A bank intelligent customer service response method, device and medium

Abstract

The application relates to a bank intelligent customer service response method, equipment and medium, a bank intelligent customer service response method comprises the following steps: collecting interactive service data in multiple interactive channels, preprocessing the interactive service data to generate corresponding feature vectors, performing clustering calculation on the interactive service data to generate corresponding customer behavior characteristics, inputting the feature vectors and customer portrait information into a response process to generate corresponding response content; obtaining customer feedback data of the response content, selecting corresponding target models in each decision model according to the customer feedback data and the customer portrait information, and training the target models according to training samples. By constructing a full-process closed-loop path of collection-understanding-generation-feedback-retraining, the system can continuously optimize the response strategy in a real scene, continuously improve the response accuracy, customer satisfaction and adaptability, and meet the efficient, accurate and evolvable requirements of bank intelligent service.

Description

Technical Field

[0001] This invention relates to the technical field of intelligent customer service response for banks, and in particular to a method, device and medium for intelligent customer service response for banks. Background Technology

[0002] Currently, existing intelligent customer service systems in banks are gradually integrating various forms of interactive data, including text, voice, and images, to obtain more comprehensive customer consultation information through multimodal information in order to automatically answer customer questions. However, in practical applications, most of these systems are still at the rudimentary stage of "multimodal data collection" and have not yet established a dynamic feedback mechanism and closed-loop learning capability that can truly leverage the value of multimodal data.

[0003] For example, when customers are dissatisfied with the response or their problems are not effectively resolved, the system typically lacks the ability to proactively perceive customer feedback. It cannot effectively correlate unsatisfactory responses with the original input data, nor can it automatically identify flaws in the response or pinpoint the specific decision-making model based on this feedback, thus hindering targeted model optimization. Consequently, even if the system collects abundant interaction data at the front end, without an effective feedback loop, this data cannot improve the response logic. This results in responses that remain inaccurate, have limited coverage, and lack personalization, leading to a poor user experience, low customer satisfaction, and ultimately, constrained bank service quality and operational efficiency. Summary of the Invention

[0004] To address the problem of poor customer experience and low bank service quality caused by the lack of an effective dynamic feedback mechanism in traditional customer service systems, this application provides a bank intelligent customer service response method, device, and medium.

[0005] The above-mentioned objective of this application is achieved through the following technical solution: a bank intelligent customer service response method, wherein the bank intelligent customer service response method includes:

[0006] Interactive business data is collected from multiple interactive channels, and the interactive business data is preprocessed to generate corresponding feature vectors. The interactive business data is multimodal data containing voice data, text data, and image data.

[0007] Clustering calculations are performed on the interactive business data to generate corresponding customer behavior features, thereby generating customer profile information for multi-strategy decision-making in the response process. The response process includes multiple sequential decision sub-processes, each of which applies its own trained decision model. The decision model includes at least a fusion model, an intent recognition model, and a graph matching model.

[0008] The feature vector and the customer profile information are input into the response process to generate the corresponding response content;

[0009] Obtain customer feedback data for the response content. When the customer feedback data indicates that the customer's satisfaction with the response content is lower than a preset standard, mark the response content and the corresponding interactive business data as training samples.

[0010] Based on the customer feedback data and the customer profile information, a corresponding target model is selected from each of the decision models, and the target model is trained based on the training samples.

[0011] By adopting the above technical solutions, and collecting multimodal interactive business data covering voice, text, and images, and preprocessing it into a unified feature vector format, the system can obtain customer demand information more comprehensively and accurately. Simultaneously, clustering calculations are introduced to generate customer behavior features and further construct customer profiles, enabling targeted multi-strategy decision-making in the intelligent response process, thereby improving the personalization of responses and response efficiency.

[0012] In a preferred embodiment, this application can be further configured such that the step of inputting the feature vector and the customer profile information into the response process to generate the corresponding response content includes:

[0013] Based on the feature vector, the target sub-process that the customer profile information needs to participate in is matched in each of the decision sub-processes. The target sub-process includes at least the decision sub-processes corresponding to the fusion model, the intent recognition model, and the graph matching model.

[0014] Based on the execution order and purpose of the target sub-process, a corresponding decision path is constructed. The decision path is a data sequence consisting of multiple decision node identifiers arranged according to the execution order of the process. The decision node identifiers are generated based on the purpose of the process. The data sequence is used to select the corresponding target model.

[0015] Based on the feature vector, the customer profile information, and the target sub-process, the corresponding response content is generated.

[0016] By adopting the above technical solution, during the process of generating response content, the target sub-process is matched based on feature vectors and a decision path is constructed, enabling customer profiles to selectively participate in the decision nodes most relevant to their needs. This achieves flexible matching between process customization and response strategies, effectively improving the system's process adaptability and response accuracy.

