A digital employee service method, system, device and storage medium

Abstract

The application discloses a digital employee service method, system, device and storage medium. The method comprises the following steps: acquiring a plurality of historical service information corresponding to a plurality of service targets; training a preset low-rank adaptive model based on the historical service information corresponding to each service target, to obtain a first service information prediction model corresponding to each service target; constructing a second service information prediction model; training the second service information prediction model through all the historical service information of the plurality of service targets, until a trained second service information prediction model is obtained; and inputting target service information into the second service information prediction model to obtain an output corresponding to a target service information prediction result, so that an independent low-rank adaptive module can be trained for each platform / store, and then the training is integrated to promote knowledge sharing, and the service effect, service efficiency and knowledge consistency are improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence service technology, and in particular to a service method, system, device and storage medium for digital employees. Background Technology

[0002] As the service industry evolves towards a multi-platform, multi-store operation model, merchants expand their businesses through this model. The services provided by their digital employees are crucial for improving service efficiency and reducing operating costs. Moreover, different platforms and stores exhibit significant differences in response styles, script standards, and service processes, requiring models to accurately capture the characteristics of each platform.

[0003] In existing technologies, services for digital employees across multiple platforms primarily employ solutions such as training models independently for a single platform or store, uniform fine-tuning of a general-purpose model, or static low-rank adaptation (LoRA) switching to implement digital employee functionality. However, the independent platform model solution is costly to train and maintain and cannot share knowledge, while the uniform fine-tuning solution for a general-purpose model leads to homogenized responses due to the overshadowing of niche platform characteristics and exacerbates the performance degradation in stores with sparse data. Furthermore, the static low-rank adaptation switching solution, due to the inherent defects of isolated module training and static switching, still cannot achieve cross-platform knowledge interaction and transfer, resulting in optimization bottlenecks for stores with sparse data and technical issues of overall knowledge consistency lack, ultimately leading to poor service performance and low service efficiency for digital employees. Summary of the Invention

[0004] The following is an overview of the subject matter described in detail herein. This overview is not intended to limit the scope of the claims.

[0005] The main objective of this disclosure is to propose a service method, system, device, and storage medium for digital employees, which can train independent low-rank adaptive modules for each platform / store, thereby integrating training to promote knowledge sharing and improve service effectiveness, service efficiency, and knowledge consistency.

[0006] A first aspect of this application provides a digital employee service method for a central controller, the method comprising:

[0007] Acquire multiple historical service information corresponding to multiple service targets, wherein the service targets are platforms and / or stores, and the historical service information includes at least historical customer service dialogue records, product information, and commonly used customer service scripts;

[0008] Based on the historical service information corresponding to each service target, a preset low-rank adaptive model is trained to obtain a first service information prediction model corresponding to each service target.

[0009] Construct a second service information prediction model; the second service information prediction model includes at least a plurality of first service information prediction models set in parallel, and a feature fusion model for fusing the output features corresponding to the plurality of first service information prediction models set in parallel and outputting the service information prediction result.

[0010] The second service information prediction model is trained using all the historical service information of the multiple service targets until the trained second service information prediction model is obtained.

[0011] In response to the target service information, the target service information is input into the second service information prediction model to obtain the corresponding target service information prediction result.

[0012] In some embodiments of this application, the second service information prediction model further includes a basic service model and a language service model. The step of training the second service information prediction model using all the historical service information of the plurality of service targets until a fully trained second service information prediction model is obtained includes:

[0013] Use all the historical service information of the multiple service targets as training data for the first training;

[0014] The training data from the first training iteration is input into the feature fusion model to obtain the first feature output by the feature fusion model;

[0015] The training data from the first training session is input into the basic service model to obtain the second feature output by the basic service model;

[0016] The training data from the first training session is input into the language service model to obtain the third feature output by the language service model.

[0017] The first feature, the second feature, and the third feature are weighted and calculated to obtain the fourth feature;

[0018] The fourth feature will be used as training data for the next training of the second service information prediction model.

[0019] In some embodiments of this application, the weighted calculation of the first feature, the second feature, and the third feature to obtain the fourth feature includes:

[0020] Using the third feature as the query vector and the second feature as the key vector and value vector, attention calculation is performed to obtain the calculation result.

[0021] Based on the first feature, the second feature, and the calculation result, the fourth feature is obtained by weighted summation using preset weighting coefficients.

[0022] In some embodiments of this application, the construction of the basic service model includes:

[0023] Preset basic model;

[0024] The basic service model is obtained by training the basic model using all the historical service information of the multiple service targets.

[0025] In some embodiments of this application, the training process of the second service information prediction model includes:

[0026] Freeze the model parameters of the language service model and each of the parallel first service information prediction models;

[0027] Through multiple rounds of iteration, the model parameters of the basic service model and the attention calculation parameters involved in the weighted calculation are trained so that the fourth feature meets the preset target conditions.

[0028] In some embodiments of this application, before obtaining the historical service information corresponding to multiple service targets, the method further includes:

[0029] The initial historical customer service dialogue records in each of the historical service information are cleaned to obtain multiple historical customer service dialogue records containing valid dialogues.

