Full-stack digital intelligent tax management service operation system

Abstract

The invention relates to the technical field of service management, and discloses a full-stack digital intelligence tax management service operation system, which comprises a data acquisition and integration module used for real-time access and automatic integration of multi-source data, the multi-source data comprises electronic tax bureau hotline data, ERP / CRM business data and enterprise cooperation platform data; the intelligent business processing module is used for automatic and intelligent processing of core businesses, including scheduling management, quality inspection management and asset management; the risk management and control module is used for closed-loop management of early warning, intervention and analysis of tax risk; the knowledge and cooperation module is used for team professional ability improvement and cross-department cooperation efficiency optimization; according to the cloud native micro-service architecture, a Spring Cloud framework is adopted to construct a core central layer, and service registration discovery and configuration center double engines are integrated; the technical problems of data islands and poor real-time performance existing in traditional tax management are solved.

Description

Technical Field

[0001] This invention relates to the field of service management technology, specifically to a full-stack digital intelligent tax and fee management service operation system. Background Technology

[0002] In traditional tax and fee management scenarios, enterprises generally face three core pain points: data silos, lack of real-time data, and uncontrollable data quality, which severely restrict the efficiency of tax and fee management and the accuracy of decision-making. Data silos: Tax source data is scattered across multiple heterogeneous systems such as the e-tax bureau, ERP / CRM systems, and collaboration platforms, lacking a unified integration mechanism. For example, e-tax bureau hotline data only records the type of inquiry and processing time, and cannot be linked to enterprise sales orders, contract information, and employee chat records, leading to problems such as "incomplete tax source data" and "inconsistent financial and tax data standards," forming cross-departmental data barriers.

[0003] Poor real-time performance: Traditional data collection relies on manual entry or batch synchronization, which results in significant delays. For example, hotline call data needs to be manually processed and uploaded the next day, which cannot reflect sudden surges in inquiries in real time, leading to delays in scheduling adjustments and resource allocation, and affecting service continuity. Summary of the Invention

[0004] The present invention aims to provide a full-stack digital tax management service operation system to solve the technical problems of data silos and poor real-time performance in traditional tax management.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a full-stack digital intelligent tax management service operation system, comprising: The data acquisition and integration module is used for real-time access and automatic integration of multi-source data, including data from the electronic tax bureau hotline, ERP / CRM business data, and enterprise collaboration platform data. The intelligent business processing module is used for the automation and intelligent processing of core businesses, including scheduling management, quality inspection management, and asset management. The risk management module is used for closed-loop management of tax and fee risks, including early warning, intervention, and analysis. The Knowledge and Collaboration module is used to enhance team professional capabilities and optimize cross-departmental collaboration efficiency. The cloud-native microservice architecture uses the Spring Cloud framework to build the core central layer, integrating a dual engine for service registration and discovery and a configuration center.

[0006] The principle and advantages of the scheme are: in actual application, the multi-source real-time access of hotline data of electronic tax authorities, ERP / CRM business data and enterprise collaboration platform data is realized through Kafka asynchronous high-concurrency collection technology, combined with ETL tool cleaning conversion and NLP key information extraction, the unified storage and associated integration of scattered data are completed. Break the data island, guarantee the real-time and quality controllability of data, solve the problem of "tax source bottom number not comprehensive" and "financial and tax data caliber not unified".

[0007] The intelligent business processing module improves the intelligent level of business processing, solves the problems of low efficiency and high misjudgment rate of traditional manual operation, such as response delay of sudden peak scheduling, large amount of manual review of quality inspection, etc. The risk control module realizes accurate risk identification and closed-loop control, solves the problems of traditional risk identification lag and high false alarm rate. The knowledge and cooperation module improves the team's professional ability and cross-department cooperation efficiency, solves the problems of difficult knowledge sharing and low cooperation efficiency. The cloud native micro-service architecture supports dynamic scaling and fault isolation, ensures business continuity and scalability, and solves the problems of poor scalability and high maintenance cost of traditional monolithic architecture.

[0008] Preferably, as an improvement, the data collection and integration module includes a multi-source data access architecture and automatic integration and abnormal alarm; The multi-source data access architecture realizes real-time grabbing of structured data of hotline call records, customer problem classification and processing time length of electronic tax authorities through API interface, and realizes asynchronous high-concurrency collection by using Kafka message queue; sales order, contract information and invoice data are extracted from ERP / CRM system through ETL tool, and stored in data warehouse after cleaning and conversion; employee chat records, knowledge base documents and training records of chat platform are obtained through OAuth2.0 authorization, and key information is extracted using NLP technology; Automatic integration and abnormal alarm, based on Spark, a distributed processing framework is built, and hotline data, scheduling data and accompanying training data are associated according to business scenarios, and the integration results of customer portrait and service efficiency indicators are generated and stored in Hive data warehouse; data quality rules are defined by using rule engine such as Drools, when data collection fails or integration is abnormal, messages are pushed to operation and maintenance personnel through enterprise WeChat robot, and abnormal logs are recorded to Elasticsearch.

[0009] The improvement has the beneficial effects of: multi-source data real-time access and integration, through API interface, Kafka message queue, ETL tool and NLP technology, realizing real-time collection, cleaning conversion and unified storage of hotline data of electronic tax authorities, ERP / CRM business data and enterprise collaboration platform data, breaking the data island, solving the problems of "tax source bottom number not comprehensive" and "financial and tax data caliber not unified", guaranteeing the real-time and quality controllability of data.

