Call routing method and device

By using a large language model to parse user intent and combining performance data to correct agent capability vectors in the call routing method, dynamic matching of user needs and agent capabilities is achieved, solving the problems of low service efficiency and decreased user satisfaction caused by static routing strategies, and reducing operating costs.

CN123123510APending Publication Date: 2026-11-17LENOVO (BEIJING) LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611340705.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-31
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing call routing methods are based on static strategies, which leads to mismatches between user needs and agent expertise, resulting in low service efficiency, decreased user satisfaction, and increased operating costs.

Method used

By matching the user's intent vector with the agent's capability vector, the call request is dynamically routed to the agent with the matching capability. The user's intent is deeply analyzed using a large language model, and the agent's capability label vector is corrected using performance data to achieve accurate matching.

Benefits of technology

It improved service efficiency for users, increased user satisfaction, and reduced operating costs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123123510A_ABST
    Figure CN123123510A_ABST
Patent Text Reader

Abstract

This application provides a call routing method and apparatus; wherein the method includes: in response to a user's call request, determining the user's intent vector; the intent vector is used to characterize the user's business intent and / or emotional intent; matching the intent vector with the capability vectors of each agent to determine candidate agents from the agents; the capability vector is obtained by modifying the agent's capability tag vector through the agent's performance data; the performance data is used to reflect the agent's actual processing capability for call requests; determining a target agent from the candidate agents, and routing the call request to the target agent.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to network communication technology, specifically to a call routing method and apparatus. Background Technology

[0002] Intelligent routing methods have replaced traditional manual queuing and allocation models, becoming a key technology for improving call center efficiency. However, these call routing methods are based on static policies, which may lead to mismatches between caller needs and agent expertise. For example, assigning a user's problem to an agent who is not proficient in that problem can result in low service efficiency, decreased user satisfaction, and increased operating costs. Summary of the Invention

[0003] This application provides a call routing method, a call routing device, a call center system, a computer-readable storage medium, and a computer program product, which are beneficial to improving user service efficiency, increasing user satisfaction, and reducing operating costs.

[0004] The technical solution of this application embodiment is implemented as follows: In a first aspect, embodiments of this application provide a call routing method, the method comprising: in response to a user's call request, determining the user's intent vector; the intent vector being used to characterize the user's business intent and / or emotional intent; matching the intent vector with the capability vectors of each agent to determine candidate agents from the agents; the capability vector being obtained by modifying the agent's capability tag vector using the agent's performance data; the performance data being used to reflect the agent's actual processing capability for call requests; determining a target agent from the candidate agents, and routing the call request to the target agent.

[0005] Secondly, embodiments of this application provide a call routing device, the device comprising: a first determining module configured to determine the user's intent vector in response to a user's call request; the intent vector being used to characterize the user's business intent and / or emotional intent; a second determining module configured to match the intent vector with the capability vectors of each agent to determine candidate agents from the agents; the capability vector being obtained by correcting the agent's capability label vector using the agent's performance data; the performance data being used to reflect the agent's actual processing capability for call requests; and a routing module configured to determine a target agent from the candidate agents and route the call request to the target agent.

[0006] Thirdly, embodiments of this application provide a call center system, which includes a call routing device and at least one agent terminal provided in embodiments of this application.

[0007] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program or computer-executable instructions for implementing the call routing method provided in embodiments of this application when executed by a processor.

[0008] Fifthly, embodiments of this application provide a computer program product, including a computer program or computer executable instructions, which, when executed by a processor, implement the call routing method provided in embodiments of this application.

[0009] In this embodiment, upon receiving a user's call request, an intent vector representing the user's business intent and / or emotional intent is determined. This intent vector is then matched with the capability vectors of each agent, candidate agents are selected from among them, and a target agent is identified. The call request is then routed to the target agent. Since the capability vector is obtained by modifying the agent's capability label vector using the agent's performance data, this method effectively routes user call requests dynamically to agents with matching capabilities, thereby improving service efficiency, increasing user satisfaction, and reducing operating costs. Attached Figure Description

[0010] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the specification, serve to explain the technical solutions of this application. Obviously, the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0011] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0012] Figure 1 A schematic diagram of the implementation process of a call routing method provided in this application embodiment. Figure 1 ; Figure 2 A schematic diagram illustrating an implementation process for obtaining a capability vector, provided in an embodiment of this application; Figure 3 A schematic diagram illustrating an implementation process for obtaining capability tag vectors, provided in an embodiment of this application; Figure 4 A schematic diagram illustrating the implementation process of determining a target agent, provided in an embodiment of this application; Figure 5A schematic diagram of the implementation process of a call routing method provided in this application embodiment. Figure 2 ; Figure 6 This is a schematic diagram of the structure of a call routing device provided in an embodiment of this application; Figure 7 This is a schematic diagram of the structure of a call center system provided in an embodiment of this application. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0014] In related technologies, rule-based static routing strategies are commonly used to route customer call requests, mainly including the following four methods: Skill group routing involves manually assigning preset tags (such as "financial business" or "technical support") to agents, and then matching them based on these tags. Round-robin allocation: Distribute seats sequentially from all available seats. Prioritize idle time and assign requests to the agent with the longest waiting time; Prioritize regular customers and assign them to agents who have previously served them.

[0015] In summary, the inventors of this application, during their research and analysis of related technologies, discovered the following shortcomings in the relevant technologies: (1) Lack of semantic understanding ability, unable to deeply analyze the true intentions and context expressed by customers in the dialogue; for example, if the customer’s question is “My payment failed”, the system cannot make accurate routing allocation according to the customer’s needs; (2) The agent's ability profile is rough and static, relying on manually preset skill tags, which cannot reflect the dynamic abilities shown by the agent in actual conversations (such as communication skills, proficiency in a certain type of sub-question), and is not updated in a timely manner. (3) Low matching accuracy. The shortcomings in (1) and (2) above directly lead to blind route matching, which may cause a series of problems such as high transfer rate, long average processing time and low first-time resolution rate. (4) The text vector is disconnected from the actual performance. The ability vector generated solely from historical dialogues cannot reflect the actual service performance of the agents. There may be situations where the vector shows strengths and / or poor actual performance. (5) When the best seat is busy, the decision is unreasonable. The difference in complaint risk between waiting for the best seat and being assigned a second-best seat is not considered. Blindly assigning seats may cause customer dissatisfaction.

