A Knowledge Graph-Based Intelligent Customer Service Question-Answering Precision Recommendation and Processing Method

Abstract

This invention relates to a knowledge graph-based intelligent customer service question-and-answer precision recommendation and processing method, belonging to the field of electronic digital data technology. The method includes: constructing a conversation prediction knowledge graph based on the historical conversation counts from multiple channels; predicting the future conversation counts from multiple channels based on the conversation prediction knowledge graph and the current conversation counts from multiple channels; setting the number of robot agents and human agents based on the future conversation counts from multiple channels; accessing conversations from multiple channels, using robot agents to perform semantic understanding of user questions in the conversation, constructing a user demand profile, and determining whether to assign a human agent; if so, assigning the optimal human agent based on the status data of the human agent; if not, generating answers based on the user demand profile, conducting a dialogue, and generating recommended questions based on a question association graph. The question association graph records the relationships between multiple questions, which has the advantage of improving the quality and efficiency of customer service.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data technology, and in particular to a method for accurate recommendation and processing of intelligent customer service questions and answers based on knowledge graphs. Background Technology

[0002] With the rapid development of internet technology, communication channels between businesses and customers have become increasingly diversified, encompassing online customer service websites, social media platforms, mobile applications, and more. To improve customer service quality and efficiency, businesses need to efficiently manage large volumes of conversation messages from different channels and respond to customer needs quickly and accurately. Against this backdrop, intelligent conversation access systems have emerged, aiming to achieve centralized management of conversation accounts, real-time reception and processing of conversation messages, intelligent responses, and collaborative work with human responses.

[0003] While existing intelligent session access systems have achieved centralized session management and basic response functions to a certain extent, they still have many shortcomings. In terms of session volume management, there is a lack of effective prediction mechanisms. This puts enterprises in a dilemma regarding resource allocation, making it difficult to rationally allocate customer service personnel and server resources in advance. This either leads to idle customer service personnel and wasted resources, or insufficient resources during peak session periods prevent timely responses to customers, affecting customer satisfaction.

[0004] In the intelligent response phase, the existing systems' robot service capabilities are limited. On the one hand, their understanding of user questions is not precise enough. Although some systems can perform simple word segmentation and semantic analysis on user questions, they lack deep understanding capabilities, resulting in inaccurate matching answers and an inability to effectively solve user problems.

[0005] Therefore, there is a need to provide a knowledge graph-based intelligent customer service question-and-answer accuracy recommendation and processing method to improve the quality and efficiency of customer service. Summary of the Invention

[0006] This invention provides a method for accurate recommendation and processing of intelligent customer service questions and answers based on knowledge graphs, including: obtaining the number of historical conversations from multiple channels; constructing a conversation prediction knowledge graph based on the number of historical conversations from multiple channels, wherein the conversation prediction knowledge graph is used to record the changing correlation features of multiple channels; predicting the number of future conversations from multiple channels based on the conversation prediction knowledge graph and the current number of conversations from multiple channels; setting the number of robot agents and human agents based on the number of future conversations from multiple channels; accessing conversations from multiple channels, performing semantic understanding of user questions in the conversation through robot agents, constructing a user demand profile, and determining whether to assign a human agent; if so, assigning the optimal human agent based on the status data of the human agent; if not, generating answers based on the user demand profile, conducting dialogue, and generating recommended questions based on a question association graph, wherein the question association graph is used to record the correlation relationships between multiple questions.

[0007] Furthermore, the historical session count across multiple channels includes the session count across multiple historical time points within multiple historical periods. Based on the historical session count across multiple channels, a session prediction knowledge graph is constructed, including: determining the average processing time period length and the overlap between adjacent time periods for any two channels based on the historical session count across multiple channels; for any two channels, calculating the average association parameter of the session count between the two channels based on the historical session count, average processing time period length, and overlap between adjacent time periods; and determining the change association features of multiple channels based on the average association parameter of the session count between any two channels, and constructing a session prediction knowledge graph for multiple channels.

[0008] Furthermore, based on the historical session counts from multiple channels, the average processing time period length and the overlap between adjacent time periods for the two channels are determined, including: S11, determining the initial range of length values ​​and the range of overlap values; S12, sampling the average processing time period length and the overlap between adjacent time periods from the length value range and the overlap value range for testing; S13, for each historical period, processing the historical session counts of the two channels based on the average processing time period length and the overlap between adjacent time periods for testing, generating the average session counts for the two channels in multiple time periods corresponding to the historical period, and calculating the correlation parameters between the two channels for the corresponding historical period; S14, based on the correlation parameters between the two channels for each historical period, calculating the average processing time period length and the overlap between adjacent time periods for testing. S15. Determine whether the length of the average processing time period used in this round of testing and the number of overlaps between adjacent time periods meet the preset conditions. If yes, proceed to S16; otherwise, proceed to S12. S16. Based on the length of the average processing time period used in this round of testing and the number of overlaps between adjacent time periods, optimize the length range and overlap range. S17. Based on the length range and overlap range before and after optimization, determine whether the termination condition is met. If yes, proceed to S18; otherwise, proceed to S12. S18. Based on the length of the average processing time period used in all rounds of testing and the number of overlaps between adjacent time periods, determine the length of the average processing time period and the overlap between adjacent time periods for the two channels.

[0009] Furthermore, based on the historical session counts, average processing time period lengths, and overlap between adjacent time periods of the two channels, the mean correlation parameter for the session counts of the two channels is calculated. This includes: for each historical period, processing the historical session counts of the two channels based on the average processing time period lengths and overlap between adjacent time periods of the two channels to generate the average session counts of the two channels in multiple time periods corresponding to the historical period, and calculating the correlation parameter for the two channels in the corresponding historical period; and averaging the correlation parameters for each historical period of the two channels to generate the mean correlation parameter for the session counts of the two channels.