[0010] Automatic integration and abnormal alarm, based on Spark to build a distributed processing framework, according to the business scene, the hotline data, scheduling data, and accompanying training data are associated to generate customer portraits, service efficiency indicators, and other integrated results; The data quality rules are defined by the Drools rule engine, and when data collection fails or integration is abnormal, the message is pushed to the operation and maintenance personnel through the WeChat robot, and the abnormal log is recorded to Elasticsearch, realizing the rapid response and traceability of the exception, and improving the operation and maintenance efficiency and data reliability.

[0011] Preferably, as an improvement, the intelligent business processing module comprises: The scheduling module predicts the hotline volume of each period in the next 7 days based on the LSTM time series model, generates the optimal scheme by combining the employee tax policy familiarity / skill label of system operation proficiency, and triggers the system to automatically increase the number of employees and notifies through SMS when the connection rate is lower than 90%. The quality inspection module converts the recording to text using ASR technology, performs semantic analysis through a large model, scores according to preset rules, generates a quality inspection report, and pushes it to the inspected personnel. The asset management module locates the asset state through RFID, and the tax linkage module automatically calculates the depreciation and generates the declaration data.

[0012] The improvement has the beneficial effects that the scheduling module predicts the hotline volume of each period in the next 7 days based on the LSTM time series model, generates the optimal scheduling scheme by combining the employee tax policy familiarity, system operation proficiency, and other skill labels, solves the problem of response delay in scheduling during sudden peak, and improves the service response speed and efficiency by triggering the system to automatically increase the number of employees and notifying through SMS when the connection rate is lower than 90%.

[0013] The quality inspection module converts the recording to text using ASR technology, performs semantic analysis through a large model, scores according to preset rules, generates a quality inspection report, and pushes it to the inspected personnel, reducing the manual review workload, improving the quality inspection accuracy and efficiency, and reducing the misjudgment rate.

[0014] The asset management module locates the asset state through RFID, and the tax linkage module automatically calculates the depreciation and generates the declaration data, realizing the automation and intelligentization of asset management, reducing manual operation errors, and improving the asset utilization rate and the accuracy of declaration data.

[0015] Preferably, as an improvement, the risk control module comprises: Data analysis and early warning, based on historical data to build intelligent voice diversion rate, call accuracy rate, and service specification rate, identify abnormal values through Z-Score anomaly detection algorithm, trigger emergency state SMS + phone three-level early warning; Support work order assignment to responsible departments, track processing progress, automatically archive processing results and generate risk reports.

[0016] The improvement has the beneficial effects of data analysis and early warning, construction of intelligent voice diversion rate, call accuracy rate and service specification rate based on historical data, identification of abnormal values through Z-Score abnormality detection algorithm, triggering of emergency state short message + telephone three-level early warning, realization of accurate identification and timely early warning of risks, and solution of the problems of lagging risk identification and high false alarm rate. Work order assignment and progress tracking support assignment of work orders to responsible departments, tracking of processing progress, automatic archiving of processing results and generation of risk reports, realization of closed-loop management of risk control, and improvement of risk processing efficiency and traceability.

[0017] Preferably, as an improvement, the knowledge and collaboration module comprises: The AI question and answer module interfaces with the State Administration of Taxation policy library, extracts the applicable points of tax reduction and fee reduction amount through NLP, and associates enterprise business data; The intelligent accompanying training module captures abnormal cases, generates standardized documents, and synchronizes them to the collaboration platform. Employee simulation exercise scores are associated with performance evaluation. The network disk supports file classification storage and cross-regional sharing. The Tian Gong value management converts operation volume into points and displays contribution through a visual board.

[0018] The improvement has the beneficial effects of the AI question and answer module, which interfaces with the State Administration of Taxation policy library, extracts the applicable points of tax reduction and fee reduction amount through NLP, and associates enterprise business data, providing accurate policy consultation and business guidance for employees, and improving team professional ability and business processing accuracy.

[0019] The intelligent accompanying training module captures abnormal cases, generates standardized documents, and synchronizes them to the collaboration platform. Employee simulation exercise scores are associated with performance evaluation. Through practical training, employees' business skills are improved, and knowledge sharing and cross-department collaboration efficiency are optimized.

[0020] The network disk and the Tian Gong value management support file classification storage and cross-regional sharing, enabling quick access and sharing of knowledge resources. The Tian Gong value management converts operation volume into points and displays contribution through a visual board, motivating employees, improving team collaboration efficiency, and enhancing overall performance.

[0021] Preferably, as an improvement, the scheduling module adopts a multi-modal heterogeneous graph neural network to construct a heterogeneous graph with multi-core nodes, including time nodes, weather event nodes, public opinion event nodes, invoice collection period nodes, and employee function nodes. The Graph Transformer attention mechanism is used to dynamically calculate the correlation strength between external sudden events and traffic volume, and the predicted traffic volume is output. The predicted traffic volume is combined with employee skill labels to adjust the scheduling arrangement.

[0022] The improved benefits are: multi-modal correlation modeling, through multi-core nodes, time / weather / public opinion / invoice collection period / employee function, constructing a heterogeneous graph, combining the Graph Transformer attention mechanism to dynamically calculate the correlation strength between external events and traffic volume, and realizing more accurate traffic volume prediction. For example, heavy rain may increase hotline consultation volume, and policy public opinion events may trigger concentrated consultation, avoiding the problem of delayed response to sudden events in traditional scheduling.

