Virtual simulation interaction method and system for e-commerce talent training platform
By establishing a 3D store scene model and scene semantic graph in the e-commerce talent training platform, and generating scene state vectors, the state sharing between the 3D scene and the AI customer simulator is realized, which solves the problem of training scene distortion in existing technologies and improves the effectiveness and relevance of training.
Patent Information
- Application Number
- CN202611417783.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-14
- Publication Date
- 2026-11-10
AI Technical Summary
Existing online training platforms cannot effectively synchronize the scene status of 3D scenes with AI customer simulators, resulting in distorted training scenes and an inability of the evaluation system to accurately identify trainees' weaknesses, thus affecting training effectiveness.
By establishing a 3D store scene model and scene semantic graph, scene state vectors are generated, enabling state sharing between the 3D scene and the AI customer simulator. During training, the state of scene objects and customer expressions are updated synchronously. Combined with trainee operation behavior and response content, joint evaluation is conducted to identify weaknesses and adjust training tasks.
It achieves a closed-loop state between 3D scene and AI dialogue, improves the effectiveness and relevance of training, ensures that training tasks are consistent with scene states, covers the evaluation of operation trajectory and response content, and improves training efficiency.
Smart Images

Figure CN123072256A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of virtual simulation training technology, and in particular to a virtual simulation interaction method and system for an e-commerce talent training platform. Background Technology
[0002] In e-commerce customer service training, trainees need to master the ability to communicate with customers and handle problems in a real store environment, taking into account product status, inventory information, promotional rules, order progress, and after-sales policies.
[0003] Existing online training platforms typically treat 3D store display, customer service dialogue training, and operation evaluation as independent modules: 3D store is only used for product display, dialogue training is based on a fixed question bank or independent historical conversation data, and operation evaluation records the click behavior of trainees separately.
[0004] This fragmented architecture means that during training, the customer service dialogue module cannot perceive the trainee's actions such as searching for products, verifying orders, and viewing rules in the 3D scene. The AI customer simulator also cannot dynamically adjust its questions and customer emotions based on whether the trainee has actually viewed the relevant product information or order status. For example, when a trainee answers a customer's inquiry about promotional rules, the system cannot distinguish whether the trainee's response is an accurate one based on actual verification of the rules, or a guessing response without verification. When a trainee handles a complaint about order delays, the system cannot verify whether the trainee has actually viewed the order's logistics progress.
[0005] Due to the lack of a unified synchronization mechanism between the scene state and the dialogue state, the questions expressed by the AI customer may be inconsistent with the actual product inventory, order status or activity rules in the current 3D scene, causing the training scene to be distorted. At the same time, the evaluation system can only make judgments based on the final response text, which cannot cover the trainee's observation, verification and operation trajectory in the business process, and it is also difficult to accurately identify the trainee's specific weak points in order to guide subsequent targeted training.
[0006] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention
[0007] The main purpose of this application is to provide a virtual simulation interaction method and system for an e-commerce talent training platform, aiming to solve the technical problem of how to enable the 3D interactive training terminal and the AI customer simulator to share and bidirectionally update the scene state around the same e-commerce business object, so as to keep the state of the e-commerce business object used for 3D scene rendering synchronized with the state of the e-commerce business object used for customer expression generation.
[0008] To achieve the above objectives, this application proposes a virtual simulation interaction method for an e-commerce talent training platform, the method comprising:
[0009] Based on spatial and product materials, a 3D store scene model is created;
[0010] Configure the business attributes of each business object in the 3D store scene model to form a scene semantic graph;
[0011] Based on the scene semantic graph, extract the business data corresponding to the current training scene, generate a scene state vector, and determine the training task seed based on the scene state vector;
[0012] The 3D rendering engine is invoked to load the 3D store scene model, and the 3D store scene is rendered according to the training task seed. At the same time, the AI customer simulator is invoked and the customer parameters in the training task seed are loaded to initialize the AI customer.
[0013] The system collects the student's actions in the 3D store scene, as well as the student's first response to the AI customer's input.
[0014] The business data of the business object corresponding to the operation behavior is written into the dialogue context, and the first reply content and the dialogue context are input into the AI customer simulator. The AI customer simulator generates the second reply content and outputs it according to the scene state vector and customer intent state. At the same time, the display state of the corresponding business object is updated according to the object identifier in the second reply content.
[0015] A training score is generated based on the operational behavior and the content of the first response, and at least one weak ability of the trainee is determined based on the training score.
[0016] Adjust the trainee's training task parameters based on the identified weaknesses.
[0017] In one embodiment, the step of extracting business data corresponding to the current training scenario based on the scene semantic graph, generating a scene state vector, and determining the training task seed based on the scene state vector includes:
[0018] In the scene semantic graph, the current state variables of each business object are read. The state variables include at least one of the following: product inventory, order processing progress, logistics node status, after-sales work order status, and customer sentiment value.
[0019] The state variables are combined to form the scene state vector;
[0020] Read data sources, including historical anonymized customer service conversations, product knowledge bases, platform rule bases, user feedback texts, preset teaching objectives, and / or enterprise job competency standards;
[0021] Based on the scenario state vector, match the corresponding customer profile, customer initial intent, product context, order context, marketing constraints, service risk points, and assessment targets from the data source;
[0022] The matched customer profile, the customer's initial intent, the product context, the order context, the marketing constraints, the service risk points, and the assessment objectives are combined to form the training task seed.
[0023] In one embodiment, the step of invoking the AI customer simulator and loading the customer parameters from the training task seed to initialize the AI customer includes:
[0024] From the training task seed, read the customer profile and the customer's initial intent;
[0025] The AI customer simulator is written with the customer profile's initial values for customer type, price sensitivity, patience level, and emotional state; and,
[0026] The initial customer intent is written into the AI customer simulator, and the customer intent state is set based on the initial customer intent. The customer intent state includes inquiry state, price comparison state, objection state, order reminder state, return and exchange state, complaint state, add to cart state, and transaction state.
[0027] In one embodiment, the step of the AI customer simulator generating and outputting a second response based on the scene state vector and the customer intent state includes:
[0028] Configure the customer's emotional state based on the first response;
[0029] The second response content is generated based on the customer's intent state, the customer's emotional state, and the scenario state vector.
[0030] In one embodiment, the operational behavior includes roaming behavior, point selection behavior, product query behavior, inventory query behavior, order processing behavior, promotional activity verification behavior, logistics query behavior, or after-sales work order processing behavior;
[0031] The step of writing the business data of the business object corresponding to the operation behavior into the dialogue context includes:
[0032] Identify the operation type and target business object of the operation;
[0033] When the operation type is a selection operation, a viewing operation, or a processing operation, the business data of the target business object is read from the scene semantic graph;
[0034] Write the business data into the cache of the dialogue context;
[0035] The step of updating the display status of the corresponding business object based on the object identifier in the second reply content includes:
[0036] Parse the second response content to extract the object identifier or event type;
[0037] Based on the object identifier or event type, locate the corresponding business object in the three-dimensional store scene;
[0038] The found business objects are highlighted, their status is toggled, or their status is marked.
[0039] In one embodiment, the step of generating a training score based on the operational behavior and the first response content, and determining at least one weakness of the trainee based on the training score, includes:
[0040] The scores for the following metrics were calculated: object search accuracy, product recommendation matching degree, platform rule compliance, customer problem resolution rate, dialogue coherence, emotional soothing effect, response timeliness, transaction or lead generation results, after-sales processing accuracy, and risk response hit rate.
[0041] Based on preset weights, the scores of each indicator are weighted and summed to obtain the training score;
[0042] Based on the scores of each indicator, the ability scores of each trainee in each ability dimension are determined. The ability dimensions include product knowledge ability, object finding ability, rule understanding ability, communication and expression ability, emotional soothing ability, order processing ability, and after-sales processing ability.
[0043] If the ability score is lower than a preset ability threshold, the ability dimension corresponding to the ability score is marked as the weak point.
[0044] In one embodiment, the step of adjusting the trainee's training task parameters based on the weakness includes:
[0045] Read the trainee's historical training records to obtain the business scenario type and service process type corresponding to the weak points in the skills;
[0046] Increase the weight of the occurrence of scenario states, customer intentions, or service risk points corresponding to the business scenario type and the service process type in the training task parameters;
[0047] The occurrence weights are associated with the task generation conditions of the next training task.