[0016] In view of this, embodiments of this application provide a call routing method, which is applied to a call center system. Figure 1 A schematic diagram of the implementation process of a call routing method provided in this application embodiment. Figure 1 ,like Figure 1 As shown, the call routing method includes the following steps 101 to 103: Step 101: In response to the user's call request, determine the user's intent vector; the intent vector is used to characterize the user's business intent and / or emotional intent; Step 102: Match the intent vector with the capability vectors of each agent to determine candidate agents from among the agents; the capability vector is obtained by correcting the agent's capability label vector through the agent's performance data; the performance data is used to reflect the agent's actual processing ability for call requests. Step 103: Determine the target agent from the candidate agents and route the call request to the target agent.

[0017] It is understood that, in this embodiment of the application, upon receiving a user's call request, an intent vector representing the user's business intent and / or emotional intent is determined; the intent vector is matched with the capability vectors of each agent, candidate agents are selected from among the agents, and then a target agent is determined, and the call request is routed to the target agent. In this way, since the capability vector is obtained by modifying the agent's capability label vector using the agent's performance data, it is beneficial to dynamically route the user's call request to an agent with matching capabilities, thereby improving user service efficiency, increasing user satisfaction, and reducing operating costs.

[0018] The following sections will describe further optional implementation methods for each of the above steps, as well as related terms.

[0019] In step 101, in response to the user's call request, the user's intent vector is determined; the intent vector is used to characterize the user's business intent and / or emotional intent.

[0020] It should be understood that, in this embodiment, the call request can be interpreted as a user's request to call a human agent. In this embodiment, the form of the call request is not limited. In some embodiments, the call request can be one or more of the following: text message, voice stream, and keystroke commands, etc.

[0021] In this embodiment, the intent vector is not limited. The intent vector is a mapping of the call request in a mathematical space. In some embodiments, determining the user's intent vector in response to the user's call request includes: in response to the user's call request, mapping the call request information corresponding to the call request to the user's intent vector using a second language model; the second language model is used to generate agent capability tags; and the first language model is used to generate the user's intent vector.

[0022] It is understood that in this embodiment, the second major language model is used to deeply analyze the user's call request information, thereby obtaining a more accurate user intent. This helps the routing agents to accurately match the user's needs, thus improving service efficiency, increasing user satisfaction, and reducing operating costs.

[0023] In some embodiments, the first large language model and the second large language model may be different models or different task instances of the same model.

[0024] In this embodiment, the call request information is not limited. In some embodiments, the call request information can be understood as related information about the call request; for example, the dialogue between the user and the robot, the user's description, etc. The dialogue between the user and the robot can be understood as the latest dialogue between the user and the robot (e.g., the latest N messages in the current dialogue), or it can be understood as all current dialogues between the user and the robot. The user's description can be understood as a description of the user's call request, for example, the user stating "what other data plans are available?"

[0025] Furthermore, in some embodiments, the step of responding to a user's call request by mapping the call request information corresponding to the call request to the user's intent vector using a second language model includes: responding to a user's call request by mapping the dialogue between the user and the robot to the user's intent vector using a second language model; or, mapping the user's description to the user's intent vector using a second language model.

[0026] It should be understood that the business intent described in this application embodiment is not limited. In some embodiments, the business intent includes one or more of the following: product consultation, complaint handling, technical troubleshooting, and sales conversion. In this application embodiment, the emotional intent is not limited. In some embodiments, the emotional intent includes one or more of the following: emotional empathy needs such as worry about loss, urgency, confusion, and satisfaction.

[0027] For example, in one possible implementation, when a user needs to connect to a human operator, the Large Language Model (LLM) (here, the second large language model) is used in real time to analyze the dialogue between the user and the robot. The LLM analyzes the user's statement, "My phone bill suddenly increased a lot this month. Are you charging me randomly? I've been searching on the application for a long time but still can't figure it out!" to generate a user intent vector. The values ​​at different positions in the intent vector correspond to different types of business intent and emotional intent. For example, if the LLM generates a user intent vector of [0.2, 0.9, 0.1, 0.0, 0.7], then 0.2 corresponds to an intent for product consultation, 0.9 to an intent for complaint handling, 0.1 to an intent for technical troubleshooting, 0.0 to an intent for sales conversion, and 0.7 to an intent for emotional empathy.

[0028] In some embodiments, the large language model uses product consultation, complaint handling, technical troubleshooting, sales conversion, and emotional empathy as prompt words to extract features from user expressions and identify user intent data; it then transforms the intent data into structured vector data; and normalizes the parameters in the structured vector data to the range of [0.0, 1.0] to output the final intent vector.

[0029] In step 102, the intent vector is matched with the capability vectors of each agent to determine candidate agents from among the agents; the capability vector is obtained by correcting the agent's capability label vector through the agent's performance data; the performance data is used to reflect the agent's actual processing capability for call requests.

[0030] It should be understood that, in this embodiment, the agent can be understood as a customer service representative and their agent terminal who resolves user call requests. In this embodiment, the agent's capability vector should have the same dimension and be feature-aligned with the user's intent vector; that is, the capability vector should represent the same meaning at the same position or dimension as the user's intent vector.

[0031] In this application embodiment, the actual processing capacity is not limited. In some embodiments, the actual processing capacity includes one or more of the following: First Contact Resolution (FCR), Customer Satisfaction Score (CSAT), Transfer Rate, and Timeout Processing Rate.

[0032] It is understood that in this embodiment, the agent's capability vector is obtained by modifying the agent's capability label vector using performance data reflecting the agent's actual processing capability for call requests. The intent vector is then matched with each agent's capability vector to identify candidate agents. This approach, which matches the user's intent vector with the agent's actual capability vector, is beneficial for selecting agents that match the user's needs, thereby avoiding the problems of low service efficiency, decreased user satisfaction, and increased operating costs caused by static routing strategies.

