:

[0009] Extract key elements from the first input data based on the hierarchical architecture using a large language model to generate a first key element corresponding to each layer in the hierarchical architecture;

[0010] Performing layered diffusion deduction based on the first key elements and the second input data extracted by a large language model to generate diffusion deduction results corresponding to each layer in the layered architecture;

[0011] generating a situation result by combining the diffusion deduction result with an active link parameter, wherein the active link parameter is generated by the scheduling model according to the second input data;

[0012] The hierarchical architecture is associated with the first input data.

[0013] Optionally, the method further comprises:

[0014] Modeling a hierarchical architecture according to the first input data, wherein each layer in the hierarchical architecture has a corresponding attribute range, and elements in each layer have attributes of the layer;

[0015] If the hierarchical architecture has M layers in total, the attribute range from layer 1 to layer M is reduced layer by layer, where M is an integer greater than or equal to 1.

[0016] Optionally, each X-1 layer element in the hierarchical architecture corresponds to one or more X layer elements, each X layer element belongs to an X-1 layer element, and there is no intersection between the elements, and X is a positive integer greater than or equal to 1 and less than or equal to M.

[0017] Optionally, each layer 1 to layer M-1 element in the layered architecture is a set of layer M elements;

[0018] The number of M-layer elements contained in the X-layer is a proper subset of the number of M-layer elements contained in the X-1-layer, where X is a positive integer greater than or equal to 1 and less than or equal to M.

[0019] Optionally, performing hierarchical diffusion deduction according to the first key elements and the second input data obtained by extraction through a large language model to generate diffusion deduction results corresponding to each layer in the hierarchical architecture includes:

[0020] Extracting information corresponding to the attributes of each layer in the hierarchical architecture from the second input data by means of a large language model as adjustment parameters of the layer;

[0021] For each layer in the layered architecture, the first key element corresponding to the layer is combined with the adjustment parameter of the layer to perform deduction to generate a diffusion deduction result of the layer.

[0022] Optionally, if the hierarchical architecture has M layers in total, the diffusion deduction result of the M-layer element includes the M-layer element attribute description and the 1st to M-1th layer elements to which the M-layer element belongs.

[0023] Optionally,

[0024] The key elements are extracted in the order of starting from the M layer and M decreasing;

[0025] The hierarchical diffusion deduction starts from layer X=1 and proceeds in the order of increasing X, where X is a positive integer greater than or equal to 1 and less than or equal to M.

[0026] Optionally, when the first input data is missing, the method further comprises:

[0027] Determine the layer corresponding to the missing in the layered architecture, extract key elements from the other layers in the M layers except the layer corresponding to the missing in descending order of M, and supplement the key elements of the layer corresponding to the missing according to the other layers except the layer corresponding to the missing;

[0028] In the case where the second input data is missing, the method further includes:

[0029] Determine a layer corresponding to the missing in the hierarchical architecture, and determine a supplementary parameter corresponding to the missing according to information of other layers in the hierarchical architecture through a large language model;

[0030] For the layer corresponding to the missing layer in the hierarchical architecture, the first key element of the layer is combined with the adjustment parameters of the layer and the supplementary parameters to generate a diffusion deduction result of the layer.

[0031] Optionally, the situation result includes execution data associated with the proactive link parameter, the execution data has M-layer attributes, the proactive link parameter has M-layer attributes, and the execution data and the corresponding proactive link parameter have one-to-one correspondence in attributes and timing;

[0032] If the M-layer element has character attributes, the execution data includes non-player character NPC corpus resources;

[0033] If the M-layer element has a teaching attribute, the execution data includes teaching data.

[0034] Optionally, the method further comprises:

[0035] The scheduling model schedules corresponding corpus from the corpus resources and feeds back to the user according to the non-player character NPC selected by the user and the active link parameter and in accordance with preset rules.

[0036] In a second aspect, an embodiment of the present application further provides a device for generating a simulation scenario based on a large language model, comprising:

[0037] A key element extraction module is configured to extract key elements from the first input data based on a hierarchical architecture through a large language model, and generate a first key element corresponding to each layer in the hierarchical architecture;

[0038] A hierarchical diffusion deduction module is configured to perform hierarchical diffusion deduction according to the first key elements and the second input data extracted through a large language model to generate diffusion deduction results corresponding to each layer in the hierarchical architecture;

[0039] A situation result module is configured to generate a situation result by combining the diffusion deduction result with an active link parameter, wherein the active link parameter is generated by the scheduling model according to the second input data;

[0040] The hierarchical architecture is associated with the first input data.

[0041] In a third aspect, an embodiment of the present application further provides a device for generating a simulation scenario based on a large language model, comprising a memory, a transceiver, and a processor, characterized in that:

[0042] The memory is used to store a computer program; the transceiver is used to send and receive data under the control of the processor; the processor is used to read the computer program in the memory and execute:

[0043] Extract key elements from the first input data based on the hierarchical architecture using a large language model to generate a first key element corresponding to each layer in the hierarchical architecture;

[0044] Performing layered diffusion deduction based on the first key elements and the second input data extracted by a large language model to generate diffusion deduction results corresponding to each layer in the layered architecture;

[0045] generating a situation result by combining the diffusion deduction result with an active link parameter, wherein the active link parameter is generated by the scheduling model according to the second input data;

[0046] The hierarchical architecture is associated with the first input data.

[0047] An embodiment of the present application further provides a processor-readable storage medium, wherein the processor-readable storage medium stores a computer program, and the computer program is used to enable the processor to execute the above method.

[0048] An embodiment of the present application also provides a computer program product, including computer instructions, which implement the steps of the above method when executed by a processor.

[0049] The beneficial effects of this application are:

[0050] The solution proposed in the present invention models a hierarchical architecture, where each layer has corresponding attributes and attribute ranges; extracts key elements from each layer based on the hierarchical architecture, and performs hierarchical diffusion deduction in combination with the adjustment parameters of each layer; and introduces active link parameters through an independent scheduling model to ultimately generate execution data for the underlying elements. The hierarchical architecture and scheduling model greatly improve the computing efficiency, and compared with the existing algorithms of simple application large models, the computing power consumption is low and the time consumption is short; through the characteristics of hierarchical derivative elements such as active link parameters, the simulated complex situations are highly consistent with the user situation and the user satisfaction is high. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0052] Figure 1 is a flowchart of a method for generating a simulation scenario using a large language model according to an embodiment of the present invention;

[0053] Figure 2 is a schematic diagram of a layered architecture of an embodiment of the present invention;

[0054] Figure 3 is a schematic diagram of a key element extraction process based on a layered architecture according to an embodiment of the present invention;

[0055] Figure 4 is a schematic diagram of a hierarchical diffusion deduction process based on a hierarchical architecture according to an embodiment of the present invention;

[0056] Figure 5 It is a schematic diagram of a process of combining diffusion deduction results with active link parameters to generate situational results according to an embodiment of the present invention;

[0057] Figure 6 is a schematic diagram of a large model interface layer of an embodiment of the present invention;

[0058] Figure 7 is a schematic diagram of a device for generating a simulation scenario based on a large language model according to an embodiment of the present invention;

[0059] Figure 8 Schematic diagram of a device for generating simulated scenarios based on a large language model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0060] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0061] The terms "first", "second", etc. in the specification and claims of the present application are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequential order. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present application described herein, for example, are implemented in a sequence other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, the process, method, system, product or equipment comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or equipment.

[0062] In the embodiments of the present application, the term "and / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can represent three situations: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the associated objects before and after are in an "or" relationship. In the embodiments of the present application, the term "plurality" refers to two or more than two, and other quantifiers are similar.

[0063] In the embodiments of the present application, words such as "exemplary" or "for example" are used to indicate examples, illustrations or descriptions. Any embodiment or design described as "exemplary" or "for example" in the embodiments of the present application should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary" or "for example" is intended to present related concepts in a specific way.

[0064] The relevant concepts mentioned in this application are briefly described below.

[0065] Large Language Model, or LLM for short, refers to a deep learning model trained with a large amount of text data that can generate natural language text or understand the meaning of language text.

[0066] Non-Player Character, or NPC for short, refers to a game character that is not controlled by real players.

[0067] In the present invention, a large model and a large language model have the same meaning.

[0068] Corpus, that is, language material, is the basic unit that constitutes the corpus. In the realization of complex multi-round multi-role dialogues, corpus is an important resource for realizing the dialogue function. By scheduling specific content from the corpus resources and feeding it back to the user (player), the dialogue between the user and the NPC is realized.

[0069] Corpus resources, that is, schedulable resources consisting of multiple corpora, are usually stored in computers in the form of corpora or collections.

[0070] Users and clients: Users in the present invention refer to users who use the system, AI agent, or game players. Clients can refer to the clients of users. For example, if a user is a company and wants to simulate the situation of the company's customers, then clients refer to the customers of the user.

[0071] In some applications, scenario simulation is used for business simulation. Business scenario simulation involves various industries and scenarios, such as business school course learning and corporate sales scenario simulation. They all need to conform to the characteristics of the industry and the actual business needs of the company, so they need to be highly customized and efficient. The existing simulation scenario generation method has problems such as low customization level, high computing power consumption, and low efficiency. According to experiments, the "general AI-Agent" based on a large language model in related technologies requires 145-215 prompts (1-2.5 hours) to reach the level of customer recognition. The method proposed in the present invention greatly improves the computing power level by modeling a hierarchical architecture and nesting small models. After simulation testing, this solution only needs less than 29 times (about 15 minutes) to achieve better results, greatly reducing the consumption of time and resources.

[0072] The embodiments of the present application are described below with reference to the accompanying drawings.

[0073] In a first aspect, the present application provides a method for generating a simulation scenario based on a large language model, such as Figure 1 shown.

[0074] like Figure 1 As shown, the method for generating a simulation scenario based on a large language model includes the following steps:

[0075] Step S101 , extracting key elements from first input data based on a hierarchical structure using a large language model, and generating first key elements corresponding to each layer in the hierarchical structure.