[0048] In one embodiment, the method further includes:
[0049] The training configuration instruction is received by the teacher or administrator. The training configuration instruction carries configuration parameters, including training objectives, job competency standards, training duration, product categories, platform rules, assessment weights, and / or customer service knowledge scope.
[0050] Based on the configuration parameters, update the task generation conditions of the training task seed and the preset weights of the training score.
[0051] In one embodiment, the method further includes:
[0052] Based on the playback instruction received by the teacher or management terminal, the playback instruction carries the target training record identifier;
[0053] Based on the target training record identifier, read the corresponding stored operation trajectory data and stored response content data;
[0054] On the teacher's or administrator's end, the stored operation trajectory data and the stored reply content data are played back synchronously.
[0055] Furthermore, to achieve the above objectives, this application also proposes a virtual simulation interactive system for an e-commerce talent training platform, the system comprising:
[0056] The 3D scene modeling module is used to create 3D store scene models based on spatial and product materials.
[0057] The scene semantic graph module configures the business attributes of each business object in the 3D store scene model to form a scene semantic graph.
[0058] The training task generation module extracts business data corresponding to the current training scenario based on the scene semantic graph, generates a scene state vector, and determines the training task seed based on the scene state vector.
[0059] The 3D interactive training module calls the 3D rendering engine to load the 3D store scene model and renders the 3D store scene according to the training task seed. At the same time, it calls the AI customer simulator and loads the customer parameters in the training task seed to initialize the AI customer.
[0060] The AI customer simulation module collects the trainee's operational behavior in the three-dimensional store scene, as well as the trainee's first response to the AI customer's input.
[0061] The scene dialogue synchronization module writes the business data of the business object corresponding to the operation behavior into the dialogue context, and inputs the first reply content and the dialogue context into the AI customer simulator. The AI customer simulator generates and outputs the second reply content according to the scene state vector and the customer intent state, and updates the display state of the corresponding business object according to the object identifier in the second reply content.
[0062] The training evaluation module generates a training score based on the operation behavior and the first response content, and determines at least one weak ability of the trainee based on the training score.
[0063] The personalized training module adjusts the training task parameters of the trainee based on the identified weaknesses in their abilities.
[0064] One or more technical solutions proposed in this application have at least the following technical effects:
[0065] By transforming the spatial layout and product display of a real store into a 3D interactive scene, trainees are provided with a training space that closely resembles a real business environment. This allows trainees to locate products and business information through visual observation and spatial navigation, rather than simply imagining business scenarios through textual descriptions. A scene semantic graph establishes a one-to-one mapping between visualized objects in the 3D scene and backend e-commerce business data, transforming the 3D scene from a purely visual display into a training interactive interface carrying complete business semantics. Every selection made by trainees within the scene accesses real business data, providing a data foundation for subsequent assessments of whether trainees have verified business facts. Scene state vectors enable a unified description of scene object states, allowing the 3D scene and AI dialogue to share the same state foundation. Training task seeds enable flexible configuration of training scenes, achieving the goal of automatically generating differentiated training tasks based on different training objectives. Synchronous initialization of the 3D scene and AI customer ensures that trainees start training with scene and customer states consistent with the task objectives, guaranteeing complete consistency in the initial conditions of the training tasks. Simultaneous collection of scene operation behaviors and customer service responses provides a complete data foundation for subsequent joint evaluation, enabling evaluations to cover both operation trajectories and response content. By writing business data corresponding to trainees' actions into the dialogue context, the AI client can perceive whether trainees have verified relevant information, thereby dynamically adjusting subsequent questions. By updating the display state of the 3D scene inversely based on object identifiers in the AI client's expressions, the scene display dynamically changes with the dialogue content, thus forming a two-way closed loop of scene-driven dialogue and dialogue-feedback scenario. By adjusting training task parameters based on weaknesses, subsequent training tasks are focused on the trainees' actual weaknesses, improving training efficiency.
[0066] By establishing 3D scene modeling, scene semantic graphs, generating scene state vectors, and synchronizing scene dialogue, a closed-loop state mechanism for dialogue between 3D scenes and AI was achieved. By jointly evaluating operation trajectories and response content, and by identifying weaknesses to update personalized training paths, the effectiveness and relevance of customer service training were improved. Attached Figure Description
[0067] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.
[0068] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0069] Figure 1 A flowchart illustrating the first embodiment of the virtual simulation interaction method for the e-commerce talent training platform of this application;
[0070] Figure 2 A simplified flowchart illustrating the virtual simulation interaction method of the e-commerce talent training platform provided in Embodiment 2 of this application;
[0071] Figure 3 A flowchart illustrating the third embodiment of the virtual simulation interaction method for the e-commerce talent training platform of this application;
[0072] Figure 4 A flowchart illustrating the fourth embodiment of the virtual simulation interaction method for the e-commerce talent training platform of this application;
[0073] Figure 5 A flowchart illustrating an embodiment of the virtual simulation interaction method for the e-commerce talent training platform of this application;
[0074] Figure 6 This is a schematic diagram of the module structure of the virtual simulation interaction system of the e-commerce talent training platform in this application embodiment.
[0075] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation
[0076] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.
[0077] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.
[0078] This application provides a virtual simulation interaction method and system for an e-commerce talent training platform. Its core concept is to establish semantic associations between objects in a 3D store scene and e-commerce business objects such as products, inventory, orders, logistics, after-sales service, and marketing rules. A scene state vector is used to uniformly describe the real-time state of these business objects, enabling 3D scene rendering and AI customer dialogue generation to share the same state foundation. During training, the business data corresponding to the trainee's actions in the 3D scene is automatically written into the dialogue context. The AI customer simulator dynamically adjusts the customer's expression and emotions accordingly, while object identifiers in the customer's expression drive the 3D scene state update. Training evaluation simultaneously covers the trainee's operational trajectory and response content, and identifies weaknesses based on the evaluation results to adjust subsequent training tasks.
[0079] It should be noted that the execution entity in this embodiment can be a training platform server, a local training terminal, a cloud training system, or a combination of the above devices. The following description uses the training platform server as the execution entity, combined with the collaborative operation of the student training terminal and the AI client simulator.
[0080] Based on this, this application provides a virtual simulation interaction method for an e-commerce talent training platform, applied to the virtual simulation interaction system of an e-commerce talent training platform. (Refer to...) Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the virtual simulation interaction method for the e-commerce talent training platform of this application.
[0081] In this embodiment, the virtual simulation interaction method of the e-commerce talent training platform includes steps S100 to S800:
[0082] Step S100: Based on the spatial materials and product materials, create a three-dimensional store scene model.
[0083] In this embodiment, the system first acquires spatial and product materials from a real store. Spatial materials may include at least one of the following: panoramic images, video frames, floor plans, point cloud data, and manually edited store layouts. Product materials include product images, 3D models, product parameters, inventory information, pricing information, promotional information, and after-sales policies.
[0084] Subsequently, the system builds a 3D store scene model based on the aforementioned materials. For stores with real-world data acquisition capabilities, a 3D scene can be generated through panoramic image reconstruction or point cloud reconstruction. For stores without data acquisition capabilities, a training scene can be generated by editing WebGL models or using templated store construction, with instructors dragging and dropping shelves, merchandise, customer service workstations, storage areas, and after-sales areas. The 3D store scene model includes the store entrance, shelves, display windows, merchandise display area, live streaming display area, customer service workstations, order dashboard, storage and shipping area, and after-sales work order area.
[0085] In this way, by transforming the spatial layout and merchandise display of a real store into a three-dimensional interactive scene, trainees are provided with a training space that closely resembles a real business environment. This allows trainees to locate merchandise and business information through visual observation and spatial navigation during training, rather than simply imagining business scenarios through text descriptions.
[0086] Step S200: Configure the business attributes of each business object in the three-dimensional store scene model to form a scene semantic graph.
[0087] In this embodiment, business attributes are used to associate each model object with its corresponding e-commerce business object, forming a scene semantic graph. Business objects include store space objects, shelf objects, product objects, customer service workstation objects, order dashboard objects, warehouse shipping area objects, and after-sales work order area objects; business attributes include at least one of object identifier, inventory quantity, price information, activity rules, order status, logistics nodes, or after-sales rules.