[0010] Furthermore, based on the mean association parameter of the number of sessions between any two channels, the change association features of multiple channels are determined, and a session prediction knowledge graph is constructed, including: based on the mean association parameter of the number of sessions between any two channels, determining the number of sessions associated with each channel, wherein the change association features of multiple channels include the number of sessions associated with each channel; and constructing a session prediction knowledge graph based on the number of sessions associated with each channel.

[0011] Furthermore, based on the conversation prediction knowledge graph and the current number of conversations across multiple channels, the future number of conversations across multiple channels is predicted, including: for each channel, based on the current number of conversations in the channel, the initial number of future conversations in the channel is predicted; based on the conversation prediction knowledge graph, the initial number of future conversations in each channel is corrected, and the future number of conversations across multiple channels is predicted.

[0012] Furthermore, determining whether to assign a human agent involves: calculating the matching degree between the user's questions and the response knowledge base based on the user's needs profile; performing sentiment analysis on the user's questions in the conversation to determine the user's current emotion; obtaining the user's current operation trajectory and calculating the matching degree between the user's current operation trajectory and the response knowledge base; performing sentiment analysis on the user's questions in the user's historical conversations to determine the user's emotion distribution characteristics; and determining whether to assign a human agent based on the matching degree between the user's questions and the response knowledge base, the user's current emotion, the matching degree between the user's operation trajectory and the response knowledge base, and the user's emotion distribution characteristics.

[0013] Furthermore, based on the status data of human agents, the optimal human agent is assigned, including: using an optimization algorithm to assign the optimal human agent based on the user's demand profile, the user's current mood and the distribution characteristics of the user's mood, and the status data of the human agents.

[0014] Furthermore, based on the question association graph, recommended questions are generated, including: determining candidate questions based on the user's current operation trajectory and user demand profile; and generating recommended questions based on the question association graph and candidate questions.

[0015] Furthermore, the method also includes: obtaining the voice conversation record of the optimal human agent; and performing quality inspection on the voice conversation record of the optimal human agent.

[0016] Compared to existing technologies, the knowledge graph-based intelligent customer service question-answering accuracy recommendation and processing method provided in this manual has at least the following beneficial effects:

[0017] 1. By collecting historical conversation data from multiple channels to construct a conversation prediction knowledge graph, the system can accurately predict the future number of conversations across various channels. Based on this, the number of robot and human agents can be rationally allocated to avoid resource waste or shortage. After a conversation is initiated, the robot agent first performs semantic understanding and demand profiling to determine whether to assign a human agent. If assignment is necessary, the optimal agent is allocated based on the human agent's status data using an optimization algorithm. This series of operations ensures that agent resources are dynamically allocated according to actual demand, reducing user waiting time, improving service response speed and overall service efficiency, and enabling the customer service system to handle user inquiries more efficiently.

[0018] 2. After accessing the conversation from multiple channels, the chatbot performs semantic understanding of the user's questions and constructs a user needs profile. It then uses factors such as the matching degree of the response knowledge base, user sentiment, and operational trajectory to determine whether to assign a human agent. If no assignment is needed, it generates an answer based on the user needs profile and uses a question association graph to generate recommended questions, guiding the user to engage in deeper conversation. This precise semantic understanding and personalized recommendation better meets user needs, provides answers that better match user intent, and allows users to experience more intelligent and considerate service, thereby improving user satisfaction and loyalty, and enhancing the user experience.

[0019] 3. Beyond simply handling user inquiries, we prioritize service quality assurance. By acquiring and quality-checking the voice conversations of top-performing human agents, we comprehensively analyze the agent service process using various technologies such as speaker separation, keyword detection, and emotion detection. This helps identify service issues promptly, such as inappropriate agent language, negative attitudes, or delayed responses, providing targeted training and improvement suggestions for agents. Simultaneously, the quality check results serve as a basis for evaluating agent performance, promoting service team optimization and management, and driving continuous improvement in the overall service quality of the intelligent customer service system. Attached Figure Description

[0020] This specification will be further described by way of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting; in these embodiments, the same reference numerals denote the same structures, wherein:

[0021] Figure 1 This is a flowchart illustrating a knowledge graph-based intelligent customer service question-and-answer precision recommendation and processing method in one embodiment of this application;

[0022] Figure 2 This is a flowchart illustrating, in one embodiment of this application, the determination of the average processing time length and the overlap between adjacent time periods for two channels.

[0023] Figure 3 This is a flowchart illustrating the determination of whether to assign a human agent in one embodiment of this application. Detailed Implementation

[0024] To more clearly illustrate the technical solutions of the embodiments in this specification, the accompanying drawings used in the description of the embodiments will be briefly introduced below.

[0025] Figure 1 This is a flowchart illustrating a knowledge graph-based intelligent customer service question-answering accuracy recommendation and processing method in one embodiment of this application, as shown below. Figure 1 As shown, the method for accurate recommendation and processing of intelligent customer service questions and answers based on knowledge graphs may include the following steps:

[0026] Step 110: Obtain the number of historical sessions from multiple channels.

[0027] The number of historical sessions across multiple channels includes the number of sessions at multiple historical points in multiple historical periods.

[0028] Specifically, multiple channels can include various visitor sources such as different WeChat official accounts, web pages, and mobile apps.

