Semantic clustering session intelligent scheduling system and method based on artificial intelligence

Abstract

The invention discloses a semantic clustering session intelligent scheduling system and method based on artificial intelligence, and relates to the technical field of semantic analysis. Voice call data and online customer service text streams are synchronously acquired, and an attention mechanism is adopted to fuse a joint semantic vector and a user value mapping vector, so that the semantic clustering session intelligent scheduling efficiency is improved. Generating a context coding vector in combination with the attenuation factor; constructing an initial topic node based on the context coding vector, dynamically updating a node weight through a session frequency attenuation factor, and generating a topic association graph with a weight edge; constructing a three-dimensional efficiency model for historical sessions associated with the map nodes, analyzing efficiency scores of customer service reply verbal skills, and triggering knowledge base sorting updating; and monitoring the value index change rate of the topic node in real time, and automatically triggering a fusing mechanism and dynamically allocating an exclusive seat when the sudden increment of the association session of the high-net-value user group is detected. The customer service response quality is improved, and the service experience of the user is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of semantic analysis, and particularly relates to a semantic clustering conversation intelligent scheduling system and method based on artificial intelligence. BACKGROUND

[0002] In today's era of information explosion, text data has become the main way for people to communicate and disseminate information. However, simply relying on keyword matching or surface feature analysis cannot meet the requirements of text understanding. In order to better understand the deep meaning of the text, semantic analysis technology emerges as the times require.

[0003] Traditional conversation scheduling and knowledge management rely on static rules and manual maintenance, and cannot effectively respond to business fluctuations and dynamic changes in user groups. The update of the dialogue library has the problems of long cycle and slow response. The promotion of high-quality dialogue seriously depends on manual analysis, and the best reply strategy cannot be captured in real time. In the scene of traffic surge or high-value user concentration interaction, the existing system is difficult to expand exclusive service resources in time, resulting in long user waiting time, decreased satisfaction and increased complaint rate.

[0004] Therefore, the present application discloses a semantic clustering conversation intelligent scheduling system and method based on artificial intelligence to solve the above problems. SUMMARY

[0005] The present application aims to provide a semantic clustering conversation intelligent scheduling system and method based on artificial intelligence to solve the problems in the prior art.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical scheme: a semantic clustering conversation intelligent scheduling method based on artificial intelligence, which comprises the following steps: S1: synchronously acquiring voice call data and online customer service text stream, using attention mechanism to fuse joint semantic vector and user value mapping vector, and combining decay factor to generate context encoding vector; S2: constructing an initial topic node based on the context encoding vector, dynamically updating node weight through conversation frequency decay factor, calculating topic comprehensive value index combining user level and consumption amount, and generating a topic correlation graph with weighted edges; S3: for the historical conversation associated with the graph node, a three-dimensional performance model is constructed by fusing solution time length decline rate, satisfaction improvement rate and complaint rate, the performance score of the customer service reply dialogue is analyzed, and the knowledge base sorting update is triggered; S4: real-time monitoring the value index change rate of the topic node, when detecting the sudden increase of high net worth user group associated conversation, triggering the fuse mechanism automatically and dynamically allocating exclusive agents.

[0007] According to the above scheme, in S1, the following contents are included: S101: After user authorization, collect the user's voice call data and online customer service text data; the voice call data is transcribed into a text sequence in real time using an automatic speech recognition algorithm to obtain a voice-text dialogue sequence with confidence parameters; the customer service text data is a sequence of text messages sent by the user through the online platform; the voice call data and online customer service text data are synchronized and aligned based on timestamps to form a fused text sequence; S102: Encode the fused text using a lightweight BERT model to generate a joint semantic vector; extract the user value vector for each user, which includes the normalized values ​​of user level and historical spending amount; perform a sigmoid transformation on the confidence parameter to obtain weight coefficients, and generate a user value mapping vector by weighting the user value vector based on these weight coefficients; input the joint semantic vector as a query and the user value mapping vector as a key and value into a multi-head attention mechanism to output a fused vector; perform a decay transformation on the fused vector using a time decay factor to obtain the final context encoding vector; the time decay factor γ at time t... t =1-exp(-at), where a is the attenuation coefficient, which is a preset constant, and exp() represents an exponential function with the natural number base.

[0008] The attention fusion of lightweight BERT and user value vector in this invention balances semantic understanding efficiency with the need for lightweight models, while introducing user value weights to enable personalized services for high-value users. The time decay factor ensures that the model's dialogue context retains both recent information and historical accumulation, improving the timeliness and accuracy of dialogue responses.