[0023] Intelligent scheduling optimization, based on predicted traffic volume and employee skill labels such as tax policy familiarity and system operation proficiency, dynamically adjusts the scheduling scheme to solve the problem of call connection rate below 90% during sudden peak periods, and improves service response efficiency and resource utilization.

[0024] Preferably, as an improvement, the quality inspection module adopts a dual-channel game architecture, including a large model path and a rule engine path. The rule engine channel labels each rule in the rule library with a unique number and a deduction standard. When triggered, after converting the recording to text, the rule engine scans the text to match the rules and generates a baseline result containing the rule number. The large model channel uses ERNIE X1 to perform deep semantic understanding on the recording text, identify the context, and structure the output including the deduction item, semantic evidence, and rule number prediction result. A ternary loss function is constructed, including a rule recall weight, a model misjudgment rate weight, and a dual-channel difference penalty item, to make the outputs of the rule engine channel and the large model channel converge.

[0025] The improved benefits are: dual-channel collaborative verification, the rule engine channel generates a baseline deduction result through a unique number rule library, the large model channel (ERNIE X1) performs deep semantic understanding, identifies the context, deduction items, and semantic evidence, and a ternary loss function, rule recall weight + model misjudgment rate weight + dual-channel difference penalty item, drives the outputs of the two channels to converge, ensuring strict execution of basic rules while covering complex semantic scenarios, reducing the amount of manual review and misjudgment rate. Evidence traceability, semantic evidence output by the large model is associated with the rule number, the quality inspection report contains deduction items, evidence chain, and rule basis, improving the transparency and recognition of the inspected person of quality inspection.

[0026] Preferably, as an improvement, the risk control module constructs a dual-dimensional fusion system of voice emotion and image verification, a voice emotion recognition model extracts a 128-dimensional emotion vector, and an OCR confidence and two-dimensional code verification invoice image cross-verify amount / time information. The spatiotemporal correlation graph takes the voice segment and the invoice image as nodes and builds edges according to the time adjacent relationship, the graph neural network Hetero-GNN fuses to learn the node features, outputs an abnormal score, and according to the abnormal score, real-time early warning and hierarchical response are carried out, and a rolling window and a confidence adjustment strategy are used to dynamically optimize the early warning trigger threshold.

[0027] The improved beneficial effects are: two-dimensional verification and spatiotemporal correlation, the voice emotion recognition model extracts a 128-dimensional emotion vector to capture abnormal customer emotions, the OCR confidence and the two-dimensional code verification cross-verify the invoice image amount / time information, combined with the spatiotemporal correlation graph, the voice segment and the invoice image are built according to the time adjacent relationship, and the Hetero-GNN fusion learning is output. Abnormal score and real-time early warning solve the missed judgment problems of “emotional abnormalities not captured” and “invoice information forgery” in traditional risk identification.

[0028] Through the rolling window and the confidence adjustment strategy, the early warning trigger threshold is dynamically optimized, the risk characteristics of different time periods and scenes are adapted, the false positive rate and the false negative rate are reduced, and the precision and real-time of risk closed-loop management are realized.

[0029] Preferably, as an improvement, the system uses a cloud-native technology base as an infrastructure layer to provide K8s containerized deployment, service mesh, and security protection support. The construction of the middle platform includes a data middle platform based on a Flink real-time data warehouse, a COS object storage, and an Iceberg lake warehouse integrated architecture; a business middle platform constructs a standardized service capability pool; and an AI middle platform integrates LLM / NLP / ASR intelligent algorithms. The covered business domains include the core business processes of the integration domain, the intelligent processing domain, the risk control domain, and the knowledge collaboration domain. The multiple scene micro-applications include scheduling, quality inspection, assets, network disk, and Tian Gong value lightweight business applications.

[0030] The improved beneficial effects are: the cloud-native elastic support K8s containerized deployment and service mesh support dynamic scaling and fault isolation to ensure business continuity; the data middle platform, Flink real-time data warehouse+COS object storage+Iceberg lake warehouse integration, realizes real-time data processing and lake warehouse fusion, solves the problems of “non-uniform data caliber” and “insufficient real-time performance”; the business middle platform constructs a standardized service capability pool, the AI middle platform integrates LLM / NLP / ASR algorithms, and supports the rapid iteration and capability reuse of scheduling, quality inspection, and other micro-applications. The scheduling, quality inspection, asset, network disk, and Tian Gong value micro-applications are lightweight designed, which is convenient for independent deployment and iteration, and improves the business response speed and system maintainability.

[0031] Preferably, as an improvement, the technical base layered architecture is divided into six layers from top to bottom. Unified portal layer: provides PC and mobile applet dual entry, realizes unified user authentication and access control; Collaboration layer: integrates multi-channel communication capabilities to build a global collaboration network; Business layer: four business domains correspond to Spring Cloud microservice clusters to support specific business logic implementation; Capability layer: AI middle platform provides intelligent algorithm capabilities such as speech recognition ASR and natural language processing NLP, and rule engine supports complex business rule configuration; Data middle platform layer: based on Flink real-time computing engine, COS object storage and Iceberg lake warehouse integrated architecture, build real-time data processing pipeline; Technical base layer: includes K8s container orchestration, Nacos service registration and configuration, SpringGW API gateway, Prometheus monitoring + Grafana visualization + Loki log management observable system, as well as OAuth2 authentication, SM encryption, and zero trust network security system.