[0029] The historical period can be daily, weekly, monthly, or quarterly, with any two historical periods having the same length. Within each historical period, there are multiple specific time points. For example, when analyzing by day, the number of conversations can be broken down to the hourly rate; when analyzing by week, the conversation activity at different times of the day can be analyzed. By collecting data from these different historical time points, we can gain a more accurate understanding of the conversation activity levels across various channels at different times.

[0030] Step 120: Construct a conversation prediction knowledge graph based on the number of historical conversations from multiple channels.

[0031] Among them, the conversation prediction knowledge graph is used to record the changing correlation features of multiple channels.

[0032] Specifically, it includes:

[0033] Based on the number of historical sessions from multiple channels, determine the average processing time length and the overlap between adjacent time periods for any two channels.

[0034] For any two channels, calculate the mean correlation parameter of the number of sessions between the two channels based on the number of historical sessions, the average processing time period length, and the overlap of adjacent time periods.

[0035] Based on the mean correlation parameter of the number of sessions between any two channels, the changing correlation features of multiple channels are determined, and a knowledge graph for session prediction of multiple channels is constructed.

[0036] Specifically, for any two channels, a suitable time period needs to be determined to calculate the average number of sessions. The choice of this time period is crucial, as it affects the analysis of trends in the number of sessions across channels. For example, choosing a time period that is too short may fail to capture long-term trends, while choosing a time period that is too long may mask short-term fluctuations. By reasonably determining the length of the time period for averaging, the average number of sessions across channels can be more accurately reflected within a given time frame. Considering the overlap between adjacent time periods, setting the degree of overlap helps to analyze changes in the number of sessions more smoothly. For example, if time periods are divided into hours, adjacent time periods can overlap to better observe the continuous changes in the number of sessions during time transitions, avoiding misunderstandings caused by sudden data abrupt changes due to strict time period divisions.

[0037] Figure 2 This is a flowchart illustrating, in one embodiment of this application, the determination of the average processing time length and the overlap between adjacent time periods for two channels, as shown below. Figure 2 As shown, in some embodiments, the average processing time length and overlap of adjacent time periods of two channels are determined based on the number of historical sessions from multiple channels, including:

[0038] S11. Determine the initial range of length values ​​and the range of overlap values;

[0039] S12. From the range of length values ​​and the range of overlap values, the mean processing time period length and the overlap of adjacent time periods are used for sampling tests;

[0040] S13. For each historical period, based on the length of the average processing time period used for testing and the overlap of adjacent time periods, process the number of historical sessions for the two channels, generate the average number of sessions for the two channels in multiple time periods corresponding to the historical period, and calculate the correlation parameters of the two channels for the corresponding historical period.

[0041] S14. Based on the correlation parameters of each historical period for the two channels, calculate the correlation fluctuation value corresponding to the length of the average processing time period and the overlap of adjacent time periods used for testing.

[0042] S15. Determine whether the length of the average processing time period used in this round of testing and the number of overlaps between adjacent time periods meet the preset conditions. For example, if the length of the average processing time period used in this round of testing and the number of overlaps between adjacent time periods are greater than the number threshold (e.g., 5, 10, etc.), if yes, execute S16; otherwise, execute S12.

[0043] S16. Based on the mean processing time length and the overlap of adjacent time periods used in this round of testing, optimize the range of length and overlap values.

[0044] S17. Based on the length range and overlap range before and after optimization, determine whether the termination condition is met. For example, calculate the interval distance between the length ranges before and after optimization. Specifically, consider the endpoints of the length range A=[a1,a2] before optimization and the length range B=[b1,b2] after optimization as two points P(a1,a2) and Q(b1,b2) in two-dimensional space, calculate the Euclidean distance between the two points, and use it as the interval distance between the length ranges before and after optimization. Calculate the interval distance between the overlap ranges before and after optimization. The calculation method is the same as that for the interval distance between the length ranges before and after optimization, so it will not be repeated here. The average of the interval distance between the length ranges before and after optimization and the interval distance between the overlap ranges before and after optimization is used as the distance between the length ranges before and after optimization and the overlap ranges before and after optimization. The termination condition can be that the distance between the length ranges before and after optimization and the overlap ranges before and after optimization is less than a distance threshold (e.g., 2). If yes, execute S18; if no, execute S12.

[0045] S18. Based on the correlation fluctuation value corresponding to the average processing time period length and the overlap of adjacent time periods for all rounds of testing, determine the average processing time period length and the overlap of adjacent time periods for the two channels. For example, take the average processing time period length and adjacent time periods for testing with the smallest correlation fluctuation value as the average processing time period length and the overlap of adjacent time periods for the two channels.

[0046] Specifically, based on prior knowledge and experience in analyzing changes in the number of channel conversations, a preliminary range of possible values ​​for the average processing time period and the overlap between adjacent time periods is set. For example, the initial range for the average processing time period could be 1-24 hours (when analyzing with hours as the time granularity), and the initial range for the overlap between adjacent time periods could be 0%-50%.

[0047] Using random sampling or other reasonable sampling methods, a specific combination of the mean processing time period length and the overlap of adjacent time periods is selected from the initially set range of length and overlap values ​​as the test parameters. For example, a set of parameters such as a mean processing time period length of 6 hours and an overlap of 20% between adjacent time periods might be selected.

[0048] Based on the parameters obtained from sampling, the historical session counts of the two channels within each historical period are divided and processed. The historical period is divided into multiple time periods according to the average processing time length and the overlap between adjacent time periods, and the average session count of the two channels within each time period is calculated. Then, based on these average data, correlation parameters between the two channels in that historical period are calculated. Correlation parameters can be indicators such as correlation coefficients that measure the degree of correlation between changes in the session counts of the two channels.