[0030] The initial product information in each of the historical service information is standardized to obtain multiple product information;

[0031] The customer service scripts in the historical service information are classified and processed to form a structured script library, resulting in multiple commonly used customer service scripts.

[0032] In some embodiments of this application, before training a preset low-rank adaptive model based on the historical service information corresponding to each service target to obtain a first service information prediction model corresponding to each service target, the method further includes: adding identification information of the corresponding service target to each of the historical service information, wherein the identification information is used to distinguish between the platform and the store.

[0033] To achieve the above objectives, a second aspect of the present invention provides a digital employee service system, the system comprising:

[0034] The acquisition module is used to acquire multiple historical service information corresponding to multiple service targets, wherein the service targets are platforms and / or stores, and the historical service information includes at least historical customer service dialogue records, product information, and commonly used customer service scripts.

[0035] The training module is used to train a preset low-rank adaptive model based on the historical service information corresponding to each service target, so as to obtain a first service information prediction model corresponding to each service target.

[0036] A construction module is used to construct a second service information prediction model; the second service information prediction model includes at least a plurality of first service information prediction models set in parallel, and a feature fusion model for fusing the output features corresponding to the plurality of first service information prediction models set in parallel and outputting the service information prediction result.

[0037] The prediction module is used to train the second service information prediction model using all the historical service information of the multiple service targets until the trained second service information prediction model is obtained.

[0038] The response module is used to respond to the target service information and input the target service information into the second service information prediction model to obtain the corresponding target service information prediction result.

[0039] To achieve the above objectives, a third aspect of the present invention provides an electronic device, comprising: at least one control processor and a memory for communicatively connecting to the at least one control processor; the memory stores instructions executable by the at least one control processor, the instructions being executed by the at least one control processor to enable the at least one control processor to perform the above-described digital employee service method.

[0040] To achieve the above objectives, a fourth aspect of the present invention provides a computer-readable storage medium storing computer-executable instructions for causing a computer to perform the above-described digital employee service method.

[0041] This application provides a service method for digital employees, which involves acquiring multiple historical service information corresponding to multiple service targets; training a preset low-rank adaptive model based on the historical service information corresponding to each service target to obtain a first service information prediction model for each service target; constructing a second service information prediction model; training the second service information prediction model using all historical service information of multiple service targets until a fully trained second service information prediction model is obtained; responding to target service information by inputting the target service information into the second service information prediction model to obtain the corresponding target service information prediction result. This method can train independent low-rank adaptive modules for each platform / store, thereby integrating training to promote knowledge sharing and improve service effectiveness, service efficiency, and knowledge consistency.

[0042] It is understood that the beneficial effects of the second to fourth aspects compared with the related technologies are the same as the beneficial effects of the first aspect compared with the related technologies. Please refer to the relevant description in the first aspect above, which will not be repeated here. Attached Figure Description

[0043] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0044] Figure 1 This is a flowchart illustrating a digital employee service method provided in an embodiment of this application;

[0045] Figure 2 This is a schematic diagram of a system framework provided in an embodiment of this application;

[0046] Figure 3 This is a schematic diagram of adaptive integration of a multiple low-rank adaptive model provided in an embodiment of this application;

[0047] Figure 4 This is a schematic diagram of the structure of a digital employee service system provided in an embodiment of this application;

[0048] Figure 5 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0049] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application.

[0050] In the description of this application, the use of terms such as "first," "second," etc., is for the purpose of distinguishing technical features only and should not be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated or the order of the technical features indicated.

[0051] In the description of this application, it should be understood that the orientation descriptions, such as up, down, etc., are based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this application.

[0052] In the description of this application, it should be noted that, unless otherwise explicitly defined, terms such as "setup," "installation," and "connection" should be interpreted broadly, and those skilled in the art can reasonably determine the specific meaning of the above terms in this application in conjunction with the specific content of the technical solution.

[0053] As the service industry evolves towards a multi-platform, multi-store operation model, merchants expand their businesses through this model. The services provided by their digital employees are crucial for improving service efficiency and reducing operating costs. Moreover, different platforms and stores exhibit significant differences in response styles, script standards, and service processes, requiring models to accurately capture the characteristics of each platform.

[0054] In existing technologies, services for digital employees across multiple platforms primarily employ solutions such as training models independently for a single platform or store, uniform fine-tuning of a general-purpose model, or static low-rank adaptation (LoRA) switching to implement digital employee functionality. However, the independent platform model solution is costly to train and maintain and cannot share knowledge, while the uniform fine-tuning solution for a general-purpose model leads to homogenized responses due to the overshadowing of niche platform characteristics and exacerbates the performance degradation in stores with sparse data. Furthermore, the static low-rank adaptation switching solution, due to the inherent defects of isolated module training and static switching, still cannot achieve cross-platform knowledge interaction and transfer, resulting in optimization bottlenecks for stores with sparse data and technical issues of overall knowledge consistency lack, ultimately leading to poor service performance and low service efficiency for digital employees.

