A session processing method and apparatus

Abstract

The application provides a conversation processing method and device, relates to the technical field of artificial intelligence, and is suitable for a customer service scene. Context information of a current conversation of a human customer service and a user and performance state information of a customer service system are acquired, context evaluation features are obtained based on the context information, performance evaluation features are obtained based on the performance state information, the context evaluation features include conversation scene features and user state features, the context evaluation features and the performance evaluation features are fused, a risk rating result for the current conversation is determined, and in response to the risk rating result meeting takeover conditions of intelligent customer service, intelligent customer service is triggered to take over the current conversation, so that reply information of the current consultation is generated.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a conversation processing method and device. Background Technology

[0002] With the rapid development of artificial intelligence technology, intelligent customer service systems have been widely applied in customer service scenarios to improve service efficiency, optimize user experience, and reduce labor costs. Currently, the industry generally adopts a collaborative model of "robot first, human as a backup," which means that user requests are first handled by intelligent customer service (i.e., robots, also known as customer service agents). Only when the intelligent customer service cannot handle the request or the user explicitly requests to be transferred to a human customer service representative will the request be transferred to a human. During this period, robots can be used to assist in the conversation.

[0003] However, in conversations led by human customer service representatives, the human customer service representative usually has to manually click a button to trigger the robot to send a pre-set script to answer the user's questions. This is inflexible, and the human customer service representative still has to deal with a large number of repetitive questions, making it difficult to effectively alleviate their high workload. They are also prone to manually triggering the robot in inappropriate scenarios due to fatigue or negligence, which may lead to misanswers or service risks and reduce the user experience. Summary of the Invention

[0004] In view of the above problems, this application provides the following solution:

[0005] A first aspect of this application provides a session processing method, the method comprising:

[0006] Obtain contextual information of the current conversation between human customer service representatives and users, as well as the performance status information of the customer service system;

[0007] Contextual evaluation features are obtained based on the contextual information, and performance evaluation features are obtained based on the performance status information; the contextual evaluation features include conversation scenario features and user status features.

[0008] By integrating the contextual assessment features and the effectiveness assessment features, a risk rating result is determined for the current session;

[0009] In response to the risk rating result meeting the takeover conditions of the intelligent customer service, the intelligent customer service is triggered to take over the current session to generate a response to the current inquiry.

[0010] In one possible implementation, the process of the intelligent customer service taking over the current session further includes:

[0011] In response to the intelligent customer service interruption takeover condition being met, the intelligent customer service is triggered to stop generating reply information, exit the current session, and send a status prompt information to the human customer service so that the human customer service can take over the current session;

[0012] The interruption takeover conditions include any one of the following:

[0013] The risk rating result is greater than the first risk threshold;

[0014] Received the forced takeover command input by the human customer service representative;

[0015] The context information includes predefined risk content;

[0016] The consultation intent of the current consultation in the context information does not belong to the intent database of the intelligent customer service.

[0017] In one possible implementation: the user status features include at least one of user profile features and user emotion features; the performance evaluation features include at least one of the consultation load features based on the customer service system and the historical consultation resolution rate of the intelligent customer service.

[0018] In one possible implementation, the fusion of the contextual assessment features and the effectiveness assessment features to determine the risk rating result for the current session includes:

[0019] The context risk rating result is obtained by weighting and calculating the various features included in the context assessment features;

[0020] The performance evaluation characteristics are quantified to obtain the corresponding performance risk coefficients;

[0021] Based on the contextual risk rating result and the effectiveness risk coefficient, a risk rating result for the current session is generated.

[0022] In one possible implementation, the intelligent customer service is triggered to take over the current session in response to the risk rating result meeting the takeover conditions, including any of the following:

[0023] In response to the risk rating result being less than the second risk threshold, the intelligent customer service is triggered to automatically take over the current session;

[0024] In response to the risk rating result being between the first risk threshold and the second risk threshold, a takeover confirmation request is sent to the human customer service representative; in response to the takeover confirmation instruction, the intelligent customer service representative is triggered to take over the current session.

[0025] Wherein, the first risk threshold is greater than the second risk threshold. In one possible implementation, obtaining session scenario features, user emotion features, and user profile features based on the context information includes:

[0026] Based on the current consultation and historical dialogue in the context information, relevant domain prior knowledge is retrieved;

[0027] Based on the prior knowledge of the domain and the predefined first prompt word template, the intent of the current consultation is identified through the first model to determine the scenario risk level to which the corresponding consultation intent belongs, so as to obtain the conversation scenario characteristics of the current session;

[0028] Based on the current consultation and the historical dialogue, emotion recognition is performed through the first model to obtain the user's emotional characteristics;

[0029] In response to the context information containing the user identifier of the current session, the user's identity tag in the customer service system is obtained from the customer data platform to determine the user profile features that match the identity tag;

[0030] The first prompt word template is dynamically updated based on updated domain prior knowledge.

[0031] One possible implementation also includes:

[0032] In response to the customer service system's inquiry load indicator indicating that the number of users waiting for a reply exceeds the sorting threshold, or the current inquiry is a high-concurrency transaction, the corresponding performance risk coefficient is reduced; or,

[0033] In response to the historical consultation resolution rate of the intelligent customer service being less than the resolution threshold, the corresponding efficiency risk coefficient is increased; or,

[0034] Based on the intelligent customer service's historical dialogues and user feedback service evaluations, at least one of the following will be dynamically adjusted:

[0035] Model parameters used to calculate the context evaluation features;

[0036] The weight allocation of various features in the context evaluation features;

[0037] Model parameters used to calculate the user's emotional characteristics;

[0038] The mapping relationship between predefined consultation intentions and scenario risk levels.

[0039] In one possible implementation, obtaining contextual assessment features based on the contextual information, obtaining performance assessment features based on the performance status information, and fusing the contextual assessment features and the performance assessment features to determine the risk rating for the current session is based on a risk rating strategy.

[0040] The risk rating strategy includes rating each consultation input by the user to determine the risk rating result for the current consultation in the current session.

[0041] The risk assessment strategy can be dynamically adjusted based on at least one of the user's input status information, system cache data, and current session status to change the frequency or method of risk rating execution for the current session.

[0042] In one possible implementation, the intelligent customer service agent acts as the main intelligent agent, interacting with a group of sub-intelligent agents to generate response information for the current inquiry, including:

[0043] The context information is sent to the main intelligent agent for semantic understanding, and the main intelligent agent distributes the generated multiple sub-tasks to the corresponding sub-intelligent agents for execution, thereby obtaining the corresponding execution results.

[0044] The main intelligent agent outputs the response information for the current inquiry based on the execution result and the second prompt word template;

[0045] The second prompt word template is learned by the main agent or the corresponding sub-agent based on the historical response information of the human customer service representative, and is used to constrain the output attributes of the response information, including language style.

[0046] A second aspect of this application provides a session processing device, comprising:

[0047] At least one memory, and a computer program stored in the memory;

[0048] At least one processing device executes the computer program to implement the conversation processing method provided in the first aspect of this application by interacting with human and intelligent customer service representatives of the customer service system. Attached Figure Description

[0049] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and the originals and elements are not necessarily drawn to scale.

[0050] Figure 1 This is a schematic diagram of the hardware structure of the session processing device proposed in Embodiment 1 of this application;

[0051] Figure 2 This is a schematic diagram of the hardware structure of the session processing device proposed in Embodiment 2 of this application;

[0052] Figure 3 This is a flowchart illustrating the session processing method proposed in Embodiment 1 of this application;

[0053] Figure 4 This is a flowchart illustrating the session processing method proposed in Embodiment 2 of this application;

[0054] Figure 5 This is a schematic flowchart of the session processing method proposed in Embodiment 3 of this application;

[0055] Figure 6 This is a flowchart illustrating the session processing method proposed in Embodiment 4 of this application;

[0056] Figure 7 This is a schematic diagram illustrating the system architecture applicable to the session processing method proposed in the embodiments of this application;

[0057] Figure 8 This is a schematic diagram of the structure of the session processing device proposed in the embodiments of this application. Detailed Implementation

[0058] The embodiments of this application are described below with reference to the accompanying drawings. The terminology used in the implementation section of this application is only for explaining specific embodiments and is not intended to limit the application. The embodiments of this application are described below with reference to the accompanying drawings. It will be understood by those skilled in the art that, with the development of technology and the emergence of new scenarios, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0059] The terms "first," "second," etc., used throughout this application and in the foregoing figures are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms are interchangeable where appropriate; this is merely a way of distinguishing objects with the same attributes in the embodiments of this application. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, so that a process, method, system, product, or apparatus that comprises a series of units is not necessarily limited to those units, but may include other units not explicitly listed or inherent to those processes, methods, products, or apparatuses.

[0060] It is understood that before using the technical solutions disclosed in the embodiments of this application, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in this application in an appropriate manner in accordance with relevant laws and regulations, and user authorization should be obtained. For example, in response to receiving a user's active request, a pop-up window may be used, and a textual prompt message may be presented in the pop-up window to clearly inform the user that the requested operation will require the acquisition and use of the user's personal information. The user can choose whether to provide personal information to the electronic device, application, server, or storage medium or other software or hardware that performs the operation of this application based on the prompt message. This application does not limit the prompt message and the method of user authorization implementation. The data involved in this technical solution (including but not limited to the data itself, the acquisition or use of the data) shall comply with the requirements of relevant laws and regulations and related provisions (such as corporate regulations).

[0061] Furthermore, the model involved in this application can be a general AI (Artificial Intelligence) model, employing, but not limited to, the Transformer or its architectural variants (such as using only an encoder-only / decoder-only architecture, encoder-decoder architecture, or MoE (Mixture of Experts, a neural network architecture) or other basic architectures). It learns the features and patterns of natural language by training on large amounts of diverse data, thereby enabling it to understand and generate natural language. It typically has hundreds of millions to trillions of model parameters (model parameters are variables that control the behavior of the target model) and is capable of capturing complex relationships and patterns in natural language.

[0062] The AI ​​models can include, but are not limited to, generative models and generative language models (GLMs). For example, one or more of the following: large language model (LLM), GPT (Generative Pre-trained Transformer) series models, T5 (Text to Text Transfer Transformer) model, BERT (Bidirectional Encoder Representation from Transformers) series models, large visual models, and multimodal large models.

[0063] Depending on actual needs, the models involved in the embodiments of this application can also be expert large models that are fine-tuned from general AI models based on actual business requirements. For example, proprietary models (such as domain models) trained on sample datasets of general AI models (pre-trained models) in specific scenarios (such as professional scenarios in vertical fields such as medical, financial, legal, biological, or remote sensing, or general scenarios). The models involved in this application can also be lightweight models that are compressed from general AI models or expert large models through lightweight methods such as quantization, knowledge distillation, or pruning, in order to better meet the needs of edge deployment with limited computing resources and improve data security. This application does not limit the type of model involved in the context and can be flexibly determined according to actual business needs.