[0049] The variance of the correlation parameters for each historical period of the two channels can be calculated and used as the mean value for testing. The correlation fluctuation value corresponding to the length of the time period and the degree of overlap between adjacent time periods is processed.

[0050] In determining the average processing time length and overlap between adjacent time periods for two channels, the initial ranges for both length and overlap are often quite broad. By conducting multiple rounds of sampling tests and collecting correlation fluctuation values ​​between the two channels under different parameter combinations, optimizing the range allows subsequent tests to be more focused, improving the efficiency of finding the optimal parameters and avoiding excessive testing within invalid or inefficient parameter ranges. This leads to a more accurate identification of the best parameters that precisely reflect the correlation characteristics of changes in the number of sessions between channels.

[0051] In this round of testing, the correlation fluctuation value corresponding to the length of the mean processing time period and the degree of overlap between adjacent time periods used in different tests is a key basis for optimization. The correlation fluctuation value reflects the stability of the relationship between two channels under a specific parameter combination. If the correlation fluctuation value corresponding to a certain parameter combination is small, it means that the change in the number of sessions between the two channels is relatively stable under that parameter setting, and the interval of this parameter combination may be closer to the optimal parameter interval; conversely, the interval of parameter combinations with large correlation fluctuation values ​​may not be suitable as the final parameter value range.

[0052] Based on the distribution of the correlation fluctuation values, the range of length values ​​and the range of overlap values ​​are narrowed. For example, if it is found that the parameter combinations with small correlation fluctuation values ​​are concentrated in the range of 5-10 hours for mean processing time and 15%-25% overlap between adjacent time periods, then the initial range of values ​​can be narrowed to this narrower range so that subsequent tests can explore the optimal parameters more accurately.

[0053] From the perspective of parameter selection, a broad range of length and overlap values ​​was initially set based on prior knowledge and experience. A sampling method was then used to select test parameters, comprehensively covering possible parameter combinations and providing a rich data foundation for subsequent analysis. When processing historical session counts, time periods were divided based on the sampling parameters, and session mean and correlation parameters were calculated, accurately measuring the correlation between changes in session counts across channels.

[0054] Using correlation fluctuation values ​​as the optimization basis effectively reflects the stability of channel relationships under different parameter combinations. Through multiple rounds of sampling tests and optimization of the value range based on correlation fluctuation values, subsequent tests become more focused. For example, narrowing the value range to an area with smaller correlation fluctuation values ​​avoids excessive testing in invalid or inefficient ranges, significantly improving the efficiency of finding optimal parameters and enabling a more accurate identification of the best parameters reflecting the correlation characteristics of changes in the number of sessions between channels.

[0055] The process ends when the distance is less than a distance threshold, ensuring that the final parameters are obtained under relatively stable value ranges. The test parameter with the smallest correlation fluctuation value is ultimately used as the result, guaranteeing that the selected parameters provide the most stable and accurate representation of the correlation between changes in the number of sessions between channels.

[0056] In some embodiments, based on the historical session counts of the two channels, the average processing time period length, and the overlap between adjacent time periods, a mean correlation parameter for the session counts of the two channels is calculated, including:

[0057] For each historical period, based on the average processing time length of the two channels and the overlap of adjacent time periods, the number of historical sessions of the two channels is processed to generate the average number of sessions of the two channels in multiple time periods corresponding to the historical period, and the correlation parameters of the two channels in the corresponding historical period are calculated.

[0058] The mean of the correlation parameters for each historical period for the two channels is calculated to generate the mean correlation parameter of the number of sessions for the two channels.

[0059] In some embodiments, based on the mean association parameter of the number of sessions between any two channels, the changing association features of multiple channels are determined, and a session prediction knowledge graph is constructed, including:

[0060] Based on the mean correlation parameter of the number of sessions between any two channels, determine the correlation of the number of sessions between each channel. Among them, the correlation features of changes in multiple channels include the correlation of the number of sessions between each channel.

[0061] A knowledge graph for prediction of conversations is constructed by associating the number of conversations with each channel.

[0062] Specifically, the mean correlation parameter is an important indicator for measuring the degree of correlation between changes in the number of sessions between any two channels. When the mean correlation parameter of two channels is high, it indicates that the changes in the number of sessions between these two channels have a strong synchronicity or correlation. For example, in the tourism business scenario, if the mean correlation parameter of the travel guide push channel and the online booking channel on WeChat official accounts is high, it may mean that users are more likely to make online bookings immediately after viewing travel guides, and there is a close relationship between the two in terms of changes in the number of sessions. For each channel, by comparing and analyzing the mean correlation parameter with all other channels, other channels whose mean correlation parameter reaches a certain threshold (which can be set according to the actual situation, such as a correlation coefficient greater than 0.6) are selected. These channels are the session quantity correlation channels for that channel. For example, for the web channel, after calculation and screening, it is found that its mean correlation parameter is high with the APP promotion channel. Therefore, the APP promotion channel is the session quantity correlation channel for the web channel.

[0063] The number of sessions for each channel forms part of the changing relationship characteristics of multiple channels. It clearly shows the relationship between each channel and other channels in terms of changes in session count, reflecting the mutual influence and synergy between channels. For example, knowing the related channels of a certain channel allows analysis of how a change in the number of sessions on that channel might affect which related channels and how.

[0064] In the conversation prediction knowledge graph, each channel is treated as a node in the graph. Nodes represent different visitor source channels, such as WeChat official accounts, web pages, and apps.