[0017] In a preferred embodiment, this application can be further configured such that the step of generating the corresponding response content based on the feature vector, the customer profile information, and the target sub-process includes:

[0018] The feature vector and the customer profile information are input into the fusion model, wherein the customer profile information is used to guide the multi-source feature fusion strategy of the fusion model to obtain the fused comprehensive feature representation;

[0019] The comprehensive feature representation and the customer profile information are jointly input into the intent recognition model, wherein the customer profile information is used to assist in recognizing the semantic intent information involved in the customer's current inquiry;

[0020] The semantic intent information and the customer profile information are input into the graph matching model. A matching query is performed in the knowledge graph pre-built in the graph matching model to obtain target knowledge data associated with the customer profile information based on the semantic intent information.

[0021] The target knowledge data and the customer profile information are input into the graph matching model to generate the corresponding response content.

[0022] By adopting the above technical solutions, customer profile information is guided into each model. The fusion model can use appropriate fusion strategies to integrate multimodal features, the intent recognition model can help understand user semantics, and the graph matching model can accurately retrieve target knowledge based on the association of customer profiles, forming a response output based on the deep fusion of semantics and profiles, thus comprehensively improving the cognitive ability and response quality of the intelligent customer service system.

[0023] In a preferred embodiment, this application can be further configured such that the step of inputting the target knowledge data and the customer profile information into the graph matching model to generate corresponding response content includes:

[0024] Based on the customer profile information, corresponding behavioral tags are determined, including business type preference, information focus dimension, risk tolerance, and service urgency.

[0025] Based on a preset sliding time window, the cumulative score of each behavior label is calculated;

[0026] Based on the relationship between the cumulative scores, the corresponding response strategy is determined;

[0027] Based on the set of response templates mapped by the response strategy, the corresponding response content is generated.

[0028] By adopting the above technical solution, in the process of generating response content through graph matching, behavioral tags are extracted from customer profiles, and cumulative scores are calculated by combining them with sliding time windows. Based on this, the response strategy is determined. The system can dynamically adjust the response direction according to the customer's current needs and priorities, so that the generated content is more in line with the customer's current focus and enhances the real-time performance and relevance of the response.

[0029] In a preferred embodiment, this application can be further configured such that the response template set includes at least:

[0030] Guided response templates are used to generate suggested content based on the aforementioned business type preferences;

[0031] Step-by-step response templates are used to generate operational guidance content based on the information focus dimensions;

[0032] A prudent response template for generating risk warning statements based on risk tolerance.

[0033] High-priority response templates are used to generate expedited processing prompts based on the urgency of the service.

[0034] By adopting the above technical solution, the response template set is categorized according to behavioral tag types, enabling the system to call different template types as needed during the response process, forming a distinctive and logically sound response strategy. Guided, step-by-step, deliberate, and high-priority templates correspond to different customer characteristic dimensions, ensuring that the response content is both information-rich and logically guided, greatly enhancing user convenience and trust.

[0035] In a preferred embodiment, this application can be further configured such that the step of calculating the cumulative score of each of the behavior tags based on a preset sliding time window includes:

[0036] Within a preset sliding time window, acquire records of behavioral tags identified during multiple historical interactions of the customer;

[0037] Based on the behavior tag records, the frequency of occurrence of the behavior tags is calculated, and a tag time-series weighting factor corresponding to the behavior tag is set according to the time interval between the occurrence time of the behavior tag and the current time.

[0038] Based on the time-series weighting factor, the cumulative score of each behavior label at the current time is calculated.

[0039] By adopting the above technical solution, the sliding time window mechanism combines the frequency and time factors of behavioral tags, and introduces a time-series weighting factor to adjust the score, thus taking into account both the historical importance and timeliness of tags, thereby enabling the dynamic evolution and real-time adjustment of customer profiles. This method improves the accuracy of behavioral preference modeling capabilities and helps the system continuously adapt to changes in customer status.

[0040] In a preferred embodiment, this application can be further configured such that: the customer feedback data is collected based on a feedback question window through which the customer interacts, and the steps for generating the feedback question window include:

[0041] Call upon each question dimension from the aforementioned question template set;

[0042] Determine the association score between each of the aforementioned problem dimensions and each of the aforementioned decision node identifiers in the decision path;

[0043] The question dimension with the highest association score with a decision node identifier is determined as the target dimension of the decision node identifier;

[0044] Create a corresponding feedback question window based on the target dimension.

[0045] By adopting the above technical solution and establishing association with the node identifiers in the decision-making path, the system can build a dedicated feedback window for each node according to the preset set of question templates, ensuring that the question dimension and model processing strategy are highly correlated when collecting feedback data, thereby improving the accuracy of feedback data and the relevance of subsequent training.

[0046] In a preferred embodiment, this application can be further configured such that the step of selecting the corresponding target model from each of the decision models based on the customer feedback data and the customer profile information includes:

[0047] The customer feedback data is analyzed to determine the target dimensions represented by the customer feedback;

[0048] The decision node identifier corresponding to the target dimension is determined as the target node identifier;

[0049] The decision model corresponding to the target node identifier is determined as the target model.