[0064] In a conversation led by human customer service, this application proposes that before responding to a user's current inquiry, the system first performs a risk assessment on the current conversation between the human customer service representative and the user. When the risk assessment result meets the conditions for intelligent customer service takeover, the system automatically triggers intelligent customer service to take over the current conversation and generate a response to the current inquiry. This effectively alleviates the high workload of human customer service representatives and shortens user waiting time. Compared to human customer service representatives manually triggering intelligent customer service to send preset scripts, this improves the flexibility of responding to user inquiries and eliminates the need for human customer service representatives to rely on experience to determine whether to switch to intelligent customer service. This application integrates contextual assessment features obtained based on the contextual information of the current conversation and performance assessment features obtained based on the performance status information of the customer service system to automatically and accurately determine the risk assessment result of the current conversation. This improves the accuracy and reliability of risk assessment and avoids problems such as misanswers or service risks caused by human customer service representatives switching to intelligent customer service in inappropriate scenarios due to fatigue, negligence, or lack of knowledge. This effectively reduces erroneous responses and security risks, and improves service efficiency and experience. The conversation processing device proposed in the embodiments of this application will be described below with reference to the accompanying drawings.

[0065] Reference Figure 1This is a schematic diagram of the hardware structure of the session processing device proposed in Embodiment 1 of this application. This session processing device is applicable to customer service systems in various fields and may include at least one electronic device. The electronic device may include device nodes in systems such as servers or cloud platforms, and may also include terminal devices with data processing capabilities, such as smartphones, laptops, Augmented Reality (AR) / Virtual Reality (VR) devices, in-vehicle terminals, smart healthcare / transportation equipment, smart home devices, or other business terminals, as needed. This application does not limit the product form of the session processing device; it can be determined based on factors such as the customer service system architecture in the corresponding customer service scenario. The session processing device of this application can be modularly integrated / embedded or externally accessed to customer service systems to implement the session processing method proposed in this application, without any limitations.

[0066] Based on the above analysis, such as Figure 1 As shown, the session processing device proposed in this embodiment may include, but is not limited to, at least one memory 110, a computer program 120 stored in the memory 110, and at least one processing device 130. The processing device 130 can execute the computer program 120 and, through interaction with the human customer service representative 140 and the intelligent customer service representative 150 of the customer service system, implement the relevant steps of the session processing method proposed in this application. The implementation process of this session processing method can be referred to the description in the corresponding section of the method embodiment below; it will not be detailed here.

[0067] Among them, the human customer service representative 140 corresponds to the agent end of the customer service system, also known as an agent. After a user initiates an inquiry to the customer service system through an application, webpage, or telephone, the customer service system acts as an "intelligent dispatcher," assigning the inquiry to the appropriate agent according to a preset routing strategy. For example, after recognizing the inquiry intent, it matches the agent's skill tags and busy status to assign each inquiry to the most suitable agent. If the agent's end is busy, a queuing mechanism can be used to wait for the agent to take over in order. The agent clicks the takeover button or the system automatically connects, establishing a conversation and initiating a real-time dialogue. However, this routing process is not limited to this one and can be flexibly adjusted according to the configuration requirements of the actual customer service scenario.

[0068] Intelligent customer service 150 can refer to an intelligent program (intelligent assistant / intelligent agent) or system that can simulate human customer service to interact with users in natural language and autonomously complete customer service tasks such as information provision, consultation response, and business processing. This application does not limit its form of expression, such as a dialogue robot based on a pre-trained large language model, which can be an intelligent dialogue system that can understand complex contexts and generate natural and fluent responses; a question-and-answer system based on a retrieval model, which can be a question-and-answer robot that generates standardized responses by recognizing the intent of user inquiries and combining knowledge base retrieval; a rule and process robot, which can be an interactive voice response system or text robot that guides users to complete tasks such as information collection and business processing based on preset business processes and dialogue templates; and a multimodal intelligent customer service, which can be an intelligent service system that supports multimodal input and output such as voice, image, and video on the basis of text interaction, etc., without limitation.

[0069] In practical applications, the intelligent customer service system 150 can be deployed independently within a customer service system as an assistant or replacement for human customer service representatives. Alternatively, it can collaborate with human customer service representatives, dynamically deciding whether to allow the intelligent customer service representative to take over (i.e., the process of the intelligent customer service representative gaining control of the conversation and generating response information) or relinquish control to the human customer service representative in a human-machine collaborative mode, according to the method proposed in this application. It should be understood that with the development of artificial intelligence technology, any automated system / program capable of replacing human customer service representatives in performing at least some of the customer service functions can be considered as the intelligent customer service system involved in this application; the form of the intelligent customer service system in this application is not limited.

[0070] In one possible implementation, let's take the form of intelligent customer service as an intelligent agent as an example. Figure 2 As shown, the intelligent customer service can act as a master agent, interacting with a group of work agents. The master agent can communicate with the work agents via A2A (Agent-to-Agent, direct communication between agents) or an agent network protocol (a protocol specifically designed for communication between agents, defining the rules, formats, and processes to ensure correct and efficient communication and collaborative work). This allows the master agent to generate responses for the current session. The work agents are responsible for performing specific domain tasks. Different work agents specialize in specific domains such as querying, ticket service, and knowledge-based question answering. These work agents can be called retrieval agents, ticket agents, and knowledge agents. In practical applications, tasks distributed to different work agents can be executed in parallel or sequentially according to a plan. This application does not restrict the communication methods between different work agents.

[0071] In customer service scenarios, the main intelligent agent coordinates the work of multiple sub-agents. It can predict tasks generated in the current session based on AI models (such as LLM) and distribute these tasks to the appropriate sub-agents via the aforementioned communication methods. The main agent then executes the assigned tasks and receives status updates and result feedback from the sub-agents, improving the response speed and processing capacity of the customer service system to enhance user experience. In practical applications, the main intelligent agent can provide a standard API (Application Programming Interface) to interface with existing customer service systems. During this interface process, real-time updates and sharing of data such as user information, consultation records, and solutions can be achieved, ensuring that the intelligent customer service system can obtain the latest user information and consultation history, thereby providing more personalized services. The implementation process is not detailed in this application.

[0072] Optionally, the intelligent customer service system can employ a redundant design, deploying multiple intelligent customer service instances to ensure that if one instance malfunctions, it can promptly switch to another instance to continue providing service. The primary intelligent agent can have a built-in heartbeat detection mechanism, periodically sending heartbeat requests to determine its operational status. If it fails to respond to the heartbeat request within a specified time, it can be determined that the intelligent customer service is unavailable, and appropriate measures can be taken, such as switching to a backup customer service system or human customer service, or prompting the user to try again later.

[0073] The processing device 130 may include an AI-Driver composed of a large language model and a dedicated reinforcement learning algorithm. It may be a logical module / software module that executes the conversation processing method proposed in this application, or it may be called a decision controller, to provide the application services proposed in this application to an existing customer service system (this application does not limit its architecture and field). In this way, during the conversation, an active trigger takeover / exit mode can be adopted to trigger the intelligent agent to take over / stop the current conversation. The implementation process is not described in detail in this application.

[0074] It should be noted that the customer service system can also adopt a human customer service authorization mode to take over the current conversation and generate a response to the current inquiry. That is, after the current inquiry is connected to the customer service system, a human customer service representative can determine that the user's current inquiry is a repetitive, standardized question and does not meet the conditions for intelligent customer service to take over. They can then activate the intelligent takeover mode with a single click in the system interface, sending the context information of the current conversation to the main intelligent agent to generate a response to the current inquiry. When the human customer service representative is idle or detects risks in the content generated by the intelligent agent, they can click the "Stop" button at any time to forcibly interrupt the running main and sub-intelligent agents, allowing the human customer service representative to smoothly take over the current conversation. This application does not detail the switching routing process between intelligent and human customer service.

[0075] Optionally, the processing device 130 may also run an intelligent agent to enable dialogue between the intelligent customer service representative and the user. For example, the first processor included in the processing device 130 may run a main intelligent agent to implement the configured functions of the intelligent customer service representative; the implementation process will not be detailed in this embodiment. In some embodiments, the aforementioned intelligent agent may also run on other device nodes of the customer service system. The main intelligent agent and different sub-intelligent agents may be integrated and run on one device node, or they may be deployed and run on different device nodes, depending on the actual scenario requirements; this application does not impose any limitations on this.

[0076] During the operation of any of the aforementioned intelligent agents, one or more models can be invoked to perform corresponding tasks through interface calls or other interactive methods. Of course, if an intelligent agent has one or more models, that is, if the model is part of the intelligent agent, an appropriate model can be launched to perform tasks as needed during the operation of the intelligent agent; this application does not impose any restrictions on this.

[0077] Furthermore, the session processing device may also include input / output ports to receive inquiries entered by users through terminal devices when accessing the customer service system, and output responses generated by intelligent customer service or provided by human customer service representatives. This application does not limit the type or number of ports. Optionally, the memory 110 or other storage devices of the customer service system may also be used to store session records, such as logs or other forms of structured storage, to record the inputs, outputs, decision-making processes, model versions, and response times of various intelligent agents and decision controllers. A visual viewing interface may also be provided, supporting queries based on conditions such as session ID, user ID, and risk level results. This application does not limit the implementation method.

[0078] Optionally, the conversation processing device may also include components for analyzing conversation records, summarizing cases, and further extracting SOPs (Standard Operating Procedures) from successful cases, i.e., extracting effective solution paths and scripts. For example, it may use a quality scoring model to select optimized learning samples and training sets for the main agent to form an optimization loop, continuously optimizing the main agent's dialogue capabilities and improving the accuracy and efficiency of intelligent customer service responses. Similarly, other models or logic modules running on the processing device 130 can also be optimized through this self-learning method to improve overall conversation processing capabilities. The implementation process is not detailed in this application.

[0079] It should be understood that, Figure 1 and 2 The structure of the session processing device shown does not constitute a limitation on the system of the embodiments of this application. In practical applications, the session processing device may include more than Figure 1 and Figure 2The number of components shown, or combinations thereof, are not detailed in this application. Optionally, the components of the customer service system and session processing equipment, such as the AI-Driver, main intelligent agent, sub-intelligent agent group, and self-learning module mentioned above, can be modularly developed and deployed using microservices and containerization (Docker, Kubernetes) technologies to obtain a modular system. This greatly enhances the maintainability and flexibility of the entire system, achieving high cohesion and low coupling, as each module can be independently developed, deployed, expanded, and upgraded. Resources can also be dynamically allocated based on business load, such as automatically scaling up "query-type" sub-intelligent agent instances during peak consultation periods to reduce front-end response bottlenecks and achieve elastic scaling. Furthermore, load balancing, heartbeat detection, and / or failover mechanisms can ensure that the failure of a single service instance does not affect the operation of the entire system, achieving high availability.

[0080] Based on the session processing device described in the above embodiments, the session processing method proposed in this application will be described in detail below from the processing device side with reference to the accompanying drawings.