[0088] For example, a specific product on a shelf can be associated with product number "SP10086", product category "digital products", inventory quantity "35 pieces", applicable discount "30 off for purchases over 300", frequently asked questions "is it genuine", and after-sales restriction "7-day no-reason return"; a specific order node in the order dashboard can be associated with order number "DD20260721001", order status "in transit", current location in the logistics nodes "shipped - in transit - on delivery - signed for", and customer history issue mark "order reminder"; and an after-sales work order can be associated with return reason "quality problem", required proof "photo / video", processing time limit "48 hours", and platform rule "compensation required for overdue processing".
[0089] In this way, by establishing a one-to-one mapping between visualized objects in a 3D scene and backend e-commerce business data through scene semantic graphs, the 3D scene is no longer a purely visual display, but a training interactive interface that carries complete business semantics. Every click made by the trainee in the scene can access real business data, providing a data foundation for subsequent judgments on whether the trainee has verified the business facts.
[0090] Step S300: Based on the scene semantic graph, extract the business data corresponding to the current training scene, generate a scene state vector, and determine the training task seed according to the scene state vector.
[0091] In this embodiment, the scene state vector is a multi-dimensional feature vector used to uniformly describe the real-time state of each business object in the current training scene.
[0092] Training task seeds include customer profiles, initial customer intent, product or order context, marketing constraints, service risk points, and assessment objectives. For example, for a "promotion explanation" training task, the training task seeds could include: customer profile "price-sensitive," initial customer intent "inquiring whether a product is eligible for a discount," product context "the product's current price and inventory status," marketing constraint "this product is only eligible for a single-item coupon and is not eligible for discounts," service risk point "incorrectly promising discounts leading to customer complaints," and assessment objective "correctly explaining the activity rules and providing alternative solutions."
[0093] In this way, a unified description of the state of scene objects is achieved through scene state vectors, enabling 3D scenes and AI dialogues to share the same state basis; and flexible configuration of training scenes is achieved through training task seeds, so as to automatically generate differentiated training tasks according to different training objectives.
[0094] In one feasible implementation, step S300 may include steps S310 to S350:
[0095] Step S310: Read the current state variables of each business object in the scene semantic graph. The state variables include at least one of the following: product inventory, order processing progress, logistics node status, after-sales work order status, and customer sentiment value.
[0096] Step S320: Combine the state variables to form the scene state vector;
[0097] Step S330: Read the data source, which includes historical anonymized customer service conversations, product knowledge base, platform rule base, user feedback text, preset teaching objectives, and / or enterprise job competency standards;
[0098] Step S340: Based on the scenario state vector, match the corresponding customer profile, customer initial intent, product context, order context, marketing constraints, service risk points, and assessment targets from the data source;
[0099] Step S350: Combine the matched customer profile, customer initial intent, product context, order context, marketing constraints, service risk points, and assessment targets into the training task seed.
[0100] In this embodiment, the system reads the current state variables of each business object from the scene semantic graph. The state variables include product inventory, such as the real-time inventory of a product being 35 units; order processing progress, such as an order currently in the shipped state; logistics node status, such as a logistics order currently in transit; after-sales work order status, such as an after-sales work order currently in the pending review state; and customer sentiment value, such as a current customer sentiment value of 0.6, indicating relatively dissatisfied.
[0101] The various state variables are combined into a scene state vector with a unified format. The scene state vector is a multi-dimensional feature vector, with a format such as: [product inventory state vector, order state vector, logistics state vector, after-sales state vector, customer sentiment value, activity state vector]. This vector is simultaneously provided to the trainee's training terminal to determine the display state of scene objects, and to the AI customer simulator as the context for generating customer expressions.
[0102] Subsequently, the system reads the data source used to generate training task seeds. Historical anonymized customer service conversations provide textual corpora of real customer questions and corresponding handling methods; the product knowledge base provides product parameters, selling points, common questions, and applicable scenarios; the platform rule base provides promotional rules, after-sales rules, and platform service standards; user feedback text provides customer satisfaction evaluation information; preset teaching objectives define the training objective type for this training, such as promotional explanation training; and enterprise job competency standards define the competency requirements and assessment weights for different positions.
[0103] Then, based on the current value of each state component in the scene state vector, the system matches the corresponding training elements from the data source. For example, if the scene state vector indicates that a product's inventory is "short in stock," the system matches urgent or expedited order customer profiles from historical anonymized sessions, matches the initial intent of expedited order customers, and matches the corresponding service risk point "incorrectly promised delivery time." Finally, the matched elements are combined to form a complete training task seed.
[0104] In this embodiment, by matching training elements from multiple data sources and combining them into training task seeds, the system can automatically generate suitable training tasks based on the current scenario state, ensuring that the content of the training tasks is consistent with the scenario state. For example, when the scenario state is "inventory shortage," the system automatically generates order reminder training tasks, avoiding the problem of task mismatch with scenario state that may occur when manually configuring training tasks. Furthermore, this embodiment generates training task seeds through data source matching, automating and standardizing the training task generation process while ensuring the consistency between training tasks and scenario states.
[0105] Optionally, step S320 may also include steps S321 to S324:
[0106] Step S321: Define the vector dimensions of the scene state vector and the range of values for each dimension;
[0107] Step S322: Based on the vector dimension, read the current value of each state variable sequentially from the scene semantic graph and perform normalization mapping;
[0108] Step S323: Concatenate the normalized state variable values in dimensional order to form the scene state vector;
[0109] Step S324: Synchronize the scene state vector to the shared state cache area for real-time reading by the student training terminal and the AI client simulator.
[0110] It should be noted that the system predefines the data structure for the scene state vector. The scene state vector is a numerical vector with fixed dimensions, as defined in Table 1:
[0111] Table 1: Definition of Scene State Vector Dimensions
[0112] Dimension Index Dimension Name Range of values Data source 1 Product inventory status [0, 1] Current inventory level / preset inventory threshold 2 Product listing status {0, 1} 0 - Removed from shelves, 1 - Added to shelves 3 Activity status {0, 1} 0 - Not in effect / Expired, 1 - In effect 4 The remaining time of the event is more than [0, 1] Remaining time / Total event duration 5 Order payment status {0, 1, 2} 0 - Unpaid, 1 - Paid, 2 - Cancelled 6 Order shipping status {0, 1, 2, 3} 0 - Not shipped, 1 - Shipped, 2 - In transit, 3 - Signed for 7 Logistics anomaly marking {0, 1} 0 - Normal, 1 - Abnormal (timeout / stalled) 8 After-sales work order status {0, 1, 2, 3} 0 - No work order; 1 - Pending processing; 2 - Under review; 3 - Completed. 9 Customer sentiment value [0, 1] 0 - Satisfied, 0.5 - Neutral, 1 - Angry 10 Service Risk Level {0, 1, 2} 0 - No risk, 1 - Medium risk, 2 - High risk 11 Current teaching stage {0, 1, 2, 3} 0 - Not started, 1 - In progress, 2 - Nearing end, 3 - Finished
[0113] Where {} represents discrete enumerated values, and [ ] represents continuous values.
[0114] In the above dimension definitions, the calculation method for product inventory status is as follows: actual inventory quantity divided by a preset inventory warning threshold. The value is 1 when the actual inventory quantity is greater than or equal to the threshold, and 0 when the inventory quantity is 0. Intermediate values are mapped proportionally. For example, if the inventory warning threshold for a product is 50 units, and the current inventory quantity is 35 units, then the inventory status component value is 35 / 50 = 0.7, indicating that the inventory is at a moderately low level. If the current inventory quantity is 10 units, then the inventory status component value is 10 / 50 = 0.2, indicating that the inventory is tight. The calculation method for the remaining activity time ratio is as follows: the remaining time until the end of the activity is divided by the total activity duration. For example, if the total activity duration is 7 days, and 3 days have passed with 4 days remaining, then the remaining activity time ratio is 4 / 7 ≈ 0.57, indicating that the activity is in the middle stage. If the remaining time is less than 10% of the total duration, then this component value approaches 0, indicating that the activity is about to end.
[0115] The specific definition and value range of each of the above dimensions can be configured and extended according to different training scenarios and business needs, and this application does not impose specific limitations on them.
[0116] Based on the vector dimensions defined above, the system sequentially reads the current values of the corresponding state variables from the scene semantic graph. For continuous state variables, the system performs normalization mapping, uniformly mapping them to the [0,1] interval. For discrete state variables, the system maps them to their corresponding integer encoded values.
