Business service processing method and device based on dialogue recognition, equipment and medium

Abstract

The invention relates to the technical field of semantic analysis, can be applied to business scenes in the fields of medical health, financial science and technology, intelligent interactive service and the like, and discloses a business service processing method based on dialogue recognition, which comprises the following steps: performing semantic analysis on current dialogue content, determining a key problem type, and obtaining a business processing record of a dialogue subject; under the condition that the service processing problem type is identified, obtaining a target operation record in combination with problem occurrence time and a current time interval, constructing a user operation behavior track based on an interface interaction event in the target operation record, and identifying a deviation node in the track; and generating a business processing exception report, matching the target service item in the service resource library according to the service demand keyword and the business processing record, and outputting operation guidance information. By fusing dialogue semantic analysis and user behavior trajectory modeling, automatic identification and accurate service recommendation of customer problems are realized, and the business processing efficiency and the service intelligence level are improved.

Description

Technical Field

[0001] The present invention relates to the field of semantic parsing technology, and in particular to a business service processing method, device, equipment and storage medium based on dialogue recognition. Background Art

[0002] In the current fintech business field, with the rapid growth of insurance companies' business volume, customer service scenarios have become increasingly complex, and a large number of users have generated high-frequency consultation and feedback needs in key processes such as insurance and claims. However, existing customer service systems generally have problems such as delayed information response and fragmented processing links. When customers proactively call the customer service hotline to report problems, it is difficult for the system to retrieve their historical policy information, insurance process records or claim details in the first place. As a result, customer service personnel need to go through multiple rounds of manual verification and multi-terminal system searches to initially obtain key information, which significantly reduces the overall problem handling efficiency, especially when facing time-sensitive requests (such as temporary modifications to insurance information or urging claims progress).

[0003] In the medical and health business field, user-oriented intelligent customer service systems also face technical bottlenecks such as slow service response and weak problem understanding. For example, when patients inquire about the coverage or reimbursement ratio of health insurance products, the current system often lacks the ability to analyze conversational semantics and cannot accurately understand the core issues that users are really concerned about. As a result, it is unable to provide targeted replies or guidance and needs to be transferred to manual processes, which not only increases service costs but also reduces user satisfaction. In addition, when a user encounters an operational abnormality in a certain service link (such as failure to fill out a form, page loading error, etc.), the system is usually unable to identify the specific card point link, and it is even more difficult to restore the user's operation path, resulting in customer service personnel relying on user oral narration to trace back step by step, which is not only inefficient but also has poor problem location accuracy.

[0004] Amid the growing popularity of intelligent interactive systems, traditional customer service systems still rely heavily on manual intervention for problem identification and resolution. They lack the ability to intelligently analyze conversation content, making it impossible to classify problem types and match them with resolution strategies in real time. Furthermore, the system fails to establish a unified view of customer behavior, making it difficult to infer the user's current business node based on historical behavior, thus preventing accurate problem resolution and response. These flaws not only extend service time but also magnify the risk of customer churn at critical junctures. This is particularly true for users who are applying for insurance or making claims, who are easily abandoned due to long wait times or unprofessional responses.

[0005] In addition, the existing customer service system lacks personalization capabilities in providing service recommendations. It usually makes simple product recommendations based only on static user tags, ignoring the dynamic intention expressions of users in conversations and historical service preference data, resulting in a mismatch between the recommended results and the actual needs of customers, and it is difficult to improve the customer experience and conversion rate. Therefore, in the context of coexistence of multiple business types and diverse customer demands, the existing system has significant deficiencies in aspects such as intelligent understanding, problem positioning, path restoration, and service recommendation, and there is an urgent need to conduct systematic upgrades by integrating technical means such as natural language processing, user behavior modeling, and multi-source data matching. Summary of the Invention

[0006] The main objective of the present invention is to provide a business service processing method, device, equipment, and storage medium based on dialogue recognition, aiming to solve the technical problem that the existing technology cannot intelligently identify the problem type based on the user's conversation content and combine the user's historical operation behaviors to achieve problem path restoration and anomaly positioning, resulting in untimely problem response and unclear processing links.

[0007] To achieve the above objective, the present invention provides a business service processing method based on dialogue recognition, including:

[0008] Parse the current conversation content through a semantic analysis module to generate a key problem type;

[0009] According to the key problem type, obtain the business processing records of the session subject of the current conversation content;

[0010] When the key problem type is a business processing problem type, obtain the target operation record according to the interval between the problem occurrence time and the current time;

[0011] Construct a user operation behavior trajectory based on the interface interaction events in the target operation record;

[0012] Identify the deviation nodes in the user operation behavior trajectory;

[0013] Generate a business processing anomaly report according to the deviation nodes;

[0014] Match the target service item in the service resource library according to the service demand keywords in the current conversation content and the business processing records, and generate operation guidance information for the target service item.

[0015] Furthermore, to achieve the above objective, the present invention provides a business service processing device based on dialogue recognition, including:

[0016] A semantic recognition module, configured to parse the current conversation content through a semantic analysis module to generate a key problem type;

[0017] A business data extraction module, configured to obtain the business processing records of the session subject of the current conversation content according to the key problem type;

[0018] An operation record generation module, configured to obtain target operation records according to the interval between the problem occurrence time and the current time when the key problem type is a business processing problem type;

[0019] An operation trace construction module, configured to construct a user operation behavior trace based on the interface interaction events in the target operation records;

[0020] An abnormal node identification module, configured to identify deviation nodes in the user operation behavior trace;

[0021] An abnormal report generation module, configured to generate a business processing abnormal report according to the deviation nodes;

[0022] A service matching and guidance module, configured to match a target service item in a service resource library according to the service demand keywords in the current conversation content and the business processing records, and generate operation guidance information for the target service item.

[0023] Further, to achieve the above object, the present invention further provides a computer device, where the computer device includes a memory, a processor, and a business service processing program based on dialogue recognition stored in the memory and executable on the processor. When the business service processing program based on dialogue recognition is executed by the processor, the steps of the business service processing method based on dialogue recognition as described above are implemented.

[0024] Further, to achieve the above object, the present invention further provides a computer-readable storage medium, where a business service processing program based on dialogue recognition is stored on the storage medium. When the business service processing program based on dialogue recognition is executed by a processor, the steps of the business service processing method based on dialogue recognition as described above are implemented.

[0025] Beneficial effects: The present invention relates to the technical field of semantic parsing and can be applied to business scenarios such as the fields of medical health, fintech, and intelligent interaction services. It discloses a business service processing method based on dialogue recognition, including: parsing the current dialogue content to generate a key problem type, obtaining the business processing record of the conversation subject of the current dialogue content according to the key problem type, and in the case where the key problem type is a business processing problem type, obtaining a target operation record according to the interval between the problem occurrence time and the current time, constructing a user operation behavior trajectory based on the interface interaction events in the target operation record, identifying the deviation nodes in the user operation behavior trajectory, generating a business processing exception report according to the deviation nodes, and matching the target service item in the service resource library by combining the service demand keywords and the business processing record in the current dialogue content, and generating the operation guidance information of the target service item. Through collaborative processing mechanisms such as semantic analysis, historical behavior data extraction, and behavior trajectory construction, the present invention can intelligently identify the customer problem type, accurately locate the abnormal nodes of the user in the business handling process, automatically generate a structured exception report, and complete service item matching and path guidance in combination with the service demand, so as to achieve efficient problem response and personalized service recommendation, and improve the business support ability and service efficiency of the intelligent customer service system. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The present invention will be further described below in conjunction with the drawings and embodiments. In the drawings:

[0027] Figure 1 is a schematic diagram of an application environment of the business service processing method based on dialogue recognition in an embodiment of the present invention;

[0028] Figure 2 is a schematic flowchart of an embodiment of the business service processing method based on dialogue recognition of the present invention;

[0029] Figure 3 is a schematic diagram of the functional modules of a preferred embodiment of the business service processing device based on dialogue recognition of the present invention;

[0030] Figure 4 is a schematic diagram of the structure of a computer device in an embodiment of the present invention;

[0031] Figure 5 is another schematic diagram of the structure of a computer device in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0032] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0033] The business service processing method based on dialogue recognition provided by the embodiments of the present invention can be applied in, for example Figure 1In the application environment, the client communicates with the server through the network. The server can generate the key problem type by parsing the current conversation content through the client, obtain the business processing record of the conversation subject of the current conversation content according to the key problem type, and when the key problem type is the business processing problem type, obtain the target operation record according to the interval between the problem occurrence time and the current time, construct the user operation behavior track based on the interface interaction events in the target operation record, identify the deviation nodes in the user operation behavior track, generate a business processing exception report according to the deviation nodes, match the target service item in the service resource library by combining the service demand keywords and the business processing record in the current conversation content, and generate the operation guide information of the target service item. Through the collaborative processing mechanisms such as semantic analysis, historical behavior data extraction, and behavior track construction, the present invention can intelligently identify the customer problem type, accurately locate the abnormal nodes of the user in the business handling process, automatically generate a structured exception report, and complete service item matching and path guidance in combination with the service demand, so as to achieve efficient problem response and personalized service recommendation, and improve the business support ability and service efficiency of the intelligent customer service system. Among them, the client can be, but is not limited to, various personal computers, laptop computers, smart phones, tablet computers, and portable wearable devices. The server can be implemented by an independent server or a server cluster composed of multiple servers. The present invention will be described in detail below through specific embodiments.

[0034] Please refer to Figure 2 , Figure 2 which is a schematic flowchart of an embodiment of the business service processing method based on dialogue recognition provided by the present invention. It should be noted that although the logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than here.

[0035] As Figure 2 shown, the business service processing method based on dialogue recognition proposed by the present invention includes the following steps:

[0036] S10, parse the current conversation content through the semantic analysis module to generate the key problem type;

[0037] In this embodiment, parsing the current conversation content and generating the key problem types rely on the semantic analysis module's ability to structurally process natural language. The semantic analysis module refers to a class of natural language understanding components integrated in the service system, and its core function is to convert the user's conversation content into structured data tags or classification results. The semantic analysis module generally includes a text preprocessing unit, a word segmentation module, a part-of-speech tagging module, a syntactic dependency analysis module, an intention recognition module, etc. The text preprocessing unit is used to remove redundant content in unstructured statements, such as punctuation marks, meaningless words, etc. The word segmentation module performs word segmentation on the input continuous text according to a preset user dictionary or a deep learning word vector model. The part-of-speech tagging module uses sequence tagging algorithms, such as conditional random fields (CRF) or BiLSTM-CRF structures, to complete the part-of-speech tagging of the segmented text and provide the basis for semantic roles. The syntactic dependency analysis module uses dependency tree modeling technology to identify the dependency relationships between words and form basic semantic units. The intention recognition module maps the current semantic unit to a predefined set of key problem type tags, such as "insurance application failure", "claim settlement exception", "service complaint", "data upload failure", etc., through a classifier based on the Transformer structure or BERT semantic vectors. This process requires that the tag set cover all expected business scenarios and leave room for generalization to adapt to future business expansion.

[0038] To improve accuracy, the semantic analysis module can integrate a context reasoning mechanism to understand the logical relationship between the current statement and the historical conversation through cross-turn dialogue. For example, when the user only mentions "can't open the payment page" in the current conversation, the system needs to trace back to "I have selected an insurance product" mentioned in the previous conversation and comprehensively judge that the intention is "abnormal payment process". In addition, this module supports model fine-tuning based on domain corpora, so that in fields such as finance and healthcare, industry terms can be more accurately recognized. For example, words such as "insurance application number", "physical examination report", "risk control reminder", etc. are accurately mapped to predefined key problem types.

[0039] The key problem type refers to the problem classification result identified based on the current conversation content, which has a clear business orientation and service intervention requirements, and is used to guide subsequent tasks such as business record extraction, operation track analysis, resource matching, and exception handling. The generation process of the key problem type depends on the semantic analysis module's structural understanding of the semantic units in the conversation content, and combines domain knowledge graphs, historical conversation data tags, and business process semantic templates for attribution judgment. This type is not limited to general speech classification, but is a semantic classification label abstracted for business goals, with the characteristics of strong operability and high semantic focus.