[0055] Based on this, embodiments of this application provide a service method, system, device, and storage medium for digital employees, aiming to train independent low-rank adaptive modules for each platform / store, thereby integrating training to promote knowledge sharing and improve service effectiveness, service efficiency, and knowledge consistency.

[0056] The digital employee service method, system, device, and storage medium provided in this application are specifically described through the following embodiments. First, the digital employee service method in this application embodiment is described.

[0057] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence (AI) refers to the theories, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0058] Foundational technologies for artificial intelligence generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies mainly encompass computer vision, robotics, biometrics, speech processing, natural language processing, and machine learning / deep learning.

[0059] The digital employee service method provided in this application relates to the field of artificial intelligence service technology. The digital employee service method provided in this application can be applied to a terminal, a server, or software running on either a terminal or a server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, etc.; the server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms; the software can be an application implementing the digital employee service method, but is not limited to the above forms.

[0060] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0061] It should be noted that in all specific embodiments of this application, when processing data related to user identity or characteristics, such as user information, user behavior data, user historical data, and user location information, user permission or consent is obtained first. Furthermore, the collection, use, and processing of this data comply with relevant laws, regulations, and standards. In addition, when embodiments of this application require access to sensitive personal information of users, separate permission or consent from the user is obtained through pop-ups or redirection to confirmation pages. Only after obtaining the user's separate permission or consent is the necessary user-related data required for the proper functioning of these embodiments acquired.

[0062] Therefore, referring to Figure 1 This application provides a digital employee service method. This method is applied to a central controller, which can be a server, an electronic device, or a mobile terminal, etc. There are no specific limitations here. The method includes the following steps S110 to S150.

[0063] Step S110: Obtain multiple historical service information corresponding to multiple service targets. The service targets are the platform and / or the store. The historical service information includes at least historical customer service dialogue records, product information, and commonly used customer service scripts.

[0064] Step S120: Train the preset low-rank adaptive model based on the historical service information corresponding to each service target to obtain the first service information prediction model corresponding to each service target.

[0065] Step S130: Construct a second service information prediction model; the second service information prediction model includes at least a plurality of parallel first service information prediction models, and a feature fusion model for fusing the output features corresponding to the plurality of parallel first service information prediction models and outputting the service information prediction result.

[0066] Step S140: Train the second service information prediction model using all historical service information of multiple service targets until the trained second service information prediction model is obtained.

[0067] Step S150: Respond to the target service information by inputting the target service information into the second service information prediction model to obtain the corresponding target service information prediction result.

[0068] In this step, we first acquire multiple historical service information items corresponding to various service targets, including historical customer service dialogue records, product information, and commonly used customer service scripts, for subsequent model training and optimization. Service targets can be the objects requiring digital employee services, including service platforms in the service field or stores on those platforms. These can be distinguished using platform unique identifiers or store registration information to clarify the data attribution of different service objects. Historical service information refers to past operational data related to the service targets, specifically including customer service and user dialogue records, product details, and frequently used standardized response templates. By cleaning and structuring this past operational data related to the service targets, we create training samples to capture the personalized service characteristics of different platforms or stores.

[0069] Furthermore, based on the historical service information of each service target, a pre-defined low-rank adaptive model is trained to obtain the first service information prediction model corresponding to each service target. The low-rank adaptive model refers to a machine learning model that efficiently adjusts the parameters of a pre-trained model using low-rank matrix factorization (LMF) technology. This effectively captures the characteristics of various platforms or stores while maintaining the model's lightweight nature. Specifically, the LoRA algorithm can be used to insert a trainable low-rank matrix while freezing the original model parameters. Fine-tuning with a small amount of data is sufficient to adapt to the characteristics of different service targets, reducing the resource consumption of training multiple models.

[0070] Furthermore, a second service information prediction model is constructed, which includes multiple parallel first service information prediction models and a feature fusion model for fusing the output features of the multiple parallel first service information prediction models. This allows the second service information prediction model to retain the characteristics of each platform while achieving cross-platform knowledge sharing.

[0071] Specifically, the first service information prediction model refers to a sub-model trained independently for a single service target. Preferably, it is trained by inputting the historical service information of the service target into a low-rank adaptive model, enabling the model to learn the product knowledge, communication style, and service process unique to that service target, thus forming a targeted prediction capability. The feature fusion model refers to a processing module used to aggregate the feature vectors of multiple sub-models. Specifically, it can use a multi-head attention mechanism or a fully connected network to interactively calculate the hidden layer outputs of parallel sub-models, extract common and differential features across service targets, and generate a comprehensive prediction result. The second service information prediction model refers to an integrated model composed of multiple first service information prediction models connected in parallel. Specifically, it uses a feature fusion layer to jointly analyze the outputs of each sub-model, and integrates cross-platform features using an attention mechanism or weighted summation method to enhance the generalization ability for stores with sparse data.

