Intelligent analysis of voice call content and agent reminder system

Abstract

This invention relates to the field of artificial intelligence technology, and specifically discloses an intelligent analysis and agent reminder system for voice call content, comprising: a voice conversion module that converts full-band audio data of a customer's voice call into high-precision text data; a recognition and capture module that identifies key business information in the high-precision text data and captures the changing trends of the customer's emotional tendencies; a matrix construction module that maps all similar nodes of the customer in the customer segmentation logic tree, and constructs a customer-specific multi-dimensional response factor matrix based on successful response cases of the current business processing needs of the historical customer groups of all similar nodes and the similarity between the corresponding similar nodes and the customer's own node; and a scheme generation module that generates a personalized intelligent agent reminder scheme based on a personalized service decision model and the previously obtained information, including business processing step guidance and emotional reassurance language; enabling agents to provide more accurate services to customers based on the scheme.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a system for intelligent analysis of voice call content and agent reminders. Background Technology

[0002] In today's highly competitive business environment, customer service quality is a key factor for enterprises to establish themselves in the market and win customers. Voice calls, as a vital communication channel between enterprises and customers, require in-depth analysis of their content and timely, effective reminders for agents. With the continuous development of technologies such as speech recognition, natural language processing, and big data analytics, intelligent voice call content analysis and agent reminder systems have emerged. This system can convert customer speech into text in real time, accurately identify key business information, and capture changes in customer emotional tendencies, providing agents with more comprehensive and in-depth insights into customer needs. By integrating multi-source data such as basic customer information and historical agent service data, the system can generate personalized agent reminder plans, helping agents respond to customer needs more professionally and efficiently, improving customer satisfaction and loyalty. This not only optimizes enterprise customer service processes and enhances enterprise competitiveness but also has broad application prospects in many customer service-dependent industries such as finance, e-commerce, and telecommunications, leading customer service towards intelligent and refined development.

[0003] However, existing systems struggle to accurately identify key business information and capture real-time trends in customer sentiment. Insufficient data integration and analysis ultimately prevent the generation of accurate and effective personalized intelligent agent alert solutions, hindering improvements in customer service quality and efficiency, as the aforementioned key factors cannot be fully considered.

[0004] Therefore, this invention proposes an intelligent voice call content analysis and agent reminder system. Summary of the Invention

[0005] This invention provides an intelligent voice call content analysis and agent reminder system, which generates a personalized intelligent agent reminder solution that includes business processing step guidance and emotional reassurance messages. This enables agents to serve customers quickly, accurately, and considerately based on the solution, thereby improving customer satisfaction. It also improves business processing efficiency and service quality, helping enterprises gain an advantage in market competition.

[0006] This invention provides a voice call content intelligent analysis and agent reminder system, comprising:

[0007] The voice conversion module is used to convert real-time collected full-frequency audio data of customer voice calls into high-precision text data;

[0008] The identification and capture module is used to identify key business information in high-precision text data based on the associated business knowledge graph, and to capture the changing trends of customers' emotional tendencies in real time based on the dynamic sentiment recognition model.

[0009] The matrix construction module is used to map all similar nodes of a customer in the customer segmentation logic tree based on the customer's basic information and historical agent service data. Based on the successful response cases of the current business processing needs of the historical customer groups of all similar nodes of the customer and the similarity between the corresponding similar nodes and the customer's own node, a customer's exclusive multi-dimensional response factor matrix is ​​constructed.

[0010] The solution generation module is used to generate personalized intelligent agent reminder solutions that include business processing step guidance and emotional reassurance messages, based on key business information, emotional tendency change trends, customer-specific multi-dimensional response factor matrix, and personalized service decision model.

[0011] Preferably, the voice conversion module includes:

[0012] The noise reduction processing submodule is used to perform noise reduction preprocessing on the full-band audio data of real-time acquired customer voice calls to obtain preprocessed customer audio data.

[0013] The speech conversion submodule is used to convert preprocessed customer audio data into high-precision text data.

[0014] Preferably, the identification and capture module includes:

[0015] The entity recognition submodule is used to identify all business information entities in high-precision text data that belong to the related business knowledge graph;

[0016] The keyness calculation submodule is used to calculate the keyness of each business information entity based on its grouping degree, coreness, and frequency of occurrence in the associated business knowledge graph.

[0017] The information recognition submodule is used to reason and merge all business information entities in the high-precision text data that belong to the related business knowledge graph and whose keyness is not less than the keyness threshold, in order to obtain the key business information in the high-precision text data.

[0018] The emotion capture submodule is used to analyze multi-dimensional emotional expression information of customers collected in real time based on a dynamic emotion recognition model, and to capture the trend of changes in customers' emotional tendencies.

[0019] Preferably, the keyness calculation submodule includes:

[0020] The edge length quantization unit is used to quantify the edge length between corresponding connected entities based on the association metric between adjacent entities in the related business knowledge graph.

[0021] The clustering degree calculation unit is used to take the ratio of the average length of the path edges between each business information entity and all other business information entities in the associated business knowledge graph (excluding the current business information entity) to the maximum length of the path edges between any two business information entities in the associated business knowledge graph as the outlier degree of the corresponding business information entity, and take the difference between 1 and the outlier degree as the clustering degree of the corresponding business information entity.

[0022] The first core degree calculation unit is used to take the length of the path between each business information entity and all edge entities in the associated business knowledge graph as the margin value of each business information entity, and take the average degree of all margin values ​​of each business information entity as the first core degree of the corresponding business information entity.

[0023] The second core degree calculation unit is used to take the ratio of the average value of all margin values ​​of each business information entity to half the maximum value of the path length between any two entities in the associated business knowledge graph as the second core degree of the corresponding business information entity.

[0024] The third core degree calculation unit is used to take the ratio of the number of edges connected to each business information entity in the associated business knowledge graph to the maximum number of edges connected to a single entity in the associated business knowledge graph as the third core degree of the corresponding business information entity.

[0025] The criticality calculation unit is used to calculate the criticality of each business information entity based on its grouping degree, first core degree, second core degree, third core degree, and frequency of occurrence in high-precision text data.

[0026] Preferably, the matrix construction module includes:

[0027] The priority level determination submodule is used to extract all historical business types involved in a customer from the customer's historical agent service data, and retrieve the business level quantification table based on the number of times all historical business types were involved to determine the priority level of each historical business type involved in the customer.

[0028] The conformity determination submodule is used to identify all historical business types involved in the customer as all target business types involved in the customer in the business type classification logic tree, and to identify all superior nodes of each target business type involved in the business type classification logic tree. Based on the priority of the historical business types involved in each target business type involved in the business type classification logic tree, and the ratio of the superior layer of each superior node to the total layer of the longest complete node link, the conformity of the business type interval corresponding to each superior node is calculated.

[0029] The similarity calculation submodule is used to calculate the similarity between the customer's node and each node in the customer partitioning logic tree based on the customer's basic information, all historical business types involved and the corresponding business type ranges of all superior nodes, the corresponding compliance, the basic information of all customers in the customer interval corresponding to each node in the customer partitioning logic tree, the historical business types involved and the corresponding business type ranges of all superior nodes, and the corresponding compliance.

[0030] The similarity node filtering submodule is used to filter all nodes in the customer partitioning logic tree that have a similarity score exceeding a similarity threshold as all similar nodes of the customer.

[0031] The matrix construction submodule is used to construct a customer's exclusive multidimensional response factor matrix based on successful response cases of current business processing needs of all historical customer groups of the customer's similar nodes and the similarity between the corresponding similar nodes and the customer's own node.

[0032] Preferably, the similarity calculation submodule includes:

[0033] The first similarity calculation unit is used to calculate the similarity between the basic information of a customer and the basic information of each customer in the customer interval corresponding to each node in the customer segmentation logic tree, and use it as the first similarity between the two customers.

[0034] The second similarity calculation unit is used to calculate the similarity between all the historical business types involved in a customer and all the historical business types involved in each customer in the customer interval corresponding to each node in the customer segmentation logic tree, and use it as the second similarity between the two corresponding customers.

[0035] The third similarity calculation unit is used to calculate the similarity between the business type range and corresponding compliance degree of all superior nodes of the customer and the business type range and corresponding compliance degree of each customer in the customer interval of each node in the customer partitioning logic tree, and use it as the third similarity between the two customers.