[0040] In actual implementation, the key problem types usually include but are not limited to the following categories: business processing problem types, service evaluation problem types, process lag problem types, resource access problem types, product consultation problem types, rule understanding problem types, operation guidance problem types, risk warning problem types, identity authentication problem types, etc. Each category of key problem types corresponds to different technical response mechanisms. For example, business processing problems will trigger the backtracking of user operation trajectories and the detection of abnormal nodes, while resource access problems are more involved in front-end path generation and permission verification processes.

[0041] To ensure the accuracy of problem type classification, the semantic analysis module adopts an intent recognition algorithm based on a large language model (LLM), and conducts clustering modeling by combining keywords, intent transfer features, and context entity relationships in the multi-round dialogue context. At the same time, through manual annotation and model fine-tuning of historical conversation samples, an intent label system that conforms to industry semantics is gradually established. To support scalability and configurability between types, a hybrid matching strategy based on rule templates and knowledge graphs is also supported. For example, when keywords such as "medical insurance, reimbursement, designated hospital" are mentioned in a medical scenario, it will be preferentially determined as a medical insurance-related problem, while in the financial field, when semantic units such as "payment failure, policy missing, claim submission" are recognized, they will be classified as business processing problem types.

[0042] In addition, the determination of key problem types not only depends on the recognition of explicit keywords, but also includes emotion recognition, sentence pattern intensity analysis, and goal-oriented judgment. For example, when the user uses a tone of anxiety or doubt and mentions the failure of a certain type of business node multiple times, the system will increase the weight of this problem in the risk or complaint category. In a specific scenario, there may be multiple candidate problem types for the same piece of dialogue content. The system will make a normalized judgment based on comprehensive indicators such as semantic confidence, historical behavior characteristics, and intent priority, and select the most representative problem label as the key problem type for the current dialogue.

[0043] In the actual deployment process, the semantic analysis module can be implemented based on a two-stage model architecture. In the first stage, the BERT model is used to encode the current conversation content and output a semantic representation vector. In the second stage, a multi-class Softmax layer is introduced to perform classification judgment based on the similarity between the encoded vector and the semantic vectors of various question types. The sources of training data include historical customer service conversation data, manually labeled datasets, and semi-supervised label extension corpora. To enhance the generalization ability of the model, adversarial training techniques can be used to generate approximate semantic perturbation texts. For example, replace "The payment page cannot be opened" with "The payment interface cannot be accessed" to expand the training data distribution. In terms of system implementation, the current conversation content can be passed in real-time through the front-end customer service interface, and the semantic analysis service is called through RPC or RESTAPI to obtain the returned key question type labels. To improve the response speed and service reliability, the semantic analysis module is deployed on an edge cloud node that supports GPU inference, and an asynchronous push mechanism is configured to send the labels back to the customer service processing platform. In the scenario of multi-service parallelism, this module can maintain the conversation turn state through a context caching mechanism, and cache and store the keywords and semantic labels of each round of discourse to avoid repeated parsing. At the same time, for statements that cannot be directly classified, such as "Why are you so slow", the system will call rule templates and auxiliary classification models to try to classify them into fuzzy types such as "Service response delay" or "System processing slow" to ensure that the system has the ability to recognize atypical expressions.

[0044] Example illustration: In the medical and health field, when a user expresses "The hospital refuses to prescribe medicine", "Outpatient settlement fails", "Medical insurance cannot be reimbursed", etc. during the interaction with the customer service, the semantic analysis module will extract semantic units such as "Refusing to prescribe medicine", "Settlement fails", and "Reimbursement" from the conversation. Combining the semantic classification and rule templates of the medical insurance business nodes defined in the medical domain knowledge graph, the system can identify the conversation content as a business processing problem type related to medical insurance settlement. Subsequently, the system will automatically retrieve the user's medical insurance reimbursement records and outpatient registration operation history, and determine the target operation records based on the problem occurrence time, providing basic data for subsequent behavior trajectory analysis and abnormal node diagnosis, and ultimately supporting the customer service to quickly lock the specific operation page and error message content of "Medical insurance settlement fails".

[0045] In financial business scenarios, customers may express feedback such as "I paid but the page didn't respond," "I clicked confirm but the policy didn't come up," and "the system keeps spinning and getting stuck." The system automatically identifies key issues such as "process jams" or "business processing issues" in the insurance process through keywords such as "payment," "policy," and "stuck" in the conversation, combined with part-of-speech analysis and behavioral intent models. It then further identifies the user as suspected of being stuck at the node where "jump failed after payment confirmation" based on historically successful process standard paths. At this point, the system will trace back the interface interaction event flow to confirm whether there is a page response timeout or a control failure, and then generate an interaction-level exception report to proactively push action suggestions or refresh instructions to the user.

[0046] By introducing a semantic analysis module and building an automatic recognition mechanism for key question types, we can significantly improve the efficiency of identifying pre-processing information for customer service requests. This allows us to instantly determine the type of question after user input, providing clear context for subsequent modules such as data retrieval, exception detection, and service recommendations.

[0047] S20, obtaining a business processing record of the conversation subject of the current conversation content according to the key question type;

[0048] In this embodiment, after completing semantic analysis of the current conversation content and identifying the key question type, it is necessary to further retrieve the historical transaction information of the user corresponding to the current conversation, focusing on the specific business scenario involved in the question type. This process first requires determining the identity of the subject of the current conversation. For example, through the identity authentication information extracted from the session, customer number, login session identifier, or bound mobile phone number, a mapping relationship between the user identifier and the business records in the backend system is established. This mapping relationship can be used to extract a collection of business processing records related to the user from the business data platform.

[0049] In actual operation, it is necessary to accurately match the scope and dimensions of business processing records according to the different types of key issues. For example, when the key issue is identified as "the page cannot jump during the insurance process", priority is given to retrieving the user's insurance business processing records, including insurance time, insurance type, terminal type, historical form submission status, failed page logs, etc.; and when the problem type is "slow claims progress", the user's claims business processing records should be extracted, such as application time, acceptance node, document material upload status and claims review status. In addition, during the processing process, timestamp restrictions should be combined to extract only business records related to the time before and after the current conversation to ensure that the match is the most recent or currently ongoing business process.

[0050] To protect user privacy, the retrieved data can be desensitized for sensitive information, masking personal privacy fields such as names, ID numbers, and bank card numbers, and only retaining data content such as structured fields, business status nodes, and behavior paths that are valuable for problem localization. In this way, it can provide basic data support for subsequent processes such as operation trajectory construction, problem node identification, and service resource matching, while avoiding problems such as users having to re-enter information and customer service asking multiple rounds of questions, significantly improving the conversation efficiency and customer experience.

[0051] In specific implementation, different business systems can provide standardized business record query capabilities through interface integration, and dynamically match the required fields and form data in combination with the recognition results of problem types. For example, in the field of medical and health, when it is recognized in the conversation that the user has questions about the interpretation of inspection reports, the system will retrieve structured data such as their physical examination appointment records, report archiving records, and doctor consultation records to support subsequent reasoning; in financial business, when the conversation content reflects that the user is dissatisfied with the delay in loan disbursement, the system will retrieve its approval process records, quota assessment logs, and transaction flow information for positioning, thus supporting subsequent business remediation and conversation strategy suggestion mechanisms.

[0052] In an actual system, the extraction process of business processing records is usually scheduled through a back-end integration service interface. During the implementation process, first, the identity information field in the user's current conversation context, such as the login credential, user-bound account, session ID, or the user identification value generated during the identity verification stage, is used as the only entry point for retrieving business processing records. This identification value will be used as the user's business primary key identifier throughout the processing process to accurately locate their historical operation records in multiple business systems.

[0053] The system will retrieve the corresponding business module in the mapping rule library of standardized problem types to business data tables based on the problem type generated by the pre-identification module. For example, when the problem type is "interruption of the insurance application process", the system will retrieve the operation flow record table and the insurance application process control table in the "insurance business database"; when the problem type is "claim not received", the system will automatically search for the review node table and the settlement receipt table under the "claim business data set". This mapping rule allows for automatic data path selection without manual intervention.

[0054] To control the data loading scope, the system will set a time window parameter. The default time window can be 72 hours before and after the current session time point to avoid retrieving irrelevant historical records. This time window parameter supports configuration adjustment and can be dynamically changed according to the business type. For example, for claim problems, the window can be extended to the past 30 days, while for insurance application behaviors, generally only the business records within the current period need to be processed.

[0055] The retrieved business data will be uniformly processed through the data desensitization module. All identification fields (such as name, ID number, mobile phone number) will be replaced with equal-length masked strings or completely removed, and only fields such as business status fields, operation node names, trigger timestamps, and failure log summaries will be retained. For structured data, the field mapping strategy is used to complete the desensitization conversion; for log-type texts, regular templates and keyword dictionaries are used to complete content regularization and sensitive word shielding processing.

[0056] When some business systems do not support structured queries, the image data or process recording system can also be called. For example, when the customer is in the insurance application process but the interface operation fails, the system can retrieve the page video data under the user's operation path as a supplementary form of business processing records. This operation path information can be integrated into the subsequent construction of the operation track for processing.

[0057] Example: In the field of medical and health, when the user asks the question "The report data cannot be viewed" during the interaction with the online medical guide robot. After identifying the problem type, the system immediately calls its health service record interface to obtain the physical examination report access behavior in the past three days, including the report generation time, report status code, and report download log. The system finds that the report status is "generated but not pushed". Combining the problem context, it prompts the customer service to push a copy of the report to the user's bound email and marks the push failure time point and the user's operation path.

[0058] In the financial field, when the customer feedbacks "Payment jump failed" during the insurance application process, the system identifies the problem type as "Payment exception". Through the identity identification value, the system retrieves from the insurance business database that the user has completed the form filling and health notification processes within the last hour, but the system returns an error of "Third-party payment interface timeout" when submitting the payment request. This record is extracted in real time and used as the input for subsequent generation of exception reports and operation guidelines to assist the customer service in determining whether to re-initiate the payment link or switch the payment channel.

[0059] By constructing a business processing record extraction mechanism driven by key problem types, the context understanding ability and data decision-making basis of the intelligent dialogue system can be effectively improved. With the mapping between the problem type and the business data source, targeted retrieval of the required records can be realized, avoiding resource waste caused by full-volume loading of irrelevant data; combined with time window control and desensitization processing technologies, the efficiency and compliance of data processing are further guaranteed. Without interrupting the dialogue process, the system can provide accurate data support for subsequent track construction, exception identification, and resource matching, thus greatly shortening the problem location time and improving the response intelligence level of the customer service system.

[0060] S30, when the key problem type is a business processing problem type, obtain the target operation record according to the interval between the problem occurrence time and the current time;

[0061] In this embodiment, when the problem involved in the current conversation is identified as a business processing problem, it is necessary to judge the time association between this problem and the user's previous actual operations to determine from which data source to extract the target operation records. In this process, the problem occurrence time represents the moment when the user last completed a key business process node. For example, in the insurance application business, it may be the time point of completing standard operation nodes such as identity verification, health declaration, or payment confirmation. This time point is usually automatically recorded by the business log system, and the data source path is determined by the triple of event ID, business type identifier, and user identifier value.

[0062] The current time is the system timestamp when the current session context is activated, usually generated immediately when the user asks a question or starts the dialogue recognition module. The interval between the problem occurrence time and the current time, as the key basis for judging whether to retrieve historical data or collect real-time data, is generally calculated in minutes or hours and supports dynamic adjustment. For example, in the case of a sudden failure, the system can shorten the judgment interval to respond in a timely manner.

[0063] When the interval exceeds the preset time threshold, the system will judge that this problem has deviated from the current real-time process state and is suitable for extracting the complete operation chain from the historical operation record database. Such an operation chain includes, but is not limited to, page access logs, control click logs, form submission behaviors, jump nodes, and timing information. Such data structures are complete and stable and are suitable for traceability analysis.

[0064] When the interval does not exceed the preset time threshold, it means that the user is still in the process of business process execution. At this time, the front-end event stream data is preferentially obtained in real time from the current operation flow monitoring interface. These data have the characteristics of high frequency, low latency, and lightweight structure, and are suitable for instant trajectory construction and behavior deviation detection. The event stream is usually obtained based on WebSocket, Kafka, or API listening channels, and the data items include timestamp, control ID, trigger action type, and DOM path, etc.