[0076] In step S101, the user can input the first input data in the form of a document for extracting industry characteristics. The first input data may include two parts, one part is the sales interview records, the company's success stories, the company's product introductions, the company's introductions, etc. provided by the user, and the other part is the industry public information actively obtained from public channels, not provided by the user; multiple documents as a whole are used as the first input data. If the first input data is in a non-document format, it can be converted into a document format for input. The first input data is input into the large language model by calling the API of the large language model, and the key elements are extracted according to the hierarchical architecture.

[0077] refer to Figure 2 , Figure 2 is a schematic diagram of a layered architecture according to an embodiment of the present invention.

[0078] In some embodiments, a hierarchical architecture may be modeled after receiving the first input data. The hierarchical architecture is associated with the first input data, and the first input data of different industries and different users may model different hierarchies. In some embodiments, the industry of the first input data is an important factor affecting the modeling of the hierarchical architecture. Each layer in the hierarchical architecture has a corresponding attribute range, and the elements in each layer have the attributes of the layer. If the hierarchical architecture has a total of M layers, the attribute range from layer 1 to layer M is reduced layer by layer.

[0079] For example, a four-layer architecture can be modeled according to the logic of execution layer (4 layers), project layer (3 layers), organization layer (2 layers), and environment layer (1 layer). Figure 2 As shown, the hierarchical attribute coverage of the 1st layer environment layer is the largest, the attribute coverage of the 2nd layer organization layer is smaller than that of the 1st layer, the attribute coverage of the 3rd layer project layer is smaller than that of the 2nd layer, and the attribute coverage of the 4th layer execution layer is the smallest. Each 1st layer element contains multiple 2nd layer elements, each 2nd layer element contains multiple 3rd layer elements, and each 3rd layer element contains multiple 4th layer elements. It should be noted that the execution layer (4th layer), project layer (3rd layer), organization layer (2nd layer), and environment layer (1st layer) are examples of hierarchical modeling logic, not limitations on the hierarchical architecture.

[0080] In some other embodiments, different hierarchical architectures can be modeled according to corresponding different logics for different first input data. For example, if the method and system described in the embodiments of the present invention are used to generate a physics teaching game, the first input data is a physics teaching text in a document format, and the modeled hierarchical architecture may be atomic layer (4 layers), molecular layer (3 layers), object layer (2 layers), and physical environment layer (1 layer); if the method and system described in the embodiments of the present invention are used to generate a hospital management simulation game, the first input data is a medical project plan in a document format, a hospital department division, etc., and the modeled hierarchical architecture may be a surgeon (4 layers), surgical operation (3 layers), surgery (2 layers), and a public hospital (1 layer).

[0081] The number of layers of the layered architecture is based on actual needs and the most efficient number of layers is selected. In actual applications, the 4-layer architecture is a preferred implementation method that has been tested to be relatively efficient. In one example, if a bank user inputs a bank financial product document as the first input data, the system can construct the first environmental layer as "bank financial market" (layer 1), the second organizational layer as "state-owned bank" (layer 2), "joint-stock bank" (layer 2), the third project layer as "investment banking department" (layer 3), "private bank" (layer 3), "product department" (layer 3), and the fourth execution layer as investment bank manager (layer 4) and financial advisor (layer 4).

[0082] In the method according to the embodiment of the present invention, the elements of each layer have the attributes of the layer, each X-1 layer (for example, 2 layers) element corresponds to one or more X layer (for example, 3 layers) elements, each X layer (for example, 3 layers) element belongs to an X-1 layer (for example, 2 layers) element, and there is no intersection between the elements, and X is a positive integer greater than or equal to 1. The attribute coverage of each layer is reduced layer by layer from the outside to the inside (1 to M), or from the top to the bottom (1 to M).

[0083] Among them, the M-th layer (a total of M-layer architecture) elements are relatively special compared to other layers. M-layer elements can be called bottom-level elements, terminal elements, leaf node elements, etc. The elements from 1 to M-1 layers are essentially a set of one or more M-layer elements, and the M-layer elements are the executors (NPCs) of the final generated game. The number of M-th layer (4th layer) elements contained in the X-th layer (for example, the 3rd layer) is a true subset of the number of M-th layer (4th layer) elements contained in the X-1 (2nd layer), and X is a positive integer less than or equal to M. For example, in some embodiments, there may be only one element in the 1st layer, and each of the other layers may have 1-10 elements. A 2-layer element contains two 3-layer elements, the first 3-layer element contains 4 4-layer elements, and the second 3-layer element contains 6 4-layer elements, then the 2-layer element contains 10 4-layer elements. That is, the number of 4-layer elements contained in the 3-layer element is a true subset of the number of 4-layer elements contained in the 2-layer element, and there is no intersection between the true subsets. The number of underlying elements in each layer increases from the inside to the outside (M to 1), or from the bottom to the top (M to 1).

[0084] Refer to Figure 3 , Figure 3 4 is a schematic diagram of a key element extraction process based on a layered architecture according to an embodiment of the present invention.

[0085] In some embodiments, Figure 3 The figure shows the process of extracting key elements based on the hierarchical architecture after modeling the hierarchical architecture.

[0086] S1011: Obtain input source.

[0087] The input source is the first input data, which may be, for example, customer input, success stories, product introductions, company introductions or other materials. If the input data is not in a document form, it is converted into a formatted document.

[0088] S1012: Hierarchical element extraction.

[0089] The first input data obtained by S1011 is input into the large language model to prompt the large language model according to the hierarchical modeling. In the example of the four-layer architecture, hierarchical features are extracted according to the execution layer (4 layers), project layer (3 layers), organization layer (2 layers), and environment layer (1 layer). The hierarchical features of each layer are related to the attributes of the layer, and the elements of each layer have the hierarchical features of the layer. For example, the execution layer can contain 1-10 NPC elements, and the hierarchical features of the elements are character characteristics, character traits, etc.; the project layer contains 1-10 project elements, and the hierarchical features of the elements are the characteristics of the project, such as project scale, project amount, staff, etc.; the organization layer contains 1-10 organization elements, and the hierarchical features of the elements are organizational characteristics, such as enterprise scale, equity nature, etc.; the environment layer contains only 1 element, and the hierarchical features of the elements are environmental characteristics, such as the financial industry, technology industry, etc.

[0090] S1013: Output the results of key element extraction.

[0091] After extracting the hierarchical features of each layer, the keyword description of each layer within 500 words is output as the first key element of the layer in combination with other description texts. The hierarchical features are obtained in step S1012, and the other description texts can come from the industry public information in the first input data. The key element can be a natural language text that includes the hierarchical feature description and other description texts.

[0092] It should be noted that, in some implementations, the order of extracting key elements based on the hierarchical architecture can be 4 layers → 3 layers → 2 layers → 1 layer, that is, first extract hierarchical features from the layer with the smallest attribute range, and extract key features layer by layer in the order of expanding attribute range. The advantage of this is that it is relatively easy to accurately locate elements with small-range attributes in the input data, and then expand the attribute range layer by layer with the resolved small-range attribute elements as anchors, which can greatly improve the efficiency and accuracy of element generation, and if the information of a certain layer is missing, a position can be reserved for that layer, and the element features of a larger layer can be extracted first.

[0093] As mentioned above, the number of elements in each layer (M layers) increases from the inside to the outside (M to 1), or from the bottom to the top (M to 1). The elements from 1 to M-1 layers are essentially a collection of the elements in the M layer, and each layer is also a collection of the elements in the next layer (X+1). Figure 3In the example, the elements from layer 1 to layer 3 are all sets of layer 4 elements. Specifically, in the example of generating a training / sales game scenario, a layer 4 element can be multiple employees (such as 1-10 NPCs); a layer 3 element can be a project, which is composed of multiple employees (such as multiple NPCs belonging to the project); a layer 2 element can be an organization or department, which is composed of multiple projects (such as 1-10 layer 3 project elements belonging to the department), and each project (layer 3 element) is composed of multiple employees (layer 4 elements); a layer 1 element can be an enterprise or industry, which is composed of multiple organizations or departments (layer 2 elements), each organization or department is composed of multiple projects (layer 3 elements), and each project is composed of multiple employees (layer 4 elements). Figure 3 The key feature extraction process shown generates feature descriptions of each layer of elements, rather than actually generating the elements. The output of the key feature extraction process can be any form of feature description, preferably a keyword description of less than 500 words per layer, which has been tested to be highly efficient.

[0094] For example, if the client is in the new energy vehicle industry:

[0095] Output of the first layer: Description of the social environment of new energy vehicles, such as consumption habits, business regulations, etc.

[0096] Output of the second layer: Description of the organization, such as company type (sole proprietorship, joint venture), scale, company attributes, advantages, upstream and downstream of the company, related industries, etc.

[0097] Level 3 output: Characteristics of projects within the company, such as procurement projects, transformation projects, investment projects, IT projects, etc.

[0098] The fourth layer outputs: name characteristics (Chinese name, English name, other foreign names), job characteristics (business position, R&D position, production position), decision-making type (cautious, radical, neutral), and personality characteristics (Big Five personality characteristics, etc.) of the people (NPCs) related to these projects.

[0099] It should be noted that step S101 extracts general hierarchical features based on the first input data, such as hierarchical features of a certain industry (new energy vehicle industry). The industry hierarchical features can be applicable to multiple enterprises and organizations. Therefore, step S101 does not need to be executed every time the method described in the embodiment of the present invention is executed, but can be called as a general industry feature intelligent agent (AI agent) when needed.

[0100] Back to Figure 1 , step S102, performing hierarchical diffusion deduction based on the first key feature and the second input data extracted by a large language model to generate a diffusion deduction result corresponding to each layer in the hierarchical architecture.

[0101] refer to Figure 4 , Figure 4 4 is a schematic diagram of a hierarchical diffusion deduction process according to an embodiment of the present invention.

[0102] like Figure 4 As shown, the input of the hierarchical diffusion deduction process includes the key element extraction results of each layer and the adjustment parameters of each layer. The adjustment parameters can be obtained by adjusting the AI ​​agent from the second input data.