[0032] The improved beneficial effects are: unified portal layer (improve user access convenience; collaboration layer integrates multi-channel communication capabilities to build a global collaboration network; business layer four business domains correspond to Spring Cloud microservice clusters to support business logic implementation; capability layer provides AI algorithm and rule engine configuration capability to realize capability reuse; data middle platform layer based on Flink real-time computing engine and lake warehouse integrated architecture, build real-time data processing pipeline; technical base layer includes K8s container orchestration, observable system and security system, guarantee system safety and stable operation.

[0033] The observable system realizes integrated management of monitoring, logging and visualization, and the security system covers authentication, encryption and network protection, which improves system security and operation efficiency, and solves the problems of "dispersion of monitoring" and "weakness of security protection" in traditional architecture. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 The system structure diagram of the embodiment of the present application.

[0035] Figure 2 The frame structure diagram of the embodiment of the present application. DETAILED DESCRIPTION

[0036] The following will be further described in detail through specific embodiments: EMBODIMENT Basically as shown in the attached Figure 1 Full-stack digital tax and fee management service operation system, including: Data Collection and Integration: Real-time access and automatic integration of multi-source data, including electronic tax hotline, ERP / CRM business data, enterprise collaboration platform of WeChat / WeChat Enjoy; through Kafka asynchronous high-concurrency collection, ETL tool cleaning and conversion and NLP key information extraction, realize the unified storage of scattered data to data warehouse; based on rule engine Drools to define data quality rules, trigger abnormal alarm and record log Elasticsearch, replace manual data entry, improve integration efficiency and reduce error rate.

[0037] Intelligent Business Processing: Automatic and intelligent processing of core business, including scheduling, quality inspection and asset management. The scheduling module is based on LSTM model to predict the hotline volume and generate the optimal scheduling scheme combined with employee skill labels, dynamically adjusting the number of scheduling; the quality inspection module realizes automatic quality inspection of recording files and generates reports through ASR transcription and large model semantic analysis; the asset management module realizes online management of asset life cycle and automatic generation of tax data through RFID positioning and tax linkage module, ensuring the consistency of accounts and reality.

[0038] Risk Control: For tax risk "early warning, intervention, analysis" closed-loop management. Through Z-Score algorithm to identify intelligent voice shunt rate, call accuracy and other index outliers, trigger three-level warning, warning methods include SMS+phone in emergency state, enterprise WeChat in general state, support work order assignment to responsible department and track processing progress, confirm risk→analyze reason→develop measures→complete processing, processing results are automatically archived to generate risk report.

[0039] Knowledge and Collaboration: For team professional capacity improvement and cross-department collaboration efficiency optimization. AI Q&A module interfaces with the State Administration of Taxation policy library, extracts policy core points through NLP technology and associates enterprise business data to support intelligent Q&A; intelligent coaching module automatically captures abnormal handling cases to generate standardized documents, employees improve their response ability through simulated exercises; network disk management supports file classification storage, permission control and cross-region sharing; Tian Gong value management converts employee operation volume into points and associates performance, and stimulates value creation through visual board.

[0040] Specifically, The data collection and integration module includes multi-source data access architecture and automatic integration and abnormal alarm; Multi-source data access architecture, through API interface real-time capture of electronic tax hotline call records, customer problem classification, processing time, etc. Structured data, asynchronous high-concurrency collection is realized by using Kafka message queue; From the ERP / CRM system, extract sales order, contract information, invoice data through ETL tools such as Apache NiFi, and store them in the data warehouse after cleaning and conversion; Through OAuth2.0 authorization to obtain employee chat records, knowledge base documents, and training records of enterprise WeChat / Tencent Enjoy platform, and use NLP technology to extract key information.

[0041] Automatic integration and exception alarm, based on Spark to build a distributed processing framework, associate hotline data, scheduling data, and accompanying training data according to business scenarios, such as matching "hotline data-scheduling data" through timestamp, generate customer portrait, service efficiency indicators, etc. Integration results are stored in the Hive data warehouse; Use rule engine such as Drools to define data quality rules, when data collection fails or integration is abnormal, push the message to the operation and maintenance personnel through the enterprise WeChat robot, and record the exception log to Elasticsearch, support query by time, module dimension.

[0042] The intelligent business processing module includes a layout module, a quality inspection module, and an asset management module. The scheduling module is based on LSTM time series model to analyze historical hotline data, predict hotline volume in each period in the next 7 days, and generate optimal scheduling scheme combined with employee skill labels, support manual adjustment and automatic approval; Real-time monitoring of hotline connection rate, if lower than threshold (such as 90%), trigger the system to automatically increase the number of scheduling personnel, and notify the on-duty personnel through SMS / enterprise WeChat.

[0043] On this basis, the layout module further includes: A heterogeneous graph G=(V,E) containing 5 types of core nodes is constructed, V is the vertex set of the heterogeneous graph, E is the edge set, the vertex V represents different modal entities, such as various core nodes, and the edge E represents the association relationship between entities, such as the causal relationship between various core nodes.