[0065] The target operation record refers to the time series data formed by a series of actions (including operations such as page jumps, control clicks, inputs, submissions, and swipes) performed by the user in a specific business process, as well as the original behavioral manifestation forms that can reproduce these actions. Its core function is to "playback and restore" the user's business operation process, thereby providing the original basis for behavior trajectory modeling, deviation detection, and problem analysis.

[0066] In the traditional definition, the target operation record is mostly an event stream composed of structured data items. For example, each operation event includes page ID, control ID, trigger timestamp, action type, parameter content, etc. These data are suitable for machine processing and path calculation, but have limited capabilities for customer service assistance in positioning and understanding the context of user behavior.

[0067] In modern intelligent interaction platforms, especially those with multimodal acquisition capabilities, the target operation records have been extended to a multimodal behavior set that integrates structured operation events and visual operation content. Therefore, video content, real-time recordings, page screen recordings, etc. generated during the customer operation process also belong to the components of the target operation records. These video contents can be: screen recordings during user insurance application (screen click process); screen recording playback during intelligent form interaction; user operation recordings automatically generated by the system (combining event trigger records and visual synthesis); real-time interface sharing images collected through remote sessions.

[0068] The fusion of video data and event streams in the target operation records can be achieved in the following ways:

[0069] Event-frame index binding: In the structured event record, add the timestamp position of the corresponding video to enable precise synchronization of click events and picture frames.

[0070] Track overlay playback: Overlay event node markers (such as control highlighting, path tracking lines) on the operation recording for easy review by customer service.

[0071] Multimodal behavior graph generation: Model user behavior as both a structured trajectory graph and a chain of picture snapshots to adapt to machine processing and manual interpretation.

[0072] This mechanism ensures that the source of the target operation records dynamically adapts to the business context state, ensuring both data accuracy and processing efficiency, and avoiding loading redundant historical data or missing current behavior records.

[0073] Example illustration: In the financial business scenario, a customer feedbacks the problem of "payment failure after insurance application" to the intelligent customer service. After the system identifies this problem as a business processing problem, it first locates the time point of the user's most recent payment request submission in the insurance application process, and this time is the "problem occurrence time". 22 minutes have passed since the current conversation initiation time from this time, exceeding the preset threshold. The system retrieves the user's complete insurance application process operation chain from the historical operation log and finds that no payment success event receipt has been received after jumping to the payment page. Thus, it is determined that this problem is a payment link timeout, and an exception report is generated subsequently and the customer service is prompted to resend the payment link.

[0074] In the medical and health platform scenario, a user proposes "the page freezes after uploading the imaging file". The system determines that the time since this problem occurred is only 3 minutes. By listening to the interface to obtain the real-time event data of the front-end upload control, and identifying that there is no feedback event after the upload component sends the file, it is initially determined to be an interface blockage or network problem. Based on this, the system marks this behavior as an active operation exception, recommends immediately guiding the user to refresh or switch the network, and uses the original event data to construct the behavior trajectory.

[0075] By constructing a target operation record extraction scheme based on a time interval judgment mechanism, dynamic adaptation processing of real-time problems and historical problems can be achieved. This not only avoids the problem of delayed loading of historical data for low-latency services but also reduces the repeated collection of real-time data for highly stable problems, greatly improving the accuracy and processing efficiency of problem location. By intelligently switching the source of operation records, the data channel pressure can also be reduced, the system response speed can be increased, and a more stable data foundation can be provided for subsequent construction of user operation trajectories and identification of behavior deviations.

[0076] S40, constructing a user operation behavior trajectory based on the interface interaction events in the target operation record;

[0077] In this embodiment, the construction process of the user operation behavior trajectory is carried out based on the interface interaction events included in the target operation record. This process relies on the original information such as timestamps, control identifiers, and page paths recorded in the interaction events, and restores the real operation sequence of the user in the business process through behavior reconstruction means. Interaction events refer to atomic-level operations triggered by the user on the interface, including clicks, inputs, dropdown selections, swipes, page jumps, etc. Each interaction event can include five categories of basic fields: page identifier, control identifier, trigger action type, trigger timestamp, and associated data parameters.

[0078] The construction process of the user operation behavior trajectory first sorts these interaction events in ascending order according to the timestamp to form an event sequence reflecting the order of user operations. During the sorting process, the millisecond-level trigger timestamp recorded by the server is usually used to ensure the accuracy of time sequence and solve the inconsistency problem caused by out-of-order logs in the distributed system. The data stream after event sorting constitutes an ordered event sequence.

[0079] On this basis, further extract the control hierarchical path of each interaction event in the DOM structure of the front-end page. This path can be composed of the page ID, module ID, and control instance ID, and is spliced according to the nested hierarchy to form a page path. For example, an event may be located as "Insurance Home Page → Product List Block → Health Insurance Card → Insurance Button". The parsing of the control hierarchical path helps to understand the operation context and the business steps where the user is located.

[0080] All the parsed event paths will form an event annotation set. Without changing the event order, this set supplements the page navigation information and control semantic information, enabling the subsequent construction of operation trajectories to have the ability of context expression.

[0081] Subsequently, all event nodes in the event annotation set are directed-connected in chronological order to generate an operation path topology graph. The nodes of this graph are page paths, and the edges are the event trigger order relationships. The topology graph not only visualizes the complete page access chain that users experience in the business process but also provides a basis for extracting behavioral structure data such as behavior levels, path lengths, and jump frequencies.

[0082] Finally, with the page path as the node and the trigger timestamp as the edge weight, this topology graph is mapped into a multi-dimensional behavioral trajectory structure to achieve a comprehensive modeling of the user operation path. This trajectory structure is traceable, visualizable, and analyzable, and is suitable for various intelligent tasks such as subsequent deviation identification, behavior scoring, and path comparison.

[0083] In different business systems, this process can be implemented by various technical means. Some systems use embedded front-end SDKs to collect interaction event data and send it back to the server through standardized APIs; some systems have the server deduce user operation events based on access logs and DOM snapshots. For obtaining the control level path, one implementation method is that the front-end page automatically injects control level identifiers during rendering. For example, when using the Vue or React framework, the component tree structure naturally provides the node level relationship, and the system can generate a unique control path representation by combining the component path, instance number, and page routing.

[0084] The generation of the topology graph is usually implemented through an adjacency list or an adjacency matrix. The sequential connection between events is driven by timestamps, and the data structure design can support multi-branch jumps and loop access paths to meet the representation requirements of complex page logics. When constructing the behavioral trajectory structure, the timestamp can be standardized as the edge weight to achieve the modeling of operation frequencies and path durations. It is also possible to further integrate features such as path access depth and page stay duration to enhance the expressive power of the trajectory model.

[0085] This process can also be linked with a visualization graphics library to generate a display interface for the behavioral topology graph, facilitating technical operation personnel or intelligent analysis models to intuitively identify abnormal paths. For the case of long user paths, a segmented classification strategy can be adopted to divide the operation trajectory into sub-path groups to reduce the complexity of the graph structure and improve the processing efficiency.

[0086] By constructing the user operation behavior trajectory, discrete interaction events in the target operation record can be organized into a structured and visualizable behavior chain, achieving an accurate modeling of the actual operation path of users during the business processing process. This trajectory can not only reflect the sequence of user behaviors and page jump logics but also reveal problems such as possible interruptions, deviations, and repeated accesses in user operations, providing a data basis for subsequent behavior deviation detection, problem tracing, and generation of exception reports.

[0087] S50. Identify the deviation nodes in the user operation behavior trajectory;

[0088] In this embodiment, when the user completes a certain business process, their behavior path often has strong regularity. By representing the operation behavior trajectory as a graph structure composed of nodes and edges, the page access logic and control trigger sequence can be modeled as computable and comparable data representations. Identifying the deviation nodes in the behavior trajectory means identifying and marking the nodes with abnormal, missing, or illegal jump behaviors in this path, aiming to discover the key positions where the user's operations deviate from the standard business process.

[0089] The deviation identification process first extracts all the nodes in the operation path graph in units of page paths, and at the same time identifies the edge connection relationships corresponding to the event order. The page path is the identifier of each page accessed by the user, and the jump order, access depth, and control logic can be reflected through the path structure. Subsequently, the representation form of the standard business process needs to be introduced, usually a business process reference model composed of a set of standard nodes and a set of legal jump edges.

[0090] Through the comparison between the sets of nodes, the nodes that do not appear in the user path but must exist in the standard process can be identified. These missing nodes reflect the important links bypassed, skipped, or abnormally exited by the user. In the comparison of edges, the jump pairs that do not exist in the standard process are detected, that is, illegal jump behaviors. Illegal jumps may be caused by system exceptions, control errors, or user mis-triggering, representing structural anomalies during the process execution.

[0091] To accurately measure the impact of each page path node on the overall deviation, it is necessary to numerically quantify the degree of deviation of the node. This process can calculate the node deviation degree by setting multi-dimensional indicators. A commonly used indicator system includes: the binary factor of missing important nodes, the number of illegal jump edges, abnormal jump directions, abnormal operation time consumption, etc. Each indicator forms a comprehensive deviation score through weighted combination and is compared with a preset deviation threshold. The nodes exceeding this threshold will be marked as deviation nodes and recorded in the abnormal identification output as the core focus of subsequent analysis or customer service intervention.

[0092] The entire deviation identification logic is based on the composite calculation of path structure, business rules, and behavior anomaly criteria, with structural interpretability and behavior reproduction ability, and is suitable for scenarios such as branches, jumps, and multi-paths in complex business processes.

[0093] Different systems can select different implementation methods of deviation recognition algorithms according to the complexity of their business processes and the granularity of data collection. A common implementation is to use a graph matching algorithm to compare the user path graph with the standard flow chart. The flow chart can be encoded by expert rules or learned through statistical analysis of a large number of historical normal paths. During the graph comparison process, an adjacency matrix can be used to represent the jump relationship, and missing nodes and illegal jump edges can be located and marked.

[0094] For the implementation of deviation quantification, a weight matrix can be introduced to control the importance of various abnormal factors. For example, the penalty coefficient for missing key nodes is set to 1.0, the penalty coefficient for illegal jump edges is set to 0.5, and the abnormal Z value of path duration is set to 0.3, and it can be flexibly configured according to the business model. The comprehensive deviation score can be output using methods such as weighted summation, logistic regression, or even graph neural networks.

[0095] In addition, to adapt to high-frequency call or real-time response scenarios, the standard flow chart can be pre-compiled into a state machine structure, and illegal state transitions can be dynamically detected during the inflow of behaviors, so as to identify deviation paths in real time and generate alarms. The recognition logic can also be integrated into the front-end user behavior SDK for edge computing to improve the detection speed and system response ability.

[0096] Example illustration: In the financial business, when a user applies for a loan, the system records their behavior path as "Login page → Identity verification page → Authorization consent page → Submission page". The standard process requires that the credit investigation authorization confirmation page must be entered before submission. The trajectory graph comparison finds that this node is missing, and there is also an illegal jump edge directly from the "Identity verification page" to the "Submission page". Based on this, the system determines that the "Submission page" is a deviation node and further identifies that this path may be caused by a front-end jump logic error, generating an exception report for operation and repair.

[0097] In the medical and health service, during the process of a user booking a physical examination package, the operation behavior trajectory records the path as "Package selection → Item selection → Appointment filling → Appointment confirmation", but the system standard process requires that the "Service terms confirmation" node must be passed through. This node does not appear in the trajectory, and there is an illegal edge from "Item selection" to "Appointment confirmation". Based on this, the system marks the "Appointment confirmation" as a deviation node and judges through the background that the user has not completed the terms confirmation procedure, generating a risk prompt and submitting it to the customer service platform for processing. This method effectively avoids potential business compliance risks and improves the customer service experience.

[0098] By identifying deviation nodes in the operation behavior trajectory, key behavior points that may lead to failures, blockages, or interruptions in the user process can be accurately captured. This structural identification mechanism breaks through the boundaries of traditional rule-based log analysis, not only discovering obvious jump errors, but also uncovering hidden process detours and unexpected branch operations, helping the system provide structural support in multiple downstream tasks such as intelligent customer service, exception report generation, and process optimization suggestions, thereby significantly improving the problem localization efficiency and service response accuracy.