[0009] According to the above scheme, S2 includes the following: S201: Analyze the similarity between existing node vectors in the graph and the current context encoding vector. The similarity is equal to the vector product of the node vector and the current context encoding vector divided by the vector length product of the node vector and the current context encoding vector. If the similarity between the current context encoding vector and any node vector is less than the similarity threshold, execute the creation of a new node. Use the current context encoding vector as the initial value of the new node vector. Initialize the attribute information of the new node. The attribute information includes session count, cumulative consumption amount, and average session duration. If the similarity between the current context encoding vector and one of the node vectors is greater than or equal to the similarity threshold, update the attribute information of the current context encoding vector to the node vector with the highest similarity. S202: Analyze the value index of nodes based on the attribute information of the updated node vectors, and assign the value index of the i-th node V to... i The value index is denoted as Val i ; The edge weight between nodes V and V is updated as the node appears in the cross-node session, and the edge weight is calculated according to the number of cross-node session interactions and the node value index, and the edge weight between nodes V and V is recorded as W : i and V j : （i，j） : ; Wherein, CSC （i，j） represents the number of times that nodes V i and V j appear in the same cross-node session; N j represents the total number of sessions in which node V j appears, Val i represents the value index of V j .

[0010] Online incremental construction of the session graph realizes automatic clustering of different intent theme sessions without manual intervention; the consumption scale and the session efficiency are jointly included in the index system, and the description ability of the graph node to the real business demand is improved.

[0011] According to the above scheme, in S3, the following contents are included: S301: Statistics of customer service reply scripts in each node, obtain the historical session statistical indicators of each customer service reply script in the node and the corresponding historical session statistical indicator benchmark value, the historical session statistical indicators include average solution time, average user satisfaction and complaint rate, the solution time is the time from the initiation of the session to the solved mark, the user satisfaction is that after each session ends, the system pushes the satisfaction evaluation interface to the user, and the user gives the score of the integral system, the complaint rate is the proportion of user complaints caused by customer service reply scripts in the total number of customer service reply scripts; based on the historical session statistical indicators and the historical session statistical indicator benchmark value of each customer service reply script, a three-dimensional performance model is constructed: ; Wherein, S m represents the performance score of the mth customer service reply script in the same node, TT m represents the average solution time of the mth customer service reply script, csat m represents the average user satisfaction of the mth customer service reply script, R m represents the complaint rate of the mth customer service reply script; TT base , csat base and R base respectively represent the benchmark value of the average solution time, the average user satisfaction and the complaint rate of the mth customer service reply script; α1, α2 and α3 represent the performance coefficients, which are system preset constants; S302: integrate the performance score set of all customer service reply scripts in the history of the node; if the performance score of the customer service reply script is greater than the product of the maximum value of the performance scores of the rest of the customer service reply scripts and the quality coefficient, mark the customer service reply script as a gold medal script, and promote the ranking of the customer service reply script to the top of the knowledge base.

[0012] The gold medal script automatic evaluation and knowledge base ranking updating mechanism in the application can automatically promote the optimal script, form a closed loop optimization, and continuously improve the overall response quality of the customer service and the user satisfaction.

[0013] According to the above scheme, the following contents are included in S4: S401: Real-time monitoring of the ratio of the conversation of the high net worth user group in each node to the total conversation of the node, the value index change rate and the conversation quantity acceleration, if the ratio of the conversation of the high net worth user group to the total conversation of the node, the value index change rate and the conversation quantity acceleration all satisfy greater than the corresponding threshold, then activate the exclusive agent allocation mechanism; the threshold value of the value index change rate is obtained by weighted calculation of the peak value and the historical mean value of the last monitoring period, and the threshold values of the conversation ratio and the conversation quantity acceleration are system preset; The high net worth user is a user whose total historical consumption amount is higher than a threshold value, and the conversation quantity acceleration represents the derivative of the conversation quantity change rate; S402: According to the normalized processing value of the value index mutation variable of the node, the exclusive agent N agent , ; wherein represents the upward rounding function, beta1 and beta2 are agent coefficients, the agent coefficients are system preset constants, and deltaVal norm represents the normalized processing value of the value index mutation variable of the node.