[0050] By employing the above technical solution, the collected customer feedback data is parsed into dimensional information, and the target dimension is matched to specific decision nodes and their corresponding models, thereby achieving automatic positioning of the model optimization target. This method establishes a fine correspondence between feedback and model structure, enabling the system to perform targeted training and optimization of failure response paths at minimal cost, and accelerating the self-learning iteration cycle.

[0051] The above-mentioned objective 2 of this application is achieved through the following technical solution:

[0052] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the aforementioned intelligent customer service response method for banks.

[0053] The above-mentioned objective three of this application is achieved through the following technical solution:

[0054] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the aforementioned intelligent customer service response method for banks.

[0055] In summary, this application includes at least one of the following beneficial technical effects:

[0056] This application constructs an intelligent customer service response mechanism based on multimodal data fusion. It also establishes a closed-loop optimization process oriented towards customer feedback, introducing customer feedback data as a post-hoc supervisory signal to achieve dynamic control of the system's response behavior and adaptive model evolution. The system not only collects multi-source data such as voice, text, and images in real time and extracts key information through feature engineering, but more importantly, it automatically captures negative feedback when customers express dissatisfaction with the response or the service falls short of expectations. It then establishes a mapping relationship between the response content and the original interaction data that triggered it, explicitly labeling these data to form training samples. Based on these samples and their associated customer profile information, the system further analyzes which decision model dominates the processing stage to attribute the cause of such dissatisfaction feedback, thereby accurately selecting the target model from multiple independent response sub-models for targeted training. This effectively solves the core problems of difficulty in perceiving, attributing, and optimizing feedback information in traditional systems, enabling the customer service system to not only possess proactive learning capabilities but also maintain the effectiveness and robustness of each model in actual interactions while ensuring service continuity. By constructing a closed-loop path encompassing data collection, understanding, generation, feedback, and retraining, the system can continuously optimize its response strategies in real-world scenarios. This enables the intelligent customer service system to achieve continuous improvement in response accuracy, customer satisfaction, and adaptability, meeting the bank's needs for efficient, precise, and evolving intelligent services. Attached Figure Description

[0057] Figure 1 This is a flowchart of a bank intelligent customer service response method according to an embodiment of this application.

[0058] Figure 2 This is a schematic diagram of a device according to one embodiment of this application. Detailed Implementation

[0059] The present application will be further described in detail below with reference to the accompanying drawings.

[0060] In one embodiment, such as Figure 1 As shown, this application discloses a bank intelligent customer service response method, which includes:

[0061] S10. Collect interactive business data from multiple interactive channels, and preprocess the interactive business data to generate corresponding feature vectors. The interactive business data is multimodal data including voice data, text data, and image data. The system first collects interactive business data from multiple interactive channels, including online banking, mobile banking, customer service hotlines, and physical branches. The interactive business data includes multimodal data such as voice data, text data, image data, and video data, as well as structured business data from the bank's core business system. The structured business data includes customer basic information, transaction records, and account status. After data collection, the system performs preprocessing operations on the interactive business data. The preprocessing includes: text cleaning of text data, specifically including removing irrelevant characters, performing word segmentation, and part-of-speech tagging; performing speech recognition on voice data to convert it into text format; and extracting keyframes from image and video data and performing feature analysis. The preprocessed data is uniformly converted into the expression form of feature vectors so that it can be input into the fusion model in the response process for multi-source feature fusion processing, providing a consistent and computable data foundation for generating high-quality response content.

[0062] S20. Clustering calculations are performed on the interactive business data to generate corresponding customer behavior features, thereby generating customer profile information for multi-strategy decision-making in the response process. The response process includes multiple sequential decision-making sub-processes, each of which applies its own trained decision-making model. The decision-making model includes at least a fusion model, an intent recognition model, and a graph matching model. The clustering calculation is implemented using an unsupervised learning algorithm, preferably K-means algorithm, DBSCAN density clustering, or a clustering method based on Gaussian mixture model (GMM). Specifically, the system constructs a behavior vector space based on the customer's historical behavior data in multiple interaction channels, including multi-dimensional features such as their operation path, access frequency, consultation topic type, interaction duration and time period. By performing clustering operations in this space, customers with similar behavioral characteristics are grouped into the same category, thereby extracting typical customer behavior feature labels, such as high-frequency transaction type, low-risk preference type, and proactive consultation type. Subsequently, the system further integrates the clustered customer behavior features with structured business data to form customer profiles for multi-strategy decision-making. These profiles encompass factors such as customer business category preferences, emotional tendencies, service response needs, and risk tolerance. This customer profile information serves as one of the input parameters in the calculation of various decision sub-processes in the subsequent response process, improving the accuracy and personalization of the responses. The decision models used in the response process each perform different functional tasks: the fusion model is responsible for integrating multi-source features, the intent recognition model identifies the customer's semantic intent, and the knowledge graph matching model obtains candidate answers based on knowledge graph query reasoning. These three models work together to complete the entire intelligent response process.