[0117] For example, the following read operation is performed:
[0118] Read the current inventory quantity "35 pieces" and the preset inventory threshold "50 pieces" from the product node of the scene semantic graph, and calculate the inventory status component = 35 / 50 = 0.7;
[0119] Read the listing status "listed" from the product node and map it to the product listing status component = 1;
[0120] Read the current activity status "in effect", remaining time "4 days" and total duration "7 days" from the activity node, and map them to the activity status component = 1 and the ratio of remaining time = 4 / 7 ≈ 0.57 respectively;
[0121] Read the payment status "paid" and the shipment status "in transit" from the order node, and map them to order payment status component = 1 and order shipment status component = 2, respectively;
[0122] Read the "normal" exception marker from the logistics node and map it to a logistics exception marker component of 0;
[0123] Read the work order status "Pending" from the after-sales node and map it to the after-sales work order status component = 1;
[0124] Read the customer's sentiment value "0.6" from the dialogue context and map it to a customer sentiment component of 0.6.
[0125] Read the highest risk level "medium risk" that has been triggered from the risk rule node of the scene semantic graph and map it to the service risk component = 1;
[0126] Read the current teaching stage "In Progress" from the training task seed and map it as a teaching stage component = 1.
[0127] Then, all the obtained state variable values are concatenated into a multi-dimensional vector according to the dimension index order. For example, the values of each dimension are as follows: V1=0.7 (inventory status), V2=1 (shelf status), V3=1 (activity effective status), V4=0.57 (activity remaining time ratio), V5=1 (payment status), V6=2 (shipping status), V7=0 (logistics anomaly), V8=1 (after-sales work order status), V9=0.6 (customer sentiment value), V10=1 (service risk level), V11=1 (teaching stage).
[0128] The resulting scene state vector is: This vector has 11 dimensions and fully describes the overall state of goods, inventory, activities, orders, logistics, after-sales service, customer sentiment, and service risks in the current training scenario.
[0129] Subsequently, the generated scene state vector is written to the shared state buffer. The shared state buffer is implemented using a publish-subscribe pattern or a shared memory pattern, with the student training terminal and the AI client simulator acting as subscribers, automatically receiving the latest vector values when the state vector is updated.
[0130] The training terminal determines the display status of objects in the current scene based on the values of each component in the scene state vector. For example, when the inventory status component V1=0.7 (low to moderate inventory), the rendering terminal displays a yellow inventory warning mark at the corresponding product location; when the after-sales work order status component V8=1 (pending work orders), the rendering terminal displays a red highlight in the after-sales work order area; when the customer sentiment component V9=0.6 (poor sentiment), the rendering terminal displays a yellow warning mark in the customer service workstation area.
[0131] Before generating each round of customer expressions, the AI customer simulator reads the latest scenario state vector from the shared state cache. For example, when it reads that the inventory state component V1=0.2 (inventory shortage) and the order shipment state component V6=2 (in transit), the AI customer simulator will ensure that the expression related to order expediting is consistent with the fact of inventory shortage, avoiding requests that contradict the scenario facts. When it reads that the activity active state component V3=1 and the remaining time of the activity is V4=0.1 (ending soon), the AI customer simulator will mention the fact that the activity is about to end when generating consultation expressions, thus synchronizing the customer expression with the scenario state.
[0132] In one implementation, the scene state vector is recalculated and updated to the shared state cache each time a state variable changes. Changes in state variables are driven by triggering conditions. When a student performs an operation (such as selecting a product or submitting an after-sales work order) or an AI customer expression triggers a state change rule, the system rereads the latest state value in the scene semantic graph, executes steps S321 to S323 to generate a new scene state vector, and synchronously updates it to the shared state cache.
[0133] For example, when a student submits an after-sales service order in a 3D scene, the corresponding after-sales service order status variable in the scene semantic graph is updated from "pending" to "under review," and the mapping value V8 is updated from 1 to 2. Based on this, the system generates a new scene state vector. The update is then synchronized to the shared state cache. After receiving the update, the trainee training terminal switches the display status of the after-sales work order area from "pending processing" to "under review". After receiving the update, the AI customer simulator can respond in the next round of communication based on the new fact that "the work order has been submitted for review", for example, by saying "Okay, I have seen the work order status update, thank you for your processing".
[0134] In this way, by combining the state variables of various business objects in the scene semantic graph into a scene state vector of a unified format, the scattered state information in the 3D scene is aggregated into a structured numerical representation, which facilitates rapid reading and parsing by the 3D rendering end and the AI client simulator. Normalizing the state variables to a unified value range eliminates the differences in units and scales between different business data, enabling subsequent display logic and dialogue generation logic to be configured with rules based on a unified numerical range. Real-time synchronization of the scene state vector is achieved through a shared state cache, ensuring that the 3D rendering end and the AI client simulator are based on the same scene facts for display and dialogue generation at any time during the same training task, avoiding contradictions between scene display and client expression caused by inconsistent states.
[0135] In other words, by combining the state variables of business objects into a structured scene state vector, this application embodiment achieves unified description and efficient synchronization of multi-source and heterogeneous business states in a 3D scene, providing a shared state foundation for 3D scene display and AI customer dialogue generation, and ensuring consistency between 3D scene objects, business rules, student operation trajectories and AI customer expressions in the same state space.
[0136] Step S400: Call the 3D rendering engine to load the 3D store scene model, and render the 3D store scene according to the training task seed. At the same time, call the AI customer simulator and load the customer parameters in the training task seed to initialize the AI customer.
[0137] In this embodiment, the trainee training terminal calls the 3D rendering engine to load the 3D store scene model. Based on the product context, order context and marketing constraints in the training task seed, the relevant business objects are set to visible, highlighted or prompted states in the scene, and the 3D store scene corresponding to the current training task is rendered.
[0138] Customer parameters include customer profile (customer type, price sensitivity, patience level, and initial emotional state) and initial customer intent (such as product inquiry, price comparison and negotiation, discount verification, order reminder, return / exchange, or complaint escalation). Based on these parameters, the AI customer simulator initializes the customer intent state and emotional state, generates the first round of customer expression, and outputs it to the trainee's training terminal.
[0139] In this way, by synchronously initializing the 3D scene and AI customer, the trainees are faced with a scene state and customer state consistent with the task objective when training begins, ensuring that the starting conditions of the training task are complete and consistent.
[0140] Step S500: Collect the student's operational behavior in the 3D store scene, as well as the student's first response to the AI customer's input.
[0141] In this embodiment, the training terminal continuously collects the trainee's operational behavior in the 3D store scene and the trainee's first response to the AI customer input.
[0142] Operational behaviors can include roaming, clicking, product querying, inventory querying, order processing, promotional activity verification, logistics querying, and after-sales work order processing.
[0143] The first response content can be customer service response content submitted by the student through text input, voice input, quick message selection, or multimodal input.
[0144] In this way, by simultaneously collecting scenario operation behavior and customer service response content, a complete data foundation is provided for subsequent joint evaluation, enabling the evaluation to cover two dimensions: "whether the student did the right thing" (operation trajectory) and "whether the student said the right thing" (response content).
[0145] Step S600: Write the business data of the business object corresponding to the operation behavior into the dialogue context, and input the first reply content and the dialogue context into the AI customer simulator. The AI customer simulator generates and outputs the second reply content according to the scene state vector and the customer intent state. At the same time, it updates the display state of the corresponding business object according to the object identifier in the second reply content.
[0146] For example, when a student selects a product to view its details, business data such as the product's name, price, inventory quantity, and applicable discounts are automatically written into the dialogue context; when a student opens the order dashboard to view the logistics node of an order, data such as the order's current status, logistics location, and estimated delivery time are written into the dialogue context.
[0147] In this embodiment, the system inputs the first response from the trainee along with the dialogue context into the AI customer simulator. The AI customer simulator combines the scene state vector and the customer intent state to generate a second response, which is then output to the trainee's training terminal.
[0148] While generating the second response, the AI customer simulator analyzes the object identifiers or event types involved in the second response, locates the corresponding business objects in the 3D store scene, and performs update operations such as highlighting, state switching, or state marking. For example, if the AI customer replies, "I mean the order for that red sweater, order number DD20260721001," the system highlights the corresponding order item in the order dashboard; if the AI customer replies, "Do you still have this item in stock?", the system flashes the corresponding product's location marker in the product display area.