[0072] Furthermore, the second service information prediction model is trained using historical service information from all service targets until training is complete, achieving the integration and optimization of global knowledge. When target service information is received, it is input into the trained second service information prediction model to obtain the target service information prediction result generated based on the user's current inquiry content, thereby achieving intelligent processing and prediction of new input information.

[0073] In this step, a dynamic fusion multi-level prediction model architecture is constructed. By independently training a low-rank adaptor sub-model for each service target, and then achieving cross-platform knowledge transfer and feature complementarity through a feature fusion layer, the personalized service features of different platforms and stores are preserved. At the same time, the shared parameter mechanism is used to solve the model generalization problem in data-sparse scenarios, while avoiding the resource waste caused by independent training of multiple models. This realizes the intelligent and personalized service of digital employees on multiple platforms, effectively solves the knowledge isolation problem caused by traditional independent model training, and avoids the platform feature overwhelming problem caused by fine-tuning of general large models. In this way, the intelligence level of digital employee services on multiple platforms and user satisfaction are improved.

[0074] In some embodiments, in step S140, the second service information prediction model is trained using all historical service information of multiple service targets until a trained second service information prediction model is obtained, including the following steps S210 to S260:

[0075] Step S210: Use all historical service information of multiple service targets as training data for the first training.

[0076] Step S220: Input the training data from the first training session into the feature fusion model to obtain the first feature output by the feature fusion model;

[0077] Step S230: Input the training data from the first training session into the basic service model to obtain the second feature output by the basic service model;

[0078] Step S240: Input the training data from the first training session into the language service model to obtain the third feature output by the language service model;

[0079] Step S250: Perform a weighted calculation on the first feature, the second feature, and the third feature to obtain the fourth feature;

[0080] Step S260: Use the fourth feature as the training data for the next training of the second service information prediction model.

[0081] In this step, the second service information prediction model also includes a basic service model and a language service model. The basic service model obtains general semantic representation capabilities through training on all historical service information. The language service model uses a pre-trained language model to extract deep semantic features of the text. The feature fusion model is responsible for integrating the platform-specific features of multiple first service information prediction models.

[0082] In this embodiment, all historical service information of multiple service targets is first used as training data for the first training. Then, during the training phase, the training data of the first training is input into the feature fusion model to obtain the first feature output by the second service information prediction model; the training data of the first training is input into the basic service model to obtain the second feature output by the basic service model; and the training data of the first training is input into the language service model to obtain the third feature output by the language service model. The basic service model and the language service model capture general service knowledge and language features respectively, while the feature fusion model integrates the outputs of multiple first service information prediction models, thereby extracting platform-specific features, general semantic features, and language structure features respectively.

[0083] Furthermore, the first, second, and third features are weighted and calculated to obtain the fourth feature. Specifically, in the weighted calculation process, the third feature is used as the query vector and is used in attention calculation with the key vector and value vector of the second feature to generate dynamic weight coefficients, which are then combined with the platform-specific features of the first feature for weighted fusion.

[0084] Furthermore, the fourth feature is used as training data for the next training of the second service information prediction model. In subsequent iterations, the fourth feature carries cross-platform knowledge interaction information, gradually optimizing the basic service model's ability to capture the commonalities and differences of multiple platforms. This realizes knowledge interaction and iterative optimization between models, thereby achieving collaborative training and feature fusion of multiple models. This effectively improves the model's adaptability to different platforms and store characteristics, while maintaining the consistency of overall knowledge. In turn, it enhances the effectiveness and efficiency of digital employee services, especially providing more accurate service predictions for stores with sparse data.

[0085] In some embodiments, in step S250, the first feature, the second feature, and the third feature are weighted and calculated to obtain the fourth feature, including the following steps S310 to S320:

[0086] Step S310: Using the third feature as the query vector and the second feature as the key vector and value vector, perform attention calculation to obtain the calculation result;

[0087] Step S320: Based on the first feature, the second feature, and the calculation results, a weighted sum is performed using preset weight coefficients to obtain the fourth feature.

[0088] In this embodiment, the third feature is first used as the query vector, and the second feature is used as the key vector and value vector, respectively. Attention is then calculated to obtain the result. Preferably, a scaled dot product attention mechanism can be used for the calculation. Then, based on the first feature, the second feature, and the calculation result, a weighted sum is performed using preset weight coefficients to obtain the fourth feature.

[0089] Specifically, weight coefficients can be set for the first feature, the second feature, and the attention calculation results, and then the three are multiplied by their respective weight coefficients and added together to obtain the final fourth feature. The weight coefficients can be adjusted according to actual needs, achieving effective fusion of features from different sources, helping to capture the interrelationships between different features, and enhancing the model's expressive power and generalization performance.

[0090] The following details the construction process of the basic service model:

[0091] Step S410: Preset basic model;

[0092] Step S420: Train the basic model using all historical service information of multiple service targets to obtain the basic service model.

[0093] In this embodiment, the preset base model can be a pre-trained large-scale language model, such as a Generative Pre-trained Transformer (GPT). This base model has a multi-layer Transformer structure, including a self-attention mechanism and a feedforward neural network. The base model is then trained using all historical service information from multiple service targets. This historical service information includes historical customer service dialogue records, product information, and commonly used customer service phrases.