[0103] The hierarchical diffusion deductive process is used to focus general industry-specific scenarios on a certain enterprise or organization, generate instances of that enterprise or organization, and the resulting simulated scenario game is also targeted at that enterprise or organization.

[0104] S1021: Generate adjustment parameters.

[0105] In order to focus on a specific enterprise or entity, the method described in the embodiment of the present invention needs to obtain second input data. The second input data can be relevant documents of a certain enterprise or organization, such as a project description for enterprise A, which is used to focus the industry hierarchical features obtained in the key element extraction process on a certain enterprise or entity.

[0106] In some embodiments, the second input data is input into the large language model, and the corresponding information is extracted as the adjustment parameters of each layer in the hierarchical architecture according to the attributes of each layer. The step of generating the adjustment parameters can be completed by an independent AI agent to improve the computing efficiency. For each layer in the hierarchical architecture, the first key feature corresponding to the layer is combined with the adjustment parameters of the layer for deduction to generate a hierarchical diffusion deduction result of the layer. Diffusion deduction refers to the introduction of adjustment parameters, the expansion of the characteristics of each layer or each element, and the generation of further descriptions.

[0107] Adjustment parameters are parameters extracted from the second input data and associated with each layer of attributes, and are used to further supplement the attribute characteristics of each layer related to the second input data. Adjustment parameters are inherent attributes of an element itself. For example, if an element in the 4th layer is an NPC (person), then the adjustment parameters are a series of personality parameters, such as extroversion, introversion, gentleness, impatience, etc., which will affect the behavior of the NPC. Since the adjustment parameters are extracted based on the second input data, the adjustment parameters are specific parameters for Enterprise A and are used to diffuse and deduce the characteristics of each layer.

[0108] In some embodiments, for example, step S101 extracts the first key feature and models a 4-layer architecture to describe the banking industry; then the 1st layer feature derived by combining the diffusion of the adjustment parameters of enterprise A can be a description of the environment layer to which bank A belongs, such as a joint-stock bank, which has the characteristics of large size and comprehensive business.

[0109] S1022: Layered Diffusion

[0110] like Figure 4 As shown, the input of each layer of hierarchical diffusion deduction includes at least one of the following:

[0111] 1) The feature description of the corresponding layer generated in “Key element extraction” within 500 words;

[0112] 2) The adjustment parameters corresponding to each layer;

[0113] 3) Additional parameters (optional).

[0114] The hierarchical diffusion deduction process adjusts the first key feature of each layer with adjustment parameters. If there is a missing feature, it is supplemented by generating supplementary parameters through the large language model. Specifically, the adjustment process can be to expand more features in the general features of the industry so that it can reflect the characteristics of enterprise A.

[0115] S1023: Output the results of hierarchical diffusion deduction

[0116] Output of the first layer: Specific description of the social environment related to Company A, such as consumption habits, business regulations, etc.

[0117] Output of the second layer: Specific organizational description of Company A, such as the nature of Company A (foreign or domestic capital), name, introduction, main business, organizational structure, collective consciousness, etc.

[0118] Output at the third level: Specific project descriptions within the A company organization, such as actual procurement projects, transformation projects, investment projects, collective consciousness, etc.

[0119] The fourth layer output: the name, position, decision-making type, personality traits, and which layer 3, layer 2, and layer 1 elements the people (NPCs) within Company A who are related to these projects belong to.

[0120] Layered diffusion deduction result: the collection of the above four layers of output.

[0121] It should be noted that the output of the bottom layer (Mth layer) is different from the output of other layers. In the above example, "Output of the 4th layer: the name, position, decision type, personality traits of the people (NPC) related to these projects, and which 3rd layer, 2nd layer element, and 1st layer element they belong to", it can be seen that in addition to the inherent characteristics of the element itself, the output of the bottom layer element also needs to output which upper layer element it belongs to.

[0122] In some implementations, the order of layered diffusion deduction using a large model is strictly in the order of 1st layer → 2nd layer → 3rd layer → 4th layer, that is, the layer with the largest attribute range is diffused and deduced first, and then diffused and deduced layer by layer in the order of narrowing the attribute range. Missing data is supplemented by a large model in combination with the context. The advantage of this is that, as mentioned above, in terms of attribute range, the X-layer elements belong to the X-1 layer elements. In the 4-layer architecture modeling, the range covered by the attributes of the 1st layer elements is the largest, and the 1st layer elements (such as the domestic banking and financial market) contain the largest number of 4-layer elements (NPC, people). Starting from the 1st layer elements, it is necessary to first establish the commonality of all NPCs, and then distinguish the different characteristics of the next layer layer by layer, and iterate until the characteristics of each NPC are deduced. This order of first deducing the commonality and then deducing the characteristics of each element can avoid repeated deductions and greatly improve the computing efficiency of the entire system.

[0123] In some embodiments, the second input data may have missing information, resulting in the inability to extract the adjustment parameters of a certain layer. In this case, the layer corresponding to the missing information in the hierarchical architecture can be first determined, and the supplementary parameters corresponding to the missing information can be determined by the large language model based on the information of other layers in the hierarchical architecture; for the layer corresponding to the missing information in the hierarchical architecture, the first key feature of the layer is combined with the adjustment parameters and supplementary parameters of the layer to generate the diffusion deduction result of the layer.

[0124] For example, if the user's project data (located at the 3rd layer) is missing because it is confidential or sensitive, the method to supplement the data is through the data of the 1st layer. For example, the banking industry is promoting the reform of supply chain finance, combined with the 2nd layer data, that is, the focus of a certain city where Bank A is located this year is several large syndicated orders. From this, it can be inferred that Bank A's projects this year are likely to include projects such as "supply chain finance" and "standby syndicated". If enough data can be obtained, it can also be tested between peer companies. This method of supplementation is efficient and accurate, and project information can be automatically generated without the user providing sensitive data again, greatly improving user satisfaction.

[0125] In some implementations, entities of each layer of elements may be generated in a layered diffusion deduction step and called after the game starts.

[0126] Back to Figure 1 , step S103, combining the diffusion deduction result with the active link parameter to generate a situational result, wherein the active link parameter is generated by the scheduling model according to the second input data.

[0127] refer to Figure 5 , Figure 5 It is a schematic diagram of a process of combining hierarchical diffusion deduction results with active link parameters to generate situational results according to an embodiment of the present invention.

[0128] In some embodiments, the scheduling model can be an independent AI agent. In the process of generating a simulation scenario, or in the AI ​​agent that generates the simulation scenario, the role of the scheduling model is to provide active link parameters; after the game or other execution activities start, the role of the scheduling model is to schedule the corresponding corpus or other execution data through the active link parameters to promote the progress of the game or other execution activities.

[0129] The implementation method of providing active link parameters in the process of generating simulation scenarios is that the scheduling model inputs the second input data and the user role characteristics into the large language model to obtain N active link parameters under the M-layer architecture, where M and N are integers greater than or equal to 1. When element A wants to actively link element B, what options or actions can be taken? These options are called "active link parameters". In practice, it is preferred to select a 4-layer architecture and 16-30 active link parameters.

[0130] In some examples, if A is a mobile phone and B is another mobile phone, "sending information via GSM" and "sending information via WIFI and the World Wide Web" are two different proactive link parameters. If A is a salesperson and B is a customer. "Visiting the customer's office" and "inviting the customer to dinner" are two different proactive link parameters. Each element may have its own specific proactive link parameters, which are affected by the inherent properties of the element itself. For example, 5G mobile phones can connect to other mobile phones through the 5G network, but 4G mobile phones cannot use the 5G network. Conversely, when B wants to actively contact A, it is the same. B has its own proactive link parameter list. Maybe B is a little weaker, it only has 8 parameters, unlike A which has 16, which means that A has more proactive link methods and stronger capabilities.

[0131] In the scheduling model, the M-layer architecture is represented by the attributes of each M-layer element from the inside to the outside. The logic of M-layer modeling is the same as that of the intelligent agent for generating simulation scenarios, but the perspective is that the M-layer elements are numbered in the order of observing outward from the self-center, and the order may be opposite to the M-layer described above. For example, in the scheduling model, the 4-layer architecture is represented by the attributes from the inside to the outside observed by the NPC with itself as the center, that is, layer 1 is the self-attribute layer, layer 2 is the team attribute layer, layer 3 is the organization attribute layer, and layer 4 is the environment attribute layer. In other embodiments, the order of numbering may be the same as that of the intelligent agent for generating simulation scenarios, that is, layer 1 is the environment attribute layer, layer 2 is the organization attribute layer, layer 3 is the team attribute layer, and layer 4 is the self-attribute layer. For ease of description, the embodiments of the present invention are described according to the same layer numbering.

[0132] The proactive link parameters are related to actions. In the scheduling model, actions also have M-layer attributes. M-layer actions or proactive link parameters of M-layer actions correspond to M-layer attributes of M-layer elements (such as NPCs) in logic and timing.

[0133] The active link parameter of the M-layer action corresponds to the M-layer attribute of the M-layer element (such as NPC) in a logical one-to-one correspondence, which means that the action taken by the m-th layer active link parameter to actively establish a link with the NPC based on a specific topic is within the scope represented by the m-th layer attribute of the NPC and within the scope of the specific topic, m = 1, 2, ..., M. For example, the active link parameter of the 4th layer action is "emotional resonance: expressing understanding and support", which corresponds to the 4th layer attribute of the NPC, the self-attribute, that is, expressing emotional resonance belongs to the scope of private topics and content related to the self, and does not belong to the 3rd layer team attribute scope of the NPC, nor does it belong to the 2nd layer organization attribute scope of the NPC, nor does it belong to the 1st layer environment attribute scope of the NPC. For another example, the active link parameter of the 1st layer action is "understanding the big environment: conducting in-depth research on the latest regulations on the construction of hospital PCR laboratories", which corresponds to the 1st layer environment attribute of the NPC, that is, conducting in-depth research on the latest regulations on the construction of hospital PCR laboratories is an action to understand the industry situation, which belongs to the scope of environmental attributes, and does not belong to the 4th layer self attribute scope of the NPC, nor does it belong to the 3rd layer team attribute scope of the NPC, nor does it belong to the 2nd layer organization attribute scope of the NPC. Within the scope of a specific topic, it means that the active link parameters of each level of action are related to the specific topic. For example, in the above example, the active link parameter of the first-level action is "Understand the general environment: conduct in-depth research on the latest regulations on the construction of hospital PCR laboratories", which corresponds to the first-level environmental attributes of the NPC and is related to the specific topic "Purchase PCR instruments" project, but not to another topic "Purchase printers" project.