[0063] S30. Input the feature vector and customer profile information into the response process to generate the corresponding response content;

[0064] S40. Obtain customer feedback data on the response content. When the customer feedback data indicates that the customer's satisfaction with the response content is lower than a preset standard, mark the response content and the corresponding interactive business data as training samples. The determination method for the satisfaction level being lower than the preset standard includes, but is not limited to, the following situations: First, based on the satisfaction rating result actively filled in by the customer after the dialogue, if the rating is lower than the set threshold, such as below 3 points in a 5-point system, negative sample marking is automatically triggered; Second, based on the customer's behavioral signals under the current response content, such as clicking negative feedback buttons such as "unresolved" or "dissatisfied", or repeatedly entering similar question keywords in a short period of time; Third, based on the follow-up questioning behavior pattern of the dialogue, if the system detects that the user initiates more than two follow-up questions that are highly related to the original question after one reply, and these follow-up questions fail to cause semantic changes in the response content, it is determined that the initial response did not meet the customer's true intention; Fourth, combine the negative emotional features analyzed from the customer's text or voice by the emotion recognition model, such as doubt, questioning, and complaints, and use emotional fluctuations as an auxiliary factor to enhance the confidence of negative sample judgment. The system can flexibly configure weighted combinations of multiple judgment indicators and precisely trigger the marking logic through threshold control strategies to ensure the credibility and representativeness of the feedback information.

[0065] S50. Based on customer feedback data and customer profile information, select the corresponding target model from various decision-making models, and train the target model based on the training samples.

[0066] In one embodiment, step S30, which involves inputting the feature vector and customer profile information into the response process to generate the corresponding response content, includes:

[0067] S301. Based on the feature vector, the system matches the target sub-processes that the customer profile information needs to participate in within each decision-making sub-process. The target sub-processes include at least the decision-making sub-processes corresponding to the fusion model, intent recognition model, and graph matching model. Specifically, the system first cross-compares the multimodal information dimensions contained in the feature vector (such as text complexity, voice emotion intensity, image recognition confidence, etc.) with the feature fields in the customer profile information (such as business type preference, service urgency, risk tolerance, etc.), and then matches them with predefined process triggering rules. This rule base sets triggering conditions for each decision-making sub-process. For example, if the customer profile shows a high-risk sensitive type, and the current semantic intent involves product investment consultation, then the sub-processes corresponding to the fusion model and graph matching model should be triggered. Similarly, if the text content is represented as a complex interrogative statement, the system will prioritize the intent recognition model for deep understanding. Based on rule matching, the system can further introduce a lightweight strategy selection model to score and rank the matching confidence between the feature vector and each sub-process. Sub-processes with a comprehensive score higher than a threshold are identified as target sub-processes. Finally, the decision models involved in the target sub-process are configured as the processing path of the current response task to ensure that customer profile information is accurately involved in the most relevant decision nodes, thereby improving the relevance of subsequent response content and service effectiveness.

[0068] S302. Based on the execution order and purpose of the target sub-processes, a corresponding decision path is constructed. This decision path is a data sequence composed of multiple decision node identifiers arranged according to the execution order. The decision node identifiers are generated based on the process purpose, and the data sequence is used to select the corresponding target model. The system first analyzes the preceding feature vectors and customer profile information to clarify the specific responsibilities and logical roles of each target sub-process in the response process. Each target sub-process corresponds to a specific process purpose. For example, the fusion model integrates multimodal information to form a semantic representation, the intent recognition model determines the customer's intent category, and the graph matching model is responsible for obtaining the most relevant knowledge data from the knowledge graph based on the identified intent. Based on this, the system automatically derives a sub-process execution order with a clear semantic orientation based on the functional dependencies between these sub-processes and the priority of customer question responses, such as from the fusion model to the intent recognition model and then to the graph matching model. The system further configures a unique decision node identifier for each target sub-process. This identifier not only reflects its position in the overall process but also embeds a semantic label corresponding to its process purpose. Finally, all decision node identifiers are combined according to the above process execution order to form a logically continuous and semantically clear decision path, which can be used for association mapping and path tracing in subsequent steps such as feedback question window generation and target model selection, thereby improving the interpretability of process execution and the accuracy of feedback direction.

[0069] S303. Generate the corresponding response content based on the feature vector, customer profile information, and target sub-process.

[0070] In one embodiment, step S303, namely the step of generating the corresponding response content based on the feature vector, customer profile information, and target sub-process, includes:

[0071] S3031. Input the feature vector and customer profile information into the fusion model. The customer profile information is used to guide the multi-source feature fusion strategy of the fusion model to obtain the fused comprehensive feature representation.