[0036] The total similarity calculation unit is used to perform a weighted summation of the first similarity, second similarity, and third similarity between each customer in the customer interval corresponding to each node in the customer-customer division logic tree, to obtain the total similarity between the corresponding two customers.

[0037] The final similarity calculation unit is used to take the average of the total similarity between all customers in the customer interval corresponding to each node in the customer partitioning logic tree as the similarity between the customer's own node and the corresponding node in the customer partitioning logic tree.

[0038] Preferably, the matrix construction submodule includes:

[0039] The variable response factor value determination unit is used to summarize all historical customers in the customer interval of all similar nodes of the customer to obtain the customer's reference customer group. In the reference customer group, multidimensional variable response factors are extracted from all successful response cases for the customer's current business processing needs. Based on the preset assignment method, values ​​are assigned to all single-dimensional response factors of the multidimensional variable response factors to obtain the multidimensional variable response factor value of each successful response case.

[0040] The application weight determination unit is used to determine the application weight of each single-dimensional response factor value in the multi-dimensional variable response factor value of each successful response case based on the proximity between each proximity node and the customer's node.

[0041] The matrix building unit is used to generate a client's exclusive multidimensional response factor matrix based on the multidimensional variable response factor values ​​of all successful response cases and the application weights of each corresponding single-dimensional response factor value.

[0042] Preferably, the matrix building unit includes:

[0043] The feedback judgment value determination subunit is used to assign values ​​to each historical agent response feedback result of the customer, obtain the feedback judgment value of each historical agent response feedback result of the customer, assign values ​​to a large number of historical agent response feedback results of reference customers that are consistent with the agent response results in each historical agent response feedback result of the customer, and obtain all feedback reference judgment values ​​of the corresponding historical agent response feedback result of the customer.

[0044] The feedback relative quantification subunit is used to determine the customer's feedback bias direction and feedback deviation factor based on the feedback judgment value of each customer's historical agent response feedback result and all corresponding feedback reference judgment values;

[0045] The application weight adjustment subunit is used to adjust the application weight of each single-dimensional response factor value in the multidimensional variable response factor value of each successful response case based on the customer's feedback bias direction and feedback deviation factor, so as to obtain the adjusted application weight of each single-dimensional response factor value in the multidimensional variable response factor value of each successful response case.

[0046] The reference weight matrix generation sub-unit is used to adjust the applied weights for each single-dimensional response factor value among the multidimensional variable response factor values ​​of all successful response cases to form the reference weight matrix.

[0047] The original matrix construction sub-unit is used to summarize all single-dimensional response factor values ​​in the multidimensional variable response factor values ​​of all successful response cases to obtain the customer's original multidimensional response factor matrix.

[0048] The dedicated matrix construction sub-unit is used to generate a customer's dedicated multidimensional response factor matrix based on the customer's original multidimensional response factor matrix and reference weight matrix.

[0049] Preferably, the solution generation module includes:

[0050] The model building submodule is used to build personalized service decision models based on a large number of successful service decision cases;

[0051] The solution generation submodule is used to input key business information, emotional tendency change trends, and the customer's exclusive multi-dimensional response factor matrix into the personalized service decision model to generate a personalized intelligent agent reminder solution that includes business processing step guidance and emotional reassurance language.

[0052] Preferred options also include:

[0053] The dual-loop iteration module is used to adjust the personalized intelligent agent reminder scheme based on the latest agent response feedback results in the inner loop of the dual-loop mechanism. At the same time, in the outer loop of the dual-loop mechanism, the personalized service decision model is iteratively optimized based on historical agent service data, historical agent response feedback results, and business update rules.

[0054] The beneficial effects of this invention compared to existing technologies are as follows: The voice conversion module converts the full-band audio data of customer voice calls into high-precision text data. This process provides clear and accurate basic information for subsequent in-depth analysis, ensuring a high-quality starting point for analysis. The recognition and capture module, leveraging a related business knowledge graph and a dynamic emotion recognition model, can not only accurately identify key business information but also capture real-time trends in customer emotional tendencies. This helps agents comprehensively understand customer needs and emotions, enabling them to prepare responses in advance. The matrix construction module, through customer basic information and historical agent service data, finds similar nodes in the customer segmentation logic tree and constructs a dedicated multi-dimensional response factor matrix based on relevant successful response cases and similarity, providing rich and targeted data support for developing personalized solutions. The solution generation module integrates key business information, emotional tendencies, the dedicated matrix, and the personalized service decision model to generate a personalized intelligent agent reminder solution that includes business processing step guidance and emotional reassurance scripts. This allows agents to quickly, accurately, and considerately serve customers based on the solution, improving customer satisfaction, business processing efficiency, and service quality, helping enterprises gain an advantage in market competition.

[0055] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in this application.

[0056] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0057] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0058] Figure 1 This is a design architecture diagram of the intelligent voice call content analysis and agent reminder system in an embodiment of the present invention;

[0059] Figure 2 This is a design architecture diagram of the speech conversion module in an embodiment of the present invention;

[0060] Figure 3 This is a design architecture diagram of the identification and capture module in an embodiment of the present invention;

[0061] Figure 4 This is a flowchart illustrating the implementation of the scheme generation module in this embodiment of the invention. Detailed Implementation

[0062] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0063] Example 1:

[0064] This invention provides an intelligent voice call content analysis and agent reminder system, referenced Figure 1 ,include:

[0065] The voice conversion module is used to convert real-time collected full-frequency audio data of customer voice calls into high-precision text data;

[0066] The identification and capture module is used to identify key business information in high-precision text data based on the associated business knowledge graph, and to capture the changing trends of customers' emotional tendencies in real time based on the dynamic sentiment recognition model.

[0067] The matrix construction module is used to map all similar nodes of a customer in the customer segmentation logic tree based on the customer's basic information and historical agent service data. Based on the successful response cases of the current business processing needs of the historical customer groups of all similar nodes of the customer and the similarity between the corresponding similar nodes and the customer's own node, a customer's exclusive multi-dimensional response factor matrix is ​​constructed.

[0068] The solution generation module is used to generate personalized intelligent agent reminder solutions that include business processing step guidance and emotional reassurance messages, based on key business information, emotional tendency change trends, customer-specific multi-dimensional response factor matrix, and personalized service decision model.

[0069] In this embodiment, full-band audio data refers to audio information containing all frequency ranges that is collected in real time during a customer's voice call.

[0070] In this embodiment, the high-precision text data is obtained by converting full-band audio data through a speech conversion module, resulting in text with high accuracy and complete information.

[0071] In this embodiment, the related business knowledge graph is a structured knowledge system that contains various information entities within a business domain and the relationships between them. For example, in e-commerce, this graph may cover entities such as product categories, product characteristics, common customer questions, and solutions, and clearly define the relationships between these entities, such as which common questions correspond to a certain type of product and what solutions exist for these questions.

[0072] In this embodiment, key business information is information representing the core business problems encountered by customers, identified from high-precision text data. For example, if a customer asks, "My phone charges very slowly, what's the reason, and how can I solve it?", after processing by the recognition and capture module, the information "phone charges slowly" is combined with the associated business knowledge graph and determined to be key business information.

[0073] In this embodiment, the dynamic emotion recognition model is a model that can analyze multi-dimensional emotional expression information of customers in real time and capture changes in their emotional tendencies. For example, when a customer is talking to customer service, the model analyzes multi-dimensional information such as the customer's tone of voice, intonation, speaking speed, and word choice to determine the trend of changes in the customer's emotional state.

[0074] In this embodiment, the emotional tendency change trend refers to the dynamic changes in the customer's emotional state during the interaction with the agent. For example, the emotional tendency gradually changes from mild complaint to strong dissatisfaction.

[0075] In this embodiment, the customer's basic information includes the customer's basic characteristic data, such as age, gender, region, occupation, etc.

[0076] In this embodiment, historical agent service data records various information from past interactions between agents and customers, such as business questions asked by customers, solutions provided by agents, and customer feedback on the service.

[0077] In this embodiment, the customer segmentation logic tree is a tree structure that classifies potential customer groups according to certain logical relationships (e.g., based on differences in basic information and differences in the types of business involved, with each node representing a customer group that meets certain basic information ranges and certain types of business involved).

[0078] In this embodiment, the similar node is a node in the customer partitioning logic tree that has a high degree of similarity to the current customer.