[0099] S60, generating a business processing exception report according to the deviation node;

[0100] In this embodiment, when there is a page path node in the user operation behavior trajectory that is inconsistent with the standard process and this node has been identified as a deviation node, the system needs to construct a report output focusing on the exception problem for problem review, customer service response, or subsequent process optimization intervention. The core of this exception report is to use the deviation node as the starting point for analysis, and structurally express its context path, impact level, and executable repair suggestions as internal and external common exception data entities.

[0101] During the report generation process, first, the front and back path information of the deviation node in the complete path diagram needs to be extracted to form its operation context. Usually, it includes the precursor page path, the successor page path, and the server-side trigger timestamps corresponding to each jump. This information combination can accurately locate the relative position of the problem node in the path and its trigger timing sequence. This path data can not only be used to reproduce the user behavior sequence but also serve as an auxiliary input for manual customer service or an automatic diagnosis system to locate the problem scope.

[0102] Secondly, a quantitative assessment of the impact degree brought by the deviation node is required, and it needs to be classified according to a predefined business impact strategy library. The strategy library is based on the problem severity and business impact dimensions and often uses a grading mechanism to divide deviations into blocking exceptions, high-risk exceptions, and low-risk exceptions. For example, missing the identity verification page or illegally bypassing the payment verification, etc. are classified as blocking exceptions, while multiple repeated jumps and useless page redundancy may be determined as low-risk exceptions.

[0103] Subsequently, the system matches corresponding handling suggestions according to the deviation type. The handling suggestions usually come from the standard response actions in the strategy library, such as page redirection, customer service intervention, process rollback, or user prompt templates. These suggestions, path information, and impact level are uniformly encapsulated into a structured exception report, and the output format should include fields such as node identification, path segment, exception type, risk level, handling suggestions, and generation time.

[0104] For the convenience of system integration and manual processing, the report should also support a unified semantic format, such as JSON, XML, or a nested structure data format supported by the business system, and can be accompanied by visual information such as screenshot annotations and page snapshot links to improve analysis efficiency.

[0105] In the financial service process, the system can access user behavior trajectory data through the service gateway and start the abnormal report generation module after identifying deviation nodes. This module first analyzes the precursor and successor jump information of the deviation nodes and reconstructs the local topology of the path using the timestamp sequence. Then it calls the local rule engine to determine the level of this anomaly through regular path rules. If a blocking rule is hit, it directly triggers the "interruption recovery" processing strategy, such as initiating page redirection, terminating the process, or popping up a user prompt.

[0106] In the medical service appointment scenario, the node priority and risk level can be mapped by tagging the page path nodes in combination with business rules. For example, the "personal information filling page" is marked as a mandatory node, and once it is missing, it is judged as a high-risk anomaly, triggering a process restart. The processing suggestions can be "prompt the user to fill in again", "guide to jump back to the original page", etc. The output abnormal report is uniformly recorded by the service middle platform and synchronized to the customer service visualization interface through a structured interface.

[0107] By generating a structured business processing abnormal report around the deviation node, the system can achieve rapid perception and transparent expression of user behavior anomalies. This abnormal expression mechanism not only improves the problem location efficiency but also enhances the customer service's understandability of the problem and the ability to support processing decisions. Compared with traditional error codes or single-point abnormal prompts, this mechanism can provide a complete path context, accurate impact assessment, and actionable suggestion items, thus building an intelligent response path for complex service problems. Overall, it can significantly reduce the misjudgment rate, compress the response time, and improve the business continuity guarantee ability.

[0108] S70, match the target service item in the service resource library according to the service demand keywords in the current conversation content and the business processing record, and generate the operation guide information of the target service item.

[0109] In this embodiment, in order to accurately identify the user's needs from the user's conversation and accurately map the needs to specific service items in the system, so as to provide the user with actionable guidance, the system first needs to identify keywords in the current conversation content. This process relies on the semantic understanding module to process the text input through semantic chunking, syntactic structure recognition, word vector embedding, etc., and extract words or phrases with directive, request, and service relevance to form an initial keyword set.

[0110] The above keyword set may not be directly used to match target service items in the service resource library, so unified standardization processing is required. The process of keyword standardization is usually based on a pre-constructed service requirement label dictionary, and the keywords are normalized and mapped through the label dictionary. For example, when a user expresses "I want to purchase health insurance", the system will map "health insurance" to the standard label "health insurance purchase" to uniformly connect to the subsequent service matching logic.

[0111] In the service resource library, each service item contains metadata such as service name, service function description, business type identifier, historical call records, and permission restriction conditions. When performing service matching, the system not only compares semantic similarities based on service requirement labels but also combines context information such as business type codes, access permission identifiers, and historical service item identifiers in the user's business processing records. The semantic similarity calculation can adopt the vector space model, and by calculating the cosine similarity between the label and the service description, a set of candidate service items that best match the user's intention is determined.

[0112] To control access permissions and avoid interference from irrelevant items, the system uses the business type code in the business processing record to filter out service items that are consistent with the current business type and further uses the permission identifier to screen out a subset of service items that the current user has access to. The historical service item identifier, as a behavioral data indicator, reflects the user's past preferences and operation frequencies, and is used to prioritize the candidate items. Finally, the service item with the highest ranking is determined as the target service item.

[0113] After determining the target service item, the system enters the stage of generating operation guidance information. The operation guidance information can consist of three core elements: service function description text, front-end routing address of the service item, and the user's current operation behavior trajectory. The function description text provides the logical sequence of operations that the user can perform and the expected completion goals. The front-end routing address provides a one-click access function entry for the user, and the user's operation behavior trajectory can be used to determine the user's current page position, thereby dynamically generating the optimal jump path or recommending the next action.

[0114] The finally generated operation guidance information can be output in structured data, including the name of the target service item, access link, operation step suggestions, context navigation guidance, etc., for presentation and implementation on the client or in the customer service system.

[0115] For example, in the scenario of applying for health insurance, the user enters "How to purchase critical illness insurance". The semantic analysis module identifies the keywords "purchase" and "critical illness insurance", and after standardization mapping, obtains the service requirement label "Critical illness insurance application". Combining the user's business processing records, the business type code is extracted as "Personal health category", the access permission is "Basic user", and the "Health insurance trial calculation" function has been used in previous operations. The system filters out all service items in the service resource library that contain the keyword "critical illness" and have a semantic similarity greater than 0.75 as the candidate set, further filters out non-health-related services and services beyond the permission, and retains two eligible service items.

[0116] By calculating the usage frequency of the user in the past 90 days, "One-click application for critical illness insurance" is determined as the service item with the highest ranking, which is the target service item. The function description text of this service item is "Fill in personal information → Health declaration → Upload image materials → Confirm payment", and the front-end routing address is / insurance / critical / entry. Combining the user's current page path "Home page → Health insurance", the system generates a complete guidance chain for jumping from the current page to the target page, and constructs operation guidance information including operation suggestions, access paths, and page jump expectations.

[0117] In the bank wealth management service, the user initiates "I want to find wealth management products with low risk". After keyword standardization, it is mapped to "Low-risk wealth management recommendation". The system combines the user's business records and identifies it as the "Basic wealth management business" type, extracts the permission identifier as "Ordinary investor", filters out products involving high-risk investments, and screens out candidate product recommendation services. Combining the user's frequently used service item "Query of stable income products", it is preferentially matched to the "Low-risk wealth management product navigation page", and a directly accessible page link and operation tips adapted to the user's investment preferences are generated.

[0118] By integrating service requirement keywords, business context data, and multi-dimensional meta-information in the service resource library, the system can intelligently complete the accurate matching of service items, significantly improving the service adaptation rate and recommendation accuracy. On this basis, through the integrated analysis of function text, page paths, and operation history, the generation of dynamic and personalized operation guidance information is realized. The automation, semanticization, and real-timeization of service recommendation are achieved, enabling users to quickly locate service entrances and efficiently complete operations in complex business processes, thereby improving business processing efficiency and user satisfaction.

[0119] The present invention relates to the field of semantic parsing technology and can be applied to business scenarios such as the fields of medical health, fintech, and intelligent interaction services. It discloses a business service processing method based on dialogue recognition, including: parsing the current dialogue content to generate a key problem type, obtaining the business processing record of the session subject of the current dialogue content according to the key problem type, and in the case where the key problem type is a business processing problem type, obtaining a target operation record according to the interval between the problem occurrence time and the current time, constructing a user operation behavior trajectory based on the interface interaction events in the target operation record, identifying the deviation nodes in the user operation behavior trajectory, generating a business processing exception report according to the deviation nodes, and matching the service demand keywords in the current dialogue content and the business processing record with the target service items in the service resource library to generate operation guidance information for the target service items. Through collaborative processing mechanisms such as semantic analysis, historical behavior data extraction, and behavior trajectory construction, the present invention can intelligently identify the customer problem type, accurately locate the abnormal nodes of the user in the business handling process, automatically generate a structured exception report, and complete service item matching and path guidance in combination with service demands, so as to achieve efficient problem response and personalized service recommendation, and improve the business support ability and service efficiency of the intelligent customer service system.

[0120] In one embodiment, the above step S30 includes:

[0121] S301, when the key problem type is a business processing problem type, extract the timestamp of the last successfully completed standard business process node of the session subject from the business operation log as the problem occurrence time;

[0122] S302, obtain the current timestamp for processing the current dialogue request;

[0123] S303, determine the interval duration between the current timestamp and the problem occurrence timestamp;

[0124] S304, when the interval duration exceeds the preset interval threshold, retrieve the complete operation chain data of the session subject from the historical operation record database and use the complete operation chain data as the target operation record;

[0125] S305, when the interval duration does not exceed the preset interval threshold, obtain the interface interaction event stream of the current dialogue from the current operation flow monitoring interface and use the interface interaction event stream as the target operation record.

[0126] In this embodiment, in order to identify and trace back the context behavior path where business processing problems occur, a target operation record extraction mechanism with precise timing and dynamic source selection capabilities needs to be constructed. First, after identifying that the critical problem type belongs to the business processing problem type, the system needs to judge the possible occurrence time of the problem based on the business operation log. The business operation log usually includes the structured behavior data of users at critical business nodes, such as the timestamp information of completing identity verification, submitting a health declaration form, or successful payment.

[0127] These critical nodes are defined as standard business process nodes, which are characterized by repeatable triggering, strong status certainty, and clear business stage division. The system obtains the timestamp of the last successfully completed standard business process node of the session subject from the log as the reference starting point for the occurrence of the problem, which is called the problem occurrence time.

[0128] Subsequently, the system needs to determine the trigger time of the current dialogue processing, that is, the timestamp of generating the current semantic request. By comparing the interval between the current timestamp and the problem occurrence timestamp, the time span for extracting the operation record is calculated. If the interval duration exceeds a preset time threshold (such as 10 minutes, 30 minutes, etc.), it means that the current problem may be caused by historical operations, and the system needs to retrieve the complete behavior chain from the historical operation record database. This complete chain usually includes multiple user behavior nodes, covering information such as jumps between different pages, control clicks, input operations, submission actions, etc., forming a closed-loop interaction history.

[0129] If the interval between the current time and the problem occurrence time is within the threshold range, it indicates that the problem is more likely to occur in the current active business processing flow. At this time, there is no need to trace back to the historical database, but directly extract the real-time interaction event data from the operation flow monitoring interface. This interface usually captures the operation actions of users on the page through a front-end listening mechanism and sends them to the server cache, including but not limited to DOM events (click, input, slide), route jumps, form submissions, etc., with high timeliness.

[0130] Through the above logical judgment and adaptation of the data source, the system can make a dynamic choice between time sensitivity and data integrity, ensuring that the target operation record has sufficient context information without introducing redundant and irrelevant data.

[0131] For example, in the insurance business, when a user reports a payment failure problem, the system first identifies that the problem belongs to the business processing problem type. The system retrieves the timestamp of the last successful completion of the health declaration submission operation in the business log, which is 10:12:00 on December 1, 2024, and obtains the current request time as 10:42:00 on December 1, 2024. The interval is 30 minutes, exceeding the system preset threshold of 15 minutes. Therefore, the system accesses the historical operation record database and retrieves all page jumps, input operations, and submission action data of the user from the health declaration submission to the current time period to form the target operation record.