[0014] The application more accurately identifies the sudden demand of the high value user group, realizes early warning, calculates the trigger threshold based on the weighted historical peak value and mean value, adaptively adjusts the monitoring standard, improves the sensitivity and robustness of the system to user behavior fluctuations, realizes dynamic and flexible management of resource allocation based on the exclusive agent allocation rule after the normalization of the value index mutation variable, can quickly mobilize service forces during high demand period, significantly reduces the user waiting time, and improves the service experience of high value users.

[0015] Another aspect of the application is a semantic clustering conversation intelligent scheduling system based on artificial intelligence, which is applied to the above-mentioned semantic clustering conversation intelligent scheduling method based on artificial intelligence, and comprises a feature analysis module, a graph updating module, a script ranking optimization module and a warning and scheduling module. The feature analysis module is configured to synchronously acquire voice call data and online customer service text stream, fuse joint semantic vectors using an attention mechanism, map the joint semantic vectors to user value mapping vectors, and generate context encoding vectors in combination with a decay factor; The graph updating module is configured to construct an initial topic node based on the context encoding vectors, dynamically update node weights through a conversation frequency decay factor, calculate a topic comprehensive value index in combination with user levels and consumption amounts, and generate a topic correlation graph with weighted edges; The dialogue sorting optimization module is configured to construct a three-dimensional performance model by fusing a solution duration decline rate, a satisfaction improvement rate and a complaint rate for historical conversations associated with graph nodes, analyze performance scores of customer service reply dialogues, and trigger knowledge base sorting updating; The early warning dispatching module is configured to monitor a value index change rate of a topic node in real time, automatically trigger a fuse mechanism and dynamically allocate dedicated agents when detecting a sudden increase in a high net worth user group associated conversation.

[0016] According to the above scheme, the feature analysis module includes a text fusion unit and a value fusion unit; The text fusion unit is configured to collect voice call data and online customer service text data of a user, synchronously time-align the voice call data and the online customer service text data based on timestamps, and form a fusion text sequence. The value fusion unit is configured to encode the fusion text based on a lightweight BERT model, generate joint semantic vectors, extract a user value vector of a corresponding user, weight the user value vector based on a weight coefficient to generate a user value mapping vector, input the joint semantic vectors as queries, the user value mapping vector as keys and values into a multi-head attention mechanism, and output a fusion vector; and perform decay transformation on the fusion vector in combination with a time decay factor to obtain a final context encoding vector.

[0017] According to the above scheme, the graph updating module includes a similarity analysis unit and a weight analysis unit; The similarity analysis unit is configured to analyze the similarity between an existing node vector in the graph and the current context encoding vector, and perform new node creation if the similarity between the current context encoding vector and any node vector is less than a similarity threshold; or update attribute information of the current context encoding vector to a node vector with the highest similarity if the similarity between the current context encoding vector and one of the node vectors is greater than or equal to the similarity threshold. The weight analysis unit is configured to analyze a value index of a node based on attribute information of an updated node vector, update edge weights of nodes appearing in a cross-node conversation, and calculate edge weights according to the number of cross-node conversation interactions and the node value index.

[0018] According to the scheme, the dialogue sorting optimization module comprises an efficiency analysis unit and a dialogue optimization unit. The efficiency analysis unit is configured to count the customer reply dialogues in each node, obtain historical session statistical indicators and corresponding historical session statistical indicator benchmark values of each customer reply dialogue in the node, construct a three-dimensional efficiency model based on the historical session statistical indicators and the historical session statistical indicator benchmark values of each customer reply dialogue, and analyze the efficiency score of the customer reply dialogue based on the three-dimensional efficiency model. The dialogue optimization unit is configured to integrate the efficiency score set of all customer reply dialogues in the node, mark a customer reply dialogue as a gold dialogue if the efficiency score of the customer reply dialogue is greater than the product of the maximum value of the efficiency scores of the remaining customer reply dialogues and a high-quality coefficient, and promote the sorting of the customer reply dialogue to the first position in the knowledge base.

[0019] According to the scheme, the early warning and dispatching module comprises a real-time monitoring unit and a seat allocation unit. The real-time monitoring unit is configured to monitor the ratio of the sessions of the high net worth user group to the total sessions of the node, the value index change rate and the session quantity acceleration in each node in real time, and activate a dedicated seat allocation mechanism if the ratio of the sessions of the high net worth user group to the total sessions of the node, the value index change rate and the session quantity acceleration all satisfy the condition of being greater than the corresponding threshold value. The seat allocation unit is configured to allocate a dedicated seat according to the normalized processing value of the value index mutation variable of the node.