[0065] Edges between nodes are determined based on the number of sessions associated with each channel. If channel A is associated with channel B in terms of session count, then an edge is created in the graph pointing from channel B to channel A. The attributes of the edges can include specific numerical information such as the mean association parameter to quantify the degree of association between channels.

[0066] Step 130: Based on the conversation prediction knowledge graph and the current number of conversations across multiple channels, predict the future number of conversations across multiple channels.

[0067] Specifically, it includes:

[0068] For each channel, predict the initial number of future sessions for the channel based on the current number of sessions for that channel;

[0069] Based on the conversation prediction knowledge graph, the initial number of future conversations for each channel is corrected to predict the number of future conversations for multiple channels.

[0070] Specifically, the current number of sessions on a channel can include the number of sessions at multiple points in time over a period of time (e.g., one day, one week, etc.).

[0071] As an example, the average number of current sessions on a channel can be calculated as the initial number of future sessions. As another example, different weights can be assigned to data from different time points, with data closer to the current time point having a higher weight. As yet another example, the ARIMA (AutoRegressive Integrated Moving Average) model can be used to predict the initial number of future sessions on a channel based on the current number of sessions. The initial number of future sessions can include the number of sessions at multiple future time points.

[0072] A quantity correction model, based on a session prediction knowledge graph, can correct the initial future session count for each channel and predict the future session count for multiple channels. This model, using a graph neural network, is based on the session prediction knowledge graph and employs a model capable of processing graph-structured data and performing association analysis. It captures the feature information of nodes (channels) and their neighboring nodes (associated channels). The quantity correction model extracts edge information between nodes from the knowledge graph, including association direction and mean association parameters. It aggregates the feature information of associated channels based on edge relationships and weights the initial features of each channel using the mean association parameters. For example, if channel A is associated with channels B and C, the model will fuse the initial predicted features of B and C into the features of A according to the association parameters. After multi-layer network propagation and feature transformation, the model outputs the corrected predicted future session count for each channel.

[0073] Step 140: Based on the number of future conversations from multiple channels, set the number of robot agents and human agents.

[0074] Specifically, based on the ratio coefficients corresponding to the number of robot seats and the number of human seats, the number of robot seats and the number of human seats are configured by multiplying the future number of conversations from multiple channels by the ratio coefficients corresponding to the number of robot seats and the number of human seats. For example, if the ratio coefficients corresponding to the number of robot seats and the number of human seats are 50% and 30% respectively, the average number of conversations at multiple future time points is calculated and multiplied by 50% and 30% respectively to obtain the number of robot seats and the number of human seats.

[0075] The particle swarm optimization (PSO) algorithm is used to determine the ratio of robot seats to human seats that satisfies saturation and reduces user queuing time. Each possible combination of ratios is treated as a particle, and a fitness function is constructed that comprehensively considers two key indicators: saturation and queuing time. Saturation is measured by calculating the matching degree between the seat processing capacity and the future number of sessions after allocating seats according to the current ratio across all channels. If the processing capacity is much greater than the number of sessions, the saturation is low but may lead to resource waste; conversely, it may fail to meet demand. Queuing time can be estimated based on queuing theory models, combined with parameters such as the number of seats and session arrival rate. The fitness function combines these two indicators with certain weights, aiming to minimize the function value, i.e., achieving the optimal state of reasonable saturation and short queuing time. During particle swarm initialization, a set of ratio combinations is randomly generated as the initial solution. Each particle has a position (ratio value) and velocity (search direction and step size). Then, an iterative optimization process begins, where each particle updates its velocity and position based on its own historical best position (individual extreme value) and the historical best positions of other particles in the swarm (global extreme value). The velocity update formula considers the particle's own inertia, its awareness of individual extrema, and its pursuit of global extrema, enabling particles to both explore new solution spaces and move closer to better solutions. Position update adjusts the scaling factor value based on the new velocity. In each iteration, the fitness function value of each particle is calculated; if a better individual or global extremum is found, it is updated. As the number of iterations increases, the particle swarm gradually converges to near the optimal solution, finding a set of scaling factors that ensures that allocating robot and human agent seats according to this ratio satisfies saturation requirements for each channel, avoiding idle or insufficient seats, while minimizing user waiting time and improving service quality and user experience. Through continuous iterative optimization, the particle swarm algorithm can effectively search for the optimal scaling factor that satisfies multiple objectives.

[0076] Step 150: Access the conversation from multiple channels, use the robot agent to perform semantic understanding of the user's questions in the conversation, build a user demand profile, and determine whether to assign a human agent.

[0077] Specifically, the text input by the user is cleaned and standardized. For example, if a user inputs "I want to go to Sanya, are there any recommended attractions?", punctuation and irrelevant characters are removed, and then word segmentation is performed to obtain "I want to go to Sanya, are there any recommended attractions?". A trained machine learning model, such as Support Vector Machine (SVM) or a deep learning model (such as BERT), is used to classify the intent of the processed text. In this example, the model can identify that the user's intent is to inquire about tourist attraction recommendations. Named entity recognition technology is used to extract key entities from the text. In the above example, "Sanya" is identified as a tourist destination entity. The intent and entity information obtained from semantic understanding are combined with the user's historical interaction records and personal information. If the user previously inquired about hotels in Sanya and is now asking about attractions, and their personal information shows they are a young couple, it indicates that the user plans to travel to Sanya and is interested in attractions and accommodations. Based on the integrated information, the user is tagged, such as "Sanya travel intention," "attraction recommendation need," "couple travel," and "interested in accommodations." When users subsequently raise new requests, such as "What are some good places for couples to take photos in Sanya?", the robot agent will promptly update the user's request profile and add a "couples' photo spot request" tag.