[0044] The core nodes include time nodes, weather event nodes, public opinion event nodes, invoice collection period nodes, and employee function nodes. The time node is a specific time period, such as 9:00-10:00; The weather event node refers to weather labels such as heavy rain / high temperature / typhoon; The public opinion event node refers to hot search on microblog / policy interpretation / sudden news, etc. Text vector; The invoice collection period node refers to the monthly / quarterly declaration cycle mark; The employee skill node refers to the familiarity with tax policy / flueness in system operation.

[0045] The attention mechanism is used to dynamically calculate the association strength between nodes, and the adjacency matrix H of the Graph Transformer is used to learn the implicit relationship, the formula is: MultiModal-GT(H, X) = softmax(QK^T / √d)V where X is the multi-modal input matrix, including traffic sequence, weather one-hot encoding, public opinion TF-IDF vector, and tax period flag. Q / K / V are query / key / value matrices, respectively.

[0046] softmax(QK^T / √d) is a calculation of attention score matrix, and √d is a scaling factor. The dimension d is used to avoid gradient explosion caused by excessive dot product. After softmax normalization, the score represents the contribution weight of each modality to the target.

[0047] H is an adjacency matrix that adjusts the attention weight as prior knowledge. For example, in the "heavy rain-invoice loss" scenario, H strengthens the correlation strength between weather events and public opinion events.

[0048] Historical traffic data for 3 years, real-time weather data from the meteorological bureau API, and public opinion data from microblog and Zhihu are collected. NLP is used to extract the public opinion event feature vector.

[0049] The mean absolute percentage error (MAPE) is used as the loss function to complete the gray test of a certain area within a predetermined period.

[0050] The API input includes the current period, real-time weather data, the latest public opinion heat value, and the tax period flag. The output is the traffic volume prediction value and confidence interval for the next 15 minutes.

[0051] The employee skill label is combined, such as senior agents prioritizing complex consultations. Genetic algorithm is used to optimize shift allocation.

[0052] For example, In the scenario of a surge in invoice loss consultations during heavy rain, a sudden heavy rain in a certain area leads to an increase in the physical damage rate of enterprise invoices, and 12366 hotline receives a large number of consultations on "how to handle invoice loss."

[0053] The weather node triggers the "heavy rain" label, and the public opinion node detects the hot search #Chongqing heavy rain leads to invoice loss# on microblog. The tax period node identifies the current month as the quarterly tax period. The traffic data real-time feedback shows that the consultation volume in the current period (14:00-14:15) has surged to 200 calls (historical average 120 calls).

[0054] The Graph Transformer captures the correlation weight of "heavy rain-invoice loss-tax period" through the attention mechanism and predicts that the traffic volume will reach 250 calls in the next 15 minutes.

[0055] MAPE verification shows that the prediction error is only 6.8%, and the recall rate of the sudden peak is 92%.

[0056] System automatically triggers the recruitment process, informs the idle senior agent through SMS / WeChat, and adds the skill label "invoice processing expert" for temporary overtime.

[0057] After the scheduling plan is automatically approved, the shift adjustment is completed within 10 minutes to ensure high connection rate Through the "traffic-weather-public opinion" multi-modal Graph Transformer, the system breaks through the limitations of traditional LSTM relying only on historical traffic, realizes dynamic response to external sudden factors, greatly improves the scheduling accuracy and peak handling capacity, and ensures the continuity and efficiency of tax service.

[0058] The quality inspection module converts the recording to text using ASR technology, performs semantic analysis using a large model (such as ERNIE X1), scores according to pre-set rules, generates a quality inspection report and pushes it to the inspected personnel, and supports online appeal and re-inspection processes.

[0059] On this basis, the quality inspection module also includes: Adopting a dual-channel game architecture, including a large model path and a rule engine path; The rule engine channel labels each rule in the rule library with a unique number and a deduction standard. When triggered, after converting the recording to text, the rule engine scans the text to match the rules and generates a baseline result containing the rule number, for example, detecting "can't solve" triggers R001 deduction 20 points.

[0060] The large model channel uses ERNIE X1 to perform deep semantic understanding on the recording text, identify the context, such as "can't solve" followed by "transfer to a special seat" does not deduct points, and "refuse to handle" deducts points. And structured output includes deduction items, semantic evidence, and rule numbers, such as "R001 deduct 20 points, evidence: 00:45-00:52 recording segment".

[0061] And design the loss function: ; Among them, α is the rule recall weight, ensuring 100% coverage of key rules, such as fatal items must trigger rules, usually set to 0.6-0.8. β is the model misjudgment rate weight, controlling the mis-deduction rate, usually set to 0.2-0.4. γ is the dual-channel difference penalty item, triggered when the rule and the large model results are inconsistent, usually set to 0.1-0.3. If the rule recall is insufficient, increase α; if the misjudgment rate is too high, increase β; if the difference between the two channels is large, increase γ. ⊕ is the exclusive or operation, when the results of the two channels are inconsistent, count 1, when they are consistent, count 0, forcing the two outputs to converge.

[0062] Rule Recall: The rule recall rate is the proportion of cases that the rule engine correctly identifies as triggering rules in quality inspection, reflecting the rule coverage capability. The calculation formula is rule_recall = (number of cases correctly triggering rules) / (total number of cases that should trigger rules). It can be achieved by using a gold standard dataset, such as a manually annotated violation case library, and comparing the rule engine output with the manual annotation results.