[0079] In this embodiment, the historical customer group refers to the set of customers that the agents have served in the past, whose basic information range and business types are similar to those of the current customer.

[0080] In this embodiment, the current business processing requirement is the business problem that the customer raises and needs to be solved during the communication with the agent. For example, "solving the network outage problem" is the customer's current business processing requirement, a task that the agent needs to respond to and handle promptly.

[0081] In this embodiment, a successful response case refers to an example in history of effective solutions provided to agents and customer satisfaction feedback.

[0082] In this embodiment, the customer's belonging node is the node in the customer segmentation logic tree that represents the current category position of the customer.

[0083] In this embodiment, the customer's personalized multidimensional response factor matrix is ​​a matrix constructed by integrating various factors, including successful response cases from historical customer groups with similar nodes. For example, in tourism services, the matrix may include dimensions such as travel destination, travel time, budget, and service preferences. Each dimension is assigned different factor values ​​and application weights based on successful response cases. For instance, for a particular customer, the matrix shows that they have a high preference for seaside tourist destinations and tend to travel during the summer.

[0084] In this embodiment, the personalized service decision model is built upon a large number of successful service decision cases and is used to generate personalized service solutions. For example, in a customer service center, by collecting and analyzing thousands of successful customer service cases from the past, including customer problems, service processes, and results, the model learns the optimal service decision patterns for different situations. When relevant information about a new customer is input, it can output service decision suggestions suitable for that customer.

[0085] In this embodiment, the business processing step guidance is a personalized intelligent agent reminder solution that provides agents with specific operational steps based on the customer's current business processing needs. For example, if a customer reports a paper jam in their printer, the business processing step guidance might include: first, asking the customer for their printer model; second, instructing the customer to open the printer cover and check the location of the paper jam; and third, informing the customer to carefully remove the paper jam using a specific method.

[0086] In this embodiment, emotional reassurance language is the communication language used in the personalized intelligent agent reminder solution to alleviate customer emotions and enhance customer satisfaction.

[0087] In this embodiment, the personalized intelligent agent reminder solution is generated by combining key business information, emotional trend changes, a dedicated multi-dimensional response factor matrix, and a personalized service decision model. It provides agents with guidance on business processing steps and emotional reassurance phrases. For example, for a customer dissatisfied with a delayed delivery, the solution guides the agent through business processing steps such as checking the package's location and communicating with the courier about the estimated delivery time. The emotional reassurance phrases instruct the agent to express understanding and apology to the customer, comprehensively helping agents provide high-quality service to the customer.

[0088] Example 2:

[0089] Based on Example 1, the speech conversion module, refer to Figure 2 ,include:

[0090] The noise reduction processing submodule is used to perform noise reduction preprocessing on the full-band audio data of real-time acquired customer voice calls to obtain preprocessed customer audio data.

[0091] The speech conversion submodule is used to convert preprocessed customer audio data into high-precision text data.

[0092] In this embodiment, noise reduction preprocessing is performed on the full-band audio data of real-time customer voice calls to obtain preprocessed customer audio data. This means, for example, removing noise signals in a specific frequency range through filtering technology, or using an adaptive noise cancellation algorithm to automatically adjust filter parameters according to the characteristics of environmental noise to reduce the impact of noise.

[0093] In this embodiment, the preprocessed customer audio data is converted into high-precision text data by using speech conversion technology to transform the noise-reduced audio signal into accurate text information. For example, an advanced speech recognition engine is used to extract features from the preprocessed audio data, analyze the frequency, amplitude, and other features of the audio, and then match and recognize them with a pre-stored speech model to convert the speech content in the audio into text word by word.

[0094] The beneficial effects of the above technologies are as follows: The noise reduction submodule first performs noise reduction preprocessing on the real-time acquired full-frequency audio data, which effectively reduces the impact of external noise interference on the audio data. The speech conversion submodule converts the noise-reduced audio data into high-precision text data. Under the premise of high-quality audio input, it can complete the conversion work more reliably, improving the stability and reliability of the entire speech conversion module. This lays a solid foundation for the subsequent analysis of text data by the recognition and capture module and for the entire system to generate accurate and effective personalized intelligent agent reminder solutions, comprehensively improving the quality of the system's analysis and processing of customer voice call content.

[0095] Example 3:

[0096] Based on Example 1, the identification and capture module is referenced. Figure 3 ,include:

[0097] The entity recognition submodule is used to identify all business information entities in high-precision text data that belong to the related business knowledge graph;

[0098] The keyness calculation submodule is used to calculate the keyness of each business information entity based on its grouping degree, coreness, and frequency of occurrence in the associated business knowledge graph.

[0099] The information recognition submodule is used to reason and merge all business information entities in the high-precision text data that belong to the related business knowledge graph and whose keyness is not less than the keyness threshold, in order to obtain the key business information in the high-precision text data.

[0100] The emotion capture submodule is used to analyze multi-dimensional emotional expression information of customers collected in real time based on a dynamic emotion recognition model, and to capture the trend of changes in customers' emotional tendencies.

[0101] In this embodiment, a business information entity refers to business information that is identical to a single entity (each entity corresponds to a business information) within the scope of the associated business knowledge graph in high-precision text data.

[0102] In this embodiment, the clustering degree, core degree, and frequency of occurrence of a business information entity in the associated business knowledge graph are as follows: the clustering degree reflects the degree of association between the entity and other entities in the graph; the core degree reflects the core position of the entity in the graph from different perspectives; and the frequency of occurrence is the number of times the business information entity appears in the high-precision text data.

[0103] In this embodiment, the criticality of a business information entity is calculated by combining its cohesion, coreness, and frequency of occurrence, and is used to quantify the importance of the business information entity.

[0104] In this embodiment, the criticality threshold is a pre-set numerical standard, such as setting the criticality threshold to 0.6.

[0105] In this embodiment, among all business information entities belonging to the associated business knowledge graph in the high-precision text data, those with a keyness not less than the keyness threshold are inferred and merged to obtain the key business information in the high-precision text data. This involves selecting the important parts from numerous business information entities and integrating them. For example, in customer feedback text about a software product, there are business information entities such as "software functions," "user interface," "running speed," and "update frequency." After calculating the keyness, it is found that the keyness of "software functions" and "running speed" is not less than the keyness threshold. By inferring and merging the information represented by these two entities, it is concluded that the customer is mainly concerned about software functions and running speed, which is the key business information.

[0106] In this embodiment, the customer's multidimensional emotional expression information refers to the emotional information conveyed by the customer in various ways during the interaction with the agent. It includes the customer's tone of voice, such as anger, calmness, or anxiety; changes in pitch, such as a rising pitch indicating emotional excitement; the speed of speech, such as speaking quickly suggesting urgency or dissatisfaction; and the emotional coloring of words used, such as using words with obvious emotional bias, like "terrible" or "satisfied".

[0107] The beneficial effects of the above technologies are as follows: Firstly, they define the scope for extracting key business information. By calculating the criticality of each business information entity based on its sociability, coreness, and frequency of occurrence, and by quantitatively evaluating business information from multiple dimensions, important information can be filtered more scientifically, improving the accuracy of information processing. Secondly, by merging business information entities with a criticality of at least a threshold, key business information is obtained. This process further refines the information, ensuring that the system focuses on content truly valuable to business processing. Thirdly, by using dynamic sentiment recognition models to analyze multi-dimensional customer emotional expressions and capture trends in emotional tendencies, agents can better understand customer emotional fluctuations, adjust service strategies promptly, provide more attentive service, enhance customer experience, and comprehensively improve the system's ability to analyze and respond to customer needs and emotions, thereby improving service quality and efficiency.

[0108] Example 4:

[0109] Based on Example 3, the keyness calculation submodule, referencing Figure 3 ,include:

[0110] The edge length quantization unit is used to quantify the edge length between corresponding connected entities based on the association metric between adjacent entities in the related business knowledge graph.

[0111] The clustering degree calculation unit is used to take the ratio of the average length of the path edges between each business information entity and all other business information entities in the associated business knowledge graph (excluding the current business information entity) to the maximum length of the path edges between any two business information entities in the associated business knowledge graph as the outlier degree of the corresponding business information entity, and take the difference between 1 and the outlier degree as the clustering degree of the corresponding business information entity.