[0081] Reference Figure 3 This is a flowchart illustrating the session processing method proposed in Embodiment 1 of this application. Combined with the hardware structure of the session processing device described above, the method proposed in this application can be applied to a session processing device. The processing unit therein interacts with the human customer service representative (agent) and the intelligent customer service representative (main agent) of the customer service system to execute the method of this application. Figure 3 As shown, the session processing method executed by the processing device may include:

[0082] Step S31: Obtain the context information of the current conversation between the human customer service representative and the user, as well as the performance status information of the customer service system;

[0083] Based on the above description of the customer service system, a user accesses the customer service system using a terminal device and uploads the current inquiry (Query) directly collected or preprocessed by the terminal device to the customer service system. According to the customer service system's own allocation strategy, the task for the current session is assigned to a suitable agent (human customer service). Before the task is actually taken over by a human customer service (i.e., in a state of waiting to be taken over), the execution of the method in this application can be triggered directly.

[0084] The current consultation can be expressed in the user's natural language, which can be obtained by the terminal device responding to the user's input operation on at least one input component (such as an audio collector, touch screen, finger, stylus, mouse and keyboard or joystick, etc.). It can be natural language input by the user in the form of voice, such as voice information input by the user through any input interface or in any state of the terminal device; it can also be natural language in the form of text, such as natural language in the form of text input (text information) in the input box displayed on the interactive interface after waking up the intelligent program (such as intelligent agent or intelligent assistant, etc.) of the terminal device; of course, the current consultation can also be natural language in the form of text converted from the collected voice signal, etc. This application does not limit the information form of the consultation and its input implementation method.

[0085] In practical applications of customer service systems, multiple users often submit inquiries simultaneously or within a short period. When agents are busy, they can queue for responses. Each user's inquiry is assigned to a human agent, establishing a session between the agent and the user. This session can be configured with a unique identifier (such as a session ID) to distinguish different user sessions during subsequent processing and avoid response errors. This application embodiment only illustrates the current session established between any single user and a human agent.

[0086] During a live customer service conversation, the context information between the customer service representative and the user can be obtained in real time through methods such as monitoring or responding to trigger requests. This context information may include, but is not limited to, the user's latest input query and previous rounds of dialogue records (including dialogue content, timestamps, or other attribute tags, which can be referred to as the history of the dialogue to understand the background of the current inquiry and the flow of the conversation). Depending on actual needs, it may also include other dimensions of data such as the duration of the current conversation and user profile tags (such as VIP level, historical complaint rate, etc.). This application does not restrict the multi-dimensional data content included in the context information or the methods of obtaining it, such as obtaining context information from the customer service system's conversation cache / database or real-time dialogue stream (such as multimodal data such as text, voice, images, or video) / message queue, etc. It should be understood that as the number of dialogue rounds between the live or intelligent customer service representative and the user increases, the context information of the current conversation will be dynamically updated; the implementation process is not detailed in this application.

[0087] In addition, this application can also obtain the performance status information of the customer service system by calling monitoring interfaces or data services. This performance status information can reflect or characterize the overall or partial operating status and performance of the customer service system. Therefore, the performance status information may include, but is not limited to, the consultation load status monitored in real time by the customer service system, such as at least one load indicator such as the total number of online customer service sessions / number of queued users / busyness level, average waiting / response time, agent occupancy rate, and peak concurrent sessions, as well as the historical resolution rate information of intelligent customer service for user consultations, such as querying the average problem resolution rate of such scenarios or such users in the past preset time window from the historical work order database based on the scenario or user profile tag (user identifier) ​​initially identified in the current session. This application does not limit the content of the performance status information or the method of obtaining it. It is understood that as time goes by, some of the established user sessions may have ended, and new user sessions may have been established, thus causing the performance status information of the customer service system to change dynamically. This application can periodically or in real time monitor or respond to trigger requests to obtain performance status information.

[0088] For example, suppose a user is inquiring about after-sales service for a product. In previous conversations, the user mentioned the product purchase date, the specific symptoms of the malfunction, etc., which are all contextual information. Meanwhile, the customer service system's average response time to a simple inquiry is 2 seconds, the number of simultaneous sessions being handled by the system is 50, and there are 3 sessions being handled by human agents. These are all performance status information, but not limited to these.

[0089] As can be seen, this application, in the active trigger takeover / exit mode, obtains contextual information and performance status information through methods including but not limited to those described above. This provides a real-time, multi-source data foundation for subsequent risk rating of the corresponding session, achieving synchronous perception of the status of a single session and the overall system. However, the information content is not limited to that listed in this application and can be flexibly adjusted according to the business requirements of the actual customer service scenario. In some customer service scenarios, contextual information or performance status information may be used as the data foundation for the current session risk rating, or it may be combined with information from other dimensions to form the data foundation for the current session risk rating, etc., which will not be detailed here.

[0090] Step S32: Obtain contextual evaluation features based on contextual information and obtain performance evaluation features based on performance status information; contextual evaluation features include conversation scene features and user status features;

[0091] Before performing risk rating on the current session, raw data such as contextual information and performance status information can be transformed into computable structured feature vectors. For the contextual information of the current session, semantic analysis or business keyword extraction can be performed using natural language processing techniques (such as intent recognition based on large language models, retrieving relevant knowledge fragments from domain knowledge bases, etc.) to identify the business category to which the current session belongs (such as bill inquiry, complaint handling, business process consultation, or casual conversation, etc.) and the progress stage of the current session (such as the current consultation has just been initiated, is under investigation, or is about to end, etc.). This allows for scenario risk rating of the current consultation, yielding corresponding session scenario features. This application does not restrict the implementation method of obtaining session scenario features from contextual information; different implementation methods can be adopted for different categories of session scenario features.

[0092] Furthermore, this application can also identify the user's current state information from contextual information. This can include one or more of the user's emotional state / polarity (which can be determined by analyzing tone words and emotional tendency words in the user's input of each consultation, such as positive, neutral, negative / angry, anxious, calm, etc., and can also include intensity values ​​obtained based on behavioral data such as user waiting time and number of repeated questions), user knowledge level (also known as user profile, which can be determined based on the user's understanding of the consultation object, such as whether they can accurately describe the object's problem, etc.), thereby generating vectorized user state features. This application can determine the feature vector obtained by fusing (e.g., splicing) multi-dimensional features such as conversation scene features and user state features obtained based on contextual information as contextual evaluation features. The feature dimensions included are not limited to those listed above.

[0093] Meanwhile, this application can also map the efficiency status information of the customer service system into efficiency evaluation features. For example, "queue size" and "agent utilization rate" can be converted into a system busy index; "average response time" can be converted into a service efficiency index, etc., which can be efficiency values ​​between 0 and 1. This efficiency evaluation feature reflects whether the current agent has the capacity to handle the current conversation, whether there is an urgent need for intelligent customer service assistance to alleviate pressure, and whether the intelligent customer service is capable of resolving the current inquiry to improve the utilization rate and experience of intelligent customer service taking over the current conversation. This application does not limit the method for extracting efficiency evaluation features.

[0094] Step S33: Integrate contextual assessment features and efficacy assessment features to determine the risk rating result for the current conversation;

[0095] In this embodiment, the multi-dimensional feature vector obtained at the moment can be fused using a predefined fusion strategy. For example, after concatenating or weighting the context assessment features and the effectiveness assessment features, the data can be input into a pre-trained decision model (also known as a risk assessment model, which can be a rule-based decision tree model, a neural network-based classification model, or a weighted scoring function, etc. This application does not limit the model form) to combine the risk of the current conversation content with the risk of the system effectiveness for a comprehensive evaluation, generating a unified and quantitative risk rating result for the current conversation (which can be any expression such as risk score, risk ratio, or risk level, etc., this application does not limit this). This result can serve as a key decision-making basis for deciding whether the intelligent customer service should take over the current conversation. This application does not limit the implementation method of step S33.

[0096] In determining the aforementioned risk rating results, corresponding weights can be configured for various comprehensive features. These weights can be determined and optimized based on actual application scenarios, historical dialogues, and user service evaluations to control the impact of different types of features on the risk rating of the current session. Features with higher impact can be assigned larger weights, while features with lower impact can be assigned relatively smaller weights. This application does not impose restrictions on the weight values ​​corresponding to each type of feature. It is understood that as knowledge of customer service scenarios evolves, and different users or enterprises change their standards for classifying risk scenarios, the impact of the same type of feature used to determine the risk rating results may change. Detecting this change allows for dynamic adjustment of the weights assigned to that type of feature in the decision-making model to improve the accuracy and reliability of risk rating.

[0097] As the above analysis shows, the risk rating result can characterize the degree of risk if the current conversation were taken over by intelligent customer service. It can be determined according to the risk level classification standards for customer service scenarios, such as low risk, medium risk, and high risk, with different risk scores / proportion ranges (multiple threshold intervals) corresponding to different risk levels. Different strategies are then implemented to ensure the service security and consultation response efficiency of the current conversation, without any restrictions. For example, if the contextual features indicate "user anger" and the efficiency assessment features indicate "extremely slow agent response," the determined risk rating result could be high risk; if the contextual features indicate "simple consultation" and the efficiency assessment features indicate "system overload," the determined risk rating result might be medium risk, and so on.

[0098] Step S34: In response to the risk rating result meeting the takeover conditions of the intelligent customer service, the intelligent customer service is triggered to take over the current session to generate a response to the current inquiry.

[0099] In this embodiment, conditions for intelligent customer service to take over a session (denoted as takeover conditions) can be pre-configured according to business needs and actual application scenarios. The pre-configured takeover conditions can be adaptively adjusted for different risk rating results to determine whether intelligent customer service is suitable to take over the current session by comparing the two. Optionally, multiple threshold ranges can be set for the risk rating results, each corresponding to a different decision action. At least one threshold range corresponding to the intelligent customer service takeover conditions is determined, and the decision on whether to allow intelligent customer service to take over the current session is made by detecting whether the currently determined risk rating result falls within that threshold range. This application does not limit the values ​​of each threshold range and their corresponding decision actions; they can be determined as needed. Furthermore, as knowledge of the customer service scenario changes, and different users or enterprises change their risk scenario classification standards, the threshold ranges can be dynamically adjusted. The implementation process is not detailed in this embodiment.

[0100] Therefore, it can be seen that when the dynamically determined risk rating result meets the takeover conditions of the intelligent customer service, and the current conversation belongs to a low-risk scenario, if the risk rating result is in the lowest threshold range, it can be considered that the intelligent customer service will take over the current conversation with virtually no risk of misanswers or security issues. This will automatically trigger the intelligent customer service to take over the current conversation, achieving seamless switching between the conversation channels (the conversation channels corresponding to human customer service and the intelligent customer service). If the current conversation belongs to a medium-risk scenario, if the risk rating result is in the higher or lower threshold range, it indicates that there is a certain risk in the intelligent customer service taking over the current conversation. This can be triggered after confirmation by human customer service, but it is not limited to this.