[0094] Preferably, during the training phase, the base model receives historical service information from all service targets as input. Batch gradient descent is used to optimize the loss function, gradually converging the model parameters to a state capable of representing common service patterns across multiple platforms. During training, user intent recognition from dialogue records, key attribute extraction of product information, and response pattern generation are simultaneously optimized, forming a unified semantic encoding space. When the second feature output by the base service model participates in attention calculation as a key and value vector, its cross-platform general knowledge complements the third feature generated by the language service model, thus balancing platform characteristics and common service patterns during weighted fusion and improving the prediction performance for data-sparse stores. Once the trained model parameters are fixed, they serve as a stable feature source for subsequent multi-model fusion, avoiding knowledge conflicts caused by dynamic adjustments.

[0095] In this embodiment, by constructing the aforementioned basic service model, all historical service data can be effectively utilized to train the basic service model, enabling the model to learn common knowledge across different platforms and stores. As a crucial component of the second service information prediction model, the basic service model provides a solid foundation for subsequent personalized predictions, avoiding the high costs of training multiple models individually, while retaining sensitivity to the characteristics of different platforms, thereby improving overall service quality and efficiency.

[0096] The training process of the second service information prediction model is described in detail below:

[0097] Step S510: Freeze the model parameters of the language service model and each of the parallel first service information prediction models;

[0098] Step S520: Through multiple rounds of iteration, train the model parameters of the basic service model and the attention calculation parameters involved in the weighted calculation so that the fourth feature meets the preset target conditions.

[0099] In this embodiment, the model parameters of the language service model and each of the parallel first service information prediction models are first frozen to retain their already trained platform or store characteristic representation capabilities and avoid knowledge forgetting caused by parameter updates. Then, through multiple rounds of iteration, the model parameters of the basic service model and the attention calculation parameters involved in the weighted calculation are trained. Preferably, gradient descent can be used for parameter optimization.

[0100] Furthermore, in each iteration, the loss function is calculated using batch data, and the gradient is calculated through backpropagation. The parameters of the base service model and the attention calculation parameters are then updated until the fourth feature meets the preset target condition. Through continuous iteration, the fourth feature gradually approaches the preset target condition, ultimately achieving a balance between cross-platform knowledge transfer and sparse data optimization. The preset target condition can be that the loss function value is below a certain threshold, or that the performance metrics on the validation set reach the expected level, used to control the timing of training termination.

[0101] In this embodiment, by training the second service information prediction model described above, the basic service model and feature fusion process can be optimized while maintaining the characteristics of each specialized model. This achieves effective integration and transfer of cross-platform knowledge, overcomes the limitations of isolated module training in static multi-low-rank adaptive switching schemes, improves the model's ability to capture features of different platforms and stores, and thus improves the overall effect and efficiency of digital employee services.

[0102] In some embodiments, before obtaining the historical service information corresponding to multiple service targets in step S110, the method further includes the following steps S610 to S620:

[0103] Step S610: Clean the initial historical customer service dialogue records in each historical service information to obtain multiple historical customer service dialogue records containing valid dialogues.

[0104] Step S620: Standardize the initial product information in each historical service information to obtain multiple product information;

[0105] Step S630: Classify and process the customer service scripts in each historical service information to form a structured script library, and obtain multiple commonly used customer service scripts.

[0106] In this embodiment, the initial historical customer service dialogue records in each historical service information are cleaned to obtain multiple historical customer service dialogue records containing valid dialogues, which helps the model capture key information. Preferably, the initial historical customer service dialogue records can be analyzed using natural language processing technology, and the initial historical customer service dialogue records can be cleaned by removing data fragments containing invalid characters, duplicate content, or non-dialogue text, while retaining dialogue content with complete question-and-answer logic.

[0107] Furthermore, the initial product information from each historical service record is standardized to obtain multiple product information entries, facilitating cross-platform and cross-store product comparison and analysis by the model. Preferably, the standardization of initial product information involves unifying naming rules for product names, units of measurement for specifications, and field formats for attribute descriptions. For example, product names, specifications, prices, and other information can be uniformly formatted to ensure that product information from different sources has a consistent structure and representation.

[0108] Furthermore, the customer service scripts in the historical service information are categorized to form a structured script library, resulting in multiple commonly used customer service scripts. These scripts are then classified according to factors such as usage scenarios and purposes, helping the model to more accurately understand and generate responses suitable for different scenarios, such as greetings, explanations, and reassurances. This structured script library facilitates subsequent model training and use. Preferably, the categorized customer service scripts utilize natural language processing technology to perform semantic clustering of the script content, classifying them into preset categories based on consultation type, business scenario, or emotional tendency.

[0109] In some embodiments, before training a preset low-rank adaptive model based on the historical service information corresponding to each service target in step S120 to obtain a first service information prediction model corresponding to each service target, the method further includes: adding identification information of the corresponding service target to each historical service information, wherein the identification information is used to distinguish between the platform and the store.