[0033] In some embodiments, the capability vector is used to characterize the corresponding agent's processing capacity and / or the number of call requests processed. Further, in some embodiments, Figure 2 This application provides a schematic diagram of an implementation process for obtaining a capability vector, as shown in the embodiment. Figure 2 As shown, the agent's capability vector can be obtained through the following steps 201 to 204: Step 201: Using the first language model, analyze the historical dialogue data of each agent to obtain the capability tag vector of each agent. Step 202: Obtain the performance data of each agent within the first time period; Step 203: Based on the performance data corresponding to each agent, determine the correction parameters of the capability tag vector for each agent. Step 204: Correct the corresponding capability tag vectors based on the correction parameters to obtain the capability vectors of each agent.

[0034] It is understood that in this embodiment, the historical dialogue data of each agent is analyzed using a first language model to obtain the capability tag vector of each agent; based on the performance data of each agent within a first time period, correction parameters for the capability tag vector of each agent are determined; and the corresponding capability tag vector is corrected based on these correction parameters to obtain the capability vector of each agent. In this way, the capability tag vector of an agent is corrected using the agent's actual performance data, ensuring that the obtained capability vector matches the agent's actual capabilities. This achieves accurate matching of call requests and avoids the defect of mismatch between capability tag vectors and actual service performance caused by generating capability tag vectors solely based on historical dialogue data.

[0035] It should be understood that the historical dialogue data is not limited in the embodiments of this application. In some embodiments, the historical dialogue data includes historical dialogue data within a second time period; the second time period can be one year, one month, one week, one day, or all historical dialogue data.

[0036] In this application embodiment, the first time period is not limited. In some embodiments, the first time period may be historical performance data from one year, one month, one week, one day, or all historical performance data.

[0037] In this embodiment, the performance data of each agent within the first time period can reflect the dynamic capabilities demonstrated by the agent in actual conversations (e.g., communication skills, proficiency in a certain type of sub-question, etc.). In this embodiment, obtaining the performance data of each agent within the first time period is for the purpose of correcting the capability label vector.

[0038] In this embodiment of the application, the first language model is used to analyze the historical dialogue data of each agent to obtain the capability tag vector of each agent. This can be understood as analyzing the historical dialogue data of each agent using the first language model to obtain the capability tag vector of the same agent.

[0039] In this embodiment of the application, determining the correction parameters of the capability tag vector of each of the seats based on the performance data corresponding to each of the seats can be understood as determining the correction parameters of the capability tag vector of the same seat based on the performance data of the seat for each seat.

[0040] In this embodiment of the application, the capability tag vectors of the corresponding seats are modified based on the modification parameters to obtain the capability vectors of each seat. This can be understood as modifying the capability tag vectors of the same seat based on the modification parameters corresponding to the seat for each seat to obtain the capability vector of the seat.

[0041] Furthermore, in some embodiments, Figure 3 This is a schematic diagram illustrating an implementation process for obtaining capability tag vectors, as provided in an embodiment of this application. Figure 3 As shown, the first language model can be used to analyze the historical dialogue data of each agent through steps 301 to 303 to obtain the capability tag vector of each agent: Step 301: Obtain the historical dialogue data of each agent; Step 302: Semantically segment each historical dialogue data to obtain multiple independent dialogue units corresponding to each historical dialogue data. Step 303: Using the first large language model, classify each of the multiple independent dialogue units, determine the business type label of each of the multiple independent dialogue units, and analyze each of the multiple independent dialogue units and the business type label corresponding to the multiple independent dialogue units to obtain the capability label vector of each agent.

[0042] It is understood that in this embodiment, historical dialogue data of each agent is acquired; and semantically segmented to obtain multiple independent dialogue units corresponding to each historical data; using the first large language model, the service type tags of the multiple independent dialogue units corresponding to each historical data are determined, and the multiple independent dialogue units and their corresponding service type tags are analyzed to obtain the capability tag vector of each agent. This is beneficial for obtaining more accurate capability tag vectors, and further, based on these more accurate capability tag vectors, more accurate capability vectors are determined; thereby achieving precise matching of call requests.

[0043] In some embodiments, after semantically segmenting each of the historical dialogue data to obtain multiple independent dialogue units corresponding to each of the historical dialogue data, the method further includes: cleaning or removing meaningless data in the independent dialogue units; for example: thank you, meaningless identifiers, etc.

[0044] It should be understood that the independent dialogue unit is not limited in the embodiments of this application. In some embodiments, the independent dialogue unit is a dialogue unit / slice with complete semantics; for example, a complete question-and-answer pair.

[0045] In this embodiment of the application, semantically segmenting each of the historical dialogue data to obtain multiple independent dialogue units corresponding to each of the historical dialogue data can be understood as semantically segmenting the historical dialogue data of each agent to obtain multiple independent dialogue units corresponding to the historical dialogue data.

[0046] In this embodiment of the application, the first large language model is used to classify each of the plurality of independent dialogue units, determine the business type labels of each of the plurality of independent dialogue units, and analyze each of the plurality of independent dialogue units and the corresponding business type labels to obtain the capability label vector of each agent. This can be understood as follows: for each agent, the first large language model is used to classify the plurality of independent dialogue units corresponding to the agent, determine the business type labels of the plurality of independent dialogue units, and analyze the plurality of independent dialogue units and the corresponding business type labels to obtain the capability label vector of each agent.

[0047] In some embodiments, determining the correction parameters of the capability tag vector of each agent based on the performance data corresponding to each agent includes: mapping each business indicator in the performance data to a correction coefficient according to business standards; and determining the correction parameters of the capability tag vector of each agent based on the correction coefficients corresponding to each agent and the weights of the business indicators of the correction coefficients.

[0048] It is understood that, in this embodiment of the application, according to business standards, each business indicator in the performance data of each agent is mapped to a correction coefficient; based on the correction coefficients corresponding to each agent and the weights of the business indicators corresponding to the correction coefficients, the correction parameters of the capability label vector are determined. In this way, dynamically calibrating the capability label vector using performance data is beneficial to ensuring that the capability vector is highly consistent with the actual service capabilities of the corresponding agents; thereby achieving accurate matching of call requests.

[0049] In some embodiments, the correction coefficients corresponding to the different business metrics may be the same or different.

[0050] In some embodiments, obtaining the historical dialogue data of each of the agents includes: obtaining the historical dialogue data of each of the agents from a historical dialogue database.