[0072] S3032. Input the comprehensive feature representation and customer profile information into the intent recognition model, wherein the customer profile information is used to assist in recognizing the semantic intent information involved in the customer's current consultation.

[0073] S3033. Input semantic intent information and customer profile information into the graph matching model, perform matching query in the knowledge graph pre-built in the graph matching model, and obtain target knowledge data related to customer profile information based on semantic intent information;

[0074] S3034. Input the target knowledge data and customer profile information into the graph matching model to generate the corresponding response content.

[0075] In this embodiment, the system first invokes the multimodal feature fusion and semantic intent recognition module to comprehensively process the preprocessed feature vectors. Specifically, the system employs a fusion model to fuse the input multimodal feature vectors, including text, speech, and images. This fusion model is a multimodal fusion neural network incorporating an attention mechanism, capable of automatically identifying the importance weights of each modality in the current interactive context and adjusting the fusion strategy based on these weights to obtain a comprehensive feature representation that includes the customer's potential intent and emotional inclination. Next, the system inputs this comprehensive feature representation along with customer profile information into the intent recognition model. This intent recognition model is a hybrid deep neural network combining recurrent neural networks and convolutional neural networks, used to perform intent classification on the fused input to identify key information such as the specific question category, service demand objective, and emergency response level of the customer's current inquiry.

[0076] After acquiring semantic intent information, the system enters the graph matching model processing stage. This graph matching model operates based on a pre-built banking knowledge graph, which integrates banking business knowledge, industry-wide common knowledge, and customer consultation cases accumulated through historical response processes. Nodes in the graph include banking business entities such as accounts, credit cards, and loan products, while edges represent process relationships, conditional constraints, or semantic dependencies between entities. Based on the semantic intent information output by the intent recognition model, the system performs a graph query operation through the graph matching model to quickly retrieve target knowledge data associated with the current customer intent, obtaining highly relevant knowledge nodes and knowledge paths that match the customer profile information features, which serve as the data foundation for subsequent response generation.

[0077] Subsequently, the system invokes the response generation module, combining the aforementioned target knowledge data with the customer's historical interaction information. Through a sequence-to-sequence natural language generation model and a template library-driven mechanism, it generates accurate, natural, and context-appropriate response content. After the response content is generated, the system further optimizes it. This includes automatically verifying the semantic coherence of the generated text, conducting compliance reviews to ensure the response content conforms to banking business norms and regulatory policies, and selecting an appropriate language style and tone based on the customer's service style preferences and semantic intent type from the customer profile information. This results in a personalized final response content that is both business-accurate and appropriate.

[0078] In one embodiment, step S3034, which involves inputting the target knowledge data and customer profile information into the graph matching model to generate the corresponding response content, includes:

[0079] S30341. Based on customer profile information, determine corresponding behavioral tags. These tags include business type preference, information focus dimensions, risk tolerance, and service urgency. Customer profile information refers to a structured feature set constructed by analyzing multi-dimensional information such as customer behavior characteristics, business preferences, risk attitudes, and service needs across multiple interaction channels. This data serves as the foundation for the intelligent customer service system to execute differentiated response strategies at different stages. Behavioral tags are core characteristic indicators extracted from customer profile information to express customers' performance in specific service dimensions. These primarily include the customer's preference for specific business types, focus of information attention, risk tolerance when facing financial services, and assessment of the urgency of the current inquiry.

[0080] S30342. Based on a preset sliding time window, calculate the cumulative score for each behavioral tag. Customer profile information refers to a structured feature set constructed by analyzing multi-dimensional information such as customer behavior characteristics, business preferences, risk attitudes, and service needs across multiple interaction channels. This data serves as the foundation for the intelligent customer service system to execute differentiated response strategies at different stages. Behavioral tags are core characteristic indicators extracted from customer profile information to express customers' performance in specific service dimensions. These mainly include the customer's preference for specific business types, focus of information attention, risk tolerance when facing financial services, and assessment of the urgency of responding to the current inquiry.

[0081] S30343. Based on the cumulative scores, determine the corresponding response strategy. Customer profile information refers to a structured feature set constructed by analyzing multi-dimensional information such as customer behavior characteristics, business preferences, risk attitudes, and service needs across multiple interaction channels. This data serves as the foundation for the intelligent customer service system to execute differentiated response strategies at different stages. Behavioral tags are core characteristic indicators extracted from customer profile information to express customers' performance in specific service dimensions. These mainly include the customer's preference for specific business types, the focus of their information attention, their risk tolerance when facing financial services, and their assessment of the urgency of responding to the current inquiry.

[0082] S30344. Generate corresponding response content based on the response template set mapped by the response strategy. The response template set is a set of pre-designed and categorized structured response statement templates. Each type of template is adapted to different response strategy logic, which is used to quickly match and construct language output with personalized style and business accuracy when generating response content.