[0110] In this embodiment, the identification information is added by labeling each historical customer service dialogue record, product information, and commonly used customer service phrases with category tags for the platform or store during the data preprocessing stage. Furthermore, the identification information can be embedded in the dataset in the form of unique codes or category tags.

[0111] Specifically, during model training, identification information and historical service information are jointly input into the low-rank adaptive model, enabling the model to dynamically adjust feature extraction weights based on the identification information. For example, when training the first service information prediction model, the identification information, as part of the input features, guides the model to learn the differentiated service models of the platform and stores.

[0112] Specifically, during the data cleaning, standardization, and classification stages, identification information is appended to the metadata of each historical service data entry. When training a low-rank adaptive model based on historical service information, the model uses this identification information to identify the service target type to which the current data belongs. This strengthens the capture of specific platform or store characteristics during parameter updates, effectively distinguishing historical service information from different platforms and stores. This improves the model's ability to identify the characteristics of each platform and store, helping to generate more personalized and accurate service information predictions for different platforms and stores, enhancing customer service quality and efficiency, and ultimately optimizing the overall service outcome.

[0113] In some embodiments, the algorithm flow of this embodiment mainly includes three aspects: a data acquisition module, a store LoRA training module, and an adaptive LoRA ensemble training module, as detailed below:

[0114] Step 1, such as Figure 2 As shown in Step 1 (red box area), the data collection module mainly collects historical customer service dialogue records, product information, and commonly used customer service scripts from the merchant's various service platform stores (such as store 1 on platform A, store 2 on platform A, store 1 on platform B, etc.).

[0115] Furthermore, the collected data undergoes cleaning (removing dialogues lacking product information), a question-and-answer knowledge base is constructed, and special phrases are filtered (such as transfer phrases and payment reminders). The processed dialogue data is then processed according to OpenAI's format. Subsequently, merchants are configured to collect customer service-related data from each store across multiple designated platforms, resulting in a total of N stores. The dialogue data, i.e. Figure 2 In ,in, Figure 2 In this example, LLM stands for Large Language Model, and a model similar to the Qwen2.5-7B specification is used.

[0116] Step Two, as follows Figure 2 As shown in Step 2 (yellow box area), the LoRA module for each store is trained. Specifically, for each store that needs to maintain its own style and process, the Qwen2.5-7B model is used. When building the training data, it is not necessary to add platform and store information in the generation layer (prompt), and the prompt for each store's data is consistent.

[0117] This step involves learning knowledge about the current store based on the dialogue. The current LoRA module represents the information of the current store, utilizing the dialogue data output in step one. For each store Train a dedicated LoRA module independently to obtain a store-specific LoRA module. .

[0118] Specifically, these LoRA modules are responsible for capturing the unique response styles, common phrases, and specific process requirements of their respective platforms / stores. In this step, each LoRA module is fine-tuned to adjust its parameters during training. The LoRA parameters used are as follows: the target modules for training are the query projection operator (q_proj), output projection operator (o_proj), key projection operator (k_proj), and value projection operator (v_proj), with a rank dimension of 8 and lora_alpha = 16. Here, lora_alpha represents the core parameter of LoRA training, used to adjust the influence of the low-rank LoRA matrix on the output of the base model.

[0119] Step 3, as follows Figure 2 As shown in Step 3 (green box area), the adaptive LoRA integration training module is trained, specifically in the store-specific LoRA module. After independent training is completed, an integrated dataset will be constructed, which is about to... The prompt in the dataset includes platform and store information, and then the datasets are mixed to obtain a new training dataset. ,use These LoRA modules Collaborative and adaptive overall training is performed, enabling different LoRA modules to share and interact with common product knowledge from various vendors during training, while retaining their individual specificities. The result is a fully trained and integrated Qwen3-8B model, which includes the trained... In this step, all LoRA modules are in a frozen state. Additionally, the Qwen3-8B model also includes... This module is in a fine-tunable state.

[0120] like Figure 3 As shown, in the adaptive multi-LoRA module, the hidden state vector is the input vector of each layer in the decoder layer of the large model (transformer). , It is a vector dimension. Based on the LoRA target modules (target_modules) operators in step two, the corresponding four operations on q_proj, o_proj, k_proj, and v_proj are performed as follows: Figure 3 The adaptive LoRA operation is shown.

[0121] Specifically, for input After performing operations on the qwen layers such as q_proj, o_proj, k_proj, and v_proj, the following is obtained: ,at the same time After multiple LoRA modules The module outputs the corresponding data. ,in At the same time, a new LoRA module is built. , to obtain the output vector Subsequently For query vector (query) Attention is calculated using a dot product attention mechanism for the key vector and value vector, resulting in a weighted output of multiple LoRAs. This weighted integration of information from each store is then obtained, as follows: , ,in, This represents the output vector after weighted integration of LoRA information from multiple stores. This represents the activation function. The query vector represents a linear transformation and can be represented by a trainable matrix. Adjustment The expression, This represents a trainable weight matrix used for adjusting... The interaction weights with multiple LoRA outputs, with dimension 1. (Must match the vector dimension). This represents the query vector obtained after processing the "Transformer decoder layer hidden state H" by the q_proj operator of the basic large model (such as Qwen), with dimension 1. , It is the vector dimension. This represents the output set of all store-specific LoRA modules. The dimension that represents the scaling factor.