[0134] The one-to-one correspondence between the proactive link parameters of the M-layer action and the M-layer attribute of the NPC in terms of time sequence means that the proactive link parameters of the M-layer action represent the sequence of actions taken by the NPC to proactively establish links based on a specific topic, which conforms to the sequence of affairs development of the specific topic. For example, the sequence of sales action development is: calling unfamiliar customers, meeting with potential customers, writing proposals, bidding, negotiation, and signing contracts. The sequence of customer procurement is: project research period, planning period, evaluation period, bidding period, and implementation period. The first action taken by the player user as a salesperson is to call and visit, which corresponds to the first-layer environmental attribute of the NPC. Therefore, when the player user calls the customer, the customer will only talk about the information about the general environment that everyone knows during the project research period, that is, the information within the conversation range of the environmental attribute. But if the player user writes a proposal at the beginning, the proposal will fall into oblivion, which will result in a mismatch in timing.

[0135] It should be noted that the embodiment of the present invention describes a method for generating a simulation scenario, which can be encapsulated into an AI Agent for the simulation scenario in a specific implementation. The AI ​​Agent for the simulation scenario is called in applications such as games and teaching. For example, before the game starts, the scheduling model will cooperate with the simulation scenario agent to generate the simulation environment and all the materials required for the game. When the game starts, the scheduling model will play the role of the game engine to complete the entire specific game interaction and the scheduling of the materials.

[0136] In some implementations, the inputs to the process of generating a situational result are the result of the hierarchical diffusion deduction and the N proactive link parameters generated by the scheduling model.

[0137] S1031: Extract features from the results of hierarchical diffusion deduction

[0138] The results of layered diffusion deduction are extracted to form a feature description of less than 1,000 words.

[0139] The extraction process is to simplify the feature description to improve the operation efficiency. In some implementations, a large language model can be used to compress the word count of the layered diffusion deduction results without losing key features. 1,000 words contain the feature descriptions of all layers. Other implementations can also be used to implement extraction, such as performing semantic analysis according to user needs, and building an industry extraction AI agent to call when needed.

[0140] S1032: Scheduling active link parameters.

[0141] The active link parameters are generated by the scheduling model, which extracts the specific characteristics of the A company from the second input data, and combines the characteristics of the player / user role to generate corresponding N active link parameters based on the large language model as the preset input of S1033. Preferably, there can be 16-30 active link parameters, which is the number of parameters with better performance after actual testing.

[0142] It should be noted that the 16-30 active link parameters have a common purpose, such as sales, that is, these 16 parameters are designed around how to sell. The scheduling model AI agent can generate one or more corpora (for example, 1-5 corpora) for each layer of the NPC according to the scheduling weight and the M-layer mapping rules based on at least one or more of the following:

[0143] 1. “Sales” purpose;

[0144] 2.NPC positions;

[0145] 3. Specific companies, departments, and projects on the 1st to 3rd levels to which the NPC belongs;

[0146] 4. The role played by the NPC in the project (generated by AI, such as key decision-making / participating in decision-making / non-decision-making).

[0147] In some embodiments, a corresponding corpus can be generated for each active connection parameter, but doing so will generate a large amount of corpus, which may also be repeated and inefficient. Preferably, efficiency can be improved through the hierarchical modeling of the present invention. For example, the hierarchy of active link parameters can be matched with the hierarchy of scheduling weights when generating corpora. When running the game, the user actively selects the active link parameters, and the active link parameters only affect the scheduling weights, and the corpus is retrieved by changing the scheduling weights. In this way, the decoupling of active link parameters and corpora is achieved, and the number of required corpora is reduced exponentially. The scheduling weight is a parameter used in the scheduling model and is used to retrieve the corpus.

[0148] It should be noted that there is no restriction on the execution order of S1031 and S1032, as long as S1033 can obtain two inputs.

[0149] S1033: Output situation results

[0150] The extraction results are combined with the N proactive link parameters and / or related texts preset in the scheduling model to generate situational results as output. Optionally, the output can be generated in combination with constraints to further improve accuracy. Constraints may include but are not limited to logical consistency, timing alignment, consistency in description of things, etc. Among them, the related text is a specification description of the output. For example, if the output execution data is NPC discourse, the specifications output through the related text description include: text format, text quantity, special requirements and other specifications. If the output execution data is teaching materials such as case study learning cases, the specifications output through the related text description include: case study requirements, whether the output is a student manual or exercises, the format of the student manual, the number of exercises and other specifications.

[0151] In some embodiments, the output context result includes the attributes of the M-layer element and the execution data related to the attribute. For example, if the M-layer element is an NPC, the execution data can be the speech of the NPC, which can be called the corpus resource or corpus of the NPC. Specifically, the execution data (such as corpus resources) also has hierarchical attributes, which are used for scheduling models according to rules. For example, by configuring the scheduling model rules, the corpus resources of layer 1 can be scheduled through the active link parameters of layer 1, and the corpus resources of layer 2 can be scheduled through the active link parameters of layer 2. It can also be configured that the corpus resources of layer 1 are scheduled through the active link parameters of layer 2, and the corpus resources of layer 4 are scheduled through any active link parameters. In some embodiments, examples of 16 active link parameters generated by the scheduling model under the 4-layer architecture are as follows:

[0152] A. Self-attribute layer (4 layers) - personal active linking method

[0153] 1. Professional consultation: proactively share professional knowledge

[0154] 2. Emotional empathy: expressing understanding and support

[0155] 3. Resource exchange: provide your own resources in exchange for cooperation

[0156] 4. Experience sharing: sharing past successful cases

[0157] Example corpus with 4 layers of proactive link parameter scheduling:

[0158] "I finally understood the key to this professional problem. Everything I had thought was wrong."

[0159] "I'm glad you understand that we are indeed facing unprecedented difficulties"

[0160] "I also have some relevant resources to share, maybe we can exchange resources"

[0161] "I have encountered similar situations before, and your sharing is very helpful for us."

[0162] B. Team attribute layer (3 layers) - active cross-team linking method

[0163] 1. Project collaboration: proactively seek opportunities for project collaboration

[0164] 2. Resource integration: Propose resource sharing solutions

[0165] 3. Team mutual assistance: providing team-level support

[0166] 4. Cross-departmental activities: launching joint activities

[0167] Example corpus with 3 layers of proactive linking parameter scheduling:

[0168] "Thanks for your team's support, it is very helpful for us. We currently have 15 developers in our team and are working on performance optimization issues"

[0169] "Regular meetings are a good idea. Our testing team is a bit short on staff recently and we are recruiting for three positions."

[0170] "You are very welcome to join us for the training. Our team has just completed the technical architecture upgrade and it is time to learn."

[0171] "Thank you for sharing. These best practices are very inspiring to us. Our team's employee satisfaction rate reached 95% this quarter."

[0172] C. Organizational attribute layer (2nd layer) - Organizational level active linking method

[0173] 1. Strategic docking: Seeking cooperation based on company strategy

[0174] 2. Process optimization: Propose cross-departmental process improvements

[0175] 3. Cultural construction: initiate and organize cultural integration activities

[0176] 4. System improvement: Propose suggestions for system coordination

[0177] Example corpus with 2-layer proactive link parameter scheduling:

[0178] "It is indeed in line with the company's strategy and we are happy to participate. Our department's innovation project budget has increased by 30% this year."

[0179] "This process optimization suggestion is very good. We have just completed the department system reorganization, so we can adjust it together."

[0180] "The cultural exchange activities are very helpful. Our department will move to a new office area next month, which will provide a better space for interaction."

[0181] "A unified incentive mechanism is indeed necessary. The cross-team collaboration KPI weight of our department this year has been raised to 40%."

[0182] D. Environmental attribute layer (layer 1) - external environment active link mode

[0183] 1. Market opportunities: Seek cooperation based on market opportunities

[0184] 2. Regulatory response: Joint response to regulatory changes

[0185] 3. Competitive response: Joint response to competitive pressure

[0186] 4. Innovative cooperation: jointly seize opportunities for technological innovation

[0187] Example corpus with 1-level proactive link parameter scheduling:

[0188] "This market opportunity is great. Our recent market research shows that user demand in this area has increased by 60%"

[0189] "Cooperation in developing new products is a wise choice. We have observed that all three competitors are laying out related technologies"

[0190] "It is necessary to jointly deal with competition. As far as we know, the largest competitor has invested 20 million in research and development."

[0191] "This innovative project is very promising. We have applied for 5 related patents and can collaborate on innovation"

[0192] In some implementations, the scheduling model has only 16-30 text parameters and 1 model weight (scheduling weight, implemented as favorability in some implementations), which greatly saves system resource consumption while meeting the customer's satisfaction level. It can be run on web pages, mobile phones, and even various low-end devices, and its effect is comparable to that of a large model (because the text it interprets is extracted from the large model). The scheduling model is an independent AI Agent with few input parameters and high fidelity of decision results, which is easy to understand and explain, making it easy for trained students to transfer the skills they have learned from practicing with the model to actual work.

[0193] This step has different outputs depending on the usage scenario of the system, and the output can be various execution data for subsequent use. For example, in the scenario of simulating a business / sales training game for users, the output of this step can include NPC lines, which are used to execute NPC dialogues with users in the game. In the scenario of generating training courses for users, the output of this step can be corresponding teaching materials or courses, which are used to execute course generation in the training course scenario. The situational result is that each element at the bottom (such as NPC) describes their image of the upper layer from a first-person perspective based on the affairs they are engaged in in the "scheduling model", referring to the business environment (1st layer), company (2nd layer), department (3rd layer), position (4th layer) and other information in the hierarchical diffusion deduction results. Each bottom-level element outputs execution data, which must be aligned and consistent in logic, description of affairs, and timing.