[0101] After triggering the intelligent customer service to take over the current session, context information can be transmitted to the intelligent customer service (main agent) to generate a response to the current inquiry based on the context information and provide it to the user, or to recommend response suggestions to the human customer service for their adoption. It can be seen that before the response information is sent to the user, the content of the response information can be displayed on the human customer service's interactive interface in a streaming output form, which can be monitored, edited or intercepted in real time, to further ensure security and response quality. Through the human-machine collaborative security buffer mechanism, the takeover process of the current session can be ensured to be both efficient and controllable, but it is not limited to this.

[0102] Optionally, the main agent can invoke a large model to perform semantic understanding of the contextual information and call appropriate sub-agents to generate response information. For example, the main agent can use prompt word engineering to understand the current inquiry, historical dialogues, and the user's long / short-term memory information, rewrite the request content of the current inquiry, maintain dialogue coherence, and improve the accuracy of subsequent intent recognition. This application does not restrict the implementation method of how the main agent generates the response information for the current inquiry.

[0103] In some embodiments, during the operation of the customer service system, a self-learning module can learn from each conversation, continuously optimize the accuracy of decisions on taking over / stopping the conversation and the dialogue flow of the main intelligent agent, forming a virtuous cycle of becoming smarter with use, and better meeting personalized service needs. This application does not describe the self-learning implementation process in detail.

[0104] In summary, in user conversations dominated by human customer service representatives, multi-dimensional risk assessment is conducted during the routing process before the human representative takes over the current conversation. This assessment includes evaluating the contextual risk of the conversation and the current system efficiency of the customer service system. This allows for more intelligent and reliable decisions based solely on the conversation content, avoiding misjudgments or omissions caused by single-dimensional judgments and improving the comprehensiveness and accuracy of conversation risk assessment. Thus, even if a consultation is simple (low contextual risk), if the intelligent customer service system is currently under extremely high load (high efficiency risk), intelligent agent takeover may be temporarily postponed, allowing for a more rational allocation of limited human and intelligent resources and improving overall system throughput. However, in cases of low efficiency risk, intelligent customer service can promptly take over parts of the conversation, reducing the service pressure on human customer service representatives, optimizing human resource allocation, and ensuring the stable operation of the customer service system.

[0105] Furthermore, compared to traditional customer service systems where a human customer service representative needs to initiate a conversation transfer to allow the AI ​​customer service to take over the current conversation during a dialogue with a human supervisor, this application constructs a proactive and preventative secure takeover mechanism. Through pre-emptive risk assessment, it sets a higher confirmation threshold for AI customer service takeover when potential risks (such as escalating user emotions or involvement in sensitive operations) emerge. This prevents service escalation or security incidents caused by misinterpretations from the AI ​​customer service, ensuring service quality and security. Therefore, this secure takeover mechanism is equivalent to an "airbag," effectively preventing the AI ​​customer service from forcibly responding in inappropriate scenarios, reducing the risk of misoperation, and improving service security.

[0106] Furthermore, when the conditions for intelligent customer service takeover are met (such as low risk), timely and automatic intelligent customer service takeover can quickly respond to user needs, fill the service gaps caused by untimely responses from human customer service, and this seamless switching of conversation channels or intelligent assistance mechanism shortens user waiting time, effectively reduces the complaint rate caused by service delays, improves overall service quality and user satisfaction, and frees human customer service from a large number of simple and repetitive questions and answers (which are transferred to intelligent customer service for response), allowing them to focus more on handling complex and emotional needs, thereby increasing the number of conversations handled per person. Depending on the consultation distribution of different industries and businesses, the number of conversations handled per person is expected to increase by 15%-50% (but not limited to this) or more, thus allowing human customer service to focus on high-value issues, alleviating the reception pressure during peak periods, and reducing the rigid demand for human customer service caused by factors such as holidays and shopping festivals.

[0107] In some embodiments, based on the above description of user state characteristics, user state characteristics can be used to characterize a user's personalized attributes and psychological state in the current session. Therefore, user state characteristics can include at least one of user profile characteristics and user emotional characteristics. User profile characteristics can be static or dynamic attribute tags extracted from historical user data, reflecting the user's long-term attributes and historical behavioral patterns to achieve differentiated service strategies. These can include, but are not limited to: user registration information, membership level (e.g., ordinary user, VIP user), historical consultation records, historical complaint frequency, spending power, etc., which can be stored in external data sources such as the enterprise's customer data platform (e.g., CRM (Customer Relationship Management) system, membership management system, etc.) and can exist in the form of structured tags; there are no restrictions on this. This application uses user screen characteristics to help session processing devices identify high-value users (e.g., VIP users) or high-risk users (e.g., repeat complainers) to give them higher weight in decision-making, ensuring that these users receive priority human customer service and avoiding user churn or escalation of complaints due to improper handling by intelligent customer service.

[0108] User emotional characteristics reflect a user's emotional state / polarity in the current conversation, with corresponding intensity values ​​configured. This characteristic can be a key indicator for assessing conversation risk and predicting whether a problem will escalate into a complaint. It can capture irrational user states. When a user is detected to be emotionally agitated, for example, when the user's current inquiry is: "Why hasn't this been resolved yet! I've been waiting for ages!", natural language processing technology is used to analyze the frequency of emotional words, punctuation, speech rate, and tone in the current inquiry to determine that the user's emotional characteristic is "anger" or "anxiety," and the intensity value of this emotional state is high. This application tends to assign a high-risk rating to the current conversation and promptly transfer it to human customer service for intervention and reassurance to prevent escalation of conflict. Conversely, conversations with stable emotions can be more confidently handled by intelligent customer service.

[0109] As can be seen, this application upgrades the risk rating of the current session from a "two-dimensional" assessment to a "three-dimensional" profile. By introducing user profiling, it ensures the differentiation and security of service strategies, adopting a more conservative and human-intervention-oriented service strategy for high-value customers (VIPs) or high-risk customers (those with a high number of historical complaints), thereby improving customer experience, mitigating service risks, and protecting brand reputation. Simultaneously, user emotional characteristics serve as early warning signals, helping the system proactively switch to human customer service for emotional reassurance before customer emotions escalate completely, preventing escalation and enhancing service security.

[0110] Based on the above analysis, and referring to Figure 4This is a flowchart illustrating the conversation processing method proposed in Embodiment 2 of this application. This embodiment describes a possible implementation method of how to obtain conversation scene features, user emotion features, and user profile features based on context information in the conversation processing method proposed above. Figure 4 The implementation method may include, but is not limited to:

[0111] Step S41: Based on the current consultation and historical dialogue in the context information, retrieve relevant domain prior knowledge;

[0112] Step S42: Based on domain prior knowledge and predefined first prompt word templates, the intent of the current consultation is identified through the first model to determine the scenario risk level to which the corresponding consultation intent belongs, so as to obtain the conversation scenario characteristics of the current session;

[0113] In this embodiment, during the process of acquiring conversation scene features, a first model (such as a large language model, a multimodal large model, or a domain fine-tuning model, etc.) can be combined with RAG (Retrieval-Augmented Generation) technology to identify scene risk levels, etc. For example, the acquired context information can be combined with relevant knowledge fragments retrieved from a domain knowledge base (such as an internal enterprise knowledge base or an open-source knowledge base within the domain, etc.) (i.e., domain prior knowledge that is highly relevant to the current conversation content, to ensure that the model correctly understands the semantics in a specific domain scene), and a first prompt word can be constructed according to a predefined first prompt word template. The first prompt word is then input into the first model, and the consultation intent is understood through the first model to determine the scene risk level to which the current consultation intent belongs (such as high risk / medium risk / low risk, but not limited to these three levels. More granular risk levels can be divided based on business scenario requirements. This application only uses these three risk levels as examples for illustration, and this application does not limit the representation of each risk level). Afterwards, the determined scenario risk level can be identified as the scenario feature of the current session, or the scenario risk level can be mapped to the scenario feature according to predefined mapping rules, etc., without any restrictions.

[0114] For example, high-risk scenarios include complaints, policy disputes, and brand crises; medium-risk scenarios include complex process issues and inquiries involving personal safety; and low-risk scenarios include information inquiries and order status confirmations. These scenarios are generally consistent with the domain standards in customer service scenarios. As the domain prior knowledge is updated, the classification of scenario risk levels may change, causing conflicts or incompleteness with the knowledge in the first prompt word template. This application can dynamically update the first prompt word template based on the updated domain prior knowledge to update the description of the identification rules or methods for scenarios with different risk levels in the first prompt word template, so as to ensure the reliability and accuracy of scenario risk level identification. This application does not limit the content of the first prompt word template or the method of updating it.

[0115] Therefore, this application, by combining RAG technology with dynamic prompts, significantly improves the accuracy and domain compliance of consultation intent recognition, reliably outputting a risk level feature with business significance, namely, a conversation scenario feature. However, it is not limited to the conversation scenario feature recognition method described above in this application.

[0116] Step S43: Based on the current consultation and historical dialogue, perform emotion recognition through the first model to obtain the user's emotional characteristics;

[0117] This application can also utilize the emotion recognition / dialogue analysis capabilities of the first model to analyze the emotional tendencies in the current consultation and historical dialogues. Combined with emotional trends in historical dialogues (such as the accumulation of consecutive negative statements), it can determine the user's emotional state in real time and quantify the analysis results into user emotional characteristics, such as emotion scores and emotion tags. This application does not limit the implementation process. Optionally, other models, such as pre-trained emotion recognition models, can also be used for user emotion recognition; the implementation process is similar and will not be detailed in this application.

[0118] As can be seen, this application provides an emotional dimension input for risk rating by capturing the user's emotional state in the current session, which helps predict user behavior, such as whether the complaint will be escalated.

[0119] Step S44: In response to the context information containing the user identifier of the current session, obtain the user's identity tag in the customer service system from the customer data platform to determine the user profile features that match the identity tag.

[0120] The user identifier can be a unique identifier for a user in the current session, used to distinguish different users entering the customer service system, such as a user ID or mobile phone number. Through the enterprise CRM or customer service system API, the user's historical behavioral data, such as historical interaction records, consumption levels, and complaint history, can be queried based on the user identifier. This allows for the determination of the user's identity tags (also known as user profile tags), such as VIP customers, designated identities, or groups. These tags are then converted into matching user profile features, such as determining a user profile score according to predefined mapping rules. This application does not limit the content or representation of user profile features. Therefore, this application achieves refined risk rating based on individual user differences by introducing historical user behavior data.