[0149] In this way, by writing the business data corresponding to the student's actions into the dialogue context, the AI customer can perceive whether the student has verified the relevant information, thereby dynamically adjusting subsequent questions. By updating the display state of the 3D scene in reverse according to the object identifier in the AI customer's expression, the scene display changes dynamically with the dialogue content, thus forming a two-way closed loop of scene-driven dialogue and dialogue feedback scene.
[0150] Step S700: Generate a training score based on the operation behavior and the first response content, and determine at least one weak ability of the trainee based on the training score.
[0151] It should be noted that the evaluation metrics include the accuracy of scene object search, compliance with platform rules, customer problem resolution rate, dialogue coherence, emotional reassurance effect, response timeliness, and the hit rate of risk-based communication.
[0152] Among these metrics, the accuracy of scene object search can be determined by whether the student selected the correct scene object related to the question before replying; platform rule compliance can be determined by whether the rule explanation in the reply is consistent with the platform rules; customer problem resolution rate can be determined by whether the customer's intent shifts from an inquiry / complaint state to a transaction / satisfaction state; dialogue coherence can be determined by whether the reply is coherent with the context; emotional reassurance effect can be determined by the amount of change in the customer's emotional state after the student's reply; response timeliness can be determined by the student's response time and task completion time; and the hit rate of risk-based communication can be determined by whether absolute promises such as "refund anytime" or "definitely arrive" are made.
[0153] The system determines the trainee's competency score for each dimension based on the scores of each sub-item, including product knowledge, object identification, rule comprehension, communication and expression, emotional management, order processing, and after-sales service. Competency dimensions below a preset threshold are marked as weak areas.
[0154] Thus, by comprehensively evaluating both the operational trajectory and the response content, it is possible to identify problem types that are difficult to distinguish using traditional text evaluations, such as "correct wording but unverified facts" and "correct search but inappropriate expression." For example, both trainees correctly answered the return and exchange period, but one trainee checked the order receipt time and after-sales rules before answering, while the other trainee simply recited the rule wording. The former scored higher in the "object search ability" dimension, a distinction that cannot be achieved solely based on the response text.
[0155] Step S800: Adjust the training task parameters of the trainee according to the identified weaknesses.
[0156] In this embodiment, the training task parameters are used to determine the scenario state, customer intent, or service risk points in the next training task.
[0157] Optionally, for trainees who are weak in understanding rules, the system increases the probability of scenarios such as marketing rule conflicts and rule ambiguities appearing in subsequent training; for trainees who are weak in emotional soothing, the system increases the probability of complaint-type customer profiles appearing; and for trainees who are weak in object search, the system increases the probability of complex shelf and multi-order scenarios appearing.
[0158] In this way, by adjusting the training task parameters according to the trainees' weaknesses, subsequent training tasks can be focused on the trainees' actual weaknesses, thereby improving training efficiency.
[0159] In the technical solution provided in this embodiment, a closed loop of state between the 3D scene and the AI dialogue is realized through 3D scene modeling, scene semantic graph establishment, scene state vector generation, and scene dialogue synchronization; the effectiveness and pertinence of customer service training are improved by jointly evaluating the operation trajectory and response content and by identifying weaknesses to achieve personalized training path updates.
[0160] As an optional implementation, step S400 may include steps S410 to S430:
[0161] Step S410: Read the customer profile and the customer's initial intent from the training task seed;
[0162] Step S420: Write the customer type, price sensitivity, patience level, and initial emotional value from the customer profile into the AI customer simulator; and,
[0163] Step S430: Write the customer's initial intent into the AI customer simulator, and set the customer intent status based on the customer's initial intent. The customer intent status includes inquiry status, price comparison status, objection status, order reminder status, return / exchange status, complaint status, add to cart status, and transaction status.
[0164] It should be noted that customer profiles can include customer type, such as "price-sensitive", "quality-focused", "urgent order", "complainant", and "hesitant"; price sensitivity, for example, a value range of 0-1, where 0 means insensitive and 1 means extremely sensitive; patience level, for example, a value range of 0-1, where 0 means extremely impatient and 1 means very patient; and initial emotional value, for example, a value range of 0-1, where 0 means very satisfied and 1 means very angry.
[0165] Initial customer intent may include product inquiries, such as "How is the battery life of this phone?"; price comparison and negotiation, such as "Other stores are 50 yuan cheaper than you, can you lower the price?"; discount verification, such as "Can I use this discount coupon?"; order follow-up, such as "Why hasn't my order been shipped after 3 days?"; returns and exchanges, such as "I want to return the item, the color of the clothes is wrong"; and escalation of complaints, such as "I want to file a complaint against your customer service!"
[0166] In this embodiment, the system writes various parameters of the customer profile into the customer profile unit of the AI customer simulator. These parameters determine the AI customer simulator's expression style, sensitivity to emotional changes, and decision-making logic in subsequent conversations. For example, customers with high price sensitivity are more likely to express objections when recommended high-priced items, and customers with low patience experience a faster rise in emotions while waiting for a response.
[0167] Subsequently, the customer's initial intent is written into the intent state unit of the AI customer simulator, and the initial intent state is set to the corresponding state. Customer intent states include: inquiry state, such as a customer asking for product information; price comparison state, such as a customer comparing prices of different products or stores; objection state, such as a customer questioning a product or rule; order expediting state, such as a customer urging delivery or logistics updates; return / exchange state, such as a customer requesting a return or exchange; complaint state, such as a customer expressing dissatisfaction and requesting a solution; add to cart state, such as a customer considering a purchase and on the verge of making a decision; and transaction state, such as a customer deciding to purchase or the problem being resolved.
[0168] In this embodiment, by writing customer profiles and initial intentions into the AI customer simulator, the AI customer has clear personality traits and behavioral goals at the start of training. Different customer profiles exhibit different expression styles and emotional change logics in subsequent dialogues, improving the realism and diversity of the simulated customer. Furthermore, through multi-dimensional configuration of customer profiles and initialization of intention states, this embodiment enables the AI customer simulator to simulate different types of customer behavior patterns, providing trainees with diverse training scenarios.
[0169] Furthermore, in step S600, the step of generating and outputting the second response content based on the scene state vector and the customer intent state by the AI customer simulator may include: configuring the customer emotional state based on the first response content; generating the second response content based on the customer intent state, the customer emotional state, and the scene state vector, thereby realizing the generation of dynamic AI customer responses.
[0170] In this embodiment, the system performs multi-dimensional analysis on the first response input by the student and updates the customer's emotional state accordingly. Specifically, the system analyzes the following aspects: rule correctness, such as whether the rule explanation in the response is consistent with the platform rules; solution effectiveness, such as whether the response provides a feasible solution; response time, such as the time interval from the customer's question to the student's response, which is considered a response timeout if it exceeds 60 seconds; and whether platform risk rules are triggered, such as absolute promises, incorrect price promises, or sending private links.
[0171] Based on the above analysis results, the system updates the customer's emotion value according to preset emotion transfer rules. For example, if the learner correctly explains the rule and provides an effective solution, the customer's emotion value decreases by 0.2; if the learner makes an incorrect promise without verifying the scenario, the customer's emotion value increases by 0.3; if the learner's response timeout exceeds 60 seconds, the customer's emotion value increases by 0.1. The customer's emotion value ranges from 0 to 1, where 0 represents satisfaction and 1 represents anger.
[0172] Then, the system integrates the current customer intent state, the updated customer emotional state, and the scenario state vector to generate a second response through the script generation unit. Script generation can be implemented using a large language model, a retrieval-enhanced generation model, a state machine model, or a template model.
[0173] For example, if a customer's intention is in a "consultation" state and their sentiment score rises to 0.7, the AI customer can express, "How come you don't even know the rules of the activity? This is very disappointing." If a customer's intention is in a "price comparison" state and their sentiment score remains at 0.3, the AI customer can express, "Then let me compare with other companies; your prices are indeed not competitive." If a customer's intention is in a "complaint" state and their sentiment score is as high as 0.9, the AI customer can express, "I'm going to complain about you on the platform! Your service is terrible."