[0194] Specifically, the context results include the execution data of the M-layer elements,

[0195] If the M-layer element has character attributes, the situational result includes the NPC's corpus resources;

[0196] If the M-layer elements have instructional attributes, the situational outcomes include instructional materials;

[0197] If the M-layer element has other execution attributes, the context result may include other execution data of one or more M-layer elements.

[0198] The extracted features (e.g., 1,000 words) and the execution data of the M-layer elements can be packaged into applications such as games and teaching plans, and then scheduled for subsequent players to play games or users to teach.

[0199] In some embodiments, after the situation results are generated according to the method of the embodiment of the present invention, the situation results can be used as inputs of the scheduling model. When the scheduling model starts the game, the scheduling model schedules the corresponding corpus from the corpus resources and feeds it back to the user according to the non-player character NPC and the active link parameter selected by the user and the preset rules.

[0200] In summary, in one embodiment, the first input data is industry data, and the second input data is business entity data;

[0201] The layered architecture includes M = 4 layers. After key feature extraction, layers 1 to M have the industry's environmental attributes, organizational attributes, project attributes, and execution attributes respectively;

[0202] The diffusion deduction result includes M=4 layers. After layered diffusion deduction, layers 1 to M respectively have the environmental attributes of the business entity, the organizational attributes of the business entity, the project attributes of the business entity, and the execution attributes of the business entity.

[0203] It should be noted that each of the above steps S101, S102, and S103 can be used as an AI Agent, or each sub-step in steps S101, S102, and S103 can be used as an AI Agent.

[0204] The following uses banking customers as an example to illustrate the process of key factor extraction and hierarchical diffusion deduction.

[0205] First input data: Customer sales case in banking industry

[0206] Year-on-year business expansion for client (Bank A) 2018

[0208] Demand: Research services;

[0209] Key services: Signing of research service agreement;

[0210] Income generated: about 100,000;

[0211] Difficulty: Market sentiment was poor at the time and the client only had a single business need for research. 2020

[0213] New customer demand: bulk transaction, successfully matching a 5% equity transfer agreement of a customer;

[0214] Bulk transaction matching business: the total revenue for the year was about 1.2 million yuan;

[0215] Difficulty: The agreement transfer business involves a long exit time. The customer's share reduction period after receiving the ticket is about one and a half years, and the subsequent willingness to participate in such projects is reduced. 2022

[0217] New customer needs: Participate in option business to serve customers' needs for private placement and bulk arbitrage, and place both long and short positions for hedging on the OTC trading desk for trading;

[0218] New revenue: approximately RMB 1 million in revenue from over-the-counter derivatives;

[0219] Difficulty: The client's long and short positions jointly call for insurance, requiring a reduction in the margin ratio. ......

[0221] Other description text:

[0222] Based on the collected public information, such as the excerpt from "Money and Banking"

[0223] Chapter 1 Overview of Global Banking

[0224] The global banking industry is an important pillar of the modern financial system and plays a key role in economic operations. From the perspective of functional positioning, the modern banking industry has formed a multi-level and multi-type operating system. As the most basic type of bank, commercial banks mainly provide traditional financial services such as deposits and loans, payment and settlement, and are an important link between savers and investors. Investment banks focus on capital market businesses such as securities underwriting, mergers and acquisitions, and restructuring, and play a unique role in corporate financing. With the deepening of financial innovation, the multifunctional banking model has begun to emerge, operating commercial banking and investment banking businesses at the same time, and providing a full range of financial services. In recent years, with the development of Internet technology, purely online Internet banks have gradually emerged. ......

[0226] The method according to the embodiment of the present invention first obtains the first input data, combines other publicly collected descriptive texts, and models the four-layer architecture of the banking industry where Bank A is located based on the first input data and the purpose of use (generating banking business / sales training games) based on the large language model. For example,

[0227] The first layer: the environmental layer, has a 1-layer element named “domestic banking and financial market” (1-layer);

[0228] The second layer: the organizational layer, under the "domestic banking and financial market" (1st layer), is the organizational layer with four elements, namely "state-owned banks" (2nd layer), "joint-stock banks" (2nd layer), "urban industry banks" (2nd layer), and "rural commercial banks" (2nd layer);

[0229] The third layer: project layer, for example, under the 2nd layer element "state-owned bank" (2nd layer), there are "investment banking department" (3rd layer), "private banking" (3rd layer), and "product department" (3rd layer);

[0230] The fourth level: the executive level. Under the “Investment Banking Department” (3rd level) are investment banking managers (4th level) and financial advisors (4th level).

[0231] At the same time, because the "domestic banking and financial market" (1st layer) contains not only banks, but also bank customers, such as securities companies (2nd layer), listed companies (2nd layer), small and medium-sized enterprises (2nd layer), individuals (2nd layer), etc., these together constitute a first-layer environment. When focusing on analyzing the bank's customers "listed companies (2nd layer)" at the second layer, the corresponding elements of the 3rd and 4th layers corresponding to the "listed companies (2nd layer)" will be traversed and decomposed again.

[0232] Next, the key feature extraction process is performed through the large language model. According to the four-layer architecture of the banking industry, the key features of each layer of the banking industry are extracted in the order of "investment bank manager" (4th layer) → "investment bank department" (3rd layer) → "state-owned bank" (2nd layer) → "domestic bank financial market" (1st layer), and the keyword description of each layer within 500 words is output as the first key feature. Under "state-owned bank" (2nd layer), there is not only "investment bank department" (3rd layer), but also "private bank" (3rd layer), so iterative extraction is performed. For "private bank" (3rd layer), it is extracted again in the order of "investment bank manager" (4th layer) → "private bank" (3rd layer) → "state-owned bank" (2nd layer) → "domestic bank financial market" (1st layer). The key features of each layer and each element in the four-layer architecture of the banking industry are extracted in the order of 4th layer → 3rd layer → 2nd layer → 1st layer. It should be noted that the first key feature is the industry description, which can be reused by different companies many times. Each banking enterprise user needs to generate a training game for his own enterprise through hierarchical diffusion deduction, and can directly reuse the key features of the banking industry without the need for each user to re-execute the key feature extraction process. When a player user wants to generate a business / sales training game for Bank A, the method and system described in the embodiment of the present invention are used to simulate the complex business environment of Bank A. It is necessary to further refine the general banking model that has completed the key element extraction into a simulated business environment of Bank A. This step is completed through hierarchical diffusion deduction.

[0233] In the hierarchical diffusion deduction process, the second input data related to Bank A input by users of Bank A is obtained, such as documents related to Bank A, for example, an introduction to Bank A's wealth management products, an introduction to Bank A's investment projects, etc.

[0234] Second input data: Introduction to wealth management products of Bank A (data is only for example)

[0235] Basic product information:

[0236] Product code / sales code: 1234567

[0237] Product name: A financial management type A US dollar net value financial management product

[0238] Establishment date: 2021-03-24

[0239] Expiration date: 2030-03-01

[0240] Duration: 3264 days

[0241] Product Type: Open-ended Net Value

[0242] Product income and net asset value performance (as of 2023-03-22):

[0243] Net value per unit: 1.00

[0244] Cumulative net value of shares: 1.00

[0245] Investment currency: USD

[0246] Product Manager: Bank A

[0247] Product investment strategy: The product adopts high quality and short duration as the main operating strategy. ......

[0249] Second input data: Introduction of Bank A (numbers are only examples)

[0250] A Bank Co., Ltd. is a leading large commercial bank in China, headquartered in City A and established in January 1950. The bank was listed on the B Stock Exchange in January 2010 (stock code 654321). The bank's market capitalization at the end of 2023 is approximately RMB 1,234.5 billion, ranking 20th among listed banks worldwide. In terms of Tier 1 capital, the Group ranks second among global banks.

[0251] The Bank provides customers with comprehensive financial services including corporate finance, personal finance, and fund asset management, serving 1 billion individual customers and 100,000 corporate customers, and has subsidiaries in multiple industries including funds, leasing, trusts, insurance, futures, pensions, and investment banking.

[0252] Based on multiple second input data files of Bank A, the adjustment parameters for each layer in the four-layer architecture are generated respectively, and the adjustment parameters of each layer are used to adjust the key elements of each layer of the general banking industry and generate the hierarchical diffusion deduction results for Bank A. Among them, the 4-layer element output in the 4-layer architecture is the name, position, decision type, personality characteristics, etc. of the person (NPC) related to these projects, as well as which 3-layer and 2-layer element (set) the person (NPC) belongs to. The following scheduling process generates specific game texts based on this information, that is, lines ranging from 5 to 30 sentences.

[0253] For example,

[0254] The first-level diffusion deduction results: The environment layer in which Bank A is located, the first-level element "domestic banking and financial market" (1st level), has the characteristics of being domestic and listed;

[0255] The second-level diffusion deduction results: At the organizational level of Bank A, the second-level element of Bank A is “state-owned bank” (second level), which has the characteristics of large market capitalization and being in the forefront of the world;

[0256] The third level: the project level of Bank A. The three-level elements of Bank A are “Investment Banking” (3 levels), “Private Banking” (3 levels), and “Products” (3 levels), which are characterized by multiple clients and subsidiary management.

[0257] The fourth layer: the executive layer of Bank A. Under the "Investment Banking Department" (3rd layer), there are Investment Banking Manager NPC (4th layer) and Financial Advisor NPC (4th layer). Each NPC has his or her own personality and business characteristics.

[0258] In some implementations, the hierarchical diffusion deduction is performed in the order of "domestic banking financial market" (1st layer) → "state-owned banks" (2nd layer) → "investment banking department" (3rd layer) → "investment banking manager" (4th layer). Assuming that the 3rd layer project data cannot be obtained from the input data and is missing due to confidentiality or sensitivity reasons, the data of the 1st layer is analyzed based on the large language model, such as the banking industry is promoting supply chain finance reform; then the 2nd layer data is analyzed, such as the focus of the work of a certain city where Bank A is located this year is large syndicated orders. Therefore, the AI ​​agent is used to infer that Bank A's projects this year are likely to include projects such as "supply chain finance" and "standby syndicated bank", so as to supplement the 3rd layer data for active simulation.