[0112] The first core degree calculation unit is used to take the length of the path between each business information entity and all edge entities in the associated business knowledge graph as the margin value of each business information entity, and take the average degree of all margin values ​​of each business information entity as the first core degree of the corresponding business information entity.

[0113] The second core degree calculation unit is used to take the ratio of the average value of all margin values ​​of each business information entity to half the maximum value of the path length between any two entities in the associated business knowledge graph as the second core degree of the corresponding business information entity.

[0114] The third core degree calculation unit is used to take the ratio of the number of edges connected to each business information entity in the associated business knowledge graph to the maximum number of edges connected to a single entity in the associated business knowledge graph as the third core degree of the corresponding business information entity.

[0115] The criticality calculation unit is used to calculate the criticality of each business information entity based on its grouping degree, first core degree, second core degree, third core degree, and frequency of occurrence in high-precision text data.

[0116] In this embodiment, the correlation between adjacent entities describes the tightness of the relationship between two directly connected entities in the related business knowledge graph. This correlation reflects the strength of the inherent connection between business information and can be determined through pre-setting or analyzed using big data technology.

[0117] In this embodiment, the edge length between adjacent entities is quantified based on the correlation between adjacent entities in the related business knowledge graph. That is, the product of the correlation degree between adjacent entities and a preset multiple (e.g., 5 times) is used as the edge length between adjacent entities.

[0118] In this embodiment, the path length between two business information entities in the associated business knowledge graph refers to the path length traversed along the edge in the graph from one business information entity to another.

[0119] In this embodiment, an edge entity refers to an entity that is located at a relatively edge position in the related business knowledge graph.

[0120] In this embodiment, the average margin value of all margin values ​​for each business information entity is the absolute value of the difference between the difference between all margin values ​​and the average of all margin values ​​and the ratio of the average of all margin values ​​to 1.

[0121] In this embodiment, the number of edges connected to a business information entity in the associated business knowledge graph is the number of edges that directly connect the entity to other entities in the graph.

[0122] In this embodiment, the criticality of each business information entity is calculated based on its sociability, first coreity, second coreity, third coreity, and frequency of occurrence in high-precision text data.

[0123] Assume that the weights of the grouping degree, first core degree, second core degree, third core degree, and frequency of occurrence in high-precision text data for each business information entity are 0.3, 0.1, 0.1, 0.1, and 0.4, respectively.

[0124] Based on the above weights, the grouping degree, first core degree, second core degree, third core degree, and frequency of occurrence in high-precision text data of each business information entity are weighted and summed to obtain the keyness of the corresponding business information entity.

[0125] The beneficial effects of the above technologies are as follows: Based on the quantified edge length of adjacent entity associations in the related business knowledge graph, a foundational metric is provided for subsequent calculations of clustering and coreness. Through a unique calculation method, clustering is defined as the reciprocal of outlier, clearly reflecting the relative aggregation degree of business information entities in the graph, helping to determine their close connections with other entities, and facilitating the filtering of information closely related to the overall business. The first coreness calculation unit measures the distance relationship between an entity and edge entities using the average edge distance value, reflecting its tendency to be central in the graph. The second coreness calculation unit compares the average edge distance with a specific ratio of the longest path edge length in the graph, highlighting the entity's core position in the overall structure. The third coreness calculation unit measures the connection activity of an entity in the graph using the proportion of connected edges. By comprehensively considering clustering, the three corenesses, and frequency of occurrence to calculate keyness, the importance of business information entities is comprehensively assessed from multiple dimensions, making the filtering of key business information more accurate and reasonable, and providing more valuable information support for the subsequent generation of personalized intelligent agent reminder solutions.

[0126] Example 5:

[0127] Based on Example 1, the matrix construction module includes:

[0128] The priority level determination submodule is used to extract all historical business types involved in a customer from the customer's historical agent service data, and retrieve the business level quantification table based on the number of times all historical business types were involved to determine the priority level of each historical business type involved in the customer.

[0129] The conformity determination submodule is used to identify all historical business types involved in the customer as all target business types involved in the customer in the business type classification logic tree, and to identify all superior nodes of each target business type involved in the business type classification logic tree. Based on the priority of the historical business types involved in each target business type involved in the business type classification logic tree, and the ratio of the superior layer of each superior node to the total layer of the longest complete node link, the conformity of the business type interval corresponding to each superior node is calculated.

[0130] The similarity calculation submodule is used to calculate the similarity between the customer's node and each node in the customer partitioning logic tree based on the customer's basic information, all historical business types involved and the corresponding business type ranges of all superior nodes, the corresponding compliance, the basic information of all customers in the customer interval corresponding to each node in the customer partitioning logic tree, the historical business types involved and the corresponding business type ranges of all superior nodes, and the corresponding compliance.

[0131] The similarity node filtering submodule is used to filter all nodes in the customer partitioning logic tree that have a similarity score exceeding a similarity threshold as all similar nodes of the customer.

[0132] The matrix construction submodule is used to construct a customer's exclusive multidimensional response factor matrix based on successful response cases of current business processing needs of all historical customer groups of the customer's similar nodes and the similarity between the corresponding similar nodes and the customer's own node.

[0133] In this embodiment, the types of services involved in history refer to the various types of services involved in the customer's past service interactions with the agent. For example, in the historical agent service data of a telecommunications customer, the customer may have inquired about services such as package changes, bill inquiries, and broadband fault reporting. These "package changes," "billing inquiries," and "broadband fault reporting" are considered the types of services involved in the customer's history.

[0134] In this embodiment, the business level quantification table is a pre-defined table that assigns corresponding level quantification values ​​to different business types according to certain standards, in order to measure the importance or priority of the business. For example, in the financial services field, different businesses such as "account loss reporting," "wealth management product consultation," and "fixed deposit business processing" are assigned different quantification values ​​in the business level quantification table based on factors such as their impact on customer fund security and business complexity. For example, "account loss reporting" is set to a quantification value of 5 (assuming the maximum is 5) because it is directly related to customer fund security, while the quantification value for "wealth management product consultation" may be 3, thus clarifying the priority level of different businesses.

[0135] In this embodiment, the priority of historically involved business types is determined by a business level quantification table, combined with the number of times the customer has been involved in a particular business type in the past, which ranks the importance of that business in the customer's overall business.

[0136] In this embodiment, the business category classification logic tree is a model that uses a tree structure to display the hierarchical relationship and classification logic between business categories. For example, in e-commerce, the top-level node may be "shopping-related business," and the next level of nodes can be divided into "product purchase," "product after-sales service," etc. "Product purchase" can be further subdivided into nodes such as "search for products" and "place an order and pay." This tree structure clearly presents the classification and hierarchy of business categories, which facilitates the systematic analysis and management of the business.

[0137] In this embodiment, all the superior nodes of the target business node refer to all the nodes that are traced upwards from a target business node (i.e., the node corresponding to the customer's historical business type) along the tree structure to the root node in the business type classification logic tree.

[0138] In this embodiment, based on the priority of the historical business types involved in each target business node, and the ratio of the number of upper-level layers corresponding to each upper-level node to the total number of layers of the longest complete node link, the conformity degree of the business type interval corresponding to each upper-level node is calculated, which is:

[0139] For example, the priority of the historical business types involved in each target business node, and the weight of the ratio of the number of upper-level layers of each upper-level node to the total number of layers of the longest complete node link are 0.6 and 0.4 respectively.

[0140] Based on the above weights, the priority of the historical business types involved in each target business node, the ratio of the number of upper-level layers of each upper-level node to the total number of layers of the longest complete node link are weighted and summed to obtain the conformity of the business type range corresponding to each upper-level node.

[0141] In this embodiment, the upper-level layer of the upper-level node refers to the number of layers between the target business node and the upper-level node in the business type classification logic tree.

[0142] In this embodiment, the total number of layers of the longest complete node link refers to the maximum number of layers traversed from the root node of the business type partitioning logic tree to the target business node (because there may be more than one path).

[0143] In this embodiment, the similarity threshold is a pre-set numerical standard used to filter nodes that are similar to the current customer in the customer partitioning logic tree. For example, the similarity threshold is set to 0.6.