[0020] Compared with the prior art, the advantages of the present application are as follows: in the present application, the attention fusion of the lightweight BERT and the user value vector can take into account the semantic understanding efficiency and the lightweight model demand on the one hand, and the user value weight is introduced to realize the personalized service inclination for high-value users on the other hand; the time decay factor ensures that the model retains recent information and takes into account historical accumulation, thereby improving the timeliness and accuracy of the dialogue response. The online incremental construction of the conversation graph realizes the automatic clustering of different intent theme conversations without human intervention; the consumption scale and the conversation efficiency are jointly included in the index system to improve the ability of the graph node to depict real business demands. The gold dialogue automatic evaluation and knowledge base sorting updating mechanism in the present application can automatically promote the optimal dialogue, form a closed loop optimization, and continuously improve the overall response quality of the customer service and the user satisfaction. The present application more accurately identifies the sudden demands of the high-value user group and realizes early warning; the trigger threshold is calculated based on the weighted historical peak value and the mean value, the monitoring standard is adaptively adjusted, the sensitivity and robustness of the system to user behavior fluctuations are improved; based on the dedicated seat allocation rule of the normalized value index mutation variable, dynamic and flexible management of resource allocation is realized, service forces can be quickly mobilized during high demand periods, user waiting time is significantly reduced, and the service experience of high-value users is improved. BRIEF DESCRIPTION OF DRAWINGS

[0021] The accompanying drawings are included to provide a further understanding of the application and are incorporated in and constitute a part of the specification, illustrate embodiments of the application and are used to explain the application, but are not intended to limit the application. In the drawings: Figure 1 A flowchart of a semantic clustering conversation intelligent scheduling method based on artificial intelligence according to the present application; Figure 2 A structural diagram of a semantic clustering conversation intelligent scheduling system based on artificial intelligence according to the present application. DETAILED DESCRIPTION

[0022] The technical solutions in the embodiments of the present application will be described in detail below with reference to the accompanying drawings of the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0023] Please refer to Figure 1 The present application provides a technical solution: a semantic clustering conversation intelligent scheduling method based on artificial intelligence, which comprises the following steps: S1: synchronously acquiring voice call data and online customer service text stream, using attention mechanism to fuse joint semantic vector and user value mapping vector, and combining decay factor to generate context encoding vector; In S1, the following contents are included: S101: After authorization by the user, the user's voice call data and online customer service text data are collected; the voice call data is converted into a text sequence in real time by an automatic speech recognition algorithm, obtaining a voice text dialogue sequence with a confidence parameter; the customer service text data is a sequence of text messages sent by the user through an online platform; the voice call data and online customer service text data are synchronized and time-aligned based on the time stamp to form a fusion text sequence; S102: encode the fusion text based on a lightweight BERT model to generate a joint semantic vector; extract the user value vector corresponding to the user, which includes the user level and the normalized value of the user's historical consumption amount; perform Sigmoid transformation on the confidence parameter to obtain a weight coefficient, and weight the user value vector based on the weight coefficient to generate a user value mapping vector; input the joint semantic vector as a query, the user value mapping vector as a key and a value into a multi-head attention mechanism, and output a fusion vector; perform decay transformation on the fusion vector in combination with a time decay factor to obtain a final context encoding vector; the time decay factor γ at time t t= 1 - exp(-at), where a is an attenuation coefficient, the attenuation coefficient is a preset constant, and exp() represents an exponential function with a natural number as a base.

[0024] S2: constructing an initial topic node based on the context encoding vector, dynamically updating the node weight through a session frequency attenuation factor, calculating a topic comprehensive value index combining a user level and a consumption amount, and generating a topic association graph with a weighted edge; In S2, the following contents are included: S201: analyzing the similarity between the existing node vector in the graph and the current context encoding vector, the similarity being equal to the vector product of the node vector and the current context encoding vector divided by the product of the vector length of the node vector and the vector length of the current context encoding vector; if the similarity of the current context encoding vector with any node vector is less than a similarity threshold, performing new node creation; taking the current context encoding vector as the initial value of the new node vector; initializing the attribute information of the new node; the attribute information including session count, cumulative consumption amount, and average session duration; if the similarity of the current context encoding vector with one of the node vectors is greater than or equal to the similarity threshold, updating the attribute information of the current context encoding vector to the node vector with the highest similarity; Embodiment 1: S202: analyzing the value index of the node based on the attribute information of the updated node vector, and taking the value index of the i-th node V i as Val i ; in this embodiment, the calculation formula of the value index Val i is as follows: ; wherein N i represents the total number of sessions in which the node V i appears, n ∈ [1, N i ], n being a positive integer; c n represents the consumption amount of the user in the n-th session in which the node V i appears, T i represents the average session duration of the sessions in which the node V i appears; the session duration is the duration from the initiation of the session to the formal closure of the session, and the session duration includes the duration of the review, quality inspection, and user confirmation links of the session record marked as solved; the review link indicates that a review process is triggered automatically by a reviewer with customer service management authority or a system; the quality inspection link is a link in which the quality of the session is evaluated by index after the review passes; and the user confirmation link is a link in which the user is initiated by the system for a second confirmation to seek the final approval of the user on the problem solving situation after the quality inspection passes; updating the edge weight of the node appearing in the cross-node session, calculating the edge weight according to the number of cross-node session interactions and the node value index, and calculating the edge weight of the node Vi and V j between V （i，j） : ; wherein, CSC （i，j） represents the number of times that the nodes V i and V j appear in the same cross-node session in sequence; N j represents the total number of sessions in which the node V j appears, Val i represents the value index of V j .