[0259] In some embodiments, the results of the hierarchical diffusion deduction are then extracted to generate a 1,000-word feature description. The joint scheduling model generates a scenario result as output, encapsulating the scenario of a business / sales training game with the business of Bank A as the background.

[0260] In the example of Bank A, NPC 1 (layer 4) of the "Private Bank" (layer 3) and NPC 2 (layer 4) of the "Investment Banking Department" (layer 3) respectively describe their impressions of the 1st, 2nd, 3rd and 4th layers from a first-person perspective based on the affairs they are engaged in in the scheduling model, referring to the environment (layer 1), department (layer 2), project (layer 3) and position (layer 4) in the hierarchical diffusion deduction results. Each NPC outputs 5-30 sentences of corpus discourse, which must be aligned and consistent in logic, description of affairs and timing.

[0261] In some examples, NPCs with different personality attributes can output different situational results corresponding to their attributes. For example, NPC (4 layers) self-attribute layer (1 layer) - personal active linking method includes the following active linking parameters:

[0262] 1. Professional consultation: proactively share professional knowledge

[0263] 2. Emotional empathy: expressing understanding and support

[0264] 3. Resource exchange: provide your own resources in exchange for cooperation

[0265] 4. Experience sharing: sharing past successful cases

[0266] NPC1 has the personality attributes of empathy and gentleness, and its output corpus may include "I fully understand the difficulties you are facing";

[0267] NPC2 has serious and unsympathetic personality attributes, and its output corpus may include "The difficulties you are facing are caused by previous mistakes."

[0268] The following is an example of generating NPC corpus using active link parameters for a client in the medical industry.

[0269] The player user is an employee of B Medical Technology Co., Ltd., and the client of B Medical Technology Co., Ltd. is Hospital A. According to the method of the embodiment of the present invention, the generated game provides a simulation scenario of Hospital A for the players of Company B to train the employees of Company B to implement sales simulation in the game. In the game, the NPC is an employee of Hospital A.

[0270] Example of generating NPC speech data

[0271] Customer Specific Company:

[0272] Company Name: Hospital A

[0273] Company size: 3,200 beds, 3,000 employees

[0274] Company nature: Class A tertiary public hospital

[0275] Company business: As a large comprehensive hospital in the region, Hospital A has more than 30 medical and technical departments. Key specialties include cardiovascular medicine, neurosurgery, oncology, pediatrics, etc. It also has emergency departments, ICU wards and other critical care departments. The hospital has high-end medical equipment such as PET-CT and linear accelerators.

[0276] Company Background: Hospital A was established in 1995. After more than 20 years of development, it has become the largest, most advanced and most technologically advanced tertiary hospital in the region. In recent years, while maintaining its traditional strengths in cardiovascular and oncology departments, Hospital A has stepped up its informatization efforts and focused on creating new cross-disciplinary areas such as emergency medicine, rehabilitation medicine and geriatrics to achieve comprehensive development of the hospital.

[0277] Company vision: To become a leading modern comprehensive hospital in China, providing safe, professional and humane medical services to patients.

[0278] Player company basic information:

[0279] Company Name: B Medical Technology Co., Ltd.

[0280] Company size: 300 employees, RMB 500 million in assets

[0281] Company business: R&D, production and sales of medical devices and medical consumables. Main products include large medical equipment such as electrocardiographs, hemodialysis machines, intraoperative navigation systems, artificial joints, and medical consumables such as gloves, masks, and syringes.

[0282] Company Background: B Medical was founded in 2005 and is headquartered in the city where Hospital A is located. It is a high-tech enterprise specializing in the research and development and production of medical equipment and consumables. Over the years, B Medical has maintained a good cooperative relationship with Hospital A, providing equipment and consumables support for multiple departments such as cardiovascular disease, neurosurgery, and operating rooms.

[0283] Company vision: To become a domestic supplier of medical equipment and consumables with core competitiveness, providing hospitals and patients with high-quality medical device products and solutions.

[0284] Company advantages: professional R&D and production capabilities, in-depth understanding of hospital needs, and efficient sales and service system.

[0285] Active link parameters:

[0286] A. Explore the client’s business problem

[0287] B. Invite customers to participate in marketing activities

[0288] C. Share the company's research

[0289] D. Meet with the client to confirm the existence and scope of the business opportunity

[0290] E. Discuss requirements with individual customers

[0291] F. Discuss requirements with multiple customers

[0292] G. Demonstrate solutions to customers

[0293] H. Establish an internal sales team

[0294] I. Determine the client's budget

[0295] J. Organize a meeting to showcase the company's capabilities and invite customers to attend

[0296] K. Design

[0297] L. Modification plan

[0298] M. Internal meeting to prepare for negotiations

[0299] N. Participate in customer negotiations

[0300] O. Meetings between implementation consultants and clients

[0301] P. Participate in customer bidding

[0302] Customer's project information (3 layers):

[0303] Overall business opportunity: A hospital medical equipment and consumables procurement project

[0304] Total budget: 250 million yuan

[0305] Sub-opportunity 1:

[0306] Name: Emergency Department Information Transformation Project

[0307] Budget: 50 million yuan

[0308] Content: Create a smart emergency treatment platform for the emergency department through medical information systems, mobile medical care, smart wearable devices, etc. to optimize the emergency process.

[0309] Sub-opportunity 2:

[0310] Name: Laboratory Intelligent Transformation Project

[0311] Budget: 20 million yuan

[0312] Content: Improve the efficiency of laboratory workflow through intelligent equipment and information management, and realize comprehensive intelligent management of the laboratory.

[0313] NPC information:

[0314] NPC affiliation information:

[0315] Hospital A plans to invest heavily in upgrading medical equipment and consumables, with a total investment of 250 million yuan. This purchase aims to further improve the hospital's diagnosis and treatment level and service quality.

[0316] Sub-opportunity 1 public information:

[0317] Optimizing the emergency process is the key to improving emergency efficiency. This project will use information technology to improve the efficiency of the emergency department.

[0318] Key decision maker information for sub-opportunity 1:

[0319] The budget is 50 million yuan and the priority is high.

[0320] Sub-opportunity 1 decision-making process information:

[0321] The project's director, Zhang San, is responsible for key decision-making, with information center director Wang Wu and emergency department director Zhao Liu participating in the decision-making.

[0322] Sub-opportunity 2 public information:

[0323] Improve the intelligence level of the laboratory: save testing time and improve work efficiency.

[0324] Key decision maker information for sub-opportunity 2:

[0325] The budget is 20 million yuan and the priority is low.

[0326] Sub-opportunity 2 decision-making process information:

[0327] The project's financial director Li Si is responsible for key decisions, and clinical nutritionist Chen Yi participates in the decision-making.

[0328] NPC Corpus:

[0329] Character name: Zhang San

[0330] Position: Dean

[0331] Decision type: key decision maker

[0332] Business opportunity: Sub-business opportunity 1 (Emergency Department Information Transformation Project)

[0333] Corpus:

[0334] txt1: I am Zhang San, the director of Hospital A. Our hospital needs to improve the efficiency of the emergency department, transform the emergency process through information technology, and establish a smart emergency platform. Because the emergency department is responsible for treating critically ill patients, optimizing the emergency process is very important to improve the quality and efficiency of emergency treatment.

[0335] txt2: Hospital A plans to invest 50 million yuan in the emergency department information transformation project. The project will introduce medical information systems, mobile medical and smart wearable devices, etc., and use information technology to reshape the emergency process, realize the intelligence and precision of emergency treatment, and improve the work efficiency of the emergency department. This is very important for us to improve the speed and quality of emergency services.

[0336] txt3: As the dean, I have always wanted our emergency department to become a leading emergency center in the province and even the country, but due to funding and management constraints, the information construction of the emergency department has always lagged behind. I must strive for the success of this information project, otherwise it will be unacceptable.

[0337] txt4: I hope that through this project, our emergency department can achieve intelligent transformation and greatly improve work efficiency. As the dean, my goal is to make the emergency department of Hospital A a leading emergency center known throughout the province and even the country.

[0338] txt5: This information project must achieve the following goals: shorten the average waiting time for patients, increase the utilization rate of wards, reduce the misdiagnosis rate, and improve the work efficiency of emergency personnel. At the same time, it is also necessary to achieve seamless connection and resource sharing in the emergency department, which is also a management challenge for me. In short, the success of this project is directly related to the realization of my personal career development goals.

[0339] Character name: Li Si

[0340] Position: Financial Director

[0341] Decision type: key decision maker

[0342] Business opportunity: Sub-business opportunity 2 (laboratory intelligent transformation project)

[0343] Corpus:

[0344] txt1: I am Li Si, the financial director of Hospital A. The laboratory of our hospital needs to be upgraded to be intelligent, and the work efficiency of the laboratory can be improved by introducing intelligent equipment and information management. As a key link in supporting clinical diagnosis and treatment, improving laboratory efficiency has a direct impact on the diagnosis and treatment level of the hospital.

[0345] txt2: Hospital A plans to invest 20 million yuan in the laboratory intelligent transformation project. The project will introduce intelligent analytical instruments and information management systems, optimize workflows and resource allocation, realize intelligent and information management of the laboratory, and improve the overall efficiency of the laboratory. This is very necessary for us to improve our clinical support capabilities.

[0346] txt3: As the financial director, I have always felt that our hospital has not invested enough in laboratory construction, resulting in many projects not having obvious results after a long time. I must strive for this laboratory renovation project, otherwise it will be difficult to explain to the hospital leaders.

[0347] txt4: I hope that through this laboratory intelligentization project, we can shorten our testing time by 30% and improve the overall laboratory efficiency by 20%. As the financial director, my goal is to maximize the use of the hospital's financial resources.

[0348] txt5: The goals that must be achieved in this laboratory intelligentization project are: shortening the time for generating sample test reports, improving the utilization rate of equipment and space, reducing the rate of repeated testing, and realizing the information-based centralized management of laboratory resources and processes. This is directly related to the cost control and benefit improvement that I am responsible for.