[0063] Model FPR: The model false positive rate is the proportion of normal cases incorrectly identified as violations by the large model, reflecting the risk of model misjudgment. The calculation formula is model_FPR = (number of normal cases misjudged by the model) / total number of normal cases. It can be achieved by using a normal case test set, such as manually confirmed samples without violations, and calculating the proportion of cases incorrectly marked by the large model.

[0064] Rule⊕Model: The rule-model difference degree is the XOR operation of the rule engine and the large model in quality inspection results, forcing both outputs to be consistent. If the decision results of the two are inconsistent, rule⊕model is 1, and if the decision results of the two are consistent, rule⊕model is 0.

[0065] For example, In a hotline recording, the agent uses forbidden language such as "this problem we cannot solve", which needs to trigger the quality inspection rule R001 to deduct 20 points.

[0066] Rule Engine Channel: Scan the content after converting the recording to text, detect the "cannot solve" keyword, trigger the preset rule R001 to deduct 20 points. Generate baseline results: R001 deduct 20 points, rule number is clear, no context judgment.

[0067] Large Model Channel: ERNIE X1 performs deep semantic understanding on the recording text, identifies the context: if "cannot solve" is followed by "transfer to a special seat", it is determined as a reasonable scenario and no points are deducted. If "cannot solve" is followed by "refuse to handle", it is determined as a violation scenario and points are deducted and R001 is cited. Structured output: R001 deducts 20 points, evidence: 00:45-00:52 recording segment "this problem we cannot solve, you find other departments", rule number R001.

[0068] Compare the results of the two channels, the decision results are consistent, rule⊕model=0 no penalty. The final output quality inspection report includes the deduction item, semantic evidence and rule number; ensure the explanation rate of 100%, effectively reduce the amount of manual review.

[0069] The asset management module realizes online management of the whole life cycle of assets, and realizes asset positioning and state monitoring through RFID technology; the tax linkage module automatically calculates asset depreciation and amortization, generates data required for tax reporting, and ensures consistency between accounts and reality.

[0070] The risk control module includes data analysis and early warning; Data analysis and early warning, based on historical data to build intelligent voice diversion rate, call accuracy rate, service specification rate and other indicators, through Z-Score anomaly detection algorithm to identify outliers, trigger three-level early warning, including emergency: SMS + phone notification, general / prompt: enterprise WeChat message; Support work order assignment to responsible departments, track processing progress, confirm risk → analyze reasons → develop measures → complete processing, processing results are automatically archived and risk reports are generated.

[0071] On this basis, the risk control module also includes: A double-dimensional fusion intelligent risk control system of "voice emotion + image verification" is built.

[0072] Voice emotion accurate capture: the system automatically intercepts call audio segments, such as 30 seconds, analyzes the voice characteristics of the agent or customer through the voice emotion recognition model, and extracts 128-dimensional emotion vectors containing "happy, angry, anxious" and other emotions.

[0073] For example, when the customer says "can't solve the problem", the system can identify a high value of anger emotion and mark it as a potential risk signal; If the agent responds "is coordinating", it may trigger an associated analysis of anxiety emotion.

[0074] Double verification of image end: "OCR recognition + two-dimensional code verification" double verification of invoice image. The OCR engine will evaluate the recognition confidence of each character, such as clarity, error-free, and generate an overall confidence score; At the same time, scan the invoice two-dimensional code and compare it with the tax bureau database to verify its authenticity.

[0075] For example, an invoice with an OCR confidence of 85% but a two-dimensional code verification pass, the system will cross-verify its amount, time, etc. to avoid missing judgment due to single-dimensional error.

[0076] The system constructs a "spatiotemporal correlation graph" of voice segments and invoice images within the same time period: voice segments and invoice images as nodes, connected edges according to time adjacency, forming a dynamic network. Through the graph neural network Hetero-GNN, the node features are fused and learned, and the abnormal probability is comprehensively judged, and the abnormal score is real-time alarmed and graded responded.

[0077] For example, when the angry emotion segment is highly correlated with the high-amount invoice image in time, the system will determine it as a high-risk combination and generate an abnormal score; If the emotion is calm and the invoice verification is correct, it is determined as low risk.

[0078] The system uses a "rolling window + confidence adjustment" strategy to realize intelligent optimization of threshold value.

[0079] In the initial stage, the system sets a basic alarm threshold based on historical data, such as an anomaly score of 0.85.

[0080] During operation, the system real-time statistics the false positive rate of the last 7 days, such as 15% to 4.1% of the target, through the dynamic adjustment of confidence coefficient to narrow or relax the threshold range. The confidence coefficient is used to control the sensitivity of the threshold to historical data.

[0081] For example, if the false positive rate is close to the preset upper limit of 5%, the system will automatically increase the threshold, such as from 0.85 to 0.87, to reduce invalid alarms; if the false negative rate is high, the threshold will be lowered to improve the risk capture ability. This "the more you use, the smarter" mechanism ensures that the system continues to adapt to business changes and maintains a balance between low false positives and high recall.

[0082] The knowledge and collaboration module includes AI Q&A module, intelligent accompanying training module, and network disk and Tian Gong value management. The AI Q&A module connects with the State Administration of Taxation policy library, extracts policy core points, applicable enterprise types, and operation suggestions through NLP technology, and associates with enterprise business data such as the amount of tax reduction and fee reduction that the enterprise can enjoy, supports policy filtering by tax type and business scenario, and provides intelligent Q&A services.