[0025] S3: For the historical sessions associated with the graph nodes, a three-dimensional performance model is constructed by fusing the solution time length decline rate, the satisfaction improvement rate and the complaint rate, the performance score of the customer service reply script is analyzed, and the knowledge base sorting update is triggered; In S3, the following contents are included: S301: The customer service reply scripts in each node are counted, the historical session statistical indicators of each customer service reply script in the node and the corresponding historical session statistical indicator baseline values are obtained, the historical session statistical indicators include the average solution time length, the average user satisfaction and the complaint rate, the solution time length is the time length from the initiation of the session to the solved label, the user satisfaction is that after each session ends, the system pushes a satisfaction evaluation interface to the user, and the user gives a score in the integral system, the complaint rate is the proportion of the user complaints caused by the customer service reply script in the total number of the customer service reply scripts; a three-dimensional performance model is constructed based on the historical session statistical indicators of each customer service reply script and the historical session statistical indicator baseline values: ; wherein, S m represents the performance score of the mth customer service reply script in the same node, TT m represents the average solution time length of the mth customer service reply script, csat m represents the average user satisfaction of the mth customer service reply script, R m represents the complaint rate of the mth customer service reply script; TT base , csat base and R base respectively represent the baseline values of the average solution time length, the average user satisfaction and the complaint rate of the mth customer service reply script; α1, α2 and α3 represent the performance coefficients, which are system preset constants; In embodiment 2, in this embodiment, the average solution time length baseline value is obtained by P90 quantile, the average user satisfaction baseline value is a system preset constant, and the complaint rate baseline value is equal to the complaint rate multiplied by the total consumption of the current topic associated user and divided by the regional baseline consumption amount; S302: integrate the performance score set of all customer service reply scripts in the history of the node; if the performance score of the customer service reply script is greater than the product of the maximum value of the performance scores of the rest of the customer service reply scripts and the quality coefficient, mark the customer service reply script as a gold medal script, and promote the ranking of the customer service reply script to the top of the knowledge base.

[0026] S4: Real-time monitoring of the value index change rate of the topic node, when detecting a sudden increase in the high net worth user group associated conversation, automatically triggering the fuse mechanism and dynamically allocating dedicated agents.

[0027] The following is included in S4: S401: Real-time monitoring of the ratio of high net worth user group conversation to total conversation in each node, value index change rate and conversation acceleration, if the ratio of high net worth user group conversation to total conversation in the node, value index change rate and conversation acceleration all meet the threshold value, then activate the dedicated agent allocation mechanism; the threshold value of the value index change rate is obtained by weighted calculation of the peak value and the historical mean value of the last monitoring period, and the threshold values of the conversation ratio and the conversation acceleration are system preset; S402: According to the normalized value of the value index mutation of the node, allocate dedicated agents N agent , ; wherein represents the upward rounding function, β1 and β2 are agent coefficients, and the agent coefficient is a system preset constant, and △Val norm represents the normalized value of the value index mutation of the node.