[0349] Character name: Wang Wu

[0350] Position: Director of Information Center

[0351] Decision type: Participants in decision making

[0352] Business opportunity: Sub-business opportunity 1 (Emergency Department Information Transformation Project)

[0353] Corpus:

[0354] txt1: I am Wang Wu, director of the Information Center of Mingyang Hospital. Our hospital plans to carry out information transformation of the emergency department, which is very important for optimizing the emergency process and improving the efficiency of emergency treatment. Informationization can integrate resources and improve the speed of emergency response.

[0355] txt2: I participated in the decision-making of the emergency department's information transformation project. This project will realize the informationization and process reengineering of the emergency department's work and comprehensively improve the response speed of emergency services. This is very important for protecting the lives of patients.

[0356] txt3: As the director of the information center, I hope to use my professional knowledge to propose feasible information solutions in the emergency department informationization project. The key decision maker of the project is the dean of our hospital. My goal is to promote the project to achieve breakthroughs and make the hospital's emergency level reach the industry leader.

[0357] Character name: Zhao Liu

[0358] Position: Director of Emergency Department

[0359] Decision type: Participants in decision making

[0360] Business opportunity: Sub-business opportunity 1 (Emergency Department Information Transformation Project)

[0361] Corpus:

[0362] txt1: I am Zhao Liu, director of the emergency department of Mingyang Hospital. Our hospital plans to carry out information transformation of the emergency department, which is very important to improve the emergency response speed and treatment efficiency. Informationization can integrate resources and optimize the emergency process.

[0363] txt2: I participated in the decision-making of the emergency department information transformation project. This project will comprehensively promote the information construction of the emergency department, making the emergency process smoother and more efficient. This is very important to ensure the safety of patients' lives.

[0364] txt3: As the director of the emergency department, I hope to use my professional knowledge in the information project and make specific suggestions for optimizing the emergency process. The key decision maker of the project is the director of our hospital. My goal is to facilitate the smooth implementation of the project and make the hospital's emergency level reach the leading level in China.

[0365] Character name: Chen Yi

[0366] Position: Clinical Nutritionist

[0367] Decision type: Participants in decision making

[0368] Business opportunity: Sub-business opportunity 2 (laboratory intelligent transformation project)

[0369] Corpus:

[0370] txt1: I am Chen Yi, a clinical nutritionist at Mingyang Hospital. Our hospital plans to improve the intelligence level of the laboratory, which is very important for optimizing the inspection process. Intelligence can shorten the inspection time and improve work efficiency.

[0371] txt2: I participated in the decision-making of the laboratory intelligent transformation project. This project will comprehensively promote the automation transformation of the laboratory and realize the rapid output of test reports. This is very important for standardizing test operations and ensuring test quality.

[0372] txt3: As a clinical nutritionist, I hope to use my professional knowledge to make suggestions for new test indicators in this project. The key decision maker of the project is the director of the finance department of our hospital. My goal is to ensure that the intelligent upgrade of the laboratory meets the needs of nutritional indicator testing.

[0373] Figure 6 is a schematic diagram of a large model interface layer according to an embodiment of the present invention.

[0374] The main process of the method described in the embodiment of the present invention can be completed based on the large language model. The large model standard interface supports high-frequency calls by the main modules of the above main processes. The large language model used can be various large language models commonly used in the industry. The implementation of the present invention encapsulates various third-party large language model APIs as internal standard interfaces, and encapsulates the locally deployed large language model as APIs and internal standard interfaces.

[0375] Through such deployment, the efficiency of using large language models is greatly improved.

[0376] In the second aspect, based on the same inventive concept, an embodiment of the present invention provides a device for generating a simulation scenario based on a large language model, such as Figure 7 As shown, the device comprises:

[0377] A key feature extraction module 702 is configured to extract key features from the first input data based on a hierarchical architecture through a large language model, and generate a first key feature corresponding to each layer in the hierarchical architecture;

[0378] A hierarchical diffusion deduction module 703 is configured to perform hierarchical diffusion deduction according to the first key features and the second input data extracted by a large language model, and generate diffusion deduction results corresponding to each layer in the hierarchical architecture;

[0379] A situation result module 704 is configured to generate a situation result by combining the diffusion deduction result with an active link parameter, wherein the active link parameter is generated by the scheduling model according to the second input data;

[0380] The hierarchical architecture is associated with the first input data.

[0381] As an optional example, the device further includes:

[0382] Hierarchical modeling module 701: The hierarchical modeling module is configured to model a hierarchical architecture based on the first input data, wherein each layer in the hierarchical architecture has a corresponding attribute range, and the elements in each layer have the attributes of the layer; if the hierarchical architecture has M layers in total, the attribute range from layer 1 to layer M is reduced layer by layer, and M is an integer greater than or equal to 1.

[0383] As an optional example, each X-1 layer element in the hierarchical architecture corresponds to one or more X layer elements, each X layer element belongs to an X-1 layer element, and there is no intersection between the elements, and X is a positive integer greater than or equal to 1 and less than or equal to M.

[0384] As an optional example, each layer 1 to layer M-1 element in the layered architecture is a set of layer M elements;

[0385] The number of M-layer elements contained in the X-layer is a proper subset of the number of M-layer elements contained in the X-1-layer, where X is a positive integer greater than or equal to 1 and less than or equal to M.

[0386] As an optional example, the hierarchical diffusion deduction module 703 is further configured to:

[0387] Extracting information corresponding to the attributes of each layer in the hierarchical architecture from the second input data by means of a large language model as adjustment parameters of the layer;

[0388] For each layer in the layered architecture, the first key feature corresponding to the layer is combined with the adjustment parameter of the layer to perform deduction to generate a diffusion deduction result of the layer.

[0389] As an optional example, if the hierarchical architecture has M layers in total, the diffusion deduction result of the M-layer element includes the M-layer element attribute description and the 1st to M-1th layer elements to which the M-layer element belongs.

[0390] As an optional example,

[0391] The key feature extraction is performed in the order of starting from the M layer and M decreasing;

[0392] The layered diffusion deduction is performed starting from layer X=1 and in increasing order of X, where X is a positive integer greater than or equal to 1 and less than or equal to M.

[0393] As an optional example, the key feature extraction module 702 is further configured to, when the first input data is missing:

[0394] Determine the layer corresponding to the missing in the layered architecture, extract key features from the other layers in the M layers except the layer corresponding to the missing in descending order of M, and supplement the key features of the layer corresponding to the missing according to the other layers except the layer corresponding to the missing;

[0395] The hierarchical diffusion deduction module 703 is further configured to, when the second input data is missing:

[0396] Determine a layer corresponding to the missing in the hierarchical architecture, and determine a supplementary parameter corresponding to the missing according to information of other layers in the hierarchical architecture through a large language model;

[0397] For the layer corresponding to the missing layer in the hierarchical architecture, the first key feature of the layer is combined with the adjustment parameters of the layer and the supplementary parameters to generate a diffusion deduction result of the layer.

[0398] As an optional example, the situation result includes execution data associated with the proactive link parameter, the execution data has M-layer attributes, the proactive link parameter has M-layer attributes, and the execution data and the corresponding proactive link parameter have one-to-one correspondence in attributes and timing;

[0399] If the M-layer element has character attributes, the execution data includes non-player character NPC corpus resources;

[0400] If the M-layer element has a teaching attribute, the execution data includes the teaching data.

[0401] As an optional example, the apparatus further includes a scheduling model module 705, which is configured to:

[0402] The scheduling model schedules corresponding corpus from the corpus resources and feeds back to the user according to the non-player character NPC selected by the user and the active link parameter and in accordance with preset rules.

[0403] It should be noted that the device embodiment is a device that corresponds one-to-one to the above-mentioned method embodiment, and all implementation methods in the above-mentioned method embodiment are applicable to the embodiment of the device. The device embodiment can implement all method steps implemented by the above-mentioned method embodiment and can achieve the same technical effect. The parts and beneficial effects that are the same as those in the method embodiment will not be described in detail here.

[0404] In a third aspect, based on the same inventive concept, an embodiment of the present invention provides a device for generating a simulation scenario based on a large language model, such as Figure 8 As shown, the device includes: a processor 800 , a user interface 810 and a memory 820 .

[0405] The processor 800 may include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor 1000 may be implemented in at least one hardware form of DSP (Digital Signal Processing), FPGA (Field Programmable Gate Array), and PLA (Programmable Logic Array). The processor 800 may also include a main processor and a coprocessor. The main processor is a processor for processing data in an awake state, also known as a CPU (Central Processing Unit); the coprocessor is a low-power processor for processing data in a standby state. In some embodiments, the processor 800 may be integrated with a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the display screen. In some embodiments, the processor 800 may also include an AI (Artificial Intelligence) processor, which is used to process computing operations related to machine learning.

[0406] The memory 820 may include one or more computer-readable storage media, which may be non-transitory. The memory 820 may also include a high-speed random access memory, and a non-volatile memory, such as one or more disk storage devices, flash memory storage devices. In some embodiments, the non-transitory computer-readable storage medium in the memory 820 is used to store at least one program code, which is used to be executed by the processor 800 to implement the method for generating a simulation scenario based on a large language model provided in the method embodiment of the present application.

[0407] In some embodiments, the apparatus for generating simulation scenarios based on a large language model in this embodiment may optionally include: a bus interface and at least one user interface 810. The processor 800, the memory 820 and the user interface 810 may be connected via a bus interface or a signal line.

[0408] As an optional example, when the processor 800 reads and executes the program code stored in the memory 820, the following is achieved:

[0409] Extract key features from the first input data based on a hierarchical architecture using a large language model to generate first key features corresponding to each layer in the hierarchical architecture;

[0410] Performing hierarchical diffusion deduction based on the first key features and the second input data extracted by a large language model to generate diffusion deduction results corresponding to each layer in the hierarchical architecture;

[0411] generating a situation result output by combining the diffusion deduction result with an active link parameter, wherein the active link parameter is generated by the scheduling model according to the second input data;

[0412] The hierarchical architecture is associated with the first input data.