[0051] In this embodiment of the application, mapping each business indicator in the performance data to a correction coefficient according to the business standard can be understood as mapping each business indicator in the performance data of each agent to a correction coefficient according to the business standard.

[0052] In this embodiment of the application, determining the correction parameter of the capability tag vector of each of the seats based on the correction coefficients corresponding to each of the seats and the weight of the business indicators of the correction coefficients can be understood as determining the correction parameter of the capability tag vector of each seat for each seat based on the correction coefficients corresponding to the seat and the weight of the business indicators of the correction coefficients.

[0053] For example, in one possible implementation, raw dialogue records (an example of historical dialogue data) are extracted from the business database. A semantic slicing algorithm is executed to cut long dialogues into independent units with complete semantics (an example of multiple independent dialogue units). The dialogue slices (an example of multiple independent dialogue units) are input into a large language model (here, the first large language model), and problem classification labels are automatically assigned (an example of business type labels, obtained by the large language model from classifying the dialogue slices). This provides data support for generating basic capability vectors (an example of capability label vectors, generated by the large language model based on classification labels and dialogue slices). Core performance data of agents (an example of performance data) is retrieved periodically. Performance data includes the following types of indicators: first-time resolution rate, customer satisfaction, transfer rate, and timeout rate. Each indicator is mapped to a correction coefficient in the range of 0.8-1.2 according to business standards. A comprehensive correction parameter (an example of correction parameters) is calculated by weighting and fusing according to priority (e.g., FCR weight 0.4, CSAT weight 0.3, transfer rate weight 0.2, and timeout rate weight 0.1). The basic capability vector is calibrated using comprehensive correction parameters, and the final capability vector is generated and stored in the vector database. The 5-dimensional capability vector has the following dimensions: [product consultation, complaint handling, technical troubleshooting, sales conversion, and emotional empathy]. For example, Agent A's basic capability vector is [0.9, 0.7, 0.3, 0.8, 0.6]. Agent A's correction coefficients for all four indicators are 1.2. The comprehensive correction parameter is 1.2×0.4+1.2×0.3+1.2×0.2+1.2×0.1=1.2, and the calibrated final capability vector is [0.9×1.2, 0.7×1.2, 0.3×1.2, 0.8×1.2, 0.6×1.2].

[0054] In this embodiment, matching the intent vector with the capability vectors of each agent can be a real-time intent matching based on the vector similarity between the intent vector and the capability vectors of each agent. In some embodiments, the vector similarity can be determined using cosine similarity.

[0055] In some embodiments, matching the intent vector with the capability vectors of each agent to determine candidate agents from the agents includes: for each capability vector, determining the cosine similarity between the intent vector and the capability vector; and determining the agent to which the capability vector corresponding to the cosine similarity that satisfies the first selection condition belongs as a candidate agent.

[0056] It is understood that, in this embodiment of the application, the cosine similarity between the intent vector and each capability vector is determined, and the agent to which the capability vector corresponding to the cosine similarity that satisfies the first selection condition belongs is determined as a candidate agent. This is beneficial for identifying candidate agents matching the user's call request from multiple agents, and then identifying the target agent matching the user's call request from the candidate agents, thereby improving user service efficiency, increasing user satisfaction, and reducing operating costs.

[0057] It should be understood that the first selection condition is not limited in the embodiments of this application. In some embodiments, the first selection condition may be the cosine similarity being the optimal or second-best cosine similarity. In other embodiments, the first selection condition may also be the cosine similarity being greater than or equal to a first similarity threshold; wherein, the first similarity threshold may be determined based on the number of user complaints, or set as a fixed threshold; the first similarity threshold is positively correlated with the number of complaints. In still other embodiments, the first selection condition may also be the cosine similarity being the top K cosine similarities in descending order; K may be a pre-set fixed value, or it may be determined based on the number of user complaints; K is negatively correlated with the number of complaints.

[0058] For example, in one possible implementation, the similarity search interface of the vector database is called, and the matching degree is calculated using the cosine similarity algorithm, with the formula as follows: Where A is the user's intent vector and B is the agent's final capability vector. A confidence threshold of 0.9 is set, and agents with a cosine similarity higher than 0.9 are considered candidate agents.

[0059] In step 103, a target agent is determined from the candidate agents, and the call request is routed to the target agent.

[0060] It should be understood that, in this embodiment of the application, the target seat is one of the candidate seats. The target seat is determined from the candidate seats.

[0061] In some embodiments, determining the target seat from the candidate seats includes: determining the first seat as the target seat in response to the availability of the first seat to which the capability vector corresponding to the maximum cosine similarity belongs.

[0062] It is understood that in this embodiment, the agent with the highest cosine similarity best matches the user's call request. Therefore, when the first agent corresponding to the capability vector with the highest cosine similarity is available, the user's call request is routed to that first agent. This helps improve user service efficiency, increase user satisfaction, and reduce operating costs.

[0063] In some embodiments, when the cosine similarity is greater than or equal to a first similarity threshold, the maximum cosine similarity is greater than or equal to the first similarity threshold.

[0064] In this embodiment, the first agent to which the capability vector corresponding to the maximum cosine similarity belongs is the optimal agent. When the first agent to which the capability vector corresponding to the maximum cosine similarity belongs is available, the call request is routed to the first agent.

[0065] In some embodiments, Figure 4 This application provides a schematic diagram of an implementation process for determining a target agent, as shown in the embodiments below. Figure 4 As shown, the target seat can be determined from the candidate seats through the following steps 401 to 403: Step 401: If the first agent corresponding to the capability vector with the maximum cosine similarity is busy, determine the first risk value for waiting for the first agent. Step 402: Determine a second risk value for routing the call request to other candidate agents besides the first agent; Step 403: Based on the first risk value and each of the second risk values, determine the target seat from the candidate seats.

[0066] It is understood that in this embodiment, when the optimal agent is busy, a risk value is determined for the user to wait for the first agent, and a second risk value is determined for other candidate agents that meet the first selection criteria to handle the user's call request. Based on the first risk value and each of the second risk values, the target agent is determined from the candidate agents. In this way, when routing a user's call request, the complaint risk of waiting for the optimal agent and assigning other candidate agents is considered, thereby avoiding the problem of blindly assigning agents when the optimal agent is busy, which may cause user dissatisfaction, and thus balancing the agent matching degree of routing and user experience.