[0028] Please refer to Figure 2 , the present application provides a technical scheme: a semantic clustering conversation intelligent scheduling system based on artificial intelligence, which comprises a feature analysis module, a graph updating module, a script ranking optimization module and a early warning and scheduling module; The feature analysis module is used for synchronously acquiring voice call data and online customer service text stream, adopting attention mechanism to fuse joint semantic vector and user value mapping vector, and combining decay factor to generate context encoding vector; The graph updating module is used for constructing an initial topic node based on the context encoding vector, dynamically updating the node weight through the conversation frequency decay factor, calculating the topic comprehensive value index in combination with the user level and the consumption amount, and generating a topic association graph with weighted edges; The script ranking optimization module is used for constructing a three-dimensional performance model by fusing the solution time decline rate, the satisfaction improvement rate and the complaint rate for the historical conversation associated with the graph node, analyzing the performance score of the customer service reply script and triggering knowledge base ranking update; The early warning scheduling module is configured to monitor the rate of change of the value index of the topic node in real time, and automatically trigger a fuse mechanism and dynamically assign a dedicated agent when detecting a sudden increase in the high net worth user group associated conversation.

[0029] The feature analysis module includes a text fusion unit and a value fusion unit. The text fusion unit is configured to collect voice call data and online customer service text data of a user, synchronize and align the voice call data and the online customer service text data based on timestamps, and form a fusion text sequence. The value fusion unit is configured to encode the fusion text based on a lightweight BERT model, generate a joint semantic vector, extract a user value vector of the corresponding user, weight the user value vector based on a weight coefficient to generate a user value mapping vector, input the joint semantic vector as a query, the user value mapping vector as a key and a value into a multi-head attention mechanism, and output a fusion vector; and perform attenuation transformation on the fusion vector in combination with a time decay factor to obtain a final context encoding vector.

[0030] The graph updating module includes a similarity analysis unit and a weight analysis unit. The similarity analysis unit is configured to analyze the similarity between the existing node vectors in the graph and the current context encoding vector. If the similarity between the current context encoding vector and any node vector is less than a similarity threshold, a new node is created. If the similarity between the current context encoding vector and one of the node vectors is greater than or equal to the similarity threshold, the attribute information of the current context encoding vector is updated to the node vector with the highest similarity. The weight analysis unit is configured to analyze the value index of the node based on the attribute information of the updated node vector, update the edge weight of the node appearing in the cross-node conversation, and calculate the edge weight according to the number of cross-node conversation interactions and the node value index.

[0031] The script ranking optimization module includes an efficiency analysis unit and a script optimization unit. The efficiency analysis unit is configured to count the customer service reply scripts in each node, obtain the historical session statistical indicators and corresponding historical session statistical indicator benchmark values of each customer service reply script in the node; construct a three-dimensional efficiency model based on the historical session statistical indicators and historical session statistical indicator benchmark values of each customer service reply script; and analyze the efficiency score of the customer service reply script based on the three-dimensional efficiency model. The script optimization unit is configured to integrate the efficiency score set of all historical customer service reply scripts in the node. If the efficiency score of a customer service reply script is greater than the product of the maximum value of the efficiency scores of the remaining customer service reply scripts and a quality coefficient, the customer service reply script is marked as a gold medal script, and the ranking of the customer service reply script is promoted to the top of the knowledge base.

[0032] The early warning dispatching module comprises a real-time monitoring unit and a seat allocation unit; The real-time monitoring unit is configured to monitor the ratio of the sessions of the high net worth user group to the total sessions of the node, the rate of change of the value index, and the acceleration of the session volume in the node in real time, and if the ratio of the sessions of the high net worth user group to the total sessions of the node, the rate of change of the value index, and the acceleration of the session volume all satisfy the corresponding threshold, the dedicated seat allocation mechanism is activated. The seat allocation unit is configured to allocate the dedicated seat according to the normalized processing value of the value index mutation of the node.

[0033] It should be noted that, in this document, the terms such as first and second are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between such entities or operations. Moreover, the terms "comprising", "including" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or apparatus including a series of elements includes not only those elements, but also other elements not explicitly listed, or inherent to such a process, method, article or apparatus.

[0034] It is obvious to those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, and can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application. Therefore, the embodiments should be regarded as exemplary and non-limiting, and the scope of the present application is defined by the appended claims rather than the above description, and all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present application. Any reference signs in the claims should not be regarded as limiting the claims involved.

Claims

1. An intelligent scheduling method for semantic clustering sessions based on artificial intelligence, characterized in that, The method includes the following steps: S1: Simultaneously acquire voice call data and online customer service text stream, use attention mechanism to fuse joint semantic vector and user value mapping vector, and combine with decay factor to generate context encoding vector; S2: Construct initial topic nodes based on context encoding vectors, dynamically update node weights through conversation frequency decay factors, calculate the comprehensive value index of topics by combining user level and consumption amount, and generate a topic association graph with weighted edges; S3: Based on the historical conversations associated with the graph nodes, a three-dimensional performance model is constructed by integrating the resolution time reduction rate, satisfaction improvement rate, and complaint rate. The performance score of customer service response scripts is analyzed and knowledge base sorting is updated. S4: Real-time monitoring of the value index change rate of topic nodes. When a sudden increase in the number of sessions associated with high-net-worth users is detected, the circuit breaker mechanism is automatically triggered and dedicated seats are dynamically allocated.