[0078] Figure 3 This is a flowchart illustrating the determination of whether to assign a human agent in one embodiment of this application, such as... Figure 3 As shown, in some embodiments, step 150 specifically includes:

[0079] Based on user demand profiles, calculate the matching degree between user questions and the answer knowledge base;

[0080] Perform sentiment analysis on user questions during the conversation to determine the user's current sentiment;

[0081] Obtain the user's current operation trajectory and calculate the matching degree between the user's current operation trajectory and the response knowledge base;

[0082] Perform sentiment analysis on user questions in the user's historical conversations to determine the characteristics of user sentiment distribution;

[0083] Based on the matching degree between user questions and the answer knowledge base, the user's current mood, the matching degree between the user's operation trajectory and the answer knowledge base, and the user's mood distribution characteristics, it is determined whether to assign a human agent.

[0084] Specifically, by measuring the degree of fit between the user's question and the content in the response knowledge base, it is determined whether the robot agent can accurately and comprehensively answer the user's question based on the existing knowledge base. A high matching degree indicates that the robot agent is likely to solve the user's question; a low matching degree suggests that the user's question may be complex or novel, making it difficult for the robot agent to handle effectively. Text similarity algorithms, such as cosine similarity and Jaccard similarity, are used. The user's question is preprocessed (e.g., word segmentation, stop word removal) and converted into a vector form. Simultaneously, each entry in the response knowledge base undergoes the same preprocessing and vectorization. Then, the similarity between the user's question vector and the vectors of each entry in the knowledge base is calculated, and the maximum value is taken as the matching degree between the user's question and the response knowledge base. For example, in a travel consultation scenario, if a user asks "What are some lesser-known islands in Sanya?", and the response knowledge base contains detailed information about islands in Sanya, the matching degree between the question and the relevant content in the knowledge base can be determined by calculating the text similarity.

[0085] Understanding the user's emotional state during the current conversation is crucial, as emotions influence their satisfaction and acceptance of the service. If a user is highly agitated or dissatisfied, even if a chatbot can provide answers, it may not meet their needs. In such cases, assigning a human agent for reassurance and more in-depth service is more appropriate. Employing emotion analysis models, such as rule-based methods or deep learning models, is also important. Rule-based methods determine user emotions by pre-defining emotional keywords and rules; for example, the presence of words like "angry," "furious," or "dissatisfied" in the user's text indicates a negative emotion. Deep learning models, trained on large amounts of labeled emotion data, can more accurately identify complex emotional expressions. For instance, recurrent neural networks (RNNs) or convolutional neural networks (CNNs) can be used to classify user text into positive, negative, or neutral emotion categories.

[0086] A user's action trajectory reflects their behavioral path and points of interest within the system. Combined with the response knowledge base, this allows us to determine whether the user's current action is relevant to the knowledge base's content and whether the chatbot can provide effective guidance and assistance based on the action trajectory. If the action trajectory has a low match with the knowledge base, it indicates that the user may have encountered difficulties or has specific needs, requiring human intervention. The system records various user actions, such as clicked links, browsed pages, and entered keywords, converting these actions into sequences. These sequences are then matched against relevant content in the response knowledge base to calculate the match degree. For example, in a travel booking system, a user might first browse hotel pages in Sanya and then click the "Featured Restaurants" link. This sequence of actions is matched against relevant information about Sanya tourism in the response knowledge base to determine whether the chatbot can provide corresponding recommendations or answers based on these actions.

[0087] Understanding the trends and overall emotional tendencies of users during past interactions with the system helps to more comprehensively assess their current emotional state and service needs. If negative emotions are repeatedly observed in a user's historical conversations, it indicates potentially low user satisfaction with the system, requiring more careful handling in the current conversation and, if necessary, the allocation of a human agent. Collecting historical conversation data allows for sentiment analysis of user questions in each conversation, statistically analyzing the frequency of positive, negative, and neutral emotions to form a user emotion distribution characteristic. For example, analyzing all travel consultation conversations a user has had over the past month reveals that 30% of conversations were characterized by negative emotions, 20% by positive emotions, and 50% by neutral emotions, thus providing insight into the user's emotional tendencies.

[0088] Assign corresponding weights to each dimension, adjusting the weights according to actual conditions and business needs. For example, the weight for matching the user's question with the answer knowledge base is set to 0.4, the weight for the user's current emotion is set to 0.3, the weight for matching the user's operation trajectory with the answer knowledge base is set to 0.2, and the weight for the user's emotion distribution characteristics is set to 0.1. Then, calculate a comprehensive score based on the score and weight of each dimension. When the comprehensive score exceeds a certain threshold, it is determined that a human agent needs to be assigned; otherwise, the robot agent continues to handle the process. For example, if the comprehensive score is 0.7 and the threshold is set to 0.6, a human agent will be assigned.

[0089] Step 160: If yes, then allocate the optimal human agent based on the status data of the human agent.

[0090] Specifically, it includes:

[0091] The optimal human agent is assigned based on the user's needs profile, current mood, mood distribution characteristics, and agent status data using optimization algorithms (such as genetic algorithms, simulated annealing algorithms, particle swarm optimization algorithms, etc.).