[0413] Optionally, the processor 800 is further configured to:

[0414] Modeling a hierarchical architecture according to the first input data, wherein each layer in the hierarchical architecture has a corresponding attribute range, and elements in each layer have attributes of the layer;

[0415] If the hierarchical architecture has M layers in total, the attribute range from layer 1 to layer M is reduced layer by layer, where M is an integer greater than or equal to 1.

[0416] Optionally, each X-1 layer element in the layered architecture corresponds to one or more X layer elements, each X layer element belongs to an X-1 layer element, and there is no intersection between the elements, and X is a positive integer greater than or equal to 1 and less than or equal to M.

[0417] Optionally, each layer 1 to layer M-1 element in the layered architecture is a set of layer M elements;

[0418] The number of M-layer elements contained in the X-layer is a proper subset of the number of M-layer elements contained in the X-1-layer, where X is a positive integer greater than or equal to 1 and less than or equal to M.

[0419] Optionally, the processor 800 is further configured to:

[0420] Extracting information corresponding to the attributes of each layer in the hierarchical architecture from the second input data by means of a large language model as adjustment parameters of the layer;

[0421] For each layer in the layered architecture, the first key feature corresponding to the layer is combined with the adjustment parameter of the layer to perform deduction to generate a diffusion deduction result of the layer.

[0422] Optionally, if the hierarchical architecture has M layers in total, the diffusion deduction result of the M-layer element includes the M-layer element attribute description and the 1st to M-1th layer elements to which the M-layer element belongs.

[0423] Optional,

[0424] The key feature extraction is performed in the order of starting from the M layer and M decreasing;

[0425] The layered diffusion deduction is performed starting from layer X=1 and in increasing order of X, where X is a positive integer greater than or equal to 1 and less than or equal to M.

[0426] Optionally, the processor 800 is further configured to:

[0427] In the case where the first input data is missing,

[0428] Determine the layer corresponding to the missing in the layered architecture, extract key features from the other layers in the M layers except the layer corresponding to the missing in descending order of M, and supplement the key features of the layer corresponding to the missing according to the other layers except the layer corresponding to the missing;

[0429] In the case where the second input data is missing,

[0430] Determine a layer corresponding to the missing in the hierarchical architecture, and determine a supplementary parameter corresponding to the missing according to information of other layers in the hierarchical architecture through a large language model;

[0431] For the layer corresponding to the missing layer in the hierarchical architecture, the first key feature of the layer is combined with the adjustment parameters of the layer and the supplementary parameters to generate a diffusion deduction result of the layer.

[0432] Optionally, the situation result includes execution data associated with the proactive link parameter, the execution data has M-level attributes, the proactive link parameter has M-level attributes, and the execution data and the corresponding proactive link parameter have one-to-one correspondence in attributes and timing;

[0433] If the M-layer element has character attributes, the execution data includes non-player character NPC corpus resources;

[0434] If the M-layer element has a teaching attribute, the execution data includes teaching data.

[0435] Optionally, the processor 800 is further configured to:

[0436] The scheduling model schedules corresponding corpus from the corpus resources and feeds back to the user according to the non-player character NPC selected by the user and the active link parameter and in accordance with preset rules.

[0437] It should be noted that the device embodiment is a device that corresponds one-to-one to the above-mentioned method embodiment, and all implementation methods in the above-mentioned method embodiment are applicable to the embodiment of the device. The device embodiment can implement all method steps implemented by the above-mentioned method embodiment and can achieve the same technical effect. The parts and beneficial effects that are the same as those in the method embodiment will not be described in detail here.

[0438] It should be noted that the division of units in the embodiments of the present application is schematic and is only a logical function division. There may be other division methods in actual implementation. In addition, each functional unit in each embodiment of the present application may be integrated into a processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0439] The embodiment of the present application also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method for transmitting quality of service information applied to a terminal. The processor-readable storage medium may be any available medium or data storage device that the processor can access, including but not limited to magnetic storage (e.g., floppy disk, hard disk, magnetic tape, magneto-optical disk (MO), etc.), optical storage (e.g., CD, DVD, BD, HVD, etc.), and semiconductor storage (e.g., ROM, EPROM, EEPROM, non-volatile memory (NAND FLASH), solid-state drive (SSD)), etc.

[0440] The embodiment of the present application also provides a computer program product, including computer instructions. When the computer instructions are executed by a processor, the various processes in the above method embodiment are implemented and the same technical effect can be achieved. To avoid repetition, they are not described here.

[0441] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage and optical storage, etc.) that contain computer-usable program code.

[0442] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A method for generating a simulation scenario based on a large language model, the method comprising: Extract key elements from the first input data based on the hierarchical architecture using a large language model to generate a first key element corresponding to each layer in the hierarchical architecture; Performing hierarchical diffusion deduction based on the first key elements and the second input data extracted by a large language model to generate diffusion deduction results corresponding to each layer in the hierarchical architecture; generating a situation result by combining the diffusion deduction result with an active link parameter, wherein the active link parameter is generated by the scheduling model according to the second input data; The hierarchical architecture is associated with the first input data.

2. The method according to claim 1, characterized in that The method further comprises: Modeling a hierarchical architecture according to the first input data, wherein each layer in the hierarchical architecture has a corresponding attribute range, and elements in each layer have attributes of the layer; If the hierarchical architecture has M layers in total, the attribute range from layer 1 to layer M is reduced layer by layer, where M is an integer greater than or equal to 1.

3. The method according to claim 2, characterized in that In the hierarchical architecture, each X-1 layer element corresponds to one or more X layer elements, each X layer element belongs to an X-1 layer element, and there is no intersection between the elements, and X is a positive integer greater than or equal to 1 and less than or equal to M.

4. The method according to claim 2, characterized in that Each layer 1 to layer M-1 element in the layered architecture is a set of layer M elements; The number of M-layer elements contained in the X-layer is a proper subset of the number of M-layer elements contained in the X-1-layer, where X is a positive integer greater than or equal to 1 and less than or equal to M.

5. The method according to claim 1, characterized in that: Performing hierarchical diffusion deduction based on the first key elements and the second input data extracted by the large language model to generate diffusion deduction results corresponding to each layer in the hierarchical architecture includes: Extracting information corresponding to the attributes of each layer in the hierarchical architecture from the second input data by means of a large language model as adjustment parameters of the layer; For each layer in the layered architecture, the first key element corresponding to the layer is combined with the adjustment parameter of the layer to perform deduction to generate a diffusion deduction result of the layer.

6. The method according to claim 5, characterized in that If the hierarchical structure has M layers in total, the diffusion deduction result of the M-layer element includes the M-layer element attribute description and the 1st to M-1th layer elements to which the M-layer element belongs.

7. The method according to claim 2 or 5, characterized in that: The key element extraction is performed in the order of starting from the M layer and M decreasing; The layered diffusion deduction is performed starting from layer X=1 and in increasing order of X, where X is a positive integer greater than or equal to 1 and less than or equal to M.

8. The method according to claim 7, characterized in that In the case where the first input data is missing, the method further includes: Determine the layer corresponding to the missing in the layered architecture, extract key elements from the other layers in the M layers except the layer corresponding to the missing in descending order of M, and supplement the key elements of the layer corresponding to the missing according to the other layers except the layer corresponding to the missing; In the case where the second input data is missing, the method further includes: Determine a layer corresponding to the missing in the hierarchical architecture, and determine a supplementary parameter corresponding to the missing according to information of other layers in the hierarchical architecture through a large language model; For the layer corresponding to the missing layer in the hierarchical architecture, the first key element of the layer is combined with the adjustment parameters of the layer and the supplementary parameters to generate a diffusion deduction result of the layer.

9. The method according to claim 1, characterized in that: The situation result includes execution data associated with the proactive link parameter, the execution data has M-level attributes, the proactive link parameter has M-level attributes, and the execution data and the corresponding proactive link parameter have one-to-one correspondence in attributes and timing; If the M-layer element has character attributes, the execution data includes non-player character NPC corpus resources; If the M-layer element has a teaching attribute, the execution data includes teaching data.

10. The method according to claim 9, characterized in that The method further comprises: The scheduling model schedules corresponding corpus from the corpus resources and feeds back to the user according to the non-player character NPC selected by the user and the active link parameter and in accordance with preset rules.

11. A device for generating simulated situations based on a large language model, characterized in that: include: A key element extraction module is configured to extract key elements from the first input data based on a hierarchical architecture through a large language model, and generate a first key element corresponding to each layer in the hierarchical architecture; A hierarchical diffusion deduction module is configured to perform hierarchical diffusion deduction according to the first key elements and the second input data extracted through a large language model to generate diffusion deduction results corresponding to each layer in the hierarchical architecture; A situation result module is configured to generate a situation result by combining the diffusion deduction result with an active link parameter, wherein the active link parameter is generated by the scheduling model according to the second input data; The hierarchical architecture is associated with the first input data.

12. A device for generating simulated situations based on a large language model, comprising a memory, a transceiver, and a processor, characterized in that ； The memory is used to store a computer program; the transceiver is used to send and receive data under the control of the processor; the processor is used to read the computer program in the memory and execute: Extract key elements from the first input data based on the hierarchical architecture using a large language model to generate a first key element corresponding to each layer in the hierarchical architecture; Performing hierarchical diffusion deduction based on the first key elements and the second input data extracted by a large language model to generate diffusion deduction results corresponding to each layer in the hierarchical architecture; generating a situation result by combining the diffusion deduction result with an active link parameter, wherein the active link parameter is generated by the scheduling model according to the second input data; The hierarchical architecture is associated with the first input data.

13. A processor-readable storage medium, wherein: The processor-readable storage medium stores a program, and the program is used to enable the processor to execute the method according to any one of claims 1 to 10.

14. A computer program product, characterized in that The computer program product includes at least one program, which is stored in a computer-readable storage medium. A processor of a computer device reads the at least one program from the computer-readable storage medium, and the processor executes the at least one program, so that the computer device executes the corpus resource scheduling method according to any one of claims 1 to 10.

Images