[0067] It should be understood that, in the embodiments of this application, "the first agent is busy" can be understood as "the first agent is on a call," that is, the first agent is processing other users' call requests.

[0068] In this embodiment of the application, the first risk value is used to characterize the risk arising from the user waiting for the first agent; the second risk value is used to characterize the risk arising from other candidate agents handling the call request.

[0069] In some embodiments, determining the first risk value for waiting for the first agent includes: determining a first ratio of the estimated waiting time for the first agent to a first value; the first value being a baseline value; determining a first product of the first ratio and the probability of the user's complaint; and determining the first product as the first risk value for the first agent.

[0070] In some embodiments, the first value is a normalized baseline parameter; for example, using 60 minutes as the normalized baseline, which is beneficial for limiting the boundary of the first risk value, thereby reducing the subsequent calculation cost for the first risk value.

[0071] This application does not limit the method of obtaining the user complaint probability. In some embodiments, the user's complaint probability is obtained based on the user's historical complaint data. In other embodiments, the user's complaint probability is determined based on machine learning or deep learning models, utilizing the user's historical data (which can be understood as all data, consumption records, consumption categories, dialogue data, and complaint data, etc.) and the user's intent.

[0072] Furthermore, in some embodiments, the machine learning or deep learning model used to determine the probability of a user's complaint can be trained based on multiple user data in the call center system and validated based on the actual complaint probabilities corresponding to the multiple user data respectively, so that the prediction accuracy of the machine learning or deep learning model used to determine the probability of a user's complaint meets the requirements, for example, the prediction accuracy is higher than 95%.

[0073] In this embodiment of the application, the specific value of the user's complaint probability is not limited. For example, the specific value of the user's complaint probability can be [0, 1], and the larger the value, the greater the user's complaint probability.

[0074] In some embodiments, determining a second risk value for routing the call request to other candidate agents besides the first agent includes: for each candidate agent, determining a first processing score for the call request based on the candidate agent's capability vector; determining a second ratio between the first processing score and a second processing score for the call request by the first agent; determining a first difference between a second value and the second ratio; the second value being 1; determining a second product of the first difference and the user's complaint probability; and determining the second product as the second risk value for the corresponding candidate agent.

[0075] In some embodiments, determining a first processing score for the call request for each candidate agent based on the agent's capability vector includes: determining a first processing score for each candidate agent based on the update lag time of the agent's capability vector, and the cosine similarity and decay coefficient between the agent and the user's intent.

[0076] In some embodiments, the method further includes: determining a second processing score for the call request by the first agent based on the update lag time of the first agent's capability vector, and the cosine similarity and decay coefficient between the first agent and the user's intent.

[0077] It is understood that in this embodiment of the application, during the process of determining the risk value, an update lag time and a time decay coefficient for the capability vector are introduced to achieve dynamic updating of the capability vector. This is beneficial for more intelligent routing of call requests based on time awareness.

[0078] In this embodiment of the application, the update lag time can be understood as the time difference between the update time of the capability vector and the current time.

[0079] For example, in one possible implementation, the processing score = similarity × (decay coefficient ^ ((current time - vector update time) / decay period)).

[0080] In some embodiments, determining the target seat from the candidate seats based on the first risk value and each of the second risk values ​​includes: if the first risk value is less than each of the second risk values, determining the first seat as the target seat; if at least one of the second risk values ​​is less than the first risk value, determining the target candidate seat from the second seats based on the second risk value and the status of the second seat to which the second risk value belongs; the status of the second seat includes idle or busy.

[0081] It is understood that in this embodiment, if the first risk value is less than each of the second risk values, the first agent is determined as the target agent; if each of the second risk values ​​is less than the first risk value, the target candidate agent is determined from the second agents based on the status of the second agents to which the second risk values ​​belong. Thus, when routing a user's call request, the complaint risk of waiting for the optimal agent versus assigning other candidate agents, as well as the status of the candidate agents and the optimal agent, are considered, thereby balancing the agent matching degree of the routing with the user experience.

[0082] In some embodiments, the step of determining a target candidate seat from the second seats based on the second risk value and the status of the second seat to which the second risk value belongs; the status of the second seat includes idle or busy, including: if all the second seats are idle or if the third seat to which the smallest second risk value belongs is idle, determining the third seat as a target candidate seat; if the third seat to which the smallest second risk value belongs is busy, determining a third risk value to wait for the third seat; and determining the target seat based on the relationship between the third risk value and the second smallest second risk value.

[0083] In some embodiments, where the first selection condition can be the optimal and second-best cosine similarity, determining the target agent from the candidate agents includes: if the first agent to which the capability vector corresponding to the maximum cosine similarity belongs is busy, determining a first risk value for waiting for the first agent; determining a second risk value for routing the call request to the second agent to which the capability vector corresponding to the second largest cosine similarity belongs; and determining the target agent from the candidate agents based on the first risk value and the second risk value.

[0084] It should be understood that, in the embodiments of this application, the second seat to which the capability vector corresponding to the second largest cosine similarity belongs can be understood as the suboptimal seat.

[0085] Furthermore, in some embodiments, when the first selection condition can be the optimal and second-best cosine similarity, the target agent is determined from the candidate agents based on the first risk value and the second risk value. If the first risk value is less than the second risk value, the first agent is determined as the target agent; if the second risk value is less than the first risk value and the second agent is available, the second agent is determined as the target agent; if the second risk value is less than the first risk value and the second agent is busy, the target agent for the call request is determined according to the static routing strategy.

[0086] In some embodiments, when the first selection condition is a cosine similarity greater than or equal to a first similarity threshold, the method further includes: if there is no cosine similarity satisfying the first selection condition, determining candidate agents based on a static routing strategy, and determining target agents based on a static routing strategy. This facilitates timely agent allocation to users when no matching agent is available, avoiding problems such as low service efficiency, decreased user satisfaction, and increased operating costs.

[0087] The following describes an exemplary application of the embodiments of this application in a real-world application scenario.