[0121] In summary, this application's embodiments utilize techniques such as enhanced retrieval, dynamic prompts, and multi-source data fusion to construct a precise, real-time, and highly business-relevant contextual assessment feature extraction mechanism, laying a solid data foundation for the dynamic takeover decision-making of intelligent customer service. More specifically, by introducing domain prior knowledge and combining it with dynamically updated first prompt templates, the first model makes judgments based on an understanding of the business context, improving business relevance, significantly reducing the error rate of intent recognition, ensuring the accuracy of scenario risk level determination, and providing a reliable basis for subsequent conversation takeover decisions. Furthermore, it achieves multi-dimensional user state perception. Compared to traditional single emotion recognition or profile recognition methods, this application combines real-time emotion and historical profiles to establish a more comprehensive user state profile. For example, it can not only identify that a user is currently angry (emotional feature) but also recognize that the user is an important VIP (profile feature), thereby triggering higher-level risk warnings and response strategies.

[0122] In addition, the dynamic update mechanism of the first prompt word template enables the session processing device to automatically adjust the judgment criteria as prior domain knowledge changes, without the need for frequent retraining of the first model. This ensures real-time adaptability when facing new businesses and policies, reduces maintenance costs, and improves adaptability and real-time performance.

[0123] In some embodiments, the session processing method described above, obtaining performance evaluation features based on performance status information may include at least one of the following: consultation load characteristics of the customer service system and the historical consultation resolution rate of the intelligent customer service, to characterize the current operating load of the customer service system and the processing capacity of the intelligent customer service. The consultation load characteristics can be indicators reflecting the current busyness (current real-time pressure) of the customer service system, which can be determined by detecting the current consultation concurrency and queuing status of customer service and queues. Their quantified values ​​(such as high concurrency and the number of queued users, which can be normalized and converted into a score of 0-10) dynamically affect subsequent takeover decisions. For example, when queuing for high concurrency or more than a specified number of users, the weighting of consultation complexity is reduced (i.e., its quantified value is decreased to hopefully lower the risk rating result), and conversely, the weighting is increased (its quantified value is increased to hopefully improve the risk rating result), dynamically balancing the allocation of customer service manpower.

[0124] This demonstrates that the consultation load characteristics enable the session processing device to sense the overall service pressure of the customer service system. When the customer service system is under high load, the tendency for low-risk sessions to be handled by intelligent customer service should be appropriately increased to alleviate the pressure on human customer service; conversely, when the load is low, high-risk sessions can be more strictly screened and handed over to human customer service to ensure service quality.

[0125] Historical consultation resolution rate reflects the historical success rate of intelligent customer service in handling specific types of conversations (current conversation type) or specific user groups (the group to which the current user belongs). It can periodically update the resolution rate weight / coefficient based on the intelligent customer service's historical resolution performance, giving weight to inquiries with low resolution rates, raising the threshold for intelligent customer service to take over, and reducing the impact on overall service quality. Therefore, the historical resolution rate provides prior knowledge of the intelligent customer service's conversation handling capabilities. For scenarios where intelligent customer service has historically handled poorly, even if the current conversation appears to be of low risk, a low resolution rate can assign a higher risk rating, thus proactively transferring the conversation to human customer service and improving the overall service success rate.

[0126] Optionally, this application can query historical work order databases or knowledge base performance statistics platforms based on the initial session scenario (such as "after-sales complaint") or user profile tags (such as "enterprise customer") identified in the current session to obtain the problem resolution rate handled by intelligent customer service for this type of scenario or user within a preset time window (such as 30 days). For example, if the historical resolution rate for the "after-sales complaint" scenario is only 60%, it can be mapped to a resolution rate score (such as 6 points) and included in the risk rating calculation. This application does not restrict the calculation method for historical consultation resolution rates.

[0127] Therefore, this application achieves a comprehensive and precise risk perception by considering both user dimensions (profiles and sentiment) and system dimensions (load and resolution rate), thereby improving the reliability and accuracy of the risk rating results and enabling more refined takeover decisions that are more aligned with actual business needs. This significantly enhances the accuracy of risk assessment and reduces instances of erroneous takeovers or failures to take over when necessary.

[0128] Based on the descriptions of contextual assessment features and effectiveness assessment features in the above embodiments, the former represents the urgency and personalization requirements of the "demand side," while the latter represents the availability and processing capacity of the "supply side." This application, through this dynamic fusion decision-making mechanism of supply and demand matching, ensures that the risk rating results consider not only "who the user is and how urgent the user is," but also "how busy the system is and whether the intelligent customer service can perform the task." This maximizes the triage efficiency of intelligent customer service while ensuring user experience, achieving optimal collaboration between human and intelligent customer service. The following describes how to integrate these two dimensions of features to determine the risk rating results for the current session.

[0129] Reference Figure 5 This is a flowchart illustrating the session processing method proposed in Embodiment 3 of this application, as follows: Figure 5 As shown, a possible implementation method for determining the risk rating result for the current session by integrating contextual assessment features and effectiveness assessment features, as proposed in this application, may include:

[0130] Step S51: Perform weighted calculations on the various features included in the context assessment features to obtain the context risk rating result;

[0131] The process involves converting at least one of the following three features—conversation scenario features, user profile features, and user emotion features—into feature values ​​for the context assessment features of the current conversation. For example, "high-risk scenario" might be quantified as 8 points, "VIP customer" as 8 points, and "neutral emotion" as 5 points, but this is not limited to these. Then, based on the business rules of the customer service scenario, a weight coefficient can be configured for the corresponding feature categories. For instance, if the context assessment features actually include these three types of features, the weight 'a' for the conversation scenario feature is 50%, the weight 'b' for the user profile feature is 30%, and the weight 'c' for the user emotion feature is 20%. If the context assessment features actually include two of these types of features (e.g., a first feature combination of conversation scenario features and user profile features, a second feature combination of user profile features and user emotion features, or a third feature combination of conversation scenario features and user emotion features), the sum of their respective weights is 1. If the context assessment features actually include only one type of feature (any one of conversation scenario features, user profile features, or user emotion features), no weight needs to be configured, or weights can be assigned in conjunction with energy efficiency assessment features, etc., without restriction.

[0132] Therefore, this application can obtain a context risk rating result by weighted summation of the quantitative feature values ​​corresponding to the conversation scene features, user profile features, and user emotion features. Alternatively, it can obtain a context risk rating result by weighted summation of the quantitative feature values ​​corresponding to two features in the first, second, or third feature combination. Of course, it can also directly determine the context risk rating result by using the quantitative feature values ​​corresponding to the conversation scene features, user profile features, or user emotion features. This application does not limit the implementation method of step S51; it can be determined based on the feature categories included in the currently obtained context assessment features.

[0133] Step S52: Quantify the performance evaluation characteristics to obtain the corresponding performance risk coefficient;

[0134] Similarly, this application can quantify the consultation load characteristics and / or the historical resolution rate of intelligent customer service included in the performance evaluation features, such as through numerical processing, to obtain corresponding performance risk coefficients, such as corresponding quantitative feature values ​​or values ​​obtained through normalization. For example, a high load might be quantified as 1.5, and a low load as 0.8; a high historical resolution rate might be quantified as 0.9, and a low historical resolution rate as 1.2, etc., without limitation. It can be understood that the performance risk coefficient obtained in step S52 corresponds to the features actually included in the performance evaluation features. If it includes consultation load characteristics or historical resolution rate, a corresponding performance risk coefficient is obtained; if it includes consultation load characteristics and historical resolution rate, two corresponding performance risk coefficients are obtained. This application does not limit the implementation process.

[0135] As can be seen, this application transforms the dynamic state and historical performance of the customer service system into a moderating factor that influences the final decision. This performance risk coefficient typically amplifies contextual risks (such as high load and low resolution rate) when it is greater than 1, and weakens contextual risks (such as low load and high resolution rate) when it is less than 1. This reflects the moderating effect of the system's overall performance on individual conversation decisions, thereby improving the accuracy and reliability of risk decisions.

[0136] Step S53: Based on the contextual risk rating results and the effectiveness risk coefficient, generate a risk rating result for the current conversation.

[0137] Based on the contextual risk rating results and effectiveness risk coefficients described above, they are obtained in different ways (i.e., the content of the contextual risk features and the content of the effectiveness assessment features they rely on are different). The product of the two generates different risk rating results for the current conversation, which will not be detailed here by example.

[0138] Preferably, when the context assessment features actually include conversation scene features, user profile features, and user emotion features, and the effectiveness assessment features include consultation load features and historical resolution rate, the context risk rating result can be obtained by the calculation method of (weight a × conversation scene score f1 + weight c × user profile score f2 + weight c × user emotion score f3). Then, the context risk rating result, the effectiveness risk coefficient of consultation load feature (consultation load weight), and the effectiveness risk coefficient of historical resolution rate (resolution rate weight) are multiplied, and the result is determined as the risk rating result S for the current conversation.

[0139] Subsequently, a decision can be made based on S to trigger intelligent customer service to take over the current conversation. This can be achieved by comparing S with different threshold ranges to execute the decision action corresponding to the threshold range where S falls. For example, S < 3 automatically takes over, 3 ≤ S ≤ 6 prompts for takeover, and S > 6 does not take over / exits takeover and issues a warning to the agent, etc. There are no restrictions on this. It should be understood that the above-mentioned quantitative feature values ​​include, but are not limited to, numerical features such as ratings, and can also be represented by other methods such as levels or proportions. The process of generating risk rating results is similar, and this application will not provide detailed examples of each. In addition, if the feature categories included in the context assessment feature change, and / or the feature categories included in the effectiveness assessment feature change, S can still be generated according to the calculation method described above, such as removing the quantitative feature values ​​of features that are not currently available in the above formula. This application will not provide detailed examples of each.

[0140] The aforementioned risk rating result can also be referred to as the confidence level at which the intelligent customer service is suitable to take over the current session. This confidence level is used to dynamically decide whether the intelligent customer service can take over. Of course, even if the intelligent customer service has already taken over, this can still be used to decide whether to stop providing service or exit the current session. In this case, the threshold range corresponding to high risk can be a low confidence range, the threshold range corresponding to medium risk can be a medium confidence range, and the threshold range corresponding to low risk can be a high confidence range. This application does not impose upper or lower limits on the thresholds for each region; these can be dynamically updated based on actual circumstances.

[0141] Based on the above analysis, this application transforms subjective and vague business judgments (such as "this problem is a bit troublesome" or "the system is quite busy right now") into objective and unified quantitative characteristic values ​​(such as scores) through weighted calculation and coefficient adjustment. This eliminates the inconsistency and arbitrariness of manual judgment, ensuring that every takeover decision is based on a traceable and auditable standardized strategy. This improves the standardization and transparency of the service process and provides a data foundation for quality monitoring and optimization. In the risk rating process, an effectiveness risk coefficient is introduced as a dynamic adjustment factor, so that the final risk rating result is not statically dependent on the session content, but dynamically fluctuates with the system's real-time load and historical AI capabilities, enabling context-aware intelligent scheduling.