[0144] The beneficial effects of the above technologies are as follows: By extracting the types of business a customer has historically been involved in and determining priority levels based on the frequency of involvement in a business level quantification table, the importance of different customer businesses can be clearly distinguished, providing an important priority reference for subsequent analysis and helping to grasp the key points in complex business processes. In the business type classification logic tree, the target business node and its parent node are identified, and the conformity degree of each parent node's corresponding business type interval is calculated using a unique method. This quantitative analysis of the hierarchical relationship and conformity degree of business nodes makes business connections clearer and helps to comprehensively understand the position and fit of the customer's business in the overall logic tree. By comprehensively considering factors such as customer basic information, business types, and conformity degree, the similarity between the customer's node and other nodes in the logic tree is calculated, comprehensively considering the similarity between the customer and other nodes from multiple dimensions, providing a comprehensive and accurate basis for selecting similar nodes. Filtering similar nodes based on similarity thresholds can accurately locate the node group similar to the current customer, facilitating targeted analysis and learning from the handling experience of similar customers. Based on successful response cases and similarity of similar nodes, a unique multi-dimensional response factor matrix is ​​constructed, providing rich data support that is tailored to the actual situation of customers for generating personalized agent reminder solutions.

[0145] Example 6:

[0146] Based on Example 5, the similarity calculation submodule includes:

[0147] The first similarity calculation unit is used to calculate the similarity between the basic information of a customer and the basic information of each customer in the customer interval corresponding to each node in the customer segmentation logic tree, and use it as the first similarity between the two customers.

[0148] The second similarity calculation unit is used to calculate the similarity between all the historical business types involved in a customer and all the historical business types involved in each customer in the customer interval corresponding to each node in the customer segmentation logic tree, and use it as the second similarity between the two corresponding customers.

[0149] The third similarity calculation unit is used to calculate the similarity between the business type range and corresponding compliance degree of all superior nodes of the customer and the business type range and corresponding compliance degree of each customer in the customer interval of each node in the customer partitioning logic tree, and use it as the third similarity between the two customers.

[0150] The total similarity calculation unit is used to perform a weighted summation of the first similarity, second similarity, and third similarity between each customer in the customer interval corresponding to each node in the customer-customer division logic tree, to obtain the total similarity between the corresponding two customers.

[0151] The final similarity calculation unit is used to take the average of the total similarity between all customers in the customer interval corresponding to each node in the customer partitioning logic tree as the similarity between the customer's own node and the corresponding node in the customer partitioning logic tree.

[0152] In this embodiment, the similarity between the customer's basic information and the basic information of each customer in the customer interval corresponding to each node in the customer partitioning logic tree is calculated. The customer's basic information can be integrated into a basic information vector, and the cosine similarity between the two basic information vectors can be regarded as the similarity between the customer's basic information and the basic information of each customer in the customer interval corresponding to each node in the customer partitioning logic tree.

[0153] In this embodiment, the similarity between all historical business types involved in a customer and all historical business types involved in each customer within the customer interval corresponding to each node in the customer segmentation logic tree is calculated. This primarily assesses the similarity between customers from the perspective of business experience. Alternatively, all historical business types involved in a customer can be assigned values ​​and fitted into a historical business type vector. Simultaneously, all historical business types involved in each customer within the customer interval corresponding to each node in the customer segmentation logic tree can be assigned values ​​and fitted into a historical business type vector. The cosine similarity between these two historical business type vectors is then used as the similarity between all historical business types involved in a customer and all historical business types involved in each customer within the customer interval corresponding to each node in the customer segmentation logic tree.

[0154] In this embodiment, the similarity between the business type range and corresponding compliance degree corresponding to all superior nodes of a customer and the business type range and corresponding compliance degree corresponding to all superior nodes of each customer in the customer partitioning logic tree is calculated, which is:

[0155] Assign values ​​to all business types within the business type intervals corresponding to all parent nodes of the customer and generate a business type interval vector. At the same time, assign values ​​to all business types within the customer intervals corresponding to each node in the customer partitioning logic tree and generate a business type interval vector. The product of the cosine similarity of the business type interval vectors of the two, the conformity of the business type intervals corresponding to all parent nodes of the corresponding customer, and the conformity of the business type intervals corresponding to all parent nodes of each customer in the customer intervals corresponding to each node in the customer partitioning logic tree is taken as the similarity between the business type intervals corresponding to all parent nodes of the customer and their corresponding conformity and the business type intervals corresponding to all parent nodes of each customer in the customer intervals corresponding to each node in the customer partitioning logic tree.

[0156] In this embodiment, the first similarity, second similarity, and third similarity between customers in each customer interval corresponding to each node in the customer-customer division logic tree are weighted and summed to obtain the total similarity between the corresponding two customers. For example, assuming the weight of the first similarity (basic information similarity) is 0.3, the weight of the second similarity (historical business type similarity) is 0.4, and the weight of the third similarity (upper-level node business type interval and compliance similarity) is 0.3, if a customer has a first similarity of 0.7, a second similarity of 0.6, and a third similarity of 0.8 with another customer in the logic tree, then the total similarity = 0.3 × 0.7 + 0.4 × 0.6 + 0.3 × 0.8 = 0.69.

[0157] The beneficial effects of the above technologies are as follows: The first similarity calculation unit, by calculating the similarity between basic customer information, can consider the similarity between customers from the perspective of basic customer characteristics, which helps to initially screen customer groups with similar basic attributes. The second similarity calculation unit focuses on calculating the similarity of historical business types, which directly reflects the similarity of customer business experiences, enabling the system to understand the degree of overlap between different customers in business areas. The third similarity calculation unit calculates the similarity between business type intervals corresponding to higher-level nodes, deeply analyzes the similarity between customers from the perspective of business logic hierarchy, and explores deep-level business connections, which helps to discover the similar positioning of customers in the business logic system. The total similarity calculation unit sums the similarity of the three dimensions according to weights, comprehensively considering factors such as basic information, business types, and business logic hierarchy, and comprehensively and reasonably measures the overall similarity between customers, making the assessment of similarity more scientific and accurate. The final similarity calculation unit determines the similarity between its node and the corresponding node in the logic tree based on the total average similarity. This processing method further smooths and stabilizes the similarity results, avoids excessive influence of special cases of individual customers on the similarity judgment, and provides more reliable and representative similarity data for subsequent screening of similar nodes and construction of a dedicated multidimensional response factor matrix.

[0158] Example 7:

[0159] Based on Example 5, the matrix construction submodule includes:

[0160] The variable response factor value determination unit is used to summarize all historical customers in the customer interval of all similar nodes of the customer to obtain the customer's reference customer group. In the reference customer group, multidimensional variable response factors are extracted from all successful response cases for the customer's current business processing needs. Based on the preset assignment method, values ​​are assigned to all single-dimensional response factors of the multidimensional variable response factors to obtain the multidimensional variable response factor value of each successful response case.

[0161] The application weight determination unit is used to determine the application weight of each single-dimensional response factor value in the multi-dimensional variable response factor value of each successful response case based on the proximity between each proximity node and the customer's node.

[0162] The matrix building unit is used to generate a client's exclusive multidimensional response factor matrix based on the multidimensional variable response factor values ​​of all successful response cases and the application weights of each corresponding single-dimensional response factor value.

[0163] In this embodiment, multidimensional variable response factors are extracted from all successful response cases of the reference customer group addressing the customer's current business processing needs. This means identifying multiple factors that may affect business processing and are likely to change from successful cases of resolving similar business problems by a historical customer group similar to the current customer (i.e., the reference customer group). For example, in the scenario where a customer inquires about slow network speed, the multidimensional variable response factors in the successful response cases of the reference customer group may include network device model, network environment (home / office), troubleshooting steps, network optimization tools used, etc. These factors affect the resolution of the slow network speed problem from different dimensions and may have different values ​​in different cases.

[0164] In this embodiment, the preset assignment method is a set of pre-defined rules used to assign specific values ​​to each single-dimensional response factor in the multidimensional variable response factors. For example, in the network problem case mentioned above, for the single-dimensional response factor of "network device model", it may be specified that common and high-performance device models are assigned a value of 8, general models are assigned a value of 6, and old models are assigned a value of 4; for "network environment", home environment is assigned a value of 5, office environment is assigned a value of 6, and so on.