[0088] In the embodiments of this application, the inventors of this application discovered through creative work that the capability vector is generated only based on historical dialogue text (i.e., an example of historical dialogue data), which has the defect that the text inference does not match the actual service effect; and that directly assigning the second-best seat when the best seat is busy can easily cause customer complaints. Therefore, it is necessary to further optimize the accuracy of routing matching and the rationality of decision-making.

[0089] The call routing method provided in this application can solve the problem of mismatch between customer (i.e., a user example) needs and agent expertise caused by static routing strategies in the call center industry. The call routing method provided in this application can avoid assigning customer problems to agents who are not proficient in those problems, thereby preventing low service efficiency, decreased customer satisfaction, and increased operating costs.

[0090] In some embodiments, the call routing method provided in this application can be a dynamic routing method that is "data-driven, intelligently matched, dynamically calibrated, and risk-controllable," and the method includes the following steps 11 to 15: Step 11) Offline database construction and dynamic calibration: Analyze the historical dialogue records of all agents (i.e., an example of historical dialogue data) using the Large Language Model (LLM) (which can be understood as the first large language model), and generate basic capability vectors (i.e., an example of capability tag vectors) through intelligent slicing technology.

[0091] For example, assuming the 5-dimensional capability vector has dimensions of [product consultation, complaint handling, technical troubleshooting, sales conversion, emotional empathy], Agent A's basic capability vector is: [0.9, 0.7, 0.3, 0.8, 0.6].

[0092] Step 12: Dynamically adjust the agent vector using core performance data (i.e., an example of performance data) to obtain the capability vector.

[0093] Every day (i.e., an example of the first time period), retrieve the core performance data: First Contact Resolution (FCR), Customer Satisfaction Score (CSAT), Transfer Rate, and Overdue Processing Percentage, which are four operational indicators.

[0094] The calculation logic can be as follows: First, map each indicator to a correction coefficient in the range of 0.8-1.2 according to the business standard (e.g., FCR≥90% corresponds to 1.2, <60% corresponds to 0.8), and then weight and merge them according to business priority (e.g., FCR weight 0.4, CSAT 0.3, transfer rate 0.2, timeout percentage 0.1) to obtain the comprehensive correction parameter (i.e., an example of the correction parameter).

[0095] For example: The correction coefficients for all four indicators of seat A are 1.2. The comprehensive correction parameter = 1.2×0.4 + 1.2×0.3 + 1.2×0.2 + 1.2×0.1 = 1.2. Its final capability vector after calibration is [0.9×1.2, 0.7×1.2, 0.3×1.2, 0.8×1.2, 0.6×1.2].

[0096] Step 13, Online Matching: When a customer needs to be transferred to a human agent, the conversation between the customer and the robot is analyzed in real time using LLM (i.e., the second largest language model) to generate an "intent vector" for the customer's question (i.e., an example of a user's intent vector).

[0097] For example, if a customer says, "My phone bill suddenly increased a lot this month, are you deducting money randomly? I checked the application (APP) for a long time but still couldn't figure it out!", the intent vector generated by LLM (the second largest language model) is: [0.2, 0.9, 0.1, 0.0, 0.7].

[0098] Step 14, Similarity Calculation: The system quickly calculates the cosine similarity between the "intent vector" and all final capability vectors (i.e., an example of capability vectors) in the "capability vector library" to select the best and second-best seats.

[0099] Step 15, Risk Decision: Calculate the difference between the complaint risk of waiting for the best agent (i.e., an example of the first risk value) and the complaint risk of assigning the second-best agent (i.e., an example of the second risk value), and dynamically decide on the routing strategy based on the difference result.

[0100] Wherein, the optimal complaint risk (i.e., an example of the first risk value) = customer complaint probability × (optimal agent's estimated waiting time / 60); the suboptimal complaint risk (i.e., an example of the second risk value) = customer complaint probability × (1 - suboptimal score / optimal score); the suboptimal score is an example of the first processing score, and the optimal score is an example of the second processing score.

[0101] The decision-making rule could be: if the risk of waiting for the optimal complaint is less than the risk of allocating the second-optimal complaint, then wait for the optimal agent; otherwise, allocate the second-optimal agent.

[0102] In some embodiments, Table 1 shows the advantages of using the call routing method relative to related technologies provided in this application.

[0103]

[0104] In some embodiments, the call routing method provided in this application includes two stages: an offline training and calibration stage (building an accurate capability library) and an online reasoning and risk decision-making stage (real-time matching of routes).

[0105] In some embodiments, Figure 5 A schematic diagram of the implementation process of a call routing method provided in this application embodiment. Figure 2 ,like Figure 5 As shown, the first phase (construction and dynamic calibration of the agent capability library) includes the following steps 21 to 27: Step 21, extract historical dialogue data; Step 22, semantic slicing processing (each 5-10 rounds of dialogue constitutes one unit); Step 23: LLM (an example of the first major language model) is vectorized to generate basic capability vectors; Step 24: Retrieve core performance data (FCR, CSAT, transfer rate, and percentage of overdue processing). Step 25: Calculate the comprehensive correction parameters (integrate the coefficients of the four indicators according to their weights). Step 26, correct the basic capability vector (i.e., calibrate the basic capability vector to generate the final capability vector). Step 27: Store in the agent capability vector library.

[0106] The second phase (i.e., real-time routing decision-making and risk control) includes the following steps 31 to 35: Step 31: Collect real-time customer-robot dialogue (i.e., dialogue between the user and the robot). Step 32: LLM (an example of the second largest language model) parses the customer intent to generate an intent vector; Step 33: Calculate the cosine similarity; Among them, the cosine similarity calculation is performed to match the best / second best agent; step 33 can be understood as calling the agent capability vector library and calculating the cosine similarity based on the intent vector; and calling the cosine similarity algorithm in the similarity matching calculation to calculate the matching degree.

[0107] Step 34: Determine if the optimal agent is available; Optimal availability: Directly assign the best available seat; Optimal Busyness: Calculate the risk difference between the two types of complaints; Step 35, Risk difference decision; Waiting for the best or assigning the second best: If the risk of waiting for the best complaint is less than the risk of assigning the second best complaint, then wait for the best agent; otherwise, assign the second best agent.