[0132] If the user asks "Why can't I go to the next step?" on the wealth management product insurance page, the system determines that the key problem type is business processing. By checking the log, it finds that the user has just completed identity verification, and the current time is only 4 minutes different from this operation. At this time, instead of accessing the historical record library, the system obtains, through the operation flow monitoring interface, real-time operation events such as the failure to click the "Next Step" button and form verification exceptions on the current insurance page of the user, and uses them as the target operation record to input to the subsequent trajectory analysis module.

[0133] In the medical registration service scenario, when a user reports "unable to register", the system identifies the problem type as a business processing problem. By analyzing the log, it is found that the last successful operation was "selecting a registration department" at 11:00, and the current request time is 11:02. The system determines that it is still within the current business flow, so it calls the monitoring interface to extract the operation sequence from "selecting a department" to "submitting the registration" for backtracking processing.

[0134] This embodiment avoids the limitations of fixedly calling historical data or only relying on real-time caches by switching between historical records and real-time event sources. In the case of a time lag between business actions and problem occurrences, the system can accurately backtrack the complete behavior path before and after the problem occurs, improving the accuracy of problem location; while during the active business period, the system can obtain operation trajectories in real time and respond quickly, improving the backtracking efficiency. Overall, it realizes an intelligent extraction strategy for target operation records driven by time, taking into account both real-time and integrity.

[0135] In one embodiment, the above step S40 includes:

[0136] S401, sort the interface interaction events in the target operation record in ascending order of the trigger timestamp to generate an ordered event sequence;

[0137] S402, extract the control hierarchy path of each interface interaction event from the ordered event sequence, and generate an event annotation set containing the page path based on the control hierarchy path;

[0138] S403. Generate an operation path topology graph based on the triggering order of the interface interaction events in the event annotation set. The nodes in the operation path topology graph are page paths, and the edges are event triggering orders.

[0139] S404. Construct a user operation behavior track according to the page jump relationship and control triggering order in the operation path topology graph.

[0140] In this embodiment, the user operation behavior track is a time-series structure data used to dynamically depict the operation path and behavior characteristics of the user in the business system, and is used to support subsequent problem location, deviation identification, and task optimization. Its construction process depends on the interface interaction events extracted from the target operation records, usually including interface operation data such as button clicks, control inputs, and page jumps.

[0141] First, all the interface interaction events included in the target operation records need to be sorted in ascending order according to the trigger timestamps to generate an ordered event sequence. The purpose of this sorting operation is to reconstruct the time order of the user's behavior, so that the generation of subsequent structures has a causal sequence. The trigger timestamps generally come from server-side records or browser-side buried point feedback, and UTC time or relative time offset can be used as the sorting basis.

[0142] Interface interaction events refer to the operation behaviors triggered by the user during the interaction with the graphical user interface (GUI) and can be perceived and recorded by the system. These events usually represent specific actions of the user on the front-end page or application interface, reflecting how the user uses the interface to complete tasks. These events are the basic units for constructing the user operation behavior track and have attributes such as clear trigger time, operation location (control path), and operation type. Common interface interaction events include:

[0143] Click event (click): The user clicks on elements such as buttons, links, icons, etc. For example: Click the "Submit" button to submit form data.

[0144] Input event (input / change): The user fills in text in the input box, selects a dropdown box, or checks a checkbox. For example: Enter "Zhang San" in the "Name" input box, or select "Health Insurance" as the insurance type.

[0145] Page jump event (navigation): The user jumps from one page to another page, or switches modules through the menu. For example: Jump from the "Product Introduction" page to the "Policy Filling" page.

[0146] Hover event (hover): The mouse hovers over an element, triggering a tooltip or expanding content. For example: Move the mouse over the "Guarantee Details" icon to pop up the terms description.

[0147] Scroll / drag event: The user scrolls the page to view information or drags the slider to adjust parameters. For example, scroll down the long page to view more content.

[0148] Modal open / close event: Open or close modal windows such as dialog boxes and prompt windows. For example, a file selection dialog box pops up after clicking "Upload materials".

[0149] Load / ready event: Passively triggered after a certain page or module is loaded. For example, display a "Welcome back" prompt after the page is loaded.

[0150] After obtaining the ordered event sequence, the system needs to extract the control hierarchy path from each interaction event, which is used to describe the specific interaction position between the user and the interface elements. The control hierarchy path usually consists of elements such as page identifiers, component numbers, and container nesting structures, and can accurately locate the occurrence point of the user's behavior in the interface structure. By performing structure splicing and page path mapping on these control paths, an event annotation set can be generated, where each record contains metadata such as page path, event type, and trigger timing.

[0151] Based on this event annotation set, the system builds an operation path topology graph in the order of events. The nodes of this topology graph are the page paths associated with the events, and the edges represent the logical sequence of event triggers. The edges between nodes not only identify the linear progression of the user's behavior flow but also reflect the user's thinking migration and operation logic in the interface to a certain extent.

[0152] Finally, based on the above topology graph, a user operation behavior trajectory is constructed. The page path is used as the node in the trajectory graph, and the time difference of the trigger timestamps of the interface interaction events is used as the weight value of the edge, forming a directed graph structure with time attributes. This structure not only retains the execution order of the user's operations but also introduces the characteristics of operation rhythm and behavior interval, which helps to detect abnormal behavior points or operation bottleneck positions in subsequent analysis.

[0153] This entire construction process realizes the transformation from the original event data to a structured trajectory model, enabling the system to analyze the evolution path of the user's operation behavior in a graph computing manner and supporting multi-dimensional user behavior modeling and business process diagnosis tasks.

[0154] For example, in a financial service system, the user's insurance purchase operation involves the interaction process of multiple pages and controls. When the user reports that "the submission of the health declaration form fails", the system first obtains its target operation record, which records the click and input operations performed by the user on multiple pages.

[0155] The system first sorts all the interaction events in the operation record in ascending order of timestamp. For example: click "Next Step" at 10:02, fill in health information at 10:04, click "Submit" at 10:05, etc., to generate an event sequence. The page IDs and control IDs included in each event are extracted and concatenated into a control hierarchy path (such as insurance_form.step2.submit_button), thereby constructing an event annotation set.

[0156] Based on the event order, a topological structure is constructed: Page A → Page B → Page C. Each jump is connected by an edge, and each edge is attached with a time weight, such as 2 minutes, 1 minute, etc. Finally, a graph structure with page paths as nodes and time as edge weights is formed to represent the complete behavior trajectory of the user.

[0157] In the medical registration system, user behaviors include page interactions such as selecting a hospital, selecting a department, and confirming registration. When the user encounters interface lag issues during the registration process, the system can judge whether the user is repeating refreshes, making invalid jumps, or missing key nodes based on the constructed trajectory graph, thereby quickly assisting in locating abnormal behaviors.

[0158] In this embodiment, by structuring the interface interaction events and constructing them into the user operation behavior trajectory, not only the visualization expression of the original operation data is realized, but also the time sequence and page relationship of the user behavior can be revealed through the graph structure. At the same time, by introducing the trigger time as the weight of the edge, the time interval between different operations can be quantified in the trajectory, which helps to identify experience problems such as lag and high latency. This graph structure can also be directly used in high-level applications such as deviation node identification, behavior clustering, and process reconstruction, effectively improving the problem location accuracy and processing efficiency.

[0159] In one embodiment, the above step S50 includes:

[0160] S501, extract all page path nodes and trigger order edges from the operation path topology graph of the user operation behavior trajectory;

[0161] S502, compare the page path nodes with the predefined set of necessary nodes of the standard business process, and mark the necessary nodes not included in the operation path topology graph;

[0162] S503, compare the trigger order edges with the predefined set of legal jump edges, and mark the illegal jump edges not matched in the operation path topology graph;

[0163] S504, based on the necessary nodes not included and the illegal jump edges not matched, determine the node deviation degree of each page path node;

[0164] S505, mark the page path nodes with node deviation exceeding the preset deviation threshold as deviation nodes.

[0165] In this embodiment, the user operation behavior trajectory reflects each page passed by the user during the process of completing a certain service and their jump relationships, and these trajectories are modeled in the form of a graph structure composed of page paths. The core purpose of identifying deviation nodes is to discover abnormal operation paths or uncompleted key service steps in the trajectory, so as to judge the possible problems encountered by the user and perform active intervention.

[0166] Extracting page path nodes and jump edges from the constructed user operation behavior trajectory is the basic decomposition of the trajectory graph structure. Page path nodes usually consist of interface identifiers and hierarchical paths, and are the vertices of the graph structure for constructing jump relationships; jump edges represent the access order between pages, and their construction is based on the timestamp sorting result. After extracting the structure of the operation path topology graph, compare the nodes with the set of necessary nodes defined in the service specification, which is pre-configured by domain experts or obtained by inductive analysis of historical data, such as the pages where key steps such as identity verification pages, health declaration pages, and payment confirmation pages are located.

[0167] During the comparison process, if it is found that some necessary nodes do not appear in the trajectory, it means that the user has not performed the key operations that should be performed, and abnormal marks need to be made. In addition, by comparing the jump edges with the set of legal jump edges, further identify whether the user bypasses or skips intermediate pages. For example, directly jumping from the home page to the payment page constitutes an illegal jump. These illegal jump edges are potential operation anomalies in the trajectory.

[0168] To quantify the abnormality degree of different nodes, the system needs to calculate the deviation degree of each page path node. The deviation degree can be composed of two dimensions. One is the situation of missing necessary nodes. If a node belongs to the necessary nodes but does not appear in the trajectory, it is regarded as a serious anomaly. The other is the number of illegal jump edges involved in this node. Each illegal edge is assigned 0.5 to reduce the impact of single misoperations. Finally, use the addition rule to combine the two to form a comprehensive deviation degree, which is used to evaluate the overall abnormality degree of this node.

[0169] The system judges which nodes are deviation nodes according to the preset threshold. For example, nodes with a deviation degree greater than or equal to 1 may mean process breakage or frequent illegal jumps, and need to be marked with emphasis. This identification mechanism can dynamically sense and locate logical anomalies in the user operation path in the actual service flow, and has high generalization.

[0170] In actual deployment, the acquisition of page path nodes usually depends on front-end buried-point data or the SDK automatic collection mechanism. Each page load event reports a unique page identifier. The system aggregates these page load events and combines the jump events formed by the front and back click controls to construct an operation path topology graph. The set of necessary nodes can be obtained through manual rule configuration, derivation in combination with the business BPM flow chart, or induction from high-frequency historical paths through the frequent itemset mining algorithm. The set of legal jump edges is defined by the back-end system and synchronized to the analysis module through the configuration interface. The trajectory comparison uses a graph matching algorithm, such as the path matching method based on the adjacency matrix, to judge illegal edges. The deviation calculation module can batch process the node deviation information in all user trajectories through the MapReduce or stream computing framework. The configured threshold supports dynamic adjustment. For example, during peak periods, the tolerance is relaxed, and when the resources for manual access are limited, the trajectories with a high number of deviation nodes are preferentially processed. In an actual business system, additional factors such as page access density and user portraits (new or old customers) can also be introduced as weighted factors for deviation to further improve the recognition accuracy.

[0171] Example illustration: In the field of medical and health, when a patient conducts intelligent medical guidance on the hospital's mobile service platform, the preset standard process includes: filling in preliminary symptoms → selecting department suggestions → registering and making an appointment → doctor consultation page. In a typical scenario, the user inputs "recent palpitations and dizziness" through voice, and the system automatically enters the preliminary symptom filling page. However, due to unstable network or delayed interface feedback, the user repeatedly clicks the back button after inputting the symptoms and then directly jumps to the "appointment payment" page through the history function. By analyzing the operation behavior trajectory of this user, the background finds that the key node of "selecting department suggestions" is missing in its page path, and there is an illegal jump edge from "preliminary symptom filling" directly to "appointment payment". This path is identified as a seriously deviated path. Based on this, the system calculates the deviation of the page path node "appointment payment" as 1.5, where the contribution of missing necessary nodes is 1 and the contribution of illegal jumps is 0.5. Since the deviation exceeds the set threshold of 1, the system automatically marks it as a deviation node and triggers the risk control module to send a prompt to the user: "There are missing key links in your medical treatment path. It is recommended to return and supplement the diagnosis and treatment information." At the same time, the record of this deviation node is pushed to the medical guidance management background for subsequent quality tracking by medical staff.