[0083] The intelligent accompanying training module automatically captures abnormal handling cases, such as bank account freezing due to declaration failure, generates standardized documents, and synchronizes them to Tencent Enjoy, employees improve their response ability through simulation training, and the system records the training results and associates them with performance evaluation.

[0084] The network disk and Tian Gong value management support file classification storage, permission control, and cross-region sharing, and realize mobile access through WeChat Enterprise; convert employee operation volume into points, link points with performance, and display points ranking and contribution through visual board to encourage employees to actively improve work value.

[0085] As shown in the accompanying Figure 2 The system architecture adopts a "cloud native + microservice" dual-core driving mode, and builds a layered decoupling system: The core hub layer is built based on the Spring Cloud framework, integrates Nacos as a service registration and discovery and configuration center dual-engine, and realizes three core capabilities of unified scheduling, permission control, and process arrangement. Through service governance components, the system ensures high availability, supports dynamic scaling and fault isolation, and ensures business continuity and scalability.

[0086] The business module layer is vertically split into N independent micro-service modules such as shift management, examination management, credit management, message management, asset management, etc. according to business scenarios, and each module supports flexible selection of technology stacks. Synchronous calls between modules are implemented through Feign or asynchronous message communication is completed through RocketMQ, forming the architecture characteristics of "high cohesion and low coupling", supporting on-demand independent deployment and agile iteration.

[0087] The overall architecture design adopts a layered model: The cloud-native technology base serves as the infrastructure layer, providing K8s containerized deployment, service mesh, and security protection support; 3 major middle platforms: the data middle platform is based on Flink real-time data warehouse, COS object storage, and Iceberg lake warehouse integrated architecture; the business middle platform builds a standardized service capability pool; the AI middle platform integrates intelligent algorithms such as LLM / NLP / ASR; 4 major business domains: integration domain, intelligent processing domain, risk control domain, and knowledge collaboration domain cover core business processes of enterprises; N scene micro-applications: including shift, quality inspection, asset, network disk, and Tian Gong value, etc. lightweight business applications.

[0088] The technology base layered architecture is divided into six layers from top to bottom: Unified portal layer: provides PC and mobile applet dual entry, realizes user unified authentication and access control; Collaboration layer: integrates enterprise WeChat, Tencent Enjoy, SMS, email, and other multi-channel communication capabilities to build a global collaboration network; Business layer: four business domains correspond to Spring Cloud micro-service clusters, supporting specific business logic implementation; Capability layer: AI middle platform provides intelligent algorithm capabilities such as speech recognition ASR and natural language processing NLP, and rule engine supports complex business rule configuration; Data middle platform layer: based on Flink real-time computing engine, COS object storage, and Iceberg lake warehouse integrated architecture, builds real-time data processing pipeline; Technology base layer: includes K8s container orchestration, Nacos service registration and configuration, SpringGW API gateway, Prometheus monitoring + Grafana visualization + Loki log management observability system, as well as OAuth2 authentication, SM encryption, and zero-trust network security system.

[0089] The above-mentioned are only embodiments of the present application, and common technical solutions and / or common knowledge of the scheme are not described in detail. It should be pointed out that, for those skilled in the art, without departing from the technical solutions of the present application, a number of modifications and improvements can be made, which should also be considered as the protection scope of the present application, and these will not affect the effect and practicality of the patent. The protection scope claimed in the present application should be subject to the content of its claims, and the specific implementation mode and the like recorded in the specification can be used to explain the content of the claims.

Claims

1. A full-stack digital intelligent tax and fee management service operation system, characterized in that: include: The data acquisition and integration module is used for real-time access and automatic integration of multi-source data, including data from the electronic tax bureau hotline, ERP / CRM business data, and enterprise collaboration platform data. The intelligent business processing module is used for the automation and intelligent processing of core businesses, including scheduling management, quality inspection management, and asset management. The risk management module is used for closed-loop management of tax and fee risks, including early warning, intervention, and analysis. The Knowledge and Collaboration module is used to enhance team professional capabilities and optimize cross-departmental collaboration efficiency. The cloud-native microservice architecture uses the Spring Cloud framework to build the core central layer, integrating a dual engine for service registration and discovery and a configuration center.

2. The full-stack digital intelligent tax management service operation system according to claim 1, characterized in that: The data acquisition and integration module includes a multi-source data access architecture and automatic integration and anomaly alarm; The multi-source data access architecture captures structured data such as call records, customer issue classifications, and processing times from the e-tax bureau hotline in real time via API interfaces, and uses Kafka message queues to achieve asynchronous high-concurrency data collection; sales orders, contract information, and invoice data are extracted from ERP / CRM systems using ETL tools, cleaned and transformed, and then stored in the data warehouse; employee chat logs, knowledge base documents, and training records from the chat platform are obtained through OAuth2.0 authorization, and NLP technology is used to extract key information. Automatic integration and anomaly alerts are implemented using a distributed processing framework built on Spark. Hotline data, scheduling data, and tutoring data are associated according to business scenarios to generate integrated results of customer profiles and service efficiency indicators, which are then stored in a Hive data warehouse. Data quality rules are defined using a rule engine such as Drools. When data collection fails or integration anomalies occur, messages are pushed to operations and maintenance personnel via WeChat robot, and anomaly logs are recorded in Elasticsearch.