[0108] In some embodiments, this application provides a call center system, which includes the following modules: The dialogue slicing and preprocessing module is used to extract raw dialogue records (i.e., historical dialogue data) from the business database, execute the semantic slicing algorithm to cut long dialogues into independent units with complete semantics (such as a complete question-answer pair); and input the dialogue slices into the large language model (i.e., the first large language model) to automatically add question classification labels (i.e., an example of a business type) to provide data support for generating basic capability vectors.

[0109] The performance data integration and correction calculation module is used to retrieve core agent performance data daily: first-time resolution rate, customer satisfaction, transfer rate, and overdue processing percentage; map each indicator to a correction coefficient in the range of 0.8-1.2 according to business standards; calculate comprehensive correction parameters by weighting and integrating according to priority (FCR weight 0.4, CSAT weight 0.3, transfer rate weight 0.2, overdue processing percentage weight 0.1); and calibrate the basic capability vector with the comprehensive correction parameters to generate the final capability vector and store it in the vector database.

[0110] Large Language Model (LLM) Service Module: Used to call the LLM Application Programming Interface (API) or deploy open-source models; based on Transformers-based bidirectional encoder representations, such as (Bidirectional Encoder Representations from Transformers, BERT); for capability library construction, the input is dialogue slice text (i.e., an example of dialogue units from historical dialogue data), and the output is a basic capability vector. For real-time routing: the input is the latest dialogue record between the customer and the robot, and the output is the customer intent vector.

[0111] The routing decision engine receives requests from the intelligent routing gateway, containing customer intent vectors; it then calls the similarity search interface of the vector database and uses the cosine similarity algorithm to calculate the matching degree, with the formula as follows: Where A is the user's intent vector and B is the agent's final capability vector.

[0112] Risk Decision Submodule: Used to calculate the risk of waiting for the best complaint and the risk of allocating the second best complaint, and to make routing decisions based on the difference; and to apply intelligent backup and decay mechanisms: set a confidence threshold (e.g., 0.9), and if the highest cosine similarity is lower than this value, trigger the backup routing strategy (which can be a static routing strategy); and introduce a time decay factor to realize the dynamic update of the capability vector, and process the score = similarity × (decay coefficient ^ ((current time - vector update time) / decay period)).

[0113] In summary, the call routing method provided in this application addresses industry pain points, compensates for the deficiencies of related technologies, and possesses a clear technical path and high feasibility.

[0114] Based on the same inventive concept as the foregoing embodiments, this application provides a call routing device.

[0115] Figure 6 This is a schematic diagram of the structure of a call routing device provided in an embodiment of this application; as shown below. Figure 6 As shown, the call routing device 60 includes: a first determining module 601, a second determining module 602, and a routing module 603; wherein, The first determining module 601 is configured to determine the user's intent vector in response to a user's call request; the intent vector is used to characterize the user's business intent and / or emotional intent. The second determining module 602 is configured to match the intent vector with the capability vectors of each agent to determine candidate agents from among the agents; the capability vector is obtained by correcting the agent's capability tag vector through the agent's performance data; the performance data is used to reflect the agent's actual processing capability for call requests. The routing module 603 is configured to determine the target agent from the candidate agents and route the call request to the target agent.

[0116] In some embodiments, the call routing device 60 further includes a correction module; the correction module is configured to analyze the historical dialogue data of each agent using a first language model to obtain the capability tag vector of each agent; obtain the performance data of each agent within a first time period; determine the correction parameters of the capability tag vector of each agent based on the performance data corresponding to each agent; and correct the corresponding capability tag vector based on the correction parameters to obtain the capability vector of each agent.

[0117] In some embodiments, the correction module is further configured to acquire historical dialogue data of each of the agents; perform semantic segmentation on each of the historical dialogue data to obtain multiple independent dialogue units corresponding to each of the historical dialogue data; use the first large language model to classify each of the multiple independent dialogue units, determine the business type labels of each of the multiple independent dialogue units, and analyze each of the multiple independent dialogue units and the business type labels corresponding to the multiple independent dialogue units to obtain the capability label vector of each agent.

[0118] In some embodiments, the correction module is further configured to map each business indicator in each performance data into a correction coefficient according to business standards; and to determine the correction parameters of the capability tag vector of each agent based on the correction coefficients corresponding to each agent and the weights of the business indicators of the correction coefficients.

[0119] In some embodiments, the first determining module 601 is further configured to, in response to a user's call request, map the call request information corresponding to the call request into the user's intent vector using a second language model; the second language model is used to generate agent capability tags; and the first language model is used to generate the user intent vector.

[0120] In some embodiments, the second determining module 602 is further configured to determine, for each capability vector, the cosine similarity between the intent vector and the capability vector; and to determine the seat to which the capability vector corresponding to the cosine similarity that satisfies the first selection condition belongs as a candidate seat.

[0121] In some embodiments, the routing module 603 is further configured to determine the first seat as the target seat in response to the first seat belonging to the capability vector corresponding to the maximum cosine similarity being available.

[0122] In some embodiments, the routing module 603 is further configured to: if the first agent to which the capability vector corresponding to the maximum cosine similarity belongs is busy, determine a first risk value for waiting for the first agent; determine a second risk value for routing the call request to other candidate agents that are not the first agent; and determine a target agent from the candidate agents based on the first risk value and each of the second risk values.

[0123] In some embodiments, the routing module 603 is further configured to determine the first agent as the target agent if the first risk value is less than each of the second risk values; and to determine the target candidate agent from the second agents based on the second risk value and the status of the second agent to which the second risk value belongs if at least one of the second risk values ​​is less than the first risk value; the status of the second agent includes idle or busy.

[0124] In some embodiments, the routing module 603 is further configured to, for each of the other candidate seats that is not the first agent, determine the processing score of the candidate agent based on the time difference between the update time of the candidate agent's capability vector and the current time, and the cosine similarity between the candidate agent and the user's intent; and determine the second risk of each candidate agent based on the processing score of the candidate agent and the user's complaint probability, respectively.

[0125] The descriptions of the above device embodiments are similar to those of the above method embodiments, and have similar beneficial effects. For technical details not disclosed in the device embodiments of this application, please refer to the descriptions of the method embodiments of this application for understanding.