[0172] In the financial sector, an insurance company customer applies for a high-end medical insurance product on a self-service insurance platform. The system's standard process is: real-name authentication → health disclosure → product selection → electronic signature → premium payment. After completing real-name authentication, a user exits the platform due to a browser error. After reopening the platform, they are directly redirected to the "Payment Confirmation" page through the "Order Center." The system automatically fills in the previously selected policy details. This behavior appears to save operations, but it does not go through the two key business links of "health disclosure" and "electronic signature." The user operation behavior trajectory recorded by the system shows that the path node of the "Health Disclosure" page is missing, with a deviation of 1; the path node of the "Electronic Signature" page is missing, with a deviation of 1; the operation of jumping from "Real-Name Authentication" to "Payment Confirmation" does not comply with the preset process and constitutes an illegal jump. This illegal jump edge is associated with the "Payment Confirmation" node, increasing the deviation by 0.5.

[0173] Ultimately, the deviations for the "Health Notification" and "Electronic Signature" nodes were 1, and the deviation for the "Payment Confirmation" node was 0.5. If the system's default deviation threshold is 0.8, the "Payment Confirmation" node would not have reached the threshold, but the first two nodes alone would have posed a high deviation risk. The combined effect of these three would trigger a high-priority exception report and a manual review prompt.

[0174] In contrast, when another user applied for health insurance using the online insurance system, the system typically directed them to the "Insurance Clauses Summary" page after completing the "Health Declaration" page. This page displays detailed terms and conditions, with an optional reading confirmation action. However, because the user had previously previewed the terms summary through the history record, the system determined that this page should not be presented again and instead redirected them directly to the "Electronic Signature" page to continue the process. In this operation path, the "Insurance Clauses Summary" page path node did not appear, but this page is marked as a non-mandatory node in the standard business process, and its absence does not result in incomplete business data or loss of legal effect of the contract. Furthermore, this jump path falls within the system's preset tolerance jump logic and is therefore not marked as an illegal jump edge. When calculating the deviation, the system determined that the missing page path node was "Insurance Clauses Summary." Based on the process annotations, the system determined that this node was not a mandatory node, and the deviation for a required node was 0. The jump edge from "Health Disclosure" to "Electronic Signature" was a compliant jump, and the deviation associated with an illegal jump was 0. Therefore, the deviation calculation for this node was: 0 + 0 = 0. The system determined that the deviation was zero, which did not exceed the default threshold of 0.8. The node was not marked as a deviation node, and no exception reporting process was triggered.

[0175] In this embodiment, by introducing standard business path nodes and legal jump structures as control references in the user behavior trajectory, the system can accurately identify deviations in the user operation process from two dimensions: time sequence and structure. This structured identification method has stronger logical interpretation ability and fault tolerance ability. By quantifying the deviation degree, the system can judge the severity of the problem according to the weight, so as to achieve a more prioritized exception handling schedule under the condition of limited resources.

[0176] In one embodiment, step S60 described above includes:

[0177] S601, extracting the associated path data of the deviation node in the operation path topology graph of the user operation behavior trajectory, where the associated path data includes the predecessor page path, successor page path of the deviation node, and the corresponding trigger timestamp;

[0178] S602, determining the impact level of the deviation node according to the predefined business impact policy library;

[0179] S603, determining a processing solution from the processing policy library based on the impact level of the deviation node;

[0180] S604, generating a business processing exception report including the associated path data, impact level, and processing solution of the deviation node.

[0181] In this embodiment, after the deviation node in the user operation behavior trajectory has been identified, in order to achieve intelligent exception identification and processing support, the system constructs a complete exception report through multi-dimensional information to assist subsequent business response and risk control. First, the system extracts the adjacent path information of the current deviation node in the operation path topology graph from the user operation behavior trajectory, that is, the predecessor page path, successor page path of the node, and the relevant trigger timestamp, as the structured context input. This structure helps to restore the context position of the deviation node in the overall business process and supports the exception interpretation logic.

[0182] Subsequently, the system evaluates the impact level of the deviation node based on the predefined business impact policy library. The grading logic in the policy library can be constructed based on factors such as node type, the number of missing nodes, and illegal jump intensity. The current evaluation rules include: if the deviation node belongs to key process nodes such as "identity verification", "payment confirmation", etc., it is directly marked as a blocking exception; if the number of illegal jump edges associated with a single node exceeds two, it is marked as a high-risk exception; if there is only one illegal jump edge and the jump path is within the allowable range boundary of the system, it is marked as a low-risk exception. The evaluation process supports dynamic weight adjustment and can be combined with enhanced dimensions such as timestamp exceptions.

[0183] After the impact level assessment is completed, the system matches the deviation node type with the processing strategy library to generate a targeted disposal plan. The disposal plan includes, but is not limited to: for blocking exceptions, directly generate a system instruction of "re-execute the key process module" (such as restarting the identity verification); for high-risk exceptions, submit a manual review request and set the freeze flag for the current business process, waiting for manual intervention; for low-risk exceptions, generate a graphical guidance prompt message to guide the user to make up for the missing operations, while allowing the process to continue execution.

[0184] Finally, the system generates a standard exception report according to the structured template. This report is in the form of JSON or semi-structured data, and the content includes: the page path identifier of the deviation node, the impact level field, the recommended disposal plan identifier, the problem trigger timestamp, and the upstream and downstream page paths and other information. This report can be used as an input basis for the customer service system, the human review platform or the process engine to support further automatic or manual processing.

[0185] Example illustration: In the online insurance application system for medical and health insurance, after a customer fills out the health questionnaire and completes the identity verification, due to accidentally touching the page refresh button, the customer is unexpectedly redirected to the product list page. After the customer then re-selects the same insurance product, the system automatically jumps to the "Payment Confirmation" page, skipping the two necessary links of "Health Declaration Submission" and "Electronic Signature Confirmation". The system identifies that the "Payment Confirmation" node is missing two key pre-page paths in the reconstructed user operation behavior trajectory, and detects an illegal edge directly from "Product Selection" to "Payment Confirmation" in its jump path. After calculating the deviation degree, the deviation degree of this node is 2.5, exceeding the platform-set deviation threshold of 1.0. Based on this, the system marks this node as a deviation node and extracts its associated path data (the predecessor is the product selection page, the successor is the payment completion page, and the timestamp corresponds to the current trigger time). According to the business impact strategy library judgment, this behavior belongs to a blocking exception. The system then matches the blocking response measures in the processing strategy library, automatically initiates a restart request for the identity verification and health declaration modules, and pushes the operation suggestions and key screenshot information to the customer service terminal for the customer service to verify the situation and inform the customer to make up the process.

[0186] In the personal consumer finance online application system, a user attempts to apply for a small loan. After completing identity verification and risk assessment, the user intends to return to view the contract terms. However, due to page loading delays, the user repeatedly clicks the button and finally skips the "Quota Confirmation" and "Approval Feedback" pages, and the system directly guides the user to the "Contract Signing" page. In the user operation behavior trajectory, the system identifies that two key page paths are missing at the "Contract Signing" node, and there are three illegal jump edges in the path, involving undefined jump methods such as "Risk Assessment → Contract Signing" and "Contract Selection → Contract Signing". The calculated node deviation degree is 3.5. According to the policy rules, this node is marked as a high-risk anomaly. The system generates a structured anomaly report containing the deviated node path, timestamp, and jump path, freezes the user's current loan application process, and sends a manual review request to the credit review team.

[0187] Another situation with a mild deviation appears in the auto insurance renewal scenario. When viewing the insurance information, the customer does not click the "Clause Confirmation" page but directly jumps to the "Payment Confirmation" through the browser forward function. The system detects that only one illegal jump edge exists, and the deviation degree of this node is 0.5, which does not exceed the set threshold of 1.0. Therefore, it will not be marked as a deviation node, but this path will be recorded as a boundary legal path. The system generates a brief prompt to guide the user to go back and read the clause confirmation page and continue the process to ensure the integrity of the process and the user experience is not affected.

[0188] Through the above steps in this embodiment, it is possible to quickly judge the degree of process violation in user behavior without manual intervention and automatically generate a clear disposal plan. This not only improves the system's response ability to abnormal insurance application processes but also significantly reduces the labor costs of misjudgment and repeated processing. Introducing path structure, semantic-level judgment, and hierarchical strategies significantly enhances the processing accuracy and coverage, forming a stable and controllable automatic risk control closed-loop.

[0189] In one embodiment, the above step S70 includes:

[0190] S701, processing the current conversation content through a keyword extraction model based on the multi-head attention mechanism to generate an initial keyword set;

[0191] S702, mapping the initial keyword set to a preset standardized service demand label dictionary to generate a service demand label set;

[0192] S703, extracting the business type code, historical service item identifier, and resource access permission identifier from the business processing record;

[0193] S704. Determine the semantic similarity between the service item description text in the service resource library and the service requirement tags in the service requirement tag set, and filter out candidate service items with a semantic similarity higher than the preset similarity threshold;

[0194] S705. Based on the service type code and the resource access permission identifier, filter out a subset of service items that meet the current service type and whose permission requirements are not higher than the user's permission from the candidate service items;

[0195] S706. Perform a priority ranking on the subset of service items according to the relevance of the historical service item identifiers to generate a service item priority list;

[0196] S707. Determine the service item with the highest ranking in the service item priority list as the target service item;

[0197] S708. Generate operation guidance information for the target service item based on the user operation behavior trajectory, the function description text of the target service item, and the front-end routing address.

[0198] In this embodiment, when understanding the service requirements of the current user, it is necessary to accurately identify the directional keyword information from the conversation content. The keyword extraction model realizes the deep understanding of the conversation semantics through the multi-head attention mechanism. This model can model various expression methods in the conversation and map the potential service requests to a set of candidate keyword sets. Each keyword in this set represents a possible service intention fragment, forming a coarse-grained expression of the service requirements.

[0199] To ensure the interpretability and standardization of the keywords, the system will match and map the initial keyword set with the predefined service requirement tag dictionary. This dictionary is usually constructed based on the existing service classification system and covers multiple tag dimensions such as financial products, insurance services, and health plans. After mapping, a structured service requirement tag set can be generated as the basis for further semantic matching.

[0200] Subsequently, the service type code, historical service item identifier, and resource access permission identifier of the current user are extracted from the business processing records. These contents are used to clarify which business context the user is currently in, which service items have been accessed before, and which service modules can be accessed, thereby serving as important boundary conditions for service screening.

[0201] ]>The screening of candidate service items depends on the semantic similarity between the function description text of each service item in the service resource library and the tag set. The function description text can be service process descriptions, user guides, or FAQ content. The system will judge the semantic distance between the service item and the user requirements through vector embedding and similarity calculation (such as cosine similarity), and only retain the service items with a similarity higher than the set threshold as candidates.

[0202] After the initial screening is complete, the system compares candidate services with user permissions and business types. Matching business type codes ensures service applicability, while verifying resource access permissions eliminates services that the user cannot access. Only services that meet the current contextual conditions are retained to form the final service subset.

[0203] On this basis, historical service item identifiers are introduced as a relevance factor for sorting. The sorting mechanism can weight scores based on service frequency, chronological order, user preference tags, and other factors to form a priority list of services. The highest-ranked service is considered the most consistent with the current context and user intent and is ultimately selected as the target service.

[0204] Based on the content of the target service, the system combines the user's operational behavior, the service's functional description, and the front-end page routing to generate three types of guidance information: the first is the interface navigation path, which generates instructions for jumping from the user's current interface to the target interface through topological path backtracking; the second is the operation step description, which combines the service function text to generate a step-by-step guide to the operation control description, step number, and expected results; and the third is the one-click access portal, which generates a direct link by binding the front-end routing address with the current user's session parameters and status identifier. Ultimately, all user operation guidance information is encapsulated as structured data for unified output.

[0205] To ensure that customers receive problem location results, processing progress, or operational guidance information as soon as possible, the system supports not only displaying the above content through the current intelligent dialogue interface, but also simultaneously pushing it to the customer terminal through multiple information transmission channels. Push channels include SMS platforms, corporate WeChat, email service systems, etc. The system will automatically generate a multi-channel push list based on the contact information in the customer account information, and automatically select the push path based on the conversation urgency level. If the customer does not click or respond within a certain period of time after the first information push, the system will initiate a secondary reminder strategy to ensure the information delivery rate and processing completion rate. All push processes are recorded in the customer service log for subsequent complaint handling and process review.