[0122] in, Figure 3 middle , respectively corresponding Figure 2 Step 3 output , Figure 3 middle correspond Figure 2 In step 3 .

[0123] Furthermore, in order to learn information from mixed stores, this embodiment also adds a LoRA module with the same configuration. ,through Obtain vector The final input vector for the next layer , where the left side of the equation Given the input vector for the next layer, the right side of the equation... This is the hidden state vector of the current layer. In this embodiment, all layer parameters in the qwen structure are frozen and dedicated to a LoRA module. The parameters are frozen. and parameters in the weighted operator It is trainable.

[0124] In this embodiment, by training the LoRA module independently for each platform / store, its unique response style, communication habits, and specific pre-sales and after-sales processes can be accurately captured and maintained, enhancing the authenticity and relevance of the customer experience. Furthermore, the adaptive multi-LoRA integrated training mechanism promotes the interaction and sharing of common product knowledge among LoRA modules of different platforms / stores. LoRA modules in stores with less data can learn more comprehensive and accurate product knowledge from modules with more data, effectively mitigating the problem of poor model performance caused by data sparsity. In addition, integrated training not only improves the performance of individual modules but, more importantly, achieves the unification and optimization of the merchant's overall customer service knowledge system, ensuring consistency of product information across platforms and stores and reducing response errors caused by knowledge fragmentation.

[0125] like Figure 4 As shown in some embodiments of this application, a digital employee service system is provided. The system includes an acquisition module 410, a training module 420, a construction module 430, a prediction module 440, and a response module 450. Specifically:

[0126] The acquisition module 410 is used to acquire multiple historical service information corresponding to multiple service targets, wherein the service targets are platforms and / or stores, and the historical service information includes at least historical customer service dialogue records, product information, and commonly used customer service scripts.

[0127] The training module 420 is used to train a preset low-rank adaptive model based on the historical service information corresponding to each service target, so as to obtain a first service information prediction model corresponding to each service target.

[0128] The construction module 430 is used to construct a second service information prediction model; the second service information prediction model includes at least a plurality of first service information prediction models set in parallel, and a feature fusion model for fusing the output features corresponding to the plurality of first service information prediction models set in parallel and outputting the service information prediction result.

[0129] Prediction module 440 is used to train the second service information prediction model using all the historical service information of the multiple service targets until the trained second service information prediction model is obtained.

[0130] The response module 450 is used to respond to the target service information and input the target service information into the second service information prediction model to obtain the corresponding target service information prediction result.

[0131] It should be noted that the digital employee service system provided in this embodiment and the digital employee service method described above are based on the same inventive concept. Therefore, the relevant content of the digital employee service method described above also applies to the content of the digital employee service system, and will not be repeated here.

[0132] To this end, the system acquires multiple historical service information entries corresponding to multiple service targets; it trains a pre-defined low-rank adaptive model based on the historical service information for each service target to obtain a first service information prediction model for each service target; it then constructs a second service information prediction model; and finally, it trains the second service information prediction model using all historical service information for multiple service targets until a fully trained second service information prediction model is obtained; in response to target service information, the target service information is input into the second service information prediction model to obtain the corresponding target service information prediction result. In this way, independent low-rank adaptive modules can be trained for each platform / store, thereby integrating training to promote knowledge sharing and improve service effectiveness, service efficiency, and knowledge consistency.

[0133] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described digital employee service method.

[0134] like Figure 5 , Figure 5 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of this application. The electronic device includes:

[0135] At least one battery;

[0136] At least one memory;

[0137] At least one processor;

[0138] At least one program;

[0139] The program is stored in memory, and the processor executes at least one program to implement the digital employee service method described above in this disclosure.

[0140] This electronic device can be any smart terminal, including mobile phones, tablets, personal digital assistants (PDAs), and in-vehicle computers.

[0141] The electronic devices according to embodiments of this application will now be described in detail.

[0142] The processor 1600 can be implemented using a general-purpose central processing unit (CPU), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this disclosure.

[0143] The memory 1700 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 1700 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1700 and is called and executed by the processor 1600 to perform a digital employee service method according to an embodiment of this disclosure.

[0144] The input / output interface 1800 is used to implement information input and output.

[0145] The communication interface 1900 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0146] Bus 2000 transmits information between various components of the device (e.g., processor 1600, memory 1700, input / output interface 1800, and communication interface 1900);

[0147] The processor 1600, memory 1700, input / output interface 1800 and communication interface 1900 are connected to each other within the device via bus 2000.

[0148] This disclosure also provides a storage medium, which is a computer-readable storage medium storing computer-executable instructions for causing a computer to perform the aforementioned digital employee service method.