2. The method for intelligent scheduling of semantic clustering sessions based on artificial intelligence according to claim 1, characterized in that: S1 includes the following: S101: After user authorization, collect the user's voice call data and online customer service text data; the voice call data is transcribed into a text sequence in real time using an automatic speech recognition algorithm to obtain a voice-text dialogue sequence with confidence parameters; the customer service text data is a sequence of text messages sent by the user through the online platform; the voice call data and online customer service text data are synchronized and aligned based on timestamps to form a fused text sequence; S102: Encode the fused text using a lightweight BERT model to generate a joint semantic vector; extract the user value vector for each user, which includes the normalized values ​​of user level and historical spending amount; perform a sigmoid transformation on the confidence parameter to obtain weight coefficients, and generate a user value mapping vector by weighting the user value vector based on these weight coefficients; input the joint semantic vector as a query and the user value mapping vector as a key and value into a multi-head attention mechanism to output a fused vector; perform a decay transformation on the fused vector using a time decay factor to obtain the final context encoding vector; the time decay factor γ at time t... t =1-exp(-at), where a is the attenuation coefficient, which is a preset constant, and exp() represents an exponential function with the natural number base.

3. The method for intelligent scheduling of semantic clustering sessions based on artificial intelligence according to claim 2, characterized in that: S2 includes the following: S201: Analyze the similarity between existing node vectors in the graph and the current context encoding vector. The similarity is equal to the vector product of the node vector and the current context encoding vector divided by the vector length product of the node vector and the current context encoding vector. If the similarity between the current context encoding vector and any node vector is less than the similarity threshold, execute the creation of a new node. Use the current context encoding vector as the initial value of the new node vector. Initialize the attribute information of the new node. The attribute information includes session count, cumulative consumption amount, and average session duration. If the similarity between the current context encoding vector and one of the node vectors is greater than or equal to the similarity threshold, update the attribute information of the current context encoding vector to the node vector with the highest similarity. S202: Analyze the value index of nodes based on the attribute information of the updated node vectors, and assign the value index of the i-th node V to... i The value index is denoted as Val i ; Update the edge weights of nodes appearing in cross-node sessions, and calculate the edge weights based on the number of cross-node session interactions and the node value index.

4. The semantic clustering session intelligent scheduling method based on artificial intelligence according to claim 3, characterized in that: S3 includes the following: S301: Statistically analyze customer service response scripts in each node, obtain historical session statistics indicators and corresponding historical session statistics indicator baseline values ​​for each customer service response script in its respective node. The historical session statistics indicators include average resolution time, average user satisfaction, and complaint rate. The resolution time is the time from the initiation of the session to the point where it is marked as resolved. The user satisfaction rate is determined by the system pushing a satisfaction evaluation interface to the user after each session, where the user provides a point-based rating; the complaint rate is the percentage of user complaints caused by customer service response scripts out of the total number of times customer service response scripts are used; a three-dimensional performance model is constructed based on the historical session statistical indicators and the baseline values ​​of the historical session statistical indicators for each customer service response script. S302: Integrate the set of performance scores of all historical customer service response scripts under the integrated node; if the performance score of a customer service response script is greater than the product of the maximum performance score of other customer service response scripts and the quality coefficient, mark the customer service response script as a gold script and promote the ranking of the customer service response script to the first position in the knowledge base.

5. The semantic clustering session intelligent scheduling method based on artificial intelligence according to claim 4, characterized in that: S4 includes the following: S401: Real-time monitoring of the ratio of high-net-worth user group sessions to the total sessions of the node, the rate of change of value index, and the acceleration of session volume in each node. If the ratio of high-net-worth user group sessions to the total sessions of the node, the rate of change of value index, and the acceleration of session volume all meet the corresponding thresholds, the dedicated seat allocation mechanism is activated. The threshold for the rate of change of value index is obtained by weighting the peak value of the previous monitoring period with the historical average value. The thresholds for the session ratio and the acceleration of session volume are preset by the system. S402: Assign dedicated seats N based on the normalized value of the node's value index mutation. agent , ;in This represents the floor function, where β1 and β2 are the agent coefficients, which are preset system constants, and ΔVal. norm This represents the normalized value of the node's value index mutation.