[0206] Example illustration: In the field of healthcare, a user mentions through the intelligent customer service, "I want to continue the unfinished service of the previous vaccine appointment." The system extracts initial keywords such as "vaccine appointment", "continue", and "unfinished" through the keyword model and maps them to standardized service labels such as "health service appointment" and "registration reissuance". Combining the "healthcare business" type code and permission level in the user's business processing record, the system filters out multiple candidate service items related to the "vaccine appointment process" in the service resource library. Through similarity matching and historical record comparison, the "vaccine appointment progress restoration" service item is finally determined as the target service. The system generates a one-click jump link by combining its current staying page with the front-end routing of the service item and displays guiding content: "Click Next to enter the vaccine appointment page → Fill in the information of the vaccinated person → Select the appointment time period → Confirm and submit", realizing the visualization of the complete path and the guidance of operation steps.

[0207] In the field of financial services, the customer states in the conversation, "Was the loan I applied for two days ago rejected?" The system identifies keywords such as "loan application", "status query", and "rejected" and maps them to the label of "personal loan status view". Combining the historical service identifier "short-term consumer loan" extracted from its business record and the approval query permission, the system filters the "loan approval progress query" service item from the service library and finds that the user clicked on the service entry within 48 hours. Therefore, this item is ranked first in priority. The system generates a jump link through its page routing information and displays path navigation and operation instructions: "Click to enter the approval progress page → View the approval log → Download the decision letter" to help the customer quickly confirm the current status.

[0208] Through the construction of the whole process from keyword extraction to semantic screening, permission verification, and historical correlation sorting in this embodiment, the service intention in the user's current conversation content can be accurately mapped to the target service item in the service resource library. At the same time, combined with the user's historical operation path and service content definition, the collaborative guidance of path navigation, operation decomposition, and one-click direct access is realized, improving the efficiency of user problem response and the fluency of interaction, and reducing misguidance and path jump errors.

[0209] In one embodiment, after the above step S10, it further includes:

[0210] S101, when it is recognized that the current conversation content contains service complaint feature data, extract the identification information of the service personnel being complained about from the current conversation content;

[0211] S102, query the service evaluation database based on the identification information to obtain the historical complaint times and service scores of the service personnel being complained about;

[0212] S103. If the number of historical complaints exceeds a preset threshold or the service rating is lower than the passing standard score, generate a service staff replacement recommendation and push the service staff replacement recommendation and the backup service staff information to the current conversation.

[0213] S104. If a confirmation feedback signal for the service staff replacement recommendation is detected, update the service allocation record and send a service handover notice in the current conversation.

[0214] S105. If the complaint content in the current conversation contains service process defect information, generate a business process optimization recommendation and push it to the management terminal.

[0215] In this embodiment, when the current conversation content contains service complaint feature data, the text semantic recognition module can identify language fragments with negative emotions or complaint intentions, such as typical complaint expressions like "poor service attitude", "low communication efficiency", and "dissatisfied with the processing result". After identifying these feature data, further analyze the conversation context to extract the identification information of the object being complained about. This identification information is usually the service staff number, employee work number, or the service staff identity identifier associated in the session allocation record.

[0216] After obtaining the identity identifier of the service staff, the historical complaint record count, customer rating historical data, and the current service status related to this service staff can be obtained by indexing and matching with the service evaluation database. Each record in the evaluation database usually contains dimension data such as service scenario, rating score, complaint type, and timestamp. In the determination process, the system can set two key judgment criteria: one is whether the number of historical complaints exceeds the preset warning threshold, and the other is whether the average rating is lower than the service passing score set by the platform.

[0217] If either of the above conditions is met, it can be determined that there is a service quality risk for the current service staff. The system will generate a decision result including the replacement recommendation, the information of the original service staff, and the candidate information of the replacement staff. The backup service staff information can be selected from the service staff pool as an object with an available current status, a higher rating, and matching skill tags.

[0218] When the user accepts the replacement recommendation through voice or text feedback, such as replying with keywords like "agree to replace" or "can change the customer service", the system will recognize it as a confirmation feedback signal. At this time, the current service allocation record will be updated, and the user will be notified of the completion of the service handover through the current conversation channel.

[0219] If the conversation content contains descriptions of service process defects in addition to personnel issues, such as expressions like "the process is too complex" or "repeated filling many times", it can trigger the process optimization mechanism. The system will extract the structural problems that occur in the service process based on keyword matching and semantic modeling, and organize them into suggestions for nodes to be optimized, which are pushed to the management side for process managers to refer to.

[0220] When it is recognized that the current conversation content of the customer contains service complaint feature data, the system can further extract features such as emotional words, objects of complaint, and feedback methods in the customer's expression, further judge the intensity of the customer's emotion through the complaint semantic analysis model, and match the information of agents or service personnel within the company related to the customer's complaint content. Combining the historical service quality records of the complained agent and the customer label features, generate appeasement information or response suggestions, such as providing descriptions of the complaint handling progress, apology templates, prompts for temporary preferential plans, etc., and push them to the customer service personnel interface through the intelligent dialogue module. The customer service personnel can choose whether to send them to the customer immediately or submit them to the direct superior of the agent for review and response. This appeasement process can also be regarded as an intelligent intervention mechanism to reduce the risk of policy surrender or the probability of service interruption caused by customer emotional fluctuations.

[0221] Example illustration: When a user conducts an insurance purchase operation on the mobile health insurance platform, due to lack of proficiency in operation, the policy submission is not completed after multiple attempts. The system receives the voice conversation request from this user: "I have submitted the information, but the page keeps jumping and fails. Is the insurance not successfully purchased?" The intelligent customer service platform first parses this natural language input through the semantic analysis module, identifies the semantic features related to "policy submission failure" and "abnormal jump", determines that the current conversation belongs to the type of business processing problem, and then triggers the subsequent process.

[0222] The platform automatically calls the session recognition module, combines the user identity identification information, and retrieves the business processing record of the user in the most recent insurance purchase process. The system determines that the last standard process node of this user is "submission of health declaration", and extracts the timestamp of this node from the business operation log. The current processing time has passed 15 minutes since this node, exceeding the system-set 5-minute interval threshold. Therefore, the complete operation chain data of this user from real-name authentication to the current is extracted from the historical operation record database as the target operation record.

[0223] Subsequently, the platform analyzes the interface interaction events in the target operation record, sorts all events in ascending order of the trigger time, generates an ordered event sequence, extracts the control hierarchy path and page jump information among them, and generates an event annotation set containing the page path. Further construct the operation path topology graph of this user, and generate a user operation behavior trajectory graph with the page path as the node and the timestamp as the edge weight.

[0224] The system compares the user's behavior trajectory with the predefined standard insurance application process on the platform and finds that the "electronic signature" page is missing in this trajectory, and there is an illegal jump path from "health declaration" directly to "payment confirmation", constituting the absence of two key nodes and one illegal jump. The deviation value of the "payment confirmation" page is calculated to be 2.5 (the absence of two necessary nodes is counted as 2 points, and the illegal jump edge contributes 0.5 points), exceeding the set deviation threshold of 1.5, so it is marked as a deviation node.

[0225] Based on this deviation node, the platform extracts its upper and lower path information (for example, the previous page is "health declaration" and the next page is "payment confirmation"), generates associated path data in combination with the timestamp information, and determines that this problem is a blocking exception according to the business impact strategy library. The matching processing strategy is to "re-trigger the electronic signature module". Subsequently, the system generates a structured business processing exception report, displays it through the current dialogue window, and simultaneously backs it up into the customer problem handling record library.

[0226] Furthermore, the system recognizes that the user's words also contain service request features such as "I want to consult a health insurance plan suitable for me". Therefore, it synchronously calls the keyword extraction module, uses the multi-head attention mechanism to identify keywords such as "health", "insurance", "suitable for me", and "plan" from the current dialogue content, and maps them to standardized service tags such as "health insurance product consultation" and "customer portrait matching". Combining the business type code extracted from the user's business processing record as "health insurance application", the access permission level as "ordinary user", and the historical service item as the successful application of "mid-end medical insurance", the platform performs semantic similarity matching and screening on the candidate service items in the service resource library, and retains service items such as "personal health insurance" and "customized health check-up insurance".

[0227] Subsequently, the system sorts the candidate service items according to the historical service records, and finally selects "mid-end customized health check-up insurance" as the current target service item. The system automatically generates operation guide information for this service item, including: a navigation path description for jumping from the current page to the "physical examination insurance recommendation" page, an executable operation step description generated based on the serial number and position of the information filling control, and a page direct link with user identity parameters. This information is returned to the user through the intelligent customer service interface.

[0228] During the whole process, if the user expresses dissatisfaction with a certain customer service staff when reporting a problem, such as saying "Your last customer service staff always perfunctory me", the system can further extract the service staff number and obtain their historical complaint times and service scores from the service evaluation database. If the service replacement rule is met, the system will prompt the user whether to replace the service staff. If the user confirms, the background service allocation module will complete the service handover and generate a service continuation notice at the same time.

[0229] In this embodiment, through the structured understanding and processing of service complaint content, service quality risks can be identified in advance, and the dynamic adjustment of service personnel can be achieved. At the same time, combined with the service process defect identification function, the real feedback of customers is used to drive process optimization, thereby improving the overall service satisfaction and problem response efficiency.

[0230] In one embodiment, a business service processing device based on dialogue recognition is provided, and the business service processing device based on dialogue recognition corresponds one-to-one with the business service processing method based on dialogue recognition in the above embodiment. Refer to Figure 3 , Figure 3 FIG. is a schematic diagram of functional modules of a preferred embodiment of the business service processing device based on dialogue recognition of the present invention. A semantic recognition module 10, a service data extraction module 20, an operation record generation module 30, an operation track construction module 40, an abnormal node recognition module 50, an abnormal report generation module 60, and a service matching and guidance module 70. The detailed description of each functional module is as follows:

[0231] The semantic recognition module 10 is used to parse the current dialogue content through a semantic analysis module to generate a key problem type;

[0232] The service data extraction module 20 is used to obtain the service processing record of the session subject of the current dialogue content according to the key problem type;

[0233] The operation record generation module 30 is used to obtain a target operation record according to the interval between the problem occurrence time and the current time when the key problem type is a service processing problem type;

[0234] The operation track construction module 40 is used to construct a user operation behavior track based on the interface interaction events in the target operation record;

[0235] The abnormal node recognition module 50 is used to identify deviation nodes in the user operation behavior track;

[0236] The abnormal report generation module 60 is used to generate a service processing abnormal report according to the deviation nodes;

[0237] The service matching and guidance module 70 is used to match a target service item in the service resource library according to the service demand keywords in the current dialogue content and the service processing record, and generate operation guidance information for the target service item.

[0238] In one embodiment, the operation record generation module 30 is specifically used for:

[0239] When the key problem type is a service processing problem type, extract the time stamp of the last successfully completed standard service process node of the session subject from the service operation log as the problem occurrence time;

[0240] Obtain the current timestamp for processing the current conversation request;

[0241] Determine the interval duration between the current timestamp and the problem occurrence timestamp;

[0242] When the interval duration exceeds a preset interval threshold, retrieve the complete operation chain data of the session subject from the historical operation record database, and use the complete operation chain data as the target operation record;

[0243] When the interval duration does not exceed the preset interval threshold, obtain the interface interaction event stream of the current conversation from the current operation flow monitoring interface, and use the interface interaction event stream as the target operation record.

[0244] In one embodiment, the operation trajectory construction module 40 is specifically configured to:

[0245] Sort the interface interaction events in the target operation record in ascending order according to the trigger timestamp to generate an ordered event sequence;

[0246] Extract the control hierarchy path of each interface interaction event from the ordered event sequence, and generate an event annotation set including the page path based on the control hierarchy path;

[0247] Generate an operation path topology graph based on the trigger order of the interface interaction events in the event annotation set, where the nodes in the operation path topology graph are page paths and the edges are event trigger orders;

[0248] Construct the user operation behavior trajectory according to the page jump relationship and control trigger order in the operation path topology graph.