[0165] In this embodiment, all single-dimensional response factors are assigned values ​​based on a preset assignment method for multidimensional variable response factors to obtain the multidimensional variable response factor value for each successful response case. This involves assigning values ​​to each single-dimensional response factor in each successful response case according to the aforementioned preset assignment method, thereby obtaining a comprehensive numerical value to represent the response factor situation of that case. For example, in a case where a slow network speed problem was successfully resolved, the network device model is a common and high-performance model, assigned a preset value of 8; the network environment is a home environment, assigned a value of 5; the troubleshooting steps are complete and effective, assigned a value of 7; and the network optimization tools used are effective, assigned a value of 8. These single-dimensional response factor values ​​are then combined (e.g., through simple addition or other specific calculation methods) to obtain the multidimensional variable response factor value for this successful response case, allowing for subsequent comparison and analysis of different cases.

[0166] In this embodiment, based on the proximity between each proximity node and the customer's node, the application weight of each single-dimensional response factor value in the multidimensional variable response factor value of each successful response case is determined. The weight is determined according to the preset weight multiplier of the multidimensional variable response factor for each successful response case. For example, the preset weight multiplier of multidimensional variable response factor a in successful response case A is 0.8, and the corresponding proximity is 0.6. Then the application weight of multidimensional variable response factor a in successful response case A is 0.48.

[0167] The beneficial effects of the above technologies are as follows: This process comprehensively collects and quantifies key factors from historical success cases similar to the current customer situation, facilitating in-depth analysis of various factors influencing business processing and their values. The application of weighted determination units fully considers the close relationship between different similar case studies and the current customer, ensuring that factors from more similar cases have higher weights when referencing historical cases. This allows for a more precise alignment with the current customer's actual needs, improving the relevance and effectiveness of learning from historical experience. The matrix generated by the matrix construction unit integrates multi-dimensional response factors and their weight information, providing agents with a comprehensive, detailed, and targeted decision-making framework for handling current customer business. Agents can quickly understand the importance and value range of different response factors based on this matrix, formulating more personalized service strategies that better meet customer needs. This significantly improves service quality and business processing efficiency, enhances customer satisfaction, and also accumulates more valuable customer service data assets for the enterprise, facilitating continuous optimization of service processes and improvement of service levels.

[0168] Example 8:

[0169] Based on Example 7, the matrix construction unit includes:

[0170] The feedback judgment value determination subunit is used to assign values ​​to each historical agent response feedback result of the customer, obtain the feedback judgment value of each historical agent response feedback result of the customer, assign values ​​to a large number of historical agent response feedback results of reference customers that are consistent with the agent response results in each historical agent response feedback result of the customer, and obtain all feedback reference judgment values ​​of the corresponding historical agent response feedback result of the customer.

[0171] The feedback relative quantification subunit is used to determine the customer's feedback bias direction and feedback deviation factor based on the feedback judgment value of each customer's historical agent response feedback result and all corresponding feedback reference judgment values;

[0172] The application weight adjustment subunit is used to adjust the application weight of each single-dimensional response factor value in the multidimensional variable response factor value of each successful response case based on the customer's feedback bias direction and feedback deviation factor, so as to obtain the adjusted application weight of each single-dimensional response factor value in the multidimensional variable response factor value of each successful response case.

[0173] The reference weight matrix generation sub-unit is used to adjust the applied weights for each single-dimensional response factor value among the multidimensional variable response factor values ​​of all successful response cases to form the reference weight matrix.

[0174] The original matrix construction sub-unit is used to summarize all single-dimensional response factor values ​​in the multidimensional variable response factor values ​​of all successful response cases to obtain the customer's original multidimensional response factor matrix.

[0175] The dedicated matrix construction sub-unit is used to generate a customer's dedicated multidimensional response factor matrix based on the customer's original multidimensional response factor matrix and reference weight matrix.

[0176] In this embodiment, historical agent response feedback results refer to the feedback information given by customers to agents after they provided services in the past. For example, after a customer inquired about a product function, the agent provided answers and operating instructions, and the customer's evaluation of the service, such as satisfaction or dissatisfaction, or specific suggestions for improvement, all fall under the category of historical agent response feedback results.

[0177] In this embodiment, assigning values ​​to each customer's historical agent response feedback result to obtain a feedback value for each historical agent response feedback result is to quantify the customer's feedback. For example, the rules are set as follows: very satisfied is assigned a value of 5, satisfied is assigned a value of 4, neutral is assigned a value of 3, dissatisfied is assigned a value of 2, and very dissatisfied is assigned a value of 1.

[0178] In this embodiment, the agent response result refers to the specific actions taken and responses given by the agent in response to customer business inquiries and problem feedback. For example, if a customer reports a quality problem with a purchased product, the agent response result might be arranging a replacement, providing a repair solution, or informing the customer of the relevant quality inspection process. It is the direct object of the customer's feedback.

[0179] In this embodiment, assigning values ​​to a large number of historical agent response feedback results of reference customers that are consistent with the agent response results in each customer's historical agent response feedback results, and obtaining all feedback reference judgment values ​​for the corresponding historical agent response feedback results of the customer, is to provide a comparison benchmark for the current customer's feedback. For example, if the current customer receives a response result of agent arrangement for replacement due to product quality issues and provides feedback, collecting feedback from a large number of other customers with the same agent response result (replacement arrangement), and assigning values ​​according to the aforementioned assignment rules, a series of feedback reference judgment values ​​are obtained.

[0180] In this embodiment, the customer's feedback bias direction and feedback deviation factor are determined based on the feedback judgment value of each customer's historical agent response and the corresponding all feedback reference judgment values. For example, if the feedback judgment value of each customer's historical agent response is less than the average of all corresponding feedback reference judgment values, the feedback bias direction is negative; otherwise, it is positive. If the customer's feedback judgment value is 2 (dissatisfied), and the average of the corresponding feedback reference judgment values ​​is 3 (neutral), it indicates that the customer's feedback is biased negatively; this is the feedback bias direction. The feedback deviation factor is obtained by calculating the ratio of the difference between the customer's feedback judgment value and the average of all reference judgment values ​​to the average of all reference judgment values; for example, the feedback deviation factor here is 1 / 3.

[0181] In this embodiment, the application weight of each single-dimensional response factor value in the multidimensional variable response factor values ​​of each successful response case is adjusted based on the customer's feedback bias direction and feedback deviation factor. The adjusted application weight of each single-dimensional response factor value in the multidimensional variable response factor values ​​of each successful response case is thus obtained as follows:

[0182] The feedback bias direction and the feedback bias factor are combined to obtain the comprehensive feedback bias factor. For example, if the feedback bias direction is negative and the feedback bias factor is 1 / 3, then the comprehensive feedback bias factor is -1 / 3.

[0183] The product of the sum of the comprehensive feedback bias factors and the corresponding application weight is used as the adjustment application weight.

[0184] In this embodiment, the reference weight matrix is ​​constructed by adjusting the application weights of each single-dimensional response factor value among the multi-dimensional variable response factor values ​​of all successful response cases. This involves organizing the adjusted application weights from all cases into a matrix structure. For example, suppose there are three successful response cases, each involving three single-dimensional response factors: "network device model," "network environment," and "troubleshooting steps." The adjusted application weights are Case 1 (0.4, 0.3, 0.3), Case 2 (0.5, 0.2, 0.3), and Case 3 (0.3, 0.4, 0.3), respectively. Organizing these weights into a matrix form constitutes the reference weight matrix.

[0185] In this embodiment, all single-dimensional response factor values ​​in the multidimensional variable response factor values ​​of all successful response cases are aggregated to obtain the customer's original multidimensional response factor matrix. This involves gathering all single-dimensional response factor values ​​from all cases into a single matrix. Continuing with the network problem case mentioned above, the single-dimensional response factor values ​​such as "network device model," "network environment," and "troubleshooting steps" from all cases are arranged in a certain order to form a matrix.

[0186] In this embodiment, generating a customer-specific multidimensional response factor matrix based on the customer's original multidimensional response factor matrix and reference weight matrix involves comprehensively considering the original data and weight information to generate a specific matrix for the current customer. For example, the customer-specific multidimensional response factor matrix is ​​obtained by multiplying each element in the original multidimensional response factor matrix by the corresponding weight value in the reference weight matrix, and summarizing the matrix positions to obtain the customer-specific multidimensional response factor matrix.