6. An artificial intelligence-based semantic clustering session intelligent scheduling system, wherein the system is applied to the implementation of the artificial intelligence-based semantic clustering session intelligent scheduling method according to any one of claims 1-5, characterized in that, The system includes a feature analysis module, a graph update module, a speech sorting optimization module, and an early warning and scheduling module; The feature analysis module is used to simultaneously acquire voice call data and online customer service text stream, and uses an attention mechanism to fuse joint semantic vector and user value mapping vector, and combines a decay factor to generate context encoding vector. The graph update module is used to construct initial topic nodes based on context encoding vectors, dynamically update node weights through session frequency decay factors, calculate the comprehensive value index of topics by combining user level and consumption amount, and generate a topic association graph with weighted edges. The script ranking optimization module is used to construct a three-dimensional performance model by integrating the resolution duration reduction rate, satisfaction improvement rate and complaint rate of historical conversations associated with graph nodes, analyze the performance score of customer service response scripts, and trigger knowledge base ranking updates. The early warning and scheduling module is used to monitor the change rate of the value index of topic nodes in real time. When a sudden increase in the number of sessions associated with high-net-worth users is detected, the circuit breaker mechanism is automatically triggered and dedicated seats are dynamically allocated.

7. The semantic clustering session intelligent scheduling system based on artificial intelligence according to claim 6, characterized in that: The feature analysis module includes a text fusion unit and a value fusion unit; The text fusion unit is used to collect users' voice call data and online customer service text data, and synchronize the voice call data and online customer service text data based on timestamps to form a fused text sequence. The value fusion unit is used to encode the fused text based on the lightweight BERT model, generate a joint semantic vector, extract the user value vector of the corresponding user, generate a user value mapping vector by weighting the user value vector based on the weight coefficient, use the joint semantic vector as a query and the user value mapping vector as a key and value input to a multi-head attention mechanism, and output a fused vector; the fused vector is then subjected to a decay transformation by combining a time decay factor to obtain the final context encoding vector.

8. The semantic clustering session intelligent scheduling system based on artificial intelligence according to claim 6, characterized in that: The map update module includes a similarity analysis unit and a weight analysis unit; The similarity analysis unit is used to analyze the similarity between existing node vectors in the graph and the current context encoding vector. If the similarity between the current context encoding vector and any node vector is less than the similarity threshold, a new node is created. If the similarity between the current context encoding vector and one of the node vectors is greater than or equal to the similarity threshold, the attribute information of the current context encoding vector is updated to the node vector with the highest similarity. The weight analysis unit is used to analyze the value index of nodes based on the attribute information of the updated node vectors, update the edge weights of nodes appearing in cross-node sessions, and calculate the edge weights based on the number of cross-node session interactions and the node value index.

9. The semantic clustering session intelligent scheduling system based on artificial intelligence according to claim 6, characterized in that: The script sorting and optimization module includes a performance analysis unit and a script optimization unit; The performance analysis unit is used to statistically analyze customer service response scripts in each node, obtain historical conversation statistics indicators and corresponding historical conversation statistics indicator benchmark values ​​for each customer service response script in its respective node, construct a three-dimensional performance model based on the historical conversation statistics indicators and historical conversation statistics indicator benchmark values ​​for each customer service response script, and analyze the performance score of the customer service response script based on the three-dimensional performance model. The script optimization unit is used to integrate the performance scores of all historical customer service response scripts under the node; if the performance score of a customer service response script is greater than the product of the maximum performance score of other customer service response scripts and the quality coefficient, the customer service response script is marked as a gold script and the ranking of the customer service response script is promoted to the first position in the knowledge base.

10. The semantic clustering session intelligent scheduling system based on artificial intelligence according to claim 6, characterized in that: The early warning and dispatch module includes a real-time monitoring unit and a seat allocation unit; The real-time monitoring unit is used to monitor in real time the ratio of high-net-worth user group sessions to the total sessions of the node, the rate of change of value index, and the acceleration of session volume in each node. If the ratio of high-net-worth user group sessions to the total sessions of the node, the rate of change of value index, and the acceleration of session volume all meet the corresponding threshold, the dedicated seat allocation mechanism is activated. The seat allocation unit is used to allocate dedicated seats based on the normalized value of the node's value index mutation.

Images