3. The full-stack digital intelligent tax management service operation system according to claim 2, characterized in that, The intelligent business processing module includes: The scheduling module predicts hotline volume for each time period in the next 7 days based on the LSTM time series model, and generates the optimal solution by combining employees' familiarity with tax policies and proficiency in system operation skills. When the connection rate is below 90%, the system automatically adds staff and notifies them via SMS. The quality inspection module uses ASR technology to convert audio recordings into text, performs semantic analysis through a large model, scores according to preset rules, generates a quality inspection report, and pushes it to the inspected personnel. The asset management module uses RFID to locate the status of assets, while the tax linkage module automatically calculates depreciation and generates declaration data.

4. The full-stack digital intelligent tax management service operation system according to claim 3, characterized in that, The risk management module includes: Data analysis and early warning: Based on historical data, intelligent voice diversion rate, call accuracy rate and service standardization rate are constructed. The Z-Score anomaly detection algorithm is used to identify abnormal values ​​and trigger emergency state SMS + telephone three-level early warning. It supports work order assignment to responsible departments, tracking of processing progress, automatic archiving of processing results and generation of risk reports.

5. The full-stack digital intelligent tax management service operation system according to claim 4, characterized in that, The knowledge and collaboration module includes: The AI ​​Q&A module connects to the State Taxation Administration's policy database, and NLP extracts the key points of the application of tax and fee reduction amounts and associates them with enterprise business data. The intelligent tutoring module captures abnormal cases, generates standardized documents, and synchronizes them to the collaboration platform. Employee simulation practice scores are linked to performance evaluations. The cloud storage supports file categorization and cross-regional sharing. Tiangong Value Management converts operation volume into points and displays contribution through a visual dashboard.

6. The full-stack digital intelligent tax management service operation system according to claim 5, characterized in that: The scheduling module employs a multimodal heterogeneous graph neural network, constructing a heterogeneous graph with multiple core nodes. The core nodes include time nodes, weather event nodes, public opinion event nodes, invoice collection period nodes, and employee function nodes. It uses the attention mechanism of Graph Transformer to dynamically calculate the correlation strength between external emergencies and call volume, and outputs the predicted call volume. The scheduling is then adjusted based on the predicted call volume combined with employee skill tags.

7. The full-stack digital intelligent tax management service operation system according to claim 6, characterized in that: The quality inspection module adopts a dual-channel game architecture, including a large model path and a rule engine path; The rule engine channel assigns a unique number and deduction criteria to each rule in the rule base. When triggered, after the audio is transcribed into text, the rule engine scans the text to match the rules and generates a baseline result containing the rule number. The large model channel uses ERNIE X1 to perform deep semantic understanding of the recorded text, identify the context, and output structured prediction results including deduction items, semantic evidence, and rule numbers. The ternary loss function is constructed by including rule recall weights, model misclassification rate weights, and dual-channel difference penalty terms, which are used to make the outputs of the rule engine channel and the large model channel converge.

8. The full-stack digital intelligent tax management service operation system according to claim 7, characterized in that: The risk management module constructs a dual-dimensional fusion system of voice emotion and image verification. The voice emotion recognition model extracts a 128-dimensional emotion vector, and cross-verifies the amount / time information using OCR confidence and QR code verification of invoice images. The spatiotemporal correlation graph uses voice segments and invoice images as nodes and builds edges according to temporal adjacency. The graph neural network Hetero-GNN integrates and learns node features to output anomaly scores. Based on the anomaly scores, real-time warnings are issued and tiered responses are implemented. A rolling window and confidence adjustment strategy are used to dynamically optimize the warning trigger threshold.

9. The full-stack digital intelligent tax management service operation system according to claim 8, characterized in that: The system adopts a cloud-native technology foundation as its infrastructure layer, providing Kubernetes containerized deployment, service mesh, and security protection support; The middle platform is constructed as follows: the data middle platform is based on an integrated architecture of Flink real-time data warehouse, COS object storage and Iceberg lake warehouse; the business middle platform builds a standardized service capability pool; and the AI ​​middle platform integrates LLM / NLP / ASR intelligent algorithms. The business domains covered include the core business processes of the procurement and integration domain, intelligent processing domain, risk control domain, and knowledge collaboration domain; Multiple scenario-based micro-applications include lightweight business applications for scheduling, quality inspection, assets, cloud storage, and daily work points.

10. The full-stack digital intelligent tax management service operation system according to claim 9, characterized in that, The layered architecture of the technology base consists of six layers from top to bottom: Unified portal layer: Provides dual entry points on PC and mobile mini-program to achieve unified user authentication and access control; Collaboration Layer: Integrates multi-channel communication capabilities to build a global collaborative network; Business layer: The four business domains correspond to Spring Cloud microservice clusters, supporting the implementation of specific business logic; Capability Layer: The AI ​​platform provides intelligent algorithm capabilities for Automatic Speech Recognition (ASR) and Natural Language Processing (NLP), while the rule engine supports complex business rule configuration. Data platform layer: Based on the Flink real-time computing engine, COS object storage and Iceberg lake warehouse integrated architecture, a real-time data processing pipeline is built; The technical foundation layer includes an observable system consisting of K8s container orchestration, Nacos service registration configuration, SpringGW API gateway, Prometheus monitoring + Grafana visualization + Loki log management, as well as a security system including OAuth2 authentication, national cryptographic SM encryption, and zero-trust network.

Images