[0083] In one embodiment, in step S30344, the response template set includes at least:

[0084] Guided response templates are used to generate suggested content based on business type preferences. These templates are designed for customers with strong business type preferences, and their content structure tends to provide proactive recommendations and processing suggestions for a certain type of product or service. For example, for customers who prefer investment and wealth management services, the template will prioritize recommending wealth management products with high current returns and include a processing link to meet the customer's needs for independent exploration and quick access to services.

[0085] Step-by-step response templates are used to generate operational guidance content based on the information focus dimensions. These templates are primarily suitable for customers with high information focus dimensions. Their structure typically includes a breakdown of the operation path, explanations of necessary conditions, and reminders of frequently asked questions. The aim is to assist customers in completing a business operation through clear, step-by-step guidance. For example, when a customer is concerned about the loan repayment process, the system will use this type of template to sequentially output the repayment calendar, available channels, and fee explanations.

[0086] The prudent response template generates risk warning statements based on risk tolerance. This template caters to clients with lower risk tolerance, emphasizing risk control reminders, prudent information disclosure, and explanations of protection mechanisms. For example, when a client inquires about credit card installment plans or investment products, the system will prioritize using this template to provide risk warnings, including explanations of return fluctuations, principal risk, and policy changes, to guide the client in making rational decisions.

[0087] High-priority response templates are used to expedite processing based on the urgency of the service; prudent response templates serve customers with lower risk tolerance, emphasizing risk control reminders, prudent information disclosure, and explanations of protection mechanisms. For example, when a customer inquires about credit card installment plans or investment products, the system will prioritize using this template to provide risk warnings, including explanations of return fluctuations, principal risks, and policy changes, to guide the customer in making rational decisions.

[0088] In this embodiment, the response template set is categorized according to behavioral tag types, enabling the system to call different template types as needed during the response process, forming a distinctive and logically sound response strategy. Guided, step-by-step, deliberate, and high-priority templates correspond to different customer characteristic dimensions, ensuring that the response content is both informationally comprehensive and logically guided, greatly enhancing user convenience and trust.

[0089] In one embodiment, step S30342, namely the step of calculating the cumulative score of each behavior label based on a preset sliding time window, includes:

[0090] S303421. Within a preset sliding time window, acquire the behavioral tag records identified by the customer during multiple historical interactions. The behavioral tag record refers to the set of structured identifiers that the system identifies based on the content of each interaction during multiple historical customer interactions, which are associated with the customer's preferences, concerns, needs, or habits. Each behavioral tag record contains the tag type, tag value, and its timestamp, which are used to record the behavioral tendencies exhibited by the customer at a specific moment.

[0091] S303422. Based on behavioral tag records, the frequency of occurrence of behavioral tags is statistically determined, and a tag time-series weighting factor corresponding to the behavioral tag is set according to the time interval between the occurrence time of the behavioral tag and the current time. The tag time-series weighting factor is a coefficient that assigns differentiated weights to behavioral tags based on the order of their occurrence. It mainly relies on the time interval between the behavioral tag and the current time to perform exponential decay or linear scaling, so that the weight of recent behaviors is higher and the influence of long-term behaviors is weakened, thereby ensuring the sensitivity of the model to the latest intentions of customers.

[0092] S303423. Based on the time-series weighting factor, calculate the cumulative score of each behavior label at the current time.

[0093] Specifically, a customer has engaged in five inquiries related to financial products over the past three months. The most recent inquiry occurred three days ago, while the other four occurred one week, two weeks, one month, and two months ago, respectively. The system will extract all financial-related behavioral tags within the currently defined three-month sliding time window, counting the frequency of each tag as five times. Different time-series weighting factors will be assigned to each interaction based on its time interval from the current interaction, with more recent records receiving higher weights. Finally, by combining frequency and weighting factors, the system calculates a high cumulative score for the customer in financial matters, indicating sustained interest in this type of service. Therefore, the system will prioritize recommending financial services or using relevant response templates in subsequent responses.

[0094] In one embodiment, in step S40, customer feedback data is collected based on a feedback question window used for customer interaction. The feedback question window is a set of interactive questions or evaluation interfaces generated by the system based on the customer's current interaction content. Its design purpose is to guide the customer to provide clear feedback on a specific response stage, thereby establishing an effective mapping relationship between feedback information and specific processing stages or decision nodes. The steps for generating the feedback question window include:

[0095] S401. Invoke the various question dimensions from the question template set. Each question dimension represents the system's focus on a certain type of response quality indicator or interaction step, such as accuracy, timeliness, process guidance, and personalization, used to support the dynamic construction of the feedback question window. Question dimensions are structured feedback directions defined in the question template set to measure specific attributes of a decision-making step or response behavior. Each dimension represents a feedback perspective that the system focuses on, such as whether the customer finds the information clear, whether the operation instructions are clear, or whether the recommended content is appropriate.