[0142] Moreover, the risk rating decision model in this application is essentially a resource allocation optimizer, simultaneously seeking the optimal balance between two objectives: maximizing the utilization rate of intelligent customer service (improving efficiency) and minimizing service risks caused by erroneous takeover (ensuring quality). Through dynamic adjustment, the system can automatically find the optimal balance point for human-machine collaboration under different time periods and pressures. This avoids excessive idleness of human customer service during off-peak hours and prevents catastrophic errors during busy periods or at the limits of intelligent customer service capabilities, thereby improving the overall throughput, stability, and service robustness of the customer service system. It is evident that this application not only achieves objective quantification of risk assessment but, more importantly, endows the system with intelligent decision-making capabilities that perceive the environment and self-adjust, upgrading human-machine collaboration from a static division of labor based on fixed rules to dynamic collaboration based on multi-objective real-time optimization, thus achieving a significant technological breakthrough in the balance between efficiency, quality, and system stability.

[0143] The takeover decision implementation process in the above session processing method will be described below, such as... Figure 6 The flowchart shown in Embodiment 4 of this application illustrates the session processing method. The takeover decision implementation method proposed in this embodiment may include, but is not limited to:

[0144] Step S61: Determine whether the risk rating result for the current session is greater than the first risk threshold. If not, proceed to step S62; if yes, proceed to step S67.

[0145] Step S62: Determine whether the risk rating result is less than the second risk threshold. If yes, proceed to step S63; otherwise, proceed to step S64.

[0146] Step S63: Trigger the intelligent customer service to automatically take over the current session, and execute step S66;

[0147] Step S64: Send a takeover confirmation request to customer service.

[0148] Step S65: In response to the takeover confirmation command, trigger the intelligent customer service to take over the current session;

[0149] Step S66: During the process of intelligent customer service taking over the current session, determine whether the updated risk rating result meets the conditions for intelligent customer service to interrupt takeover, and then execute step S67.

[0150] Step S67: Trigger the intelligent customer service to stop generating reply information and exit the current session, and send a status prompt message to the human customer service so that the human customer service can take over the current session.

[0151] In this application, we will still take the risk level including three levels: high risk, medium risk and low risk as an example for explanation. Each level is configured with a corresponding threshold range. This third threshold range can be divided by the first risk threshold and the second risk threshold. The first risk threshold is greater than the second risk threshold. In the initial stage, it can be determined based on experience or business requirements. As domain knowledge accumulates or the system status changes, it can be dynamically updated. This application does not limit the size of each risk threshold.

[0152] Therefore, by comparing the risk rating result with the first and second risk thresholds respectively, and considering that the risk rating result for the current session is less than the second risk threshold, it indicates that the session is in a high-confidence region and is considered low-risk. This triggers the intelligent customer service to automatically take over the current session, sending the context information of the current session to the main agent. After parameter extraction and multimodal processing, combined with context engineering for intent recognition, at least one determined task is distributed to an appropriate sub-agent to generate response information. This application does not limit the implementation method of the intelligent customer service generating response information for the current inquiry. During multimodal information processing, when the user inputs information in the form of images, videos, etc., it can be processed using OCR (Optical Character Recognition) and multimodal large-scale models, but is not limited to these methods. The processing result is then updated to the context information to improve the accuracy of the response.

[0153] In some embodiments, if the risk rating result of the current session falls between the first and second risk thresholds, it indicates that the session is in the medium confidence zone and is considered medium risk. In this case, a takeover confirmation request for the intelligent customer service system can be sent to a human customer service representative. The human representative will then conduct a secondary confirmation, or the system will determine whether the intelligent customer service system can take over the current session according to its rules. If there is no response for an extended period, the process can continue, and the intelligent customer service system can take over the current session to improve response speed. It is evident that both of the above situations meet the takeover conditions for the intelligent customer service system, but the latter still requires secondary confirmation from either a human representative or the system to improve service quality.

[0154] Optionally, if the risk rating result of the current session is greater than the first risk threshold, it indicates that it is in the low confidence area and belongs to high risk. If the intelligent customer service still takes over the current session, security risks are likely to occur. At this time, the intelligent customer service can be stopped to prevent the intelligent customer service from taking over the current session. It can also remind the human customer service to take over the current session, such as sending a status prompt message to the human customer service. The system notification can be invoked to remind the human customer service in the system interface with a high priority reminder method (such as screen flashing, pop-up window, etc.) and inform the human customer service of the reason for stopping, such as the user being emotionally agitated or involving high-risk issues. This application does not restrict the content of the status prompt message or its output method.

[0155] In some embodiments, in conjunction with the session processing method described above, during the process of intelligent customer service taking over the current session, after the user re-enters the inquiry, the risk rating result of the current session can still be obtained according to the method described above. In response to the updated risk rating result satisfying the intelligent customer service's interruption takeover condition (e.g., high risk), the intelligent customer service stops service and transfers the user to a human customer service representative, thereby achieving fragmented takeover of the entire session by the intelligent customer service. The interruption takeover condition may include any of the following: the risk rating result is greater than a first risk threshold; a forced takeover instruction is received from a human customer service representative; the context information contains predefined risk content; the inquiry intent of the current inquiry in the context information does not belong to the intelligent customer service's intent database, etc., without limitation. During the process of intelligent customer service taking over the current session, the takeover decision described above can also be implemented by the main intelligent agent to obtain the updated risk rating result; the implementation process is not detailed in this application.

[0156] As can be seen, in a conversation dominated by human customer service, a portion (fragment, such as the current inquiry) is temporarily and dynamically taken over and responded to. Based on the aforementioned risk assessment, if the conditions for intelligent customer service take over are met, intelligent customer service takes over the current conversation to generate a response to the current inquiry. As the user's input inquiries are updated, the risk assessment result is updated synchronously. Once the conditions for intelligent customer service to interrupt takeover are met, the conversation will be switched back to human customer service. This fragmented takeover approach allows for precise triggering of intelligent customer service to take over the former in a complex multi-turn dialogue, while the latter remains handled by human customer service or triggers a stop mechanism, even if a user's initial inquiry is a simple order query (low risk). In other words, based on real-time and continuous risk assessment, this application allows intelligent customer service to intervene in the current conversation multiple times and briefly to handle simple, repetitive, and low-risk consultation tasks. This frees up human customer service from high-frequency, low-value dialogue sessions, allowing them to focus their energy on complex and critical issues that truly require human wisdom and emotion. This improves the handling of complex consultations and critical issues in the queue, thereby shortening user waiting time and enhancing overall service efficiency and experience.

[0157] Based on the description of the interruption and takeover conditions above, this application can continuously monitor and recalculate the risk score. If the score exceeds a preset higher safety threshold (the first risk threshold) due to escalating user emotions, increased problem complexity, or other reasons, a stop will be immediately triggered regardless of whether the main AI agent is responding. Human intervention is also possible to ensure absolute control by human customer service representatives. Optionally, if human customer service representatives find erroneous responses or unfavorable conversation trends in the monitoring interface, they can manually trigger the interruption of the intelligent customer service service by clicking the "Stop" or "Force Takeover" button at any time, but this implementation method is not limited to this.

[0158] Furthermore, considering the risk of the aforementioned decision-taking timeliness in this application, namely, when the risk rating result is greater than the first risk threshold but the service is not stopped, resulting in continuous erroneous responses from the main intelligent agent in a high-risk scenario, a rapid circuit breaker condition based on explicit rule matching can be used to detect whether the context information contains predefined risk content, such as a list of high-risk keywords or phrases. If it does, there is no need to wait for the complete risk score calculation; the intelligent customer service can be triggered to stop service or downgrade execution. It can also remind human customer service to conduct a second confirmation. This application does not restrict the implementation process.

[0159] Furthermore, considering the risk of the main agent misidentifying intents based on the second model, leading to actions exceeding the intelligent customer service's handling scope or erroneous responses, this application establishes a clearly defined database of manageable intents. It strictly defines that the main agent can only process explicitly specified intents (such as querying progress, creating work orders, etc.), directly stopping inquiries outside this database to prevent errors. More scenarios can be gradually opened as the main agent's dialogue credibility increases. The intent recognition model can determine whether the current inquiry's intent exceeds the intelligent customer service's handling scope (intent database). If it does, service is immediately stopped, and the client is transferred to a human agent.

[0160] Therefore, this application utilizes multi-dimensional data to determine the interruption and takeover conditions of intelligent customer service, covering multiple dimensions such as real-time dynamic risk assessment, proactive human intervention, rapid rule matching, and capability boundary identification. This transforms security protection from a single-point mechanism into a three-dimensional, complementary defense network. Even if the decision-making model experiences delays or misjudgments, there are still human customer service forcibly taking over and keyword rules as immediate backups; even if the rule base is not covered, the intent database boundary serves as the last line of defense. This significantly enhances the robustness and security of the entire system in high-risk scenarios, effectively preventing the risk of misoperation and decreased customer satisfaction caused by the main intelligent agent forcibly responding in incompatible scenarios.

[0161] Moreover, the aforementioned process of switching to a human customer service representative is seamless for the user and provides ample information for the human representative. The human representative can immediately understand "why the switch occurred" and "what just happened" through status prompts, allowing for quick takeover and precise service. This avoids the poor experience of users repeating questions and customer service representatives being confused, ensuring service continuity and customer experience. Optionally, in the case of switching from intelligent customer service to a human representative, in response to confirmation that the human representative has taken over the current conversation, a conversation summary generated by the intelligent customer service based on the historical dialogues it has taken over is sent to the human representative, enabling the human to quickly understand the context and avoid missing important information. This application does not restrict the implementation method of how the main intelligent agent generates the conversation summary.

[0162] Therefore, this application utilizes a multi-triggered automated circuit breaker process to delegate simple, repetitive, and low-risk fragmented consultation tasks to intelligent customer service for automatic responses. This allows for seamless intervention of the main intelligent agent, mimicking the style of human customer service representatives in generating replies. Before sending the reply, it is previewed on the human customer service representative's interactive interface via streaming output, providing the representative with a final opportunity for supervision and modification. In complex, emotional, and high-risk scenarios, an interruption mechanism ensures that control is quickly and reliably returned to the human customer service representative, making the entire process not a crude "answer-grabbing" but a collaborative process with requests, monitoring, interruptibility, and handover. The human customer service representative remains the ultimate person in charge and supervisor of the conversation, with the main intelligent agent serving as its efficient auxiliary tool. Furthermore, combined with the multi-dimensional risk rating described above, this ensures refined conversation takeover, thereby guaranteeing the accuracy and security of takeover decisions. It avoids the crude collaboration of intelligent customer service representatives "ignoring what should be answered and answering what shouldn't be answered," freeing up human customer service representatives' energy and improving the concurrent processing capabilities of the customer service system. Thus, while ensuring service security and quality, it maximizes the operational efficiency and human capital value of the overall customer service system.

[0163] Based on the conversation processing methods described in the above embodiments, the implementation process involves obtaining contextual evaluation features based on contextual information and performance evaluation features based on performance status information. The risk rating for the current conversation is determined by fusing these contextual and performance evaluation features using a risk rating strategy. This risk rating strategy may include assigning a risk rating to each user-inputted inquiry, and executing this processing flow to determine the risk rating result for the current inquiry in the current conversation. In this case, each inquiry is considered a segment of the entire conversation process. Each inquiry received by the customer service system can be subject to takeover decisions according to the method of this application, achieving dynamic and continuous risk rating of the entire conversation process.