[0249] In one embodiment, the abnormal node recognition module 50 is specifically configured to:

[0250] Extract all page path nodes and trigger order edges from the operation path topology graph of the user operation behavior trajectory;

[0251] Compare the page path nodes with a predefined set of necessary nodes in the standard business process, and mark the necessary nodes not included in the operation path topology graph;

[0252] Compare the trigger order edges with a predefined set of legal jump edges, and mark the illegal jump edges not matched in the operation path topology graph;

[0253] Determine the node deviation degree of each page path node based on the necessary nodes not included and the illegal jump edges not matched;

[0254] Mark the page path nodes with node deviation exceeding the preset deviation threshold as deviation nodes.

[0255] In one embodiment, the exception report generation module 60 is specifically configured to:

[0256] Extract the associated path data of the deviation nodes in the operation path topology graph of the user operation behavior trajectory, where the associated path data includes the predecessor page path, successor page path of the deviation nodes, and the corresponding trigger timestamps;

[0257] Determine the impact level of the deviation nodes according to the predefined business impact policy library;

[0258] Determine a processing solution from the processing policy library based on the impact level of the deviation nodes;

[0259] Generate a business processing exception report including the associated path data, impact level, and processing solution of the deviation nodes.

[0260] In one embodiment, the service matching and guidance module 70 is specifically configured to:

[0261] Process the current conversation content through a keyword extraction model based on the multi-head attention mechanism to generate an initial keyword set;

[0262] Map the initial keyword set to a preset standardized service demand label dictionary to generate a service demand label set;

[0263] Extract the business type code, historical service item identifier, and resource access permission identifier from the business processing record;

[0264] Determine the semantic similarity between the service item description text in the service resource library and the service demand labels in the service demand label set, and filter out candidate service items with semantic similarity higher than the preset similarity threshold;

[0265] Based on the business type code and resource access permission identifier, filter out a subset of service items that meet the current business type and whose permission requirements are not higher than the user's permission from the candidate service items;

[0266] Sort the subset of service items according to the relevance of the historical service item identifiers to generate a service item priority list;

[0267] Determine the service item with the highest ranking in the service item priority list as the target service item;

[0268] Generate operation guidance information for the target service item based on the user operation behavior trajectory, the function description text of the target service item, and the front-end routing address.

[0269] In one embodiment, the semantic recognition module 10 is specifically configured to:

[0270] When it is recognized that the current conversation content contains service complaint feature data, extract the identification information of the service personnel being complained about from the current conversation content;

[0271] Query the service evaluation database based on the identification information to obtain the historical complaint times and service scores of the service personnel being complained about;

[0272] If the historical complaint times exceed the preset times threshold or the service score is lower than the passing standard score, generate a service personnel replacement suggestion, and push the service personnel replacement suggestion and the information of the standby service personnel to the current conversation;

[0273] If a confirmation feedback signal of the service personnel replacement suggestion is detected, update the service allocation record and send a service handover notice in the current conversation;

[0274] If the complaint content in the current conversation contains service process defect information, generate a business process optimization suggestion and push it to the management terminal.

[0275] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as Figure 4 shown. The computer device includes a processor, a memory, a network interface, and a database connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage media. The network interface of the computer device is used to communicate with an external user terminal through a network connection. When the computer program is executed by the processor, it realizes the functions or steps on the server side of a business service processing method based on conversation recognition.

[0276] In one embodiment, a computer device is provided. The computer device may be a user terminal, and its internal structure diagram may be as Figure 5As shown in the figure. The computer device includes a processor, a memory, a network interface, a display screen, and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external server through a network connection. When the computer program is executed by the processor, it realizes the functions or steps of the client side of a business service processing method based on dialogue recognition

[0277] 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. When the processor executes the computer program, the following steps are realized:

[0278] Parse the current dialogue content through a semantic analysis module to generate a key problem type;

[0279] According to the key problem type, obtain the business processing record of the session subject of the current dialogue content;

[0280] When the key problem type is a business processing problem type, obtain a target operation record according to the interval between the problem occurrence time and the current time;

[0281] Construct a user operation behavior trajectory based on the interface interaction events in the target operation record;

[0282] Identify the deviation nodes in the user operation behavior trajectory;

[0283] Generate a business processing exception report according to the deviation nodes;

[0284] Match the target service item in the service resource library according to the service demand keywords in the current dialogue content and the business processing record, and generate operation guidance information for the target service item.

[0285] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by the processor, the following steps are realized:

[0286] Parse the current dialogue content through a semantic analysis module to generate a key problem type;

[0287] According to the key problem type, obtain the business processing record of the session subject of the current dialogue content;

[0288] When the key problem type is a business processing problem type, obtain a target operation record according to the interval between the problem occurrence time and the current time;

[0289] Construct a user operation behavior track based on the interface interaction events in the target operation record;

[0290] Identify the deviation nodes in the user operation behavior track;

[0291] Generate a business processing exception report according to the deviation nodes;

[0292] Match the target service item in the service resource library according to the service requirement keywords in the current conversation content and the business processing record, and generate the operation guide information of the target service item.

[0293] It should be noted that for the functions or steps that can be realized by the above computer-readable storage medium or computer device, reference can be made to the relevant descriptions on the server side and the user side in the foregoing method embodiments. To avoid repetition, they will not be described in detail here.

[0294] Those of ordinary skill in the art can understand that all or part of the processes of implementing the methods in the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the above method embodiments. Among them, any reference to the memory, storage, database or other media used in the various embodiments provided in the present application can include non-volatile and / or volatile memories. The non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. The 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), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0295] Those skilled in the art can clearly understand that for the convenience and simplicity of description, only the above division of each functional unit and module is used as an example. In actual applications, the above functions can be allocated to different functional units and modules according to needs, that is, the internal structure of the device is divided into different functional units or modules to complete all or part of the functions described above.

[0296] It should be noted that if software tools or components of other companies appear in the embodiments of this application, they are only used for example introduction and do not represent actual use. The above-described embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements 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 the present invention, and should all be included in the protection scope of the present invention.

Claims

1. A business service processing method based on dialogue recognition, characterized in that It includes the following steps: Parse the current conversation content through a semantic analysis module to generate a key question type; According to the key question type, obtain the business processing records of the conversation subject of the current conversation content; When the key question type is a business processing question type, obtain a target operation record according to the interval between the problem occurrence time and the current time; Construct a user operation behavior track based on the interface interaction events in the target operation record; Identify the deviation nodes in the user operation behavior track; Generate a business processing exception report according to the deviation nodes; Match the target service item in the service resource library according to the service demand keyword in the current conversation content and the business processing record, and generate the operation guide information of the target service item.

2. The method for processing a service based on dialogue recognition according to claim 1, wherein When the key question type is a business processing question type, obtaining a target operation record according to the interval between the problem occurrence time and the current time includes: When the key question type is a business processing question type, extract the timestamp of the last successfully completed standard business process node of the conversation subject from the business operation log as the problem occurrence time; Obtain the current timestamp for processing the current conversation request; Determine the interval duration between the current timestamp and the problem occurrence timestamp; When the interval duration exceeds a preset interval threshold, retrieve the complete operation chain data of the conversation subject from the historical operation record database, and use the complete operation chain data as the target operation record; When the interval duration does not exceed the preset interval threshold, obtain the interface interaction event stream of the current conversation from the current operation flow monitoring interface, and use the interface interaction event stream as the target operation record.

3. The business service processing method based on dialogue recognition according to claim 1, wherein Constructing a user operation behavior track based on the interface interaction events in the target operation record includes: Arrange the interface interaction events in the target operation record in ascending order according to the trigger timestamp to generate an ordered event sequence; Extract the control hierarchy path of each interface interaction event from the ordered event sequence, and generate an event annotation set including the page path based on the control hierarchy path; Generate an operation path topology graph based on the trigger order of the interface interaction events in the event annotation set, where the nodes in the operation path topology graph are page paths and the edges are event trigger orders; Construct a user operation behavior track according to the page jump relationship and control trigger order in the operation path topology graph.

4. The business service processing method based on dialogue recognition according to claim 1, wherein Identifying the deviation nodes in the user operation behavior track includes: Extract all page path nodes and trigger order edges from the operation path topology graph of the user operation behavior track; Compare the page path nodes with the predefined set of necessary nodes in the standard business process, and mark the necessary nodes not included in the operation path topology graph; Compare the trigger order edges with the predefined set of legal jump edges, and mark the illegal jump edges not matched in the operation path topology graph; Determine the node deviation degree of each page path node based on the necessary nodes not included and the illegal jump edges not matched; Mark the page path nodes with a node deviation degree exceeding the preset deviation threshold as deviation nodes.

5. The method for processing business services based on dialogue recognition according to claim 1, wherein, Generate a business processing exception report based on the deviation node, including: Extract the associated path data of the deviation node in the operation path topology diagram of the user operation behavior track, where the associated path data includes the predecessor page path, successor page path of the deviation node, and the corresponding trigger timestamp; Determine the impact level of the deviation node according to the predefined business impact policy library; Determine a processing solution from the processing policy library based on the impact level of the deviation node; Generate a business processing exception report including the associated path data, impact level, and processing solution of the deviation node.

6. The method for processing business services based on dialogue recognition according to claim 1, characterized in that Match the target service item in the service resource library according to the service requirement keywords in the current conversation content and the business processing record, and generate the operation guidance information of the target service item, including: Process the current conversation content through a keyword extraction model based on the multi-head attention mechanism to generate an initial keyword set; Map the initial keyword set to a preset standardized service requirement label dictionary to generate a service requirement label set; Extract the business type code, historical service item identifier, and resource access permission identifier from the business processing record; Determine the semantic similarity between the service item description text in the service resource library and the service requirement labels in the service requirement label set, and filter out candidate service items with a semantic similarity higher than the preset similarity threshold; Based on the business type code and resource access permission identifier, filter out a subset of service items that meet the current business type and whose permission requirements are not higher than the user's permission from the candidate service items; Sort the subset of service items according to the relevance of the historical service item identifier to generate a service item priority list; Determine the service item with the highest ranking in the service item priority list as the target service item; Generate the operation guidance information of the target service item based on the user operation behavior track, the function description text of the target service item, and the front-end routing address.

7. The method for processing business services based on dialogue recognition according to claim 1, wherein After generating the key problem type by parsing the current conversation content through the semantic analysis module, it further includes: When it is recognized that the current conversation content contains service complaint feature data, extract the identification information of the service personnel being complained about from the current conversation content; Query the service evaluation database based on the identification information to obtain the historical complaint times and service scores of the service personnel being complained about; If the historical complaint times exceed the preset times threshold or the service score is lower than the qualified standard score, generate a service personnel replacement suggestion, and push the service personnel replacement suggestion and the backup service personnel information to the current conversation; If a confirmation feedback signal of the service personnel replacement suggestion is detected, update the service allocation record and send a service handover notice in the current conversation; If the complaint content in the current conversation contains service process defect information, generate a business process optimization suggestion and push it to the management terminal.

8. A business service processing device based on dialogue recognition, characterized in that, The business service processing device based on dialogue recognition includes: A semantic recognition module for parsing the current conversation content through a semantic analysis module to generate a key problem type; A business data extraction module for obtaining the business processing record of the session subject of the current conversation content according to the key problem type; An operation record generation module, configured to obtain target operation records according to the interval between the problem occurrence time and the current time when the key problem type is a business processing problem type; An operation track construction module, configured to construct a user operation behavior track based on the interface interaction events in the target operation records; An abnormal node identification module, configured to identify the deviation nodes in the user operation behavior track; An abnormal report generation module, configured to generate a business processing abnormal report according to the deviation nodes; A service matching and guidance module, configured to match a target service item in a service resource library according to the service demand keywords in the current conversation content and the business processing record, and generate operation guidance information for the target service item.

9. A computer device, characterized in that, The computer device includes a memory, a processor, and a business service processing program based on dialogue recognition that is stored on the memory and can run on the processor. When the business service processing program based on dialogue recognition is executed by the processor, the steps of the business service processing method based on dialogue recognition according to any one of claims 1-7 are implemented.

10. A computer-readable storage medium, characterized in that, A business service processing program based on dialogue recognition is stored on the storage medium. When the business service processing program based on dialogue recognition is executed by a processor, the steps of the business service processing method based on dialogue recognition according to any one of claims 1-7 are implemented.

Images