[0174] In this embodiment, by dynamically updating the customer's emotional state and generating responses based on multi-dimensional states, the AI customer's expression can reflect the different emotional feedback brought about by different response quality from the learner, enhancing the realism and interactivity of the training. In other words, this embodiment, through a multi-factor update mechanism for emotional states and response generation based on state vectors, enables the AI customer's expressive style and emotional intensity to dynamically change with the learner's response quality, thereby providing a more targeted and realistic training experience and achieving the goal of improving the realism and teaching effectiveness of AI customer simulation training.
[0175] Based on the first embodiment of this application, in the second embodiment of this application, the same or similar content as the above embodiment can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 2In step S600, the step of writing the business data of the business object corresponding to the operation behavior into the dialogue context may include identifying the operation type and target business object of the operation behavior; when the operation type is a point operation, a view operation, or a process operation, reading the business data of the target business object from the scene semantic graph; and writing the business data into the cache of the dialogue context.
[0176] It should be noted that the actions can include roaming, clicking, product querying, inventory querying, order processing, promotional activity verification, logistics querying, or after-sales work order processing.
[0177] In this embodiment, the system identifies each action of the learner in real time, determines the action type, and extracts the target business object. For example, when a learner clicks on a product model in a 3D scene, the system identifies the action type as "click-and-select" and the target business object as the product number corresponding to that product.
[0178] Next, the system determines whether the operation type belongs to the business operations that need to be recorded. Among these, click operations, such as clicking on a scene object to view details; viewing operations, such as opening an information panel or order dashboard; and processing operations, such as submitting an after-sales work order or confirming an order node, all belong to the operation types that need to be recorded. For the above operations, the system reads the business data of the target business object from the scene semantic graph, such as product price, inventory quantity, order status, logistics nodes, after-sales rules, etc. Then, the read business data is written to the cache of the dialogue context.
[0179] Optionally, the dialogue context is managed on a session-by-session basis, with all dialogue rounds within the same training task sharing the same dialogue context. When the trainee subsequently responds to the customer, the AI customer simulator can read the operation trajectory recorded in the dialogue context and the viewed business data.
[0180] In this way, by writing the business data corresponding to the student's actions into the dialogue context, the AI customer simulator can perceive what the student did and what they viewed, thereby verifying whether the student has truly verified the necessary information in subsequent questions. For example, if a customer asks whether a certain product is eligible for a discount, if the AI customer reads from the dialogue context that the student has viewed the activity rules, it can ask, "Did you see the applicable scope mentioned in the rules?" when generating the next round of questions; if it does not read the viewing record, it can follow up with, "Have you confirmed the activity rules for this product? We suggest you take a look first."
[0181] Further, in step S600, the step of updating the display status of the corresponding business object according to the object identifier in the second reply content may include parsing the second reply content and extracting the object identifier or event type; searching for the corresponding business object in the three-dimensional store scene according to the object identifier or event type; and performing highlighting, status switching or status marking on the found business object.
[0182] The system parses the second response content generated by the AI customer simulator and extracts the object identifiers or event types involved. For example, when the response content contains the product name "red sweater", the system maps it to the corresponding product number; when it contains the order number "DD20260721001", the system extracts the order identifier; when it contains the event type "add to cart", the system identifies the event type as an add-to-cart operation.
[0183] In this embodiment, the system parses the second response content generated by the AI customer simulator and extracts the object identifiers or event types involved. For example, when the response content contains the product name "red sweater", the system maps it to the corresponding product number; when it contains the order number "DD20260721001", the system extracts the order identifier; when it contains the event type "add to cart", the system identifies the event type as an add-to-cart operation.
[0184] Then, based on the extracted object identifier, the system searches for the corresponding business object in the 3D store scene model. If the object identifier is a product number, the system locates the corresponding product model in the scene; if the object identifier is an order number, the system locates the corresponding order entry in the order dashboard. The system then performs a display status update operation on the found business objects. Highlighting includes highlighting the target object with a bright border or flashing light effect; status switching includes changing the product status from "Browse" to "Compare" or "Add to Cart" visual status; status labeling includes overlaying status labels on the object, such as "Customer has inquired" or "Added to Cart".
[0185] In this way, the 3D scene is updated dynamically by driving customer feedback, allowing the scene display to change dynamically with the content of the conversation. When a customer mentions a product in the conversation, that product is automatically highlighted in the 3D scene, helping trainees quickly locate the object the customer is discussing; when a customer expresses a purchase intention, the status label of the corresponding product is updated to "added to cart," providing trainees with richer contextual information.
[0186] In the technical solution provided in this embodiment, a mechanism for writing operational behavior into the dialogue context and a mechanism for AI responses to drive scene updates are used to achieve bidirectional synchronization between student operations affecting AI dialogue and AI dialogue feedback scenarios. This achieves deep coupling between the virtual simulation environment and real business logic, providing a high-fidelity, highly interactive, immersive training foundation for e-commerce talent training.
[0187] Based on the first embodiment of this application, in the third embodiment of this application, the same or similar content as the above embodiment can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 3 Step S700 may include steps S710 to S740:
[0188] Step S710: Calculate the scores for the following indicators: object search accuracy, product recommendation matching degree, platform rule compliance, customer problem resolution rate, dialogue coherence, emotional soothing effect, response timeliness, transaction or lead generation result, after-sales processing accuracy, and risk response hit rate.
[0189] Step S720: Based on preset weights, the scores of each indicator are weighted and summed to obtain the training score;
[0190] Step S730: Determine the ability score of each ability dimension of the trainee based on the scores of each indicator. The ability dimensions include product knowledge ability, object search ability, rule understanding ability, communication and expression ability, emotional soothing ability, order processing ability, and after-sales processing ability.
[0191] Step S740: If the ability score is lower than a preset ability threshold, mark the ability dimension corresponding to the ability score as the weak point.
[0192] In this embodiment, the system performs itemized evaluations on the collected scenario operation behaviors and customer service response content, specifically:
[0193] Object search accuracy is determined by whether the student selected a relevant target scene object in the 3D scene before responding to a customer's question or within a reasonable timeframe, such as within 30 seconds before responding. The accuracy rate is calculated by dividing the number of correctly found objects by the total number of objects that should have been found.
[0194] Product recommendation matching score: When trainees recommend alternative or complementary products to customers, the score is calculated based on the semantic similarity of product category, price range, and functional attributes to determine the degree of matching between the recommended products and the customer's needs.
[0195] Check the platform rules compliance. Examine the rule explanations in the response, such as return period, discount rules, and compensation standards, to see if they are consistent with the provisions in the platform's rule library.
[0196] Customer problem resolution rate is determined by the final state of the customer's intent at the end of the conversation, such as whether the consultation has progressed to a sale or satisfaction, to determine whether the customer's problem has been resolved.
[0197] Dialogue coherence is assessed by analyzing whether the response is relevant to the customer's question and whether it follows the preceding context, and scoring is based on semantic similarity or coherence models.
[0198] The effectiveness of emotional reassurance is assessed by analyzing whether the responses contain reassuring language and the extent to which the customer's emotional state decreases after the trainee's response.
[0199] Response timeliness is calculated by taking the average response time from when a customer asks a question to when a reply is sent, as well as the overall completion time of the training task, and comparing it with the preset time limit to obtain a score.
[0200] For transaction or lead generation results, if the training task includes sales targets, determine whether the conversation resulted in adding to cart, leading to a customer's contact, or a transaction.
[0201] After-sales processing accuracy rate: In after-sales training scenarios, this refers to determining whether trainees have completed the processing according to the correct after-sales procedures.
[0202] The system checks whether responses contain risky statements such as absolute promises, false promises, or privacy leaks. Points are deducted for each such statement.
[0203] It should be noted that the weighting configurations differ for different job types. For example, in after-sales customer service training, the accuracy of after-sales processing and the compliance with rules have higher weights; in pre-sales customer service training, the matching accuracy of product recommendations and emotional reassurance have higher weights; and in novice training, the accuracy of object search and the compliance with basic rules have higher weights.
[0204] Optionally, the scores of each sub-indicator can be mapped to corresponding capability dimensions. Specifically, product knowledge capability is determined based on the product rules portion of product recommendation matching and rule compliance; object search capability is determined based on object search accuracy; rule comprehension capability is determined based on platform rule compliance; communication and expression capability is determined based on dialogue coherence and the accuracy of risk-based communication; emotional reassurance capability is determined based on the effectiveness of emotional reassurance; order processing capability is determined based on order-related tasks within the order verification path and response time; and after-sales processing capability is determined based on after-sales processing accuracy.