[0096] S402. Determine the association score between each problem dimension and each decision node identifier in the decision path; the association score refers to the quantitative evaluation result of the correlation between a certain problem dimension and a certain decision node identifier. It is usually calculated through the feature mapping relationship built into the model, historical feedback sample statistics or expert experience rules, to evaluate whether the problem dimension is suitable as the content source of the feedback window of the node.

[0097] S403. The issue dimension with the highest correlation score with a decision node identifier is determined as the target dimension of that decision node identifier. The target dimension is the issue dimension selected from multiple issue dimensions based on its highest correlation score with a specific decision node identifier. It is used to construct the feedback window content that best reflects the processing quality or issue risk of that node, thereby improving the relevance and usability of the feedback.

[0098] Create corresponding feedback question windows based on the target dimensions.

[0099] Specifically, when a customer inquires about credit card installment plans, the system's decision-making path sequentially includes a fusion model to determine the service type, an intent recognition model to identify the installment needs, and a graph matching model to retrieve interest rate policy information, ultimately generating a response text. When a customer states that "the answer wasn't helpful," the system traces back to the relevant decision node based on their interaction flow. It finds that this node has the highest correlation score with the "unclear process guidance" issue dimension, so it selects this dimension as the target dimension. Based on "process guidance" related questions in the question template set, it dynamically generates a feedback question window, such as "Do you feel that the system's guidance on installment procedures is not clear enough?", guiding the customer to further clarify their dissatisfaction. This mechanism helps direct ambiguous feedback to specific process nodes, facilitating subsequent optimization efforts.

[0100] In one embodiment, step S50, which involves selecting the corresponding target model from various decision models based on customer feedback data and customer profile information, includes:

[0101] S501. Analyze customer feedback data to determine the target dimensions represented by the customer feedback;

[0102] S502. Determine the decision node identifier corresponding to the target dimension as the target node identifier;

[0103] S503. Determine the decision model corresponding to the target node identifier as the target model.

[0104] Specifically, the system first performs structured analysis on the feedback data submitted by customers after the interaction, extracting the core concerns reflected in the feedback content. For example, when a customer selects "not clear enough" as feedback, combined with the question dimension bound to the feedback question window from which the feedback originated, the system can identify that the customer's feedback represents the target dimension of "unclear operation instructions." Subsequently, the system uses a pre-set dimension-node mapping table to identify the decision node corresponding to the target dimension of "unclear operation instructions" as node number A-3 in the graph matching model responsible for service process entity retrieval. The decision model to which this node belongs is then identified as the graph matching model and recorded as the target model. This achieves an effective mapping from vague customer feedback to a specific response process model, enabling rapid and accurate location of the model's responsibility area without manual intervention. This provides a basis for subsequent targeted model optimization and training, thereby realizing closed-loop learning and autonomous optimization of the intelligent customer service system.

[0105] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 2 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a smart customer service response method for banks.

[0106] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps:

[0107] S10. Collect interactive business data from multiple interactive channels, preprocess the interactive business data to generate corresponding feature vectors, and the interactive business data is multimodal data including voice data, text data and image data.

[0108] S20. Cluster the interactive business data to generate corresponding customer behavior features, and then generate customer profile information for multi-strategy decision-making in the response process. The response process includes multiple decision sub-processes that are carried out in sequence. Each decision sub-process applies its own trained decision model. The decision model includes at least a fusion model, an intent recognition model, and a graph matching model.

[0109] S30. Input the feature vector and customer profile information into the response process to generate the corresponding response content;

[0110] S40. Obtain customer feedback data on the response content. When the customer feedback data indicates that the customer's satisfaction with the response content is lower than the preset standard, mark the response content and the corresponding interactive business data as training samples.

[0111] S50. Based on customer feedback data and customer profile information, select the corresponding target model from various decision-making models, and train the target model based on the training samples.

[0112] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0113] S10. Collect interactive business data from multiple interactive channels, preprocess the interactive business data to generate corresponding feature vectors, and the interactive business data is multimodal data including voice data, text data and image data.

[0114] S20. Cluster the interactive business data to generate corresponding customer behavior features, and then generate customer profile information for multi-strategy decision-making in the response process. The response process includes multiple decision sub-processes that are carried out in sequence. Each decision sub-process applies its own trained decision model. The decision model includes at least a fusion model, an intent recognition model, and a graph matching model.

[0115] S30. Input the feature vector and customer profile information into the response process to generate the corresponding response content;

[0116] S40. Obtain customer feedback data on the response content. When the customer feedback data indicates that the customer's satisfaction with the response content is lower than the preset standard, mark the response content and the corresponding interactive business data as training samples.

[0117] S50. Based on customer feedback data and customer profile information, select the corresponding target model from various decision-making models, and train the target model based on the training samples.