[0092] Specifically, taking a genetic algorithm as an example, all available agent seats are used as the initial population, with each seat corresponding to an individual. The individual's genetic code can contain information such as the agent's status data and professional skills. Based on factors such as user needs profiles, the user's current mood, and the distribution characteristics of user emotions, a fitness value is calculated for each individual (agent seat). The higher the fitness value, the more suitable the agent is for handling the current user's problem. For example, agents who meet the user's specific needs, can handle the user's emotions, and are currently in a good state are assigned a higher fitness value. A subset of individuals are selected to enter the next generation of the population based on their fitness values. Roulette wheel selection or tournament selection methods are commonly used to ensure that individuals with higher fitness have a greater probability of being selected. Crossover and mutation operations are performed on the selected individuals to generate new individuals. Crossover can exchange some genes between two individuals, while mutation randomly changes some genes in an individual to increase population diversity and avoid getting trapped in local optima. The fitness evaluation, selection, crossover, and mutation operations are repeated until a termination condition is met (such as reaching the maximum number of iterations or the fitness value reaching a preset threshold). The optimal assignment scheme is the human agent corresponding to the individual with the highest fitness value.

[0093] Step 170: If not, generate an answer based on the user demand profile, conduct a dialogue, and generate recommended questions based on the question association graph.

[0094] The question association graph is used to record the relationships between multiple questions. Specifically, a question association graph is a knowledge system that clearly presents the inherent relationships between multiple questions in the form of a graph structure. In this graph, each question is regarded as a node, and the relationships between questions are represented by edges. These relationships cover various types, including semantic similarity, logical order, causal relationships, and hierarchical relationships. For example, in a tourism scenario, the questions "What are the popular attractions in Chengdu?" and "Ticket prices for popular attractions in Chengdu?" can be connected by an edge, indicating that there is a close relationship between them, because after asking about popular attractions, users are likely to further inquire about the ticket prices of these attractions.

[0095] In some embodiments, step 170 specifically includes:

[0096] Based on the user's current operation trajectory and user demand profile, determine the candidate questions;

[0097] Based on the question association graph and candidate questions, a recommended question is generated.

[0098] Specifically, key information is extracted from the user's current activity path, such as clicked page titles and entered search keywords. These keywords are then used as search criteria to search the question database and identify related questions as candidate questions. Similarly, information from the user's needs profile is used as search criteria to search the question database and identify related questions as candidate questions.

[0099] The identified candidate questions are used as input, and a search and matching process is performed in the question association graph to find other questions directly related to the candidate questions. For example, if the candidate question is "The best season to visit Chengdu", the question association graph may find related questions such as "Recommendations for attractions in Chengdu during the best season" and "Accommodation options in Chengdu during the best season" as recommended questions.

[0100] The method also includes:

[0101] Obtain the optimal voice conversation recordings from human agents;

[0102] Quality control is performed on the voice conversation recordings of the best human agents.

[0103] Specifically, separating agent voice and user voice allows for targeted analysis of the language content and emotional expression of both. For example, agent responses can be extracted separately to assess their accuracy and professionalism. Based on preset quality control terms, the system detects whether agent voice contains relevant keywords. These keywords may involve service standard terminology and key business knowledge points. For example, in a credit card service scenario, keywords might include business terms such as "billing date," "payment method," and "credit limit," as well as service etiquette phrases like "Hello," "Excuse me," and "Thank you." Keyword detection allows the system to check whether the agent communicates according to service standards. Analyzing various information in the agent voice helps determine the agent's emotional state. Agent emotions significantly impact service quality; positive and patient emotions improve user satisfaction, while negative and impatient emotions may lead to user dissatisfaction. For example, through characteristics such as tone of voice, speaking speed, and pauses, the system can initially determine whether the agent maintains a good service attitude. The system also detects whether the agent fails to respond promptly or experiences awkward silences during the conversation. Prolonged silence may indicate that the agent is taking too long to think, is not proficient in operating the system, or is handling user issues improperly, which will affect the user experience. Analyze the average and real-time speech rate of agents. An appropriate speech rate allows users to clearly understand the agent's answers; too fast a speech rate may result in users not being able to hear clearly, while too slow a speech rate may make users impatient. Evaluate the agent's service attitude, proficiency, and service quality based on the speech rate detection results.

[0104] By conducting the comprehensive quality control analysis on the voice conversation records of top-performing agents, the system can promptly identify problems in the service process, providing agents with targeted training and improvement suggestions, thereby enhancing overall service quality and user satisfaction. Simultaneously, these quality control results can also serve as an important basis for evaluating agent performance, promoting the optimization and management of the service team.

[0105] Finally, it should be understood that the embodiments described in this specification are merely illustrative of the principles of the embodiments described herein. Other variations may also fall within the scope of this specification. Therefore, alternative configurations of the embodiments described herein are intended to be illustrative rather than limiting, and should be considered consistent with the teachings of this specification. Accordingly, the embodiments described herein are not limited to those explicitly introduced and described herein.

Claims

1. A knowledge graph-based intelligent customer service question-and-answer precision recommendation and processing method, characterized in that, include: Get the number of historical sessions from multiple channels; Based on the number of historical conversations from multiple channels, a conversation prediction knowledge graph is constructed, which is used to record the changing correlation features of multiple channels. Based on the conversation prediction knowledge graph and the current number of conversations across multiple channels, predict the future number of conversations across multiple channels; Set the number of robot agents and human agents based on the number of future conversations from multiple channels; By accessing conversations from multiple channels and using robot agents to perform semantic understanding of user questions in the conversations, user demand profiles are built to determine whether to assign a human agent. If so, the optimal human agent will be assigned based on the status data of the human agents. If not, then based on the user demand profile, an answer is generated, a dialogue is initiated, and a recommended question is generated based on the question association graph, which is used to record the relationship between multiple questions; Among them, the number of historical sessions across multiple channels includes the number of sessions across multiple channels at multiple historical points in time across multiple historical periods; Based on the historical conversation data from multiple channels, a conversation prediction knowledge graph is constructed, including: Based on the number of historical sessions from multiple channels, determine the average processing time length and the overlap between adjacent time periods for any two channels. For any two channels, calculate the mean correlation parameter of the number of sessions between the two channels based on the number of historical sessions, the average processing time period length, and the overlap of adjacent time periods. Based on the mean correlation parameter of the number of sessions between any two channels, the changing correlation features of multiple channels are determined, and a knowledge graph for session prediction of multiple channels is constructed.