[0205] In this embodiment, the system compares the capability scores of each capability dimension with preset capability thresholds. For example, if the capability score for object search capability is below 60 points (out of 100), that dimension is marked as a weak point. The marking results are used for training path adjustment in step S800.
[0206] In the technical solution provided in this embodiment, by calculating sub-indicators and mapping ability dimensions, the multi-dimensional evaluation results are attributed to specific ability dimensions, giving subsequent targeted training a clear direction. By mapping evaluation indicators to ability dimensions, not only is a comprehensive score provided, but the specific weaknesses of the trainees' abilities are also accurately located, providing precise input for personalized training.
[0207] Further, step S800 may include steps S810 to S830.
[0208] Step S810: Read the trainee's historical training records to obtain the business scenario type and service process type corresponding to the weak points in the ability;
[0209] Step S820: Increase the weight of the occurrence of scenario status, customer intent or service risk points corresponding to the business scenario type and the service process type in the training task parameters;
[0210] Step S830: Associate the occurrence weight with the task generation conditions of the next training task.
[0211] In this embodiment, the system reads the trainee's historical training records and extracts the business scenario type and service process type corresponding to the weak points in the records. For example, if the weak point is "rule comprehension ability," the system analyzes from the historical training records which business scenarios the trainee scored low in rule comprehension, such as "promotion rule explanation scenario" and "after-sales deadline judgment scenario." If the weak point is "emotional soothing ability," the system analyzes which service processes the trainee scored low in emotional soothing, such as "complaint escalation handling process" and "order expediting handling process."
[0212] Subsequently, based on the determined business scenario type and service process type, the parameter configuration for subsequent training tasks is adjusted. Specifically, the system increases the weight of corresponding scenario states, customer intentions, or service risk points in subsequent training tasks.
[0213] For example, if a trainee's score on rule comprehension in the "Promotion Rule Explanation Scenario" remains consistently low, the system will increase the weight of the "Marketing Rule Conflict" scenario from 1.0 to 1.5, and the weight of the "Discount Verification" customer intent scenario from 1.0 to 1.4. If a trainee's score on emotional reassurance in the "Complaint Escalation Handling Process" remains low, the system will increase the weight of the "Complaint-Type" customer profile from 1.0 to 1.6.
[0214] Then, the adjusted occurrence weights are used as the conditions for generating the next training task and saved in the trainee's personal training configuration. When the system triggers the generation of the next training task, the task generation module reads this configuration, matches the corresponding training elements from the data source according to the adjusted weights, and generates a targeted training task.
[0215] In the technical solution provided in this embodiment, by adjusting the occurrence weights in the training task parameters, the frequency of occurrence of scenario types and customer intentions corresponding to weak points in subsequent training tasks is significantly increased, enabling trainees to strengthen their weak abilities through intensive training. By identifying weak abilities and adjusting the occurrence weights of corresponding training elements, training is transformed from uniform question bank training to targeted ability improvement, achieving a personalized training effect.
[0216] Based on the first embodiment of this application, in the fourth embodiment of this application, the same or similar content as the above embodiment can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 4 The virtual simulation interaction method for e-commerce talent training platforms may also include steps S911 to S912:
[0217] Step S911: Receive training configuration instructions from the teacher's end or the management end. The training configuration instructions carry configuration parameters, including training objectives, job competency standards, training duration, product categories, platform rules, assessment weights, and / or customer service knowledge scope.
[0218] Step S912: Update the task generation conditions of the training task seed and the preset weights of the training score according to the configuration parameters.
[0219] In this embodiment, teachers or training administrators input training configuration commands through the configuration interface on the teacher's or management's end. The training objective defines the training type, such as "pre-sales consultation training," "after-sales processing training," or "comprehensive ability training"; the job competency standards define the competency requirements for different positions, such as pre-sales customer service, after-sales customer service, and live-streaming assistant; the training duration defines the maximum time for a single training session, such as 30 minutes; the product category restricts the range of products covered in the training, such as "digital products only"; the platform rules specify the applicable platform rule version; the assessment weight defines the weight allocation for each evaluation indicator; and the customer service knowledge scope restricts the range of knowledge bases that trainees can use.
[0220] Then, the system updates the generation conditions of the training task seeds based on the training objectives and product categories in the configuration parameters. For example, it restricts the range of products, customer intent types, and scene state types when matching training task seeds. The system also updates the weighted formula for the training score based on the assessment weights in the configuration parameters, making the score more aligned with the focus of the current training objective.
[0221] In the technical solution provided in this embodiment, the teacher-side configuration function allows teachers to flexibly adjust the generation logic and evaluation criteria of training tasks according to different course stages, training objectives, and job requirements, thereby improving the system's adaptability and controllability. Through the teacher-side configuration function, the training platform can adapt to the diverse training needs of different schools, enterprises, and job positions, allowing teachers to adjust training content and evaluation criteria without modifying the system code.
[0222] Reference Figure 5 Optionally, the virtual simulation interaction method of the e-commerce talent training platform may also include steps S921 to S923:
[0223] Step S921: Based on the teacher's end or the management end receiving a playback instruction, the playback instruction carries the target training record identifier;
[0224] Step S922: Based on the target training record identifier, read the corresponding stored operation trajectory data and stored response content data;
[0225] Step S923: On the teacher's end or the management end, the stored operation trajectory data and the stored reply content data are played back synchronously.
[0226] In this embodiment, the teacher or training administrator selects a student's historical training record on the teacher's or administrator's end. The system receives the playback instruction and extracts the target training record identifier. Then, it reads the student's 3D scene operation trajectory data stored in the training log database, including the timestamp of each operation, operation type, target object identifier, and state values before and after the operation; as well as the complete dialogue record between the student and the AI client.
[0227] The system synchronously displays the playback screen of the 3D scene and the timeline of the dialogue record on the playback interface of the teacher or administrator. During playback, the 3D scene reproduces each step of the trainee's operation in the time sequence of the operation trajectory, including changes in camera perspective, object selection, opening of information panels, etc.; the dialogue record is displayed synchronously in a scrolling manner according to the timeline, allowing teachers to intuitively observe the trainee's behavioral path and thought process during the training.
[0228] Thus, by synchronously replaying operation trajectories and dialogue records, this embodiment of the application allows teachers not only to see what students say, but also to observe what actions students take before or after speaking. For example, whether they answer directly or verify information first, whether they systematically search for information or randomly select items, allows teachers to identify specific problems students encounter in their business processing flow. In other words, the synchronous playback function provides teachers with an intuitive teaching diagnostic tool, enabling them to gain a deeper understanding of students' actual operating habits and thinking patterns, providing concrete evidence for teaching guidance.
[0229] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the virtual simulation interaction method of the e-commerce talent training platform of this application. Any simple modifications based on this technical concept are within the protection scope of this application.
[0230] This application also provides a virtual simulation interactive system for an e-commerce talent training platform; please refer to [reference needed]. Figure 6 The virtual simulation interactive system of the e-commerce talent training platform includes:
[0231] The 3D scene modeling module 10 is used to create a 3D store scene model based on spatial and product materials.
[0232] Scene semantic graph module 20 configures the business attributes of each business object in the three-dimensional store scene model to form a scene semantic graph;
[0233] The training task generation module 30 extracts business data corresponding to the current training scenario based on the scene semantic graph, generates a scene state vector, and determines the training task seed based on the scene state vector.
[0234] The 3D interactive training module 40 calls the 3D rendering engine to load the 3D store scene model and renders the 3D store scene according to the training task seed. At the same time, it calls the AI customer simulator and loads the customer parameters in the training task seed to initialize the AI customer.
[0235] AI customer simulation module 50 collects the student's operational behavior in the three-dimensional store scene, as well as the student's first response to the AI customer's input;
[0236] The scene dialogue synchronization module 60 writes the business data of the business object corresponding to the operation behavior into the dialogue context, and inputs the first reply content and the dialogue context into the AI customer simulator. The AI customer simulator generates and outputs the second reply content according to the scene state vector and the customer intent state, and updates the display state of the corresponding business object according to the object identifier in the second reply content.
[0237] The training evaluation module 70 generates a training score based on the operation behavior and the first response content, and determines at least one weak ability of the trainee based on the training score.