[0164] Preferably, to reduce computational pressure and shorten response time, the aforementioned risk assessment strategy can be dynamically adjusted based on at least one of the following: user input status information, system cached data, and current session status. This allows for changes in the frequency or method of risk rating execution for the current session. Thus, for high-volume, high-concurrency customer service operations, multiple inquiries can be delayed and merged for a single risk assessment. For example, if the time interval between multiple pending inquiries entered by a user is less than a time threshold (e.g., 1.5 seconds), it may indicate that the user is in a state of urgency or emotional agitation. In this case, the frequency of risk rating can be reduced, or these multiple inquiries can be merged and calculated as a single inquiry for risk rating. This avoids fragmented and potentially inaccurate assessments when the user's expression is incomplete, while also reducing system computational pressure. Conversely, when the user input interval is longer, assessments are performed at a normal or higher frequency.

[0165] Optionally, if based on system cached data, such as historical calculation results (e.g., dialogue scenarios, profiles, and key evaluation parameters such as emotions), it is determined that the contextual features of the current session (e.g., consultation, intent, user profile, etc.) highly match a cached result, the system may choose not to perform a new risk assessment calculation, but instead directly reuse the risk rating result in the cache, thereby skipping the current calculation and greatly improving response speed.

[0166] In this application, adjusting the method for determining risk ratings can involve changing the data source or calculation logic upon which the risk rating results are generated, such as based on the current session state. The session state changes as the conversation progresses. If the current session duration (the time between the input time of the first inquiry and the current system time) exceeds a preset duration (e.g., 5 minutes), the user may leave, and risk rating can be stopped to reduce unnecessary resource waste. Furthermore, this application can also set up a response circuit breaker mechanism. When excessively high computational response latency (performance bottleneck) is detected, the response circuit breaker mechanism is triggered. For example, if the response time for risk rating exceeds a preset value, the service can be downgraded, reducing the number of feature categories used in the risk rating result determination process and simplifying calculations using core features. For instance, risk rating could be downgraded to only using dialogue scenarios and emotions to alleviate computational pressure, ensure the smooth operation of core links, and provide a compromised but usable service, but this is not limited to these methods.

[0167] Therefore, the risk rating proposed in this application is not enforced on every inquiry in a fixed, high-frequency mechanical manner, but rather dynamically and intelligently adjusted according to real-time context to change the triggering timing and frequency of risk score calculation. By merging calculations, reusing caches, and simplifying calculations, the system's computational load is significantly reduced during peak hours or when dealing with fast-talking users, avoiding unnecessary resource waste, ensuring the smoothness of core links, and optimizing system performance and resource consumption. Reusing cached results enables near real-time risk assessment, thereby triggering takeover or shutdown more quickly and reducing user waiting time. Intelligent frequency adjustment also avoids interfering intermediate evaluations while the user is "typing," improving response speed and user experience. Furthermore, degradation strategies such as "response circuit breaking" and "long conversation pause" can be introduced, allowing the system to provide continuous and basic services under abnormal or high-pressure conditions, rather than completely collapsing, demonstrating the design flexibility of industrial-grade systems and improving system robustness and availability.

[0168] In some embodiments, for the risk rating method described above in this application, as analyzed above, the weights or coefficients of various features used to determine the risk rating results can be dynamically adjusted to suit different types of business scenarios, improving the utilization rate and experience of intelligent agent takeover. This can be achieved by responding to situations where the number of users waiting for a response in the customer service system's inquiry load indicator exceeds the sorting threshold, or where the current inquiry is a high-concurrency business, by reducing the corresponding efficiency risk coefficient and dynamically balancing the allocation of customer service manpower. Alternatively, in response to situations where the historical inquiry resolution rate of intelligent customer service is less than the resolution threshold, the corresponding efficiency risk coefficient can be increased to raise the threshold for automatic takeover and reduce the impact on overall service quality.

[0169] Optionally, this application can also dynamically adjust at least one of the following based on the historical dialogues and user feedback of the intelligent customer service (such as the user's satisfaction rating after the dialogue, the mark of whether the problem has been resolved, or the evaluation of the main intelligent agent's performance during this takeover by the human customer service). This includes: model parameters used to calculate context assessment features (such as some parameters of the first model; for example, when it is found that "consultation on exchange" is often misjudged as a low-risk "query" intent, the model parameters can be fine-tuned to more accurately identify it as a medium-to-high-risk "after-sales" intent); the weight allocation of various features in the context assessment features (such as the values ​​of the aforementioned weights a, b, and c, which can be determined based on the degree of influence of various features on the risk rating); model parameters used to calculate user emotional features (such as some model parameters of the first model, affecting its ability to recognize emotions, or parameters of a dedicated emotion recognition model); and the mapping relationship between predefined consultation intent and scenario risk level (which can be a business rule base, updated based on changes in domain prior knowledge, and the first prompt word template can be updated synchronously as needed). This is to dynamically optimize the decision model for achieving new risk ratings. This application does not restrict the data sources for model optimization.

[0170] Based on this, refer to Figure 7 The flowchart shown indicates that this application can record the behavior logs of each module during the implementation of the method, which may include the decision-making process, agent input, output, called models / rules, determined risk rating results, etc., to realize the above-mentioned self-learning process. It can also be used for auditing and review to meet the industry regulatory requirements for traceability and auditability of service processes.

[0171] Therefore, it can be seen that, Figure 7As shown, this application, based on the real-time dialogue stream of the current session (text, voice, and other multimodal data), sequentially performs contextual understanding, dialogue state capture, risk rating, and decision triggering, ultimately achieving seamless handover control. It can also automatically identify and correct shortcomings in current decision-making (such as inaccurate risk mapping for a certain intent or unreasonable weighting of a feature). This allows the decision model to dynamically adapt to business changes, evolving user behavior, and emerging problem types, forming a reinforced "evaluation-execution-feedback-optimization" cycle to ensure long-term stability and improvement in service quality. Furthermore, this autonomous iterative model optimization significantly reduces the workload of algorithm engineers and business experts in manually tuning parameters and maintaining rules, giving the system greater autonomy and maintainability. By continuously learning from service performance data across different business lines, user groups, and time periods, the system can gradually refine its decision-making criteria, promoting the personalization and precision of risk rating strategies.

[0172] In summary, the conversation processing methods described in the above embodiments, in the implementation process where the intelligent customer service acts as the main intelligent agent, interacting with a group of sub-intelligent agents to generate a response to the current inquiry, such as... Figure 7 As shown, contextual information (current consultation and historical dialogue, etc.) can be sent to the main agent for semantic understanding. The main agent then distributes the generated subtasks to the corresponding sub-agents for execution (multiple subtasks are executed in parallel or serially) to obtain the corresponding execution results. Based on the execution results and the second prompt word template, the main agent outputs the response information for the current consultation.

[0173] The second prompt template is learned by the main agent or its corresponding sub-agents based on historical responses from human customer service representatives. It constrains the output attributes of the responses, including language style, speaking speed, and tone, thereby mitigating the risk of disjointed user experience. In other words, to address user dissatisfaction caused by an uneven transition between intelligent and human customer service, it's crucial to ensure the lossless transmission of contextual information, achieving complete and real-time synchronization of historical dialogues, user emotions, and attempted solutions. Furthermore, it can constrain the intelligent customer service representative to mimic the response style and tone of a human representative, reducing the perceived difference in switching between the two. For users, the multiple switches between human and intelligent customer service throughout the conversation are virtually imperceptible, improving the user experience.

[0174] Therefore, this application utilizes a master agent for task planning and distribution, breaking down complex user inquiries into multiple standard sub-tasks, which are then executed by specialized sub-agents, thus achieving decoupling and reuse of processing capabilities. Compared to training a single "all-powerful" giant model (master agent), combining multiple "specialized" sub-agents to handle complex scenarios improves the accuracy, reliability, and efficiency of processing complex, multi-step inquiries, and is easily scalable (adding new sub-agents is sufficient for new business requirements). Because the second prompt word template can guide the main intelligent agent in how to organize language, state facts, and ensure the completeness and friendliness of the response information, this application, combined with the organization of the second prompt word template, ensures that the final response not only informs the user of the result but also clearly explains the subsequent steps, providing a complete service loop. This greatly improves the first-question resolution rate and user satisfaction. At the same time, by learning the second prompt word template from excellent human service records, the intelligent customer service's response can highly imitate the enterprise's unified service standards in terms of language style and service warmth. This upgrades the intelligent customer service from a simple question-and-answer robot to a virtual customer service expert capable of handling complex business processes and outputting accurate, reliable, and professional responses. This significantly reduces the user experience risks caused by inappropriate intelligent customer service response styles and is conducive to achieving high-quality, large-scale human-machine collaborative services.

[0175] In practical applications, this solution can be rolled out in a phased rollout. It can be tested first in small-scale scenarios (such as a single business line or a consultation queue), gradually expanding the scope of takeover. A / B testing can be used to compare key metrics between the traditional and intelligent takeover models, such as First Contact Resolution Rate (FCR), Average Handling Time (AHT), and Customer Satisfaction Score (CSAT), to verify effectiveness and optimize decisions. Full rollout can only proceed after successful verification; there are no restrictions on this process.

[0176] Reference Figure 8 This is a schematic diagram of the structure of the session processing device proposed in the embodiments of this application, as shown below. Figure 8 As shown, the session processing device may include, but is not limited to:

[0177] The information acquisition module 81 is used to acquire the context information of the current conversation between the human customer service representative and the user, as well as the performance status information of the customer service system;

[0178] The feature acquisition module 82 is used to obtain contextual evaluation features based on the context information and to obtain performance evaluation features based on the performance status information; the contextual evaluation features include conversation scene features and user state features;

[0179] Risk rating module 83 is used to integrate the context assessment features and the effectiveness assessment features to determine the risk rating result for the current session;

[0180] The first decision module 84 is used to trigger the intelligent customer service to take over the current session in response to the risk rating result meeting the takeover conditions of the intelligent customer service, so as to generate the reply information for the current inquiry.

[0181] Optional, may also include:

[0182] The second decision module is used to, during the process of intelligent customer service taking over the current session, respond to the satisfaction of the intelligent customer service's interruption takeover conditions, trigger the intelligent customer service to stop generating reply information and exit the current session, and send a status prompt information to the human customer service so that the human customer service can take over the current session; wherein, the interruption takeover conditions include the risk rating result being greater than a first risk threshold; receiving a forced takeover instruction input by the human customer service; the context information containing predefined risk content; and the consultation intent of the current consultation in the context information not belonging to any of the intent databases of the intelligent customer service.

[0183] In some embodiments, the user status characteristics mentioned above include at least one of user profile characteristics and user emotion characteristics; the performance evaluation characteristics include at least one of the consultation load characteristics based on the customer service system and the historical consultation resolution rate of the intelligent customer service.