[0187] The beneficial effects of the above technologies are as follows: Firstly, feedback judgment values ​​and reference judgment values ​​are obtained, which directly reflect the agent's response effectiveness. Secondly, the feedback relative quantification sub-unit determines the feedback bias direction and feedback deviation factor based on the feedback judgment value and reference judgment value, further uncovering the information behind the feedback data. The feedback bias direction reveals whether the customer feedback is more positive or negative compared to the reference customer, while the feedback deviation factor quantifies the degree of this bias, making the analysis of customer feedback more accurate and in-depth. Thirdly, the application weight adjustment sub-unit adjusts the application weights of single-dimensional response factor values ​​based on the feedback bias direction and feedback deviation factor, enabling dynamic and targeted weight optimization. The reference weight matrix generation sub-unit constructs a reference weight matrix based on the adjusted application weights, providing an important weight framework for subsequent construction of a dedicated matrix. This matrix integrates weight information considering customer feedback, making the application of weights more systematic and standardized. Finally, the original matrix construction sub-unit summarizes all single-dimensional response factor values ​​to obtain the original multi-dimensional response factor matrix, which, combined with the reference weight matrix, is used by the dedicated matrix construction sub-unit to generate a customer-specific multi-dimensional response factor matrix. This approach organically combines response factor values ​​with adjusted weights, generating a customized matrix that more accurately reflects customers' specific needs and historical feedback, providing agents with more targeted and practical decision-making support.

[0188] Example 9:

[0189] Based on Example 1, the scheme generation module, referencing Figure 4 ,include:

[0190] The model building submodule is used to build personalized service decision models based on a large number of successful service decision cases;

[0191] The solution generation submodule is used to input key business information, emotional tendency change trends, and the customer's exclusive multi-dimensional response factor matrix into the personalized service decision model to generate a personalized intelligent agent reminder solution that includes business processing step guidance and emotional reassurance language.

[0192] In this embodiment, a successful service decision case refers to a real-world example in which, during past customer service processes, an agent made an effective decision based on the customer's business needs and successfully resolved the customer's problem, while the customer expressed satisfaction with the service outcome.

[0193] In this embodiment, a personalized service decision-making model is built based on a large number of successful service decision-making cases. This involves utilizing numerous such successful case data, employing specific algorithms and analytical methods to uncover patterns and dynamics, thereby establishing a model capable of generating personalized service decisions for different customers. For example, a large number of successful resolution cases of various customer issues on e-commerce platforms (such as product quality issues, logistics issues, and after-sales service issues) are collected. The mapping relationship between key business information, sentiment trends, and the customer's unique multi-dimensional response factor matrix in these cases and the successful case data is analyzed. Machine learning algorithms are used to allow the model to learn the best response methods for different types of customer issues. When encountering a new customer issue, the personalized service decision-making model can generate suitable service decision suggestions based on what it has learned and the specific circumstances of the new customer. For example, regarding a customer's product quality feedback, the model might decide to provide a partial refund as compensation, while also promising to expedite the exchange process and other personalized service solutions.

[0194] The beneficial effects of the above technologies are as follows: They lay a solid foundation for solution generation. By aggregating experience from numerous successful cases, the model can learn effective decision-making patterns in different situations, covering various business scenarios and customer characteristics, making the generated solutions more universal and reliable. Key business information ensures that the solution closely revolves around the customer's business needs, directly addressing core issues; consideration of emotional trend changes enables the solution to not only handle business but also pay attention to customer emotions, providing appropriate emotional reassurance during the service process and enhancing customer experience; the customer-specific multi-dimensional response factor matrix incorporates successful experiences from similar points in the customer's life, further personalizing the solution. The personalized intelligent agent reminder solution generated by the combination of these three elements provides agents with guidance on business processing steps, allowing them to clearly understand each step of the operation, improving the efficiency and accuracy of business processing; at the same time, it is equipped with emotional reassurance language to help agents better communicate with customers and meet their emotional needs.

[0195] Example 10:

[0196] Based on Example 1, it also includes:

[0197] The dual-loop iteration module is used to adjust the personalized intelligent agent reminder scheme based on the latest agent response feedback results in the inner loop of the dual-loop mechanism. At the same time, in the outer loop of the dual-loop mechanism, the personalized service decision model is iteratively optimized based on historical agent service data, historical agent response feedback results, and business update rules.

[0198] In this embodiment, the inner loop of the dual-loop mechanism primarily focuses on real-time optimization of the current personalized intelligent agent reminder solution. With the inner loop at its core, the system closely monitors the latest agent response feedback. For example, if an agent serves a customer according to the current personalized intelligent agent reminder solution, and the customer reports that the problem-solving process is too cumbersome, the inner loop will immediately capture this feedback. Based on this feedback, the inner loop quickly adjusts the solution, potentially simplifying the problem-solving process, redesigning business processing steps, or optimizing emotional reassurance language to ensure that the next service better meets customer needs and improves customer satisfaction. The inner loop acts like an instant fine-tuning device, continuously optimizing the current service solution based on actual feedback.

[0199] In this embodiment, the outer loop of the dual-loop mechanism focuses on long-term, macro-level improvements to the personalized service decision-making model. It comprehensively considers historical agent service data, historical agent response feedback, and business update rules. Historical agent service data records detailed interactions between numerous agents and customers, while historical agent response feedback reflects customer evaluations of these services. For example, by analyzing long-term historical data, it can be found that the handling methods for certain business issues vary significantly in customer feedback at different times. Combined with business update rules, such as changes in industry policies and product feature updates, the outer loop iteratively optimizes the personalized service decision-making model. It adjusts the model's algorithms, parameters, or decision logic to adapt to changes in the business environment, supporting the generation of more accurate and effective personalized intelligent agent reminder solutions—similar to an upgrade and transformation of the entire service system.

[0200] In this embodiment, the personalized service decision-making model is iteratively optimized based on historical agent service data, historical agent response feedback results, and business update rules. This is a continuous learning and adaptation process. For example, historical agent service data shows that for inquiries about a certain new product, different agent responses resulted in significant differences in customer satisfaction. Historical agent response feedback further clarifies which response methods are more popular with customers. Simultaneously, business update rules indicate that the product's functionality has been upgraded. By integrating this information, the personalized service decision-making model is optimized. This may involve adjusting the decision logic for inquiries about the product, prioritizing more popular responses, and incorporating new content from the upgraded product functionality. This allows the model to generate service decisions that better meet customer needs and the current business situation when faced with similar inquiries, continuously improving the accuracy and effectiveness of the service.

[0201] The beneficial effects of the above technologies are as follows: The synergistic operation of the dual-loop mechanism forms a dynamic and continuously optimized system. The inner loop provides real-time feedback on specific solutions to the outer loop, offering more direct practical evidence for model optimization; the optimized model from the outer loop, in turn, supports the inner loop in generating higher-quality personalized intelligent agent reminder solutions. The two mutually promote each other, continuously improving the quality and adaptability of the entire service system. Fourth, through this iterative optimization approach, enterprises can continuously accumulate service experience and knowledge. Over time, the service decision-making model becomes increasingly accurate, and the personalized intelligent agent reminder solution becomes more sophisticated, thereby continuously improving customer service levels and enhancing customer satisfaction and loyalty. Fifth, the dual-loop iterative module makes the entire system more flexible and adaptable, maintaining efficient service capabilities under complex and ever-changing business scenarios and customer needs, giving enterprises an advantage in a highly competitive market environment and contributing to their long-term stable development.