[0118] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0119] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0120] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for responding to a bank intelligent customer service, characterized in that, The intelligent bank customer service response method comprises the following steps: Collecting interactive service data in multiple interactive channels, preprocessing the interactive service data to generate corresponding feature vectors, and the interactive service data being multi-modal data including voice data, text data and picture data; Carrying out cluster calculation on the interactive service data to generate corresponding customer behavior characteristics, and further generating customer portrait information for multi-strategy decision-making in the response process, the response process comprising multiple decision sub-processes carried out in sequence, wherein each decision sub-process applies a respective trained decision model, and the decision model at least comprises a fusion model, an intent recognition model and a graph matching model; Inputting the feature vectors and the customer portrait information into the response process to generate corresponding response content; Obtaining customer feedback data of the response content, and when the customer feedback data represents that the customer's satisfaction with the response content is lower than a preset standard, marking the response content and the interactive service data corresponding to the response content as training samples; According to the customer feedback data and the customer portrait information, selecting a corresponding target model in each decision model, and training the target model according to the training samples; The step of inputting the feature vectors and the customer portrait information into the response process to generate corresponding response content comprises the following steps: According to the feature vectors, matching the target sub-processes required by the customer portrait information in each decision sub-process, and the target sub-processes at least comprising the decision sub-processes corresponding to the fusion model, the intent recognition model and the graph matching model; According to the flow execution sequence and the flow purpose of the target sub-processes, constructing a corresponding decision path, wherein the decision path is a data sequence composed of multiple decision node identifiers and arranged according to the flow execution sequence, the decision node identifiers are generated based on the flow purpose, and the data sequence is used to select the corresponding target model; According to the feature vectors and the customer portrait information, and the target sub-processes, generating corresponding response content; The step of generating corresponding response content according to the feature vectors and the customer portrait information, and the target sub-processes comprises the following steps: Inputting the feature vectors and the customer portrait information into the fusion model together, wherein the customer portrait information is used to guide the multi-source feature fusion strategy of the fusion model to obtain the integrated feature representation after fusion; Inputting the integrated feature representation and the customer portrait information into the intent recognition model together, wherein the customer portrait information is used to assist in identifying the semantic intent information involved in the customer's current consultation; Inputting the semantic intent information and the customer portrait information into the graph matching model together, performing matching query in the knowledge graph pre-constructed in the graph matching model, and obtaining target knowledge data associated with the customer portrait information according to the semantic intent information; The target knowledge data and the customer portrait information are input into a graph matching model to generate corresponding response content.

2. The method of claim 1, wherein the method further comprises: The step of inputting the target knowledge data and the customer portrait information into the graph matching model to generate corresponding response content comprises: According to the customer portrait information, a corresponding behavior label is determined, and the behavior label comprises a business type preference, an information attention dimension, a risk tolerance, and a service urgency level; Based on a preset sliding time window, the cumulative score of each behavior label is calculated; According to the size relationship of the cumulative scores, a corresponding response strategy is determined; According to the response template set mapped by the response strategy, corresponding response content is generated.

3. The method of claim 2, wherein the method further comprises: The response template set at least comprises: A guided response template for generating suggestion content according to the business type preference; A step-by-step response template for generating operation guide content according to the information attention dimension; A prudent response template for generating a risk prompt according to the risk tolerance; A high-priority response template for generating an accelerated processing prompt according to the service urgency level.

4. The method of claim 2, wherein the method further comprises: The step of calculating the cumulative score of each behavior label based on the preset sliding time window comprises: Within the preset sliding time window, the behavior label records recognized by the customer in the historical multiple interaction processes are obtained; Based on the behavior label records, the occurrence frequency of the behavior label is counted, and a label time sequence weighting factor corresponding to the behavior label is set according to the time interval between the occurrence time of the behavior label and the current time; Based on the time sequence weighting factor, the cumulative score of each behavior label at the current time is calculated.

5. The method of claim 1, wherein the method further comprises: The customer feedback data is collected based on a feedback question window for interacting with the customer, and the generation step of the feedback question window comprises: Each question dimension in the question template set is called; The association score of each question dimension associated with each decision node identifier in the decision path is determined; The question dimension with the highest association score for a decision node identifier is determined as the target dimension for the decision node identifier; According to the target dimension, a corresponding feedback question window is created.

6. The method of claim 5, wherein the method further comprises: The step of selecting a corresponding target model in each decision model based on the customer feedback data and the customer portrait information comprises: The customer feedback data is analyzed to determine the target dimension represented by the customer feedback; The decision node identifier corresponding to the target dimension is determined as the target node identifier; The decision model corresponding to the target node identifier is determined as the target model.

7. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the steps of the bank intelligent customer service response method according to any one of claims 1 to 6.

8. A computer-readable storage medium storing a computer program, the computer-readable storage medium comprising: The computer program is executed by the processor to implement the steps of the bank intelligent customer service response method according to any one of claims 1 to 6.

Images