[0184] Based on this, the aforementioned risk rating module 83 may include:

[0185] The context risk rating unit is used to perform weighted calculations on the various features included in the context assessment features to obtain the context risk rating result.

[0186] The performance risk rating unit is used to quantify the performance evaluation characteristics and obtain the corresponding performance risk coefficient.

[0187] The risk rating result generation unit is used to generate a risk rating result for the current session based on the context risk rating result and the effectiveness risk coefficient.

[0188] Optionally, the first decision module 84 described above may include any of the following decision sub-units:

[0189] The first decision subunit is used to trigger the intelligent customer service to automatically take over the current session in response to the risk rating result being less than the second risk threshold; wherein the first risk threshold is greater than the second risk threshold.

[0190] The second decision subunit is used to send a takeover confirmation request to the human customer service representative in response to the risk rating result being between the first risk threshold and the second risk threshold, and to trigger the intelligent customer service representative to take over the current session in response to the takeover confirmation instruction.

[0191] Optionally, the feature acquisition module 82 may include a first feature acquisition unit for acquiring session scene features, user emotion features, and user profile features based on the context information, which may include:

[0192] The retrieval unit is used to retrieve relevant domain prior knowledge based on the current consultation and historical dialogue in the context information.

[0193] The scenario risk level determination unit is used to identify the intent of the current consultation based on the domain prior knowledge and a predefined first prompt word template, and determine the scenario risk level to which the corresponding consultation intent belongs, so as to obtain the conversation scenario features of the current session; wherein, the first prompt word template is dynamically updated based on updated domain prior knowledge.

[0194] An emotion recognition unit is used to perform emotion recognition based on the current consultation and the historical dialogue, and obtain the user's emotion characteristics through the first model.

[0195] The user profile feature determination unit is used to obtain the user's identity tag in the customer service system from the customer data platform in response to the context information containing the user identifier of the current session, so as to determine the user profile features that match the identity tag.

[0196] In some embodiments, the above-described apparatus may further include any of the following adjustment modules:

[0197] The first adjustment module is used to reduce the corresponding performance risk coefficient in response to the customer service system's inquiry load indication that the number of users waiting for a reply is greater than the sorting threshold, or the current inquiry is a high-concurrency business.

[0198] The second adjustment module is used to increase the corresponding efficiency risk coefficient in response to the historical consultation resolution rate of the intelligent customer service being less than the resolution threshold.

[0199] The third adjustment module is used to dynamically adjust at least one of the following based on the intelligent customer service's historical dialogues and user feedback service evaluations:

[0200] Model parameters used to calculate the context evaluation features;

[0201] The weight allocation of various features in the context evaluation features;

[0202] Model parameters used to calculate the user's emotional characteristics;

[0203] The mapping relationship between predefined consultation intentions and scenario risk levels.

[0204] In practical applications of this application, the aforementioned information acquisition module 81, feature acquisition module 82, and risk rating module 83 can be implemented based on a risk rating strategy. This risk rating strategy includes performing a risk rating for each query input by the user to determine the risk rating result for the current query in the current session. Optionally, the risk assessment strategy can be dynamically adjusted based on at least one of the user's input state information, system cache data, and the current session state to change the frequency or method of risk rating execution in the current session.

[0205] Optionally, when the intelligent customer service acts as the main intelligent agent, interacting with a group of sub-intelligent agents to generate a response to the current inquiry, the first decision module 84 may include:

[0206] The context information transmission unit is used to send the context information to the main intelligent agent for semantic understanding, and the main intelligent agent distributes the generated multiple sub-tasks to the corresponding sub-intelligent agents for execution to obtain the corresponding execution results;

[0207] The output unit is used to output the response information for the current inquiry based on the execution result and the second prompt word template by the main intelligent agent;

[0208] The second prompt word template is learned by the main agent or the corresponding sub-agent based on the historical response information of the human customer service representative, and is used to constrain the output attributes of the response information, including language style.

[0209] This application also provides a computer program product including computer-readable instructions. When the computer-readable instructions are executed on an electronic device (session processing device), the electronic device implements any of the session processing methods provided in this application. The computer program product can be stored in a readable storage medium, such as a computer floppy disk, USB flash drive, portable hard drive, ROM (Read-Only Memory), RAM (Random Access Memory), magnetic disk, or optical disk, and includes several instructions to cause an electronic device to execute the session processing methods described in the various embodiments of this application.

[0210] This application also provides a computer-readable storage medium that carries one or more computer programs. When the one or more computer programs are executed by an electronic device (session processing device), the electronic device can implement any of the session processing methods provided in this application.

[0211] This application also provides an intelligent program (such as an agent or intelligent assistant) to receive user input inquiries and implement the session processing method proposed in this application through interaction with a model. The implementation process can be referred to the description of the corresponding part of the method embodiment above. In this implementation process, other components of the application or operating system can also be controlled through interface calls or other interactive methods to respond to the current inquiry, without limitation.

[0212] It should also be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and 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 modules can be selected to achieve the purpose of this embodiment according to actual needs. In addition, in the accompanying drawings of the device embodiments provided in this application, the connection relationship between modules indicates that they have a communication connection, which can be implemented as one or more communication buses or signal lines.

[0213] In the above embodiments, the invention can be implemented entirely or partially by software, hardware, firmware, or any combination thereof. Through the description of the above embodiments, those skilled in the art can clearly understand that this application can be implemented using software plus necessary general-purpose hardware, or it can be implemented using dedicated hardware including dedicated integrated circuits, dedicated CPUs, dedicated memory, dedicated components, etc. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The various embodiments in this specification are described in a progressive or parallel manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to mutually. Therefore, those skilled in the art can make various modifications or variations to the embodiments of this application without departing from the technical concept scope of this application, and all modified or varied embodiments fall within the protection scope of this application. For the apparatuses and devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple, and relevant parts can be referred to in the method section.

Claims

1. A session processing method, the method comprising: Obtain contextual information of the current conversation between human customer service representatives and users, as well as the performance status information of the customer service system; Contextual evaluation features are obtained based on the contextual information, and performance evaluation features are obtained based on the performance status information; the contextual evaluation features include conversation scenario features and user status features. By integrating the contextual assessment features and the effectiveness assessment features, a risk rating result is determined for the current session; In response to the risk rating result meeting the takeover conditions of the intelligent customer service, the intelligent customer service is triggered to take over the current session to generate a response to the current inquiry.

2. The method according to claim 1, further comprising the following steps during the intelligent customer service takeover of the current session: In response to the intelligent customer service interruption takeover condition being met, the intelligent customer service is triggered to stop generating reply information, exit the current session, and send a status prompt information to the human customer service so that the human customer service can take over the current session; The interruption takeover conditions include any one of the following: The risk rating result is greater than the first risk threshold; Received the forced takeover command input by the human customer service representative; The context information includes predefined risk content; The consultation intent of the current consultation in the context information does not belong to the intent database of the intelligent customer service.

3. The method according to claim 1 or 2, wherein: The user status characteristics include at least one of user profile characteristics and user emotion characteristics; The performance evaluation features include at least one of the consultation load features of the customer service system and the historical consultation resolution rate of the intelligent customer service.

4. The method according to claim 3, wherein fusing the contextual assessment features and the effectiveness assessment features to determine the risk rating result for the current session includes: The context risk rating result is obtained by weighting and calculating the various features included in the context assessment features; The performance evaluation characteristics are quantified to obtain the corresponding performance risk coefficients; Based on the contextual risk rating result and the effectiveness risk coefficient, a risk rating result for the current session is generated.

5. The method according to claim 1 or 2, wherein the response to the risk rating result satisfying the intelligent customer service takeover conditions, triggering the intelligent customer service to take over the current session, includes any one of the following: In response to the risk rating result being less than the second risk threshold, the intelligent customer service is triggered to automatically take over the current session; In response to the risk rating result being between the first risk threshold and the second risk threshold, a takeover confirmation request is sent to the human customer service representative; in response to the takeover confirmation instruction, the intelligent customer service representative is triggered to take over the current session. in, The first risk threshold is greater than the second risk threshold.

6. The method according to claim 4, wherein, Based on the aforementioned contextual information, conversation scenario features, user emotion features, and user profile features are obtained, including: Based on the current consultation and historical dialogue in the context information, relevant domain prior knowledge is retrieved; Based on the prior knowledge of the domain and the predefined first prompt word template, the intent of the current consultation is identified through the first model to determine the scenario risk level to which the corresponding consultation intent belongs, so as to obtain the conversation scenario characteristics of the current session; Based on the current consultation and the historical dialogue, emotion recognition is performed through the first model to obtain the user's emotional characteristics; In response to the context information containing the user identifier of the current session, the user's identity tag in the customer service system is obtained from the customer data platform to determine the user profile features that match the identity tag; The first prompt word template is dynamically updated based on updated domain prior knowledge.

7. The method according to claim 4, further comprising: In response to the customer service system's inquiry load indication that the number of users waiting for a reply is greater than the sorting threshold, or that the current inquiry is a high-concurrency business, the corresponding performance risk coefficient is reduced. or, If the historical consultation resolution rate of the intelligent customer service is less than the resolution threshold, the corresponding efficiency risk coefficient will be increased. or, Based on the intelligent customer service's historical dialogues and user feedback service evaluations, at least one of the following will be dynamically adjusted: Model parameters used to calculate the context evaluation features; The weight allocation of various features in the context evaluation features; Model parameters used to calculate the user's emotional characteristics; The mapping relationship between predefined consultation intentions and scenario risk levels.

8. The method according to claim 1 or 2, wherein, The process of obtaining contextual assessment features based on the contextual information, obtaining performance assessment features based on the performance status information, and fusing the contextual assessment features and the performance assessment features to determine the risk rating for the current session is based on a risk rating strategy. The risk rating strategy includes rating each consultation input by the user to determine the risk rating result for the current consultation in the current session. The risk assessment strategy can be dynamically adjusted based on at least one of the user's input status information, system cache data, and current session status to change the frequency or method of risk rating execution for the current session.

9. The method according to claim 1 or 2, wherein, The intelligent customer service, acting as the primary intelligent agent, interacts with a group of sub-intelligent agents to generate responses to the current inquiry, including: The context information is sent to the main intelligent agent for semantic understanding, and the main intelligent agent distributes the generated multiple sub-tasks to the corresponding sub-intelligent agents for execution, thereby obtaining the corresponding execution results. The main intelligent agent outputs the response information for the current inquiry based on the execution result and the second prompt word template; The second prompt word template is learned by the main agent or the corresponding sub-agent based on the historical response information of the human customer service representative, and is used to constrain the output attributes of the response information, including language style.

10. A session processing device, comprising: At least one memory, and a computer program stored in the memory; At least one processing device executes the computer program to implement the conversation processing method according to any one of claims 1-9 by interacting with human customer service and intelligent customer service of the customer service system.

Images