[0149] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0150] The embodiments described in this disclosure are for the purpose of more clearly illustrating the technical solutions of this disclosure and do not constitute a limitation on the technical solutions provided by this disclosure. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by this disclosure are also applicable to similar technical problems.

[0151] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this disclosure, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0152] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0153] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0154] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0155] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0156] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0157] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0158] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0159] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause an electronic device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0160] The above is a detailed description of the preferred embodiments of this application. However, the embodiments of this application are not limited to the above-described implementation methods. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the embodiments of this application. All such equivalent modifications or substitutions are included within the scope defined by the claims of the embodiments of this application.

[0161] The embodiments of this application have been described in detail above with reference to the accompanying drawings. However, this application is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of this application.

Claims

1. A service method for digital employees, characterized in that, The method includes: Acquire multiple historical service information corresponding to multiple service targets, wherein the service targets are platforms and / or stores, and the historical service information includes at least historical customer service dialogue records, product information, and commonly used customer service scripts; Based on the historical service information corresponding to each service target, a preset low-rank adaptive model is trained to obtain a first service information prediction model corresponding to each service target. Construct a second service information prediction model; the second service information prediction model includes at least a plurality of first service information prediction models set in parallel, and a feature fusion model for fusing the output features corresponding to the plurality of first service information prediction models set in parallel and outputting the service information prediction result. The second service information prediction model is trained using all the historical service information of the multiple service targets until the trained second service information prediction model is obtained. In response to the target service information, the target service information is input into the second service information prediction model to obtain the corresponding target service information prediction result.

2. The digital employee service method according to claim 1, characterized in that, The second service information prediction model further includes a basic service model and a language service model. The step of training the second service information prediction model using all the historical service information of the multiple service targets until a fully trained second service information prediction model is obtained includes: Use all the historical service information of the multiple service targets as training data for the first training; The training data from the first training iteration is input into the feature fusion model to obtain the first feature output by the feature fusion model; The training data from the first training session is input into the basic service model to obtain the second feature output by the basic service model; The training data from the first training session is input into the language service model to obtain the third feature output by the language service model. The first feature, the second feature, and the third feature are weighted and calculated to obtain the fourth feature; The fourth feature will be used as training data for the next training of the second service information prediction model.

3. The digital employee service method according to claim 2, characterized in that, The step of weighting the first feature, the second feature, and the third feature to obtain the fourth feature includes: Using the third feature as the query vector and the second feature as the key vector and value vector, attention calculation is performed to obtain the calculation result. Based on the first feature, the second feature, and the calculation result, the fourth feature is obtained by weighted summation using preset weighting coefficients.

4. The digital employee service method according to claim 2, characterized in that, The construction of the basic service model includes: Preset basic model; The basic service model is obtained by training the basic model using all the historical service information of the multiple service targets.

5. The digital employee service method according to claim 2, characterized in that, The training process of the second service information prediction model includes: Freeze the model parameters of the language service model and each of the parallel first service information prediction models; Through multiple rounds of iteration, the model parameters of the basic service model and the attention calculation parameters involved in the weighted calculation are trained so that the fourth feature meets the preset target conditions.

6. The digital employee service method according to claim 1, characterized in that, Before obtaining multiple historical service information corresponding to multiple service targets, the method further includes: The initial historical customer service dialogue records in each of the historical service information are cleaned to obtain multiple historical customer service dialogue records containing valid dialogues. The initial product information in each of the historical service information is standardized to obtain multiple product information; The customer service scripts in the historical service information are classified and processed to form a structured script library, resulting in multiple commonly used customer service scripts.

7. The digital employee service method according to claim 1, characterized in that, Before training a preset low-rank adaptive model based on the historical service information corresponding to each service target to obtain a first service information prediction model corresponding to each service target, the method further includes: adding identification information of the corresponding service target to each of the historical service information, wherein the identification information is used to distinguish between the platform and the store.

8. A digital employee service system, characterized in that, The system includes: The acquisition module is used to acquire multiple historical service information corresponding to multiple service targets, wherein the service targets are platforms and / or stores, and the historical service information includes at least historical customer service dialogue records, product information, and commonly used customer service scripts. The training module is used to train a preset low-rank adaptive model based on the historical service information corresponding to each service target, so as to obtain a first service information prediction model corresponding to each service target. A construction module is used to construct a second service information prediction model; the second service information prediction model includes at least a plurality of first service information prediction models set in parallel, and a feature fusion model for fusing the output features corresponding to the plurality of first service information prediction models set in parallel and outputting the service information prediction result. The prediction module is used to train the second service information prediction model using all the historical service information of the multiple service targets until the trained second service information prediction model is obtained. The response module is used to respond to the target service information and input the target service information into the second service information prediction model to obtain the corresponding target service information prediction result.

9. An electronic device, characterized in that, It includes at least one control processor and a memory for communicatively connecting to the at least one control processor; the memory stores instructions executable by the at least one control processor to enable the at least one control processor to perform a digital employee service method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions for causing a computer to perform a service method for a digital employee as described in any one of claims 1 to 7.

Images