2. The method for accurate recommendation and processing of intelligent customer service questions and answers based on knowledge graphs according to claim 1, characterized in that, Based on the historical session count from multiple channels, the average processing time duration and overlap between adjacent time periods of the two channels are determined, including: S11. Determine the initial range of length values ​​and the range of overlap values; S12. From the range of length values ​​and the range of overlap values, the mean processing time period length and the overlap of adjacent time periods are used for sampling tests; S13. For each historical period, based on the length of the average processing time period used for testing and the overlap of adjacent time periods, process the number of historical sessions for the two channels, generate the average number of sessions for the two channels in multiple time periods corresponding to the historical period, and calculate the correlation parameters of the two channels for the corresponding historical period. S14. Based on the correlation parameters of each historical period for the two channels, calculate the correlation fluctuation value corresponding to the length of the average processing time period and the overlap of adjacent time periods used for testing. S15. Determine whether the length of the average processing time period and the number of overlaps between adjacent time periods used in this round of testing meet the preset conditions. If yes, execute S16; otherwise, execute S12. S16. Based on the mean processing time length and the overlap of adjacent time periods used in this round of testing, optimize the range of length and overlap values. S17. Based on the length range and overlap range before and after optimization, determine whether the termination condition is met. If yes, execute S18; otherwise, execute S12. S18. Based on the correlation fluctuation value corresponding to the average processing time period length and the overlap of adjacent time periods used for testing in all rounds, determine the average processing time period length and the overlap of adjacent time periods for the two channels.

3. The method for accurate recommendation and processing of intelligent customer service questions and answers based on knowledge graphs according to claim 1, characterized in that, Based on the historical session counts, average processing time period lengths, and overlap between adjacent time periods of the two channels, the mean correlation parameters for the session counts of the two channels are calculated, including: For each historical period, based on the average processing time length of the two channels and the overlap of adjacent time periods, the number of historical sessions of the two channels is processed to generate the average number of sessions of the two channels in multiple time periods corresponding to the historical period, and the correlation parameters of the two channels in the corresponding historical period are calculated. The mean of the correlation parameters for each historical period for the two channels is calculated to generate the mean correlation parameter of the number of sessions for the two channels.

4. The method for accurate recommendation and processing of intelligent customer service questions and answers based on knowledge graphs according to claim 1, characterized in that, Based on the mean association parameter of the number of sessions between any two channels, the changing association features of multiple channels are determined, and a session prediction knowledge graph is constructed, including: Based on the mean correlation parameter of the number of sessions between any two channels, determine the correlation of the number of sessions between each channel. Among them, the correlation features of changes in multiple channels include the correlation of the number of sessions between each channel. A knowledge graph for prediction of conversations is constructed by associating the number of conversations with each channel.

5. The method for accurate recommendation and processing of intelligent customer service questions and answers based on knowledge graphs according to claim 1, characterized in that, Based on the conversation prediction knowledge graph and the current number of conversations across multiple channels, predict the future number of conversations across multiple channels, including: For each channel, predict the initial number of future sessions for the channel based on the current number of sessions for that channel; Based on the conversation prediction knowledge graph, the initial number of future conversations for each channel is corrected to predict the number of future conversations for multiple channels.

6. The method for accurate recommendation and processing of intelligent customer service questions and answers based on knowledge graphs according to any one of claims 1-5, characterized in that, Determining whether to assign a human agent includes: Based on user demand profiles, calculate the matching degree between user questions and the answer knowledge base; Perform sentiment analysis on user questions during the conversation to determine the user's current sentiment; Obtain the user's current operation trajectory and calculate the matching degree between the user's current operation trajectory and the response knowledge base; Perform sentiment analysis on user questions in the user's historical conversations to determine the characteristics of user sentiment distribution; Based on the matching degree between user questions and the answer knowledge base, the user's current mood, the matching degree between the user's operation trajectory and the answer knowledge base, and the user's mood distribution characteristics, it is determined whether to assign a human agent.

7. The method for accurate recommendation and processing of intelligent customer service questions and answers based on knowledge graphs according to claim 6, characterized in that, Based on the status data of human agents, the optimal human agent is allocated, including: The optimal human agent is assigned based on the user's needs profile, current emotions, emotional distribution characteristics, and agent status data using an optimization algorithm.

8. The method for accurate recommendation and processing of intelligent customer service questions and answers based on knowledge graphs according to any one of claims 1-5, characterized in that, Based on the question association graph, recommended questions are generated, including: Based on the user's current operation trajectory and user demand profile, determine the candidate questions; Based on the question association graph and candidate questions, a recommended question is generated.

9. The method for accurate recommendation and processing of intelligent customer service questions and answers based on knowledge graphs according to any one of claims 1-5, characterized in that, Also includes: Obtain the optimal voice conversation recordings from human agents; Quality control is performed on the voice conversation recordings of the best human agents.

Images