[0202] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A system for intelligent analysis of voice call content and agent alerting, characterized in that, The method comprises the following steps: a voice conversion module is used to convert full-band audio data of real-time collected customer voice calls into high-precision text data; an identification and capture module is used to identify key business information in the high-precision text data based on an associated business knowledge graph, and to capture real-time changes in customer emotional trends based on a dynamic emotion recognition model; a matrix construction module is used to map all similar nodes of a customer in a customer division logical tree based on basic information and historical agent service data of the customer, and to construct an exclusive multi-dimensional response factor matrix of the customer based on successful response cases of current business processing needs of a historical customer group of all similar nodes of the customer and the similarity between corresponding similar nodes and a node to which the customer belongs, wherein the node to which the customer belongs is a node representing the classification position of the current customer in the customer division logical tree; a scheme generation module is used to generate an individualized intelligent agent reminding scheme containing business processing step instructions and emotional pacification techniques based on the key business information, the emotional trend changes, the exclusive multi-dimensional response factor matrix of the customer and an individualized service decision model; The matrix construction module comprises: a priority level determination submodule is used to extract all historical business categories involved by the customer from the historical agent service data of the customer, and to determine the priority level of each historical business category involved by the customer based on the number of times of involvement of all historical business categories involved; a coincidence determination submodule is used to determine all target business nodes corresponding to the historical business categories involved by the customer in a business category division logical tree as all target business nodes of the customer, and to determine all superior nodes of each target business node in the business category division logical tree, and to calculate the coincidence of the business category interval corresponding to each superior node based on the ratio of the priority of the historical business category corresponding to each target business node, the number of superior layers corresponding to each superior node and the total number of layers of the corresponding longest complete node link; a similarity calculation submodule is used to calculate the similarity between the node to which the customer belongs and each node in the customer division logical tree based on the basic information of the customer, all historical business categories involved and the business category intervals corresponding to all superior nodes, the corresponding coincidence, the basic information of all customers in the customer interval corresponding to each node in the customer division logical tree, the historical business categories involved and the business category intervals corresponding to all superior nodes, and the corresponding coincidence; a similar node screening submodule is used to screen all nodes with a similarity exceeding a similarity threshold value in the customer division logical tree as all similar nodes of the customer; a matrix construction submodule is used to construct the exclusive multi-dimensional response factor matrix of the customer based on the successful response cases of current business processing needs of the historical customer group of all similar nodes of the customer and the similarity between corresponding similar nodes and the node to which the customer belongs.

2. The intelligent voice call content analysis and agent alerting system of claim 1, wherein, The voice conversion module comprises: an anti-noise processing submodule is used to perform anti-noise preprocessing on the full-band audio data of the real-time collected customer voice calls to obtain preprocessed customer audio data; The voice conversion sub-module is configured to convert the preprocessed customer audio data into high-precision text data.

3. The intelligent voice call content analysis and agent alerting system of claim 1, wherein, The recognition capturing module comprises: The entity recognition sub-module is configured to identify all business information entities in the high-precision text data that belong to the associated business knowledge graph. The key degree calculation sub-module is configured to calculate the key degree of each business information entity based on the group degree and core degree of each business information entity in the associated business knowledge graph and the frequency of occurrence. The information recognition sub-module is configured to infer and merge all business information entities in the high-precision text data that have a key degree not less than the key degree threshold, to obtain key business information in the high-precision text data. The emotion capturing sub-module is configured to analyze the multi-dimensional emotional expression information of the customer collected in real time based on a dynamic emotion recognition model, and capture the trend of the customer's emotional inclination. The key degree calculation sub-module comprises: The edge length quantification unit is configured to quantify the edge length between the connected entities based on the correlation between adjacent entities in the associated business knowledge graph. The group degree calculation unit is configured to take the ratio of the average value of the path edge length of each business information entity and all business information entities except the current business information entity in the associated business knowledge graph to the maximum value of the path edge length of each pair of business information entities in the associated business knowledge graph as the outlying degree of the corresponding business information entity, and take the difference between 1 and the outlying degree as the group degree of the corresponding business information entity. The first core degree calculation unit is configured to take the path edge length between each business information entity and all edge entities in the associated business knowledge graph as the edge distance value of each business information entity, and take the average degree of all edge distance values of each business information entity as the first core degree of the corresponding business information entity. The second core degree calculation unit is configured to take the ratio of the average value of all edge distance values of each business information entity to half the maximum value of the path edge length between each pair of entities in the associated business knowledge graph as the second core degree of the corresponding business information entity. The third core degree calculation unit is configured to take the ratio of the number of connected edges of each business information entity in the associated business knowledge graph to the maximum value of the number of connected edges of a single entity in the associated business knowledge graph as the third core degree of the corresponding business information entity. The key degree calculation unit is configured to calculate the key degree of each business information entity based on the group degree, the first core degree, the second core degree, the third core degree of each business information entity, and the frequency of occurrence in the high-precision text data.

4. The intelligent voice call content analysis and agent alerting system of claim 1, wherein, The proximity calculation sub-module comprises: The first proximity calculation unit is configured to calculate the proximity between the basic information of the customer and the basic information of each customer in the customer interval corresponding to each node in the customer division logical tree, as the first proximity of the corresponding two customers. The second proximity calculation unit is configured to calculate the proximity between all historical involved business types of the customer and all historical involved business types of each customer in the customer interval corresponding to each node in the customer division logical tree, as the second proximity of the corresponding two customers. The third similarity calculation unit is configured to calculate a similarity between the service category interval corresponding to all the upper nodes of each customer in the customer interval corresponding to each node in the customer division logical tree and the service category interval corresponding to all the upper nodes of each customer in the customer interval corresponding to each node in the customer division logical tree, as a third similarity corresponding to two customers; The total similarity calculation unit is configured to add the first similarity, the second similarity and the third similarity between the customer and each customer in the customer interval corresponding to each node in the customer division logical tree by weight, to obtain a total similarity corresponding to two customers; The final similarity calculation unit is configured to take the average of the total similarities of all the customers in the customer interval corresponding to each node in the customer division logical tree as the similarity between the node to which the customer belongs and the corresponding node in the customer division logical tree.

5. The intelligent voice call content analysis and agent alerting system of claim 1, wherein, The matrix construction submodule includes: The variable response factor value determination unit is configured to aggregate all the historical customers in the customer interval of all the similar nodes of the customer to obtain a reference customer group of the customer, extract multi-dimensional variable response factors from all the successful response cases of the reference customer group in response to the current service processing needs of the customer, and assign values to all the single-dimensional response factors of the multi-dimensional variable response factors based on a preset assignment mode, to obtain multi-dimensional variable response factor values of each successful response case; The application weight determination unit is configured to determine an application weight of each single-dimensional response factor value in the multi-dimensional variable response factor values of each successful response case based on the similarity between each similar node and the node to which the customer belongs; The matrix construction unit is configured to generate an exclusive multi-dimensional response factor matrix of the customer based on the multi-dimensional variable response factor values of all the successful response cases and the application weight of each single-dimensional response factor value.

6. The intelligent voice call content analysis and agent alerting system of claim 5, wherein, The matrix construction unit includes: The feedback judgment value determination subunit is configured to assign values to each historical agent response feedback result of the customer to obtain a feedback judgment value of each historical agent response feedback result of the customer, and assign values to historical agent response feedback results of a large number of reference customers that are consistent with the agent response result in each historical agent response feedback result of the customer to obtain all feedback reference judgment values of the corresponding historical agent response feedback result of the customer. The feedback relative quantization subunit is configured to determine a feedback bias direction and a feedback bias factor of the customer based on the feedback judgment value of each historical agent response feedback result of the customer and all the feedback reference judgment values corresponding thereto. The application weight adjustment subunit is configured to adjust the application weight of each single-dimensional response factor value in the multi-dimensional variable response factor values of each successful response case based on the feedback bias direction and the feedback bias factor of the customer, to obtain an adjusted application weight of each single-dimensional response factor value in the multi-dimensional variable response factor values of each successful response case. The reference weight matrix generation subunit is configured to generate a reference weight matrix based on the adjusted application weight of each single-dimensional response factor value in the multi-dimensional variable response factor values of all the successful response cases. An original matrix construction subunit is configured to aggregate all single-dimension response factor values in the multi-dimension variable response factor values of all successful response cases to obtain an original multi-dimension response factor matrix of the customer; A dedicated matrix construction subunit is configured to generate a dedicated multi-dimension response factor matrix of the customer based on the original multi-dimension response factor matrix of the customer and a reference weight matrix.

7. The intelligent voice call content analysis and agent alerting system of claim 1, wherein, The scheme generation module comprises: A model construction sub-module configured to construct a personalized service decision model based on a large number of successful service decision cases; A scheme generation sub-module configured to input the key business information, the emotional tendency change trend, and the dedicated multi-dimension response factor matrix of the customer into the personalized service decision model to generate a personalized intelligent agent reminding scheme containing business processing step guidance and emotional pacification language.

8. The intelligent voice call content analysis and agent alerting system of claim 1, wherein, Further comprising: A double-loop iteration module configured to adjust the personalized intelligent agent reminding scheme based on the latest obtained agent response feedback result in an inner loop in a double-loop mechanism, and to iteratively optimize the personalized service decision model based on historical agent service data, historical agent response feedback results, and business update rules in an outer loop in the double-loop mechanism.

Images