[0238] The personalized training module 80 adjusts the training task parameters of the trainee based on the identified weaknesses in their abilities.
[0239] It should be noted that the system's data access layer is used to obtain spatial materials, business data, and teacher configuration information from real stores. The various modules work together to achieve a closed loop of state synchronization between the 3D scene and the AI dialogue.
[0240] The virtual simulation interaction system for the e-commerce talent training platform provided in this application adopts the virtual simulation interaction method of the e-commerce talent training platform in the above embodiments. It can solve the technical problem of how to enable the 3D interactive training terminal and the AI customer simulator to share and bidirectionally update the scene state around the same e-commerce business object, so as to keep the state of the e-commerce business object used for 3D scene rendering synchronized with the state of the e-commerce business object used for customer expression generation. Compared with the prior art, the beneficial effects of the virtual simulation interaction system for the e-commerce talent training platform provided in this application are the same as the beneficial effects of the virtual simulation interaction method for the e-commerce talent training platform provided in the above embodiments. Furthermore, other technical features of the virtual simulation interaction system for the e-commerce talent training platform are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.
[0241] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.
Claims
1. A virtual simulation interaction method for an e-commerce talent training platform, characterized in that, The method includes: Based on spatial and product materials, a 3D store scene model is created; Configure the business attributes of each business object in the 3D store scene model to form a scene semantic graph; Based on the scene semantic graph, extract the business data corresponding to the current training scene, generate a scene state vector, and determine the training task seed based on the scene state vector; The 3D rendering engine is invoked to load the 3D store scene model, and the 3D store scene is rendered according to the training task seed. At the same time, the AI customer simulator is invoked and the customer parameters in the training task seed are loaded to initialize the AI customer. The system collects the student's actions in the 3D store scene, as well as the student's first response to the AI customer's input. The business data of the business object corresponding to the operation behavior is written into the dialogue context, and the first reply content and the dialogue context are input into the AI customer simulator. The AI customer simulator generates the second reply content and outputs it according to the scene state vector and customer intent state. At the same time, the display state of the corresponding business object is updated according to the object identifier in the second reply content. A training score is generated based on the operational behavior and the content of the first response, and at least one weak ability of the trainee is determined based on the training score. Adjust the trainee's training task parameters based on the identified weaknesses.
2. The virtual simulation interaction method for the e-commerce talent training platform as described in claim 1, characterized in that, The steps of extracting business data corresponding to the current training scenario based on the scene semantic graph, generating a scene state vector, and determining the training task seed based on the scene state vector include: In the scene semantic graph, the current state variables of each business object are read. The state variables include at least one of the following: product inventory, order processing progress, logistics node status, after-sales work order status, and customer sentiment value. The state variables are combined to form the scene state vector; Read data sources, including historical anonymized customer service conversations, product knowledge bases, platform rule bases, user feedback texts, preset teaching objectives, and / or enterprise job competency standards; Based on the scenario state vector, match the corresponding customer profile, customer initial intent, product context, order context, marketing constraints, service risk points, and assessment targets from the data source; The matched customer profile, the customer's initial intent, the product context, the order context, the marketing constraints, the service risk points, and the assessment objectives are combined to form the training task seed.
3. The virtual simulation interaction method for the e-commerce talent training platform as described in claim 1, characterized in that, The steps of calling the AI customer simulator and loading the customer parameters from the training task seed to initialize the AI customer include: From the training task seed, read the customer profile and the customer's initial intent; The AI customer simulator is written with the customer profile's initial values for customer type, price sensitivity, patience level, and emotional state; and, The initial customer intent is written into the AI customer simulator, and the customer intent state is set based on the initial customer intent. The customer intent state includes inquiry state, price comparison state, objection state, order reminder state, return and exchange state, complaint state, add to cart state, and transaction state.
4. The virtual simulation interaction method for the e-commerce talent training platform as described in claim 3, characterized in that, The steps of generating and outputting the second response content based on the scenario state vector and the customer intent state by the AI customer simulator include: Configure the customer's emotional state based on the first response; The second response content is generated based on the customer's intent state, the customer's emotional state, and the scenario state vector.
5. The virtual simulation interaction method for the e-commerce talent training platform as described in claim 1, characterized in that, The operational behaviors include roaming, clicking, product querying, inventory querying, order processing, promotional activity verification, logistics querying, or after-sales work order processing. The step of writing the business data of the business object corresponding to the operation behavior into the dialogue context includes: Identify the operation type and target business object of the operation; When the operation type is a selection operation, a viewing operation, or a processing operation, the business data of the target business object is read from the scene semantic graph; Write the business data into the cache of the dialogue context; The step of updating the display status of the corresponding business object based on the object identifier in the second reply content includes: Parse the second response content to extract the object identifier or event type; Based on the object identifier or event type, locate the corresponding business object in the three-dimensional store scene; The found business objects are highlighted, their status is toggled, or their status is marked.
6. The virtual simulation interaction method for the e-commerce talent training platform as described in claim 1, characterized in that, The step of generating a training score based on the operational behavior and the content of the first response, and determining at least one weakness of the trainee based on the training score, includes: The scores for the following metrics were calculated: object search accuracy, product recommendation matching degree, platform rule compliance, customer problem resolution rate, dialogue coherence, emotional soothing effect, response timeliness, transaction or lead generation results, after-sales processing accuracy, and risk response hit rate. Based on preset weights, the scores of each indicator are weighted and summed to obtain the training score; Based on the scores of each indicator, the ability scores of each trainee in each ability dimension are determined. The ability dimensions include product knowledge ability, object finding ability, rule understanding ability, communication and expression ability, emotional soothing ability, order processing ability, and after-sales processing ability. If the ability score is lower than a preset ability threshold, the ability dimension corresponding to the ability score is marked as the weak point.
7. The virtual simulation interaction method for the e-commerce talent training platform as described in claim 6, characterized in that, The step of adjusting the trainee's training task parameters based on the identified weaknesses includes: Read the trainee's historical training records to obtain the business scenario type and service process type corresponding to the weak points in the skills; Increase the weight of the occurrence of scenario states, customer intentions, or service risk points corresponding to the business scenario type and the service process type in the training task parameters; The occurrence weights are associated with the task generation conditions of the next training task.
8. The virtual simulation interaction method for an e-commerce talent training platform as described in any one of claims 1 to 7, characterized in that, The method further includes: The training configuration instruction is received by the teacher or administrator. The training configuration instruction carries configuration parameters, including training objectives, job competency standards, training duration, product categories, platform rules, assessment weights, and / or customer service knowledge scope. Based on the configuration parameters, update the task generation conditions of the training task seed and the preset weights of the training score.
9. The virtual simulation interaction method for the e-commerce talent training platform as described in claim 8, characterized in that, The method further includes: Based on the playback instruction received by the teacher or management terminal, the playback instruction carries the target training record identifier; Based on the target training record identifier, read the corresponding stored operation trajectory data and stored response content data; On the teacher's or administrator's end, the stored operation trajectory data and the stored reply content data are played back synchronously.
10. A virtual simulation interactive system for an e-commerce talent training platform, characterized in that, The system includes: The 3D scene modeling module is used to create 3D store scene models based on spatial and product materials. The scene semantic graph module configures the business attributes of each business object in the 3D store scene model to form a scene semantic graph. The training task generation module extracts business data corresponding to the current training scenario based on the scene semantic graph, generates a scene state vector, and determines the training task seed based on the scene state vector. The 3D interactive training module calls the 3D rendering engine to load the 3D store scene model and renders the 3D store scene according to the training task seed. At the same time, it calls the AI customer simulator and loads the customer parameters in the training task seed to initialize the AI customer. The AI customer simulation module collects the trainee's operational behavior in the three-dimensional store scene, as well as the trainee's first response to the AI customer's input. The scene dialogue synchronization module writes the business data of the business object corresponding to the operation behavior into the dialogue context, and inputs the first reply content and the dialogue context into the AI customer simulator. The AI customer simulator generates and outputs the second reply content according to the scene state vector and the customer intent state, and updates the display state of the corresponding business object according to the object identifier in the second reply content. The training evaluation module generates a training score based on the operation behavior and the first response content, and determines at least one weak ability of the trainee based on the training score. The personalized training module adjusts the training task parameters of the trainee based on the identified weaknesses in their abilities.