[0126] It should be noted that the module division in the embodiments of this application is illustrative and only represents one logical functional division. In actual implementation, there may be other division methods. Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, exist as separate physical units, or have two or more units integrated into one unit. The integrated units can be implemented in hardware, as software functional units, or a combination of software and hardware.

[0127] It should be noted that, in the embodiments of this application, if the above-described methods are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiments of this application, or the part that contributes to related technologies, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause the call center system to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), magnetic disks, or optical disks. Thus, the embodiments of this application are not limited to any specific hardware and software combination.

[0128] This application provides a call center system. Figure 7 This is a schematic diagram of a call center system provided in an embodiment of this application. The call center system 70 includes the call routing device 60 and at least one agent terminal 80 provided in the above embodiment. In some embodiments, the agent terminal 80 may be a mobile agent terminal or a fixed agent terminal.

[0129] This application provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the method provided in the above embodiments.

[0130] This application provides a computer program product containing instructions that, when run on a computer, cause the computer to perform the steps in the method provided in the above-described method embodiments.

[0131] It should be noted that the descriptions of the storage medium and device embodiments above are similar to the descriptions of the method embodiments above, and have similar beneficial effects. For technical details not disclosed in the storage medium and device embodiments of this application, please refer to the descriptions of the method embodiments of this application for understanding.

[0132] It should be understood that phrases such as "an embodiment," "one embodiment," "some embodiments," "in one possible implementation," or "example" throughout the specification mean that a specific feature, structure, or characteristic related to an embodiment is included in at least one embodiment of this application. Therefore, phrases such as "in one embodiment," "in one embodiment," "in some embodiments," "in one possible implementation," or "example" appearing throughout the specification do not necessarily refer to the same embodiment.

[0133] Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of this application, the sequence numbers of the above processes do not imply a specific order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0134] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments. The descriptions of the various embodiments above tend to emphasize the differences between the various embodiments, and their similarities or commonalities can be referred to each other. For the sake of brevity, they will not be repeated here.

[0135] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that there can be three kinds of relationships. For example, object A and / or object B can represent three situations: object A exists alone, object A and object B exist simultaneously, and object B exists alone.

[0136] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0137] In the several embodiments provided in this application, it should be understood that the disclosed devices and apparatuses can be implemented in other ways. The embodiments described above are merely illustrative. The features disclosed in the several product embodiments provided in this application can be arbitrarily combined to obtain new product embodiments without conflict. The above descriptions are merely implementation methods of this application, but the protection scope of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the protection scope of this application.

Claims

1. A call routing method, the method comprising: In response to a user's call request, determine the user's intent vector; The intent vector is used to characterize the user's business intent and / or emotional intent; The intent vector is matched with the capability vector of each agent to determine candidate agents from the agents; the capability vector is obtained by modifying the agent's capability label vector through the agent's performance data. The performance data is used to reflect the agent's actual ability to process call requests; The target agent is determined from the candidate agents, and the call request is routed to the target agent.

2. The method according to claim 1, further comprising: Using the first major language model, the historical dialogue data of each agent is analyzed to obtain the capability tag vector of each agent; Obtain the performance data of each of the aforementioned agents within the first time period; Based on the performance data corresponding to each of the aforementioned agents, the correction parameters for the capability tag vector of each of the aforementioned agents are determined respectively; The corresponding capability tag vectors are modified based on the modification parameters to obtain the capability vectors of each agent.

3. The method according to claim 2, wherein the step of using the first large language model to analyze the historical dialogue data of each agent to obtain the capability tag vector of each agent further includes: Obtain the historical dialogue data of each of the aforementioned agents; Each of the historical dialogue data is semantically segmented to obtain multiple independent dialogue units corresponding to each of the historical dialogue data. Using the first large language model, each of the multiple independent dialogue units is classified to determine the business type label of each of the multiple independent dialogue units. Then, the multiple independent dialogue units and their corresponding business type labels are analyzed to obtain the capability label vector of each agent.

4. The method according to claim 2, wherein determining the correction parameters of the capability tag vector for each of the agents based on the performance data corresponding to each agent includes: According to business standards, each business indicator in the aforementioned performance data is mapped to a correction coefficient; Based on the respective correction coefficients and the weights of the business indicators corresponding to each of the aforementioned seats, the correction parameters of the capability tag vectors for each of the aforementioned seats are determined.

5. The method according to claim 1, wherein determining the user's intent vector in response to a user's call request comprises: In response to a user's call request, the second major language model is used to map the call request information corresponding to the call request into the user's intent vector; The second language model is used to generate agent capability labels; the first language model is used to generate user intent vectors.

6. The method according to any one of claims 1 to 5, wherein matching the intent vector with the capability vectors of each agent to determine candidate agents from the agents comprises: For each capability vector, determine the cosine similarity between the intent vector and the capability vector; The seats to which the capability vectors corresponding to the cosine similarity that meet the first selection condition belong are determined as candidate seats.

7. The method according to claim 6, wherein determining the target agent from the candidate agents comprises: In response to the availability of the first seat corresponding to the capability vector with the maximum cosine similarity, the first seat is determined as the target seat.

8. The method according to claim 6, wherein determining the target seat from the candidate seats comprises: If the first agent corresponding to the capability vector with the maximum cosine similarity is busy, determine the first risk value for waiting for the first agent. Determine a second risk value for routing the call request to other candidate agents besides the first agent; Based on the first risk value and each of the second risk values, the target seat is determined from the candidate seats.

9. The method according to claim 8, wherein determining the target agent from the candidate agents based on the first risk value and each of the second risk values ​​comprises: If the first risk value is less than each of the second risk values, the first seat will be identified as the target seat; If at least one of the second risk values ​​is less than the first risk value, a target candidate seat is determined from the second seats based on the second risk value and the status of the second seat to which the second risk value belongs; The status of the second seat includes whether it is idle or busy.

10. A call routing device, the device comprising: The first determining module is configured to determine the user's intent vector in response to the user's call request; The intent vector is used to characterize the user's business intent and / or emotional intent; The second determining module is configured to match the intent vector with the capability vectors of each agent to determine candidate agents from among the agents; the capability vector is obtained by modifying the agent's capability label vector using the agent's performance data; The performance data is used to reflect the agent's actual ability to process call requests; The routing module is configured to determine the target agent from the candidate agents and route the call request to the target agent.