Government Affairs Function Efficiency Optimization Method and System Based on Multimodal Data Fusion

Abstract

The present invention discloses a method and system for optimizing the efficiency of government functions based on multimodal data fusion, which belongs to the field of artificial intelligence and smart service technology. The technical problem to be solved by the present invention is how to optimize government functions from the data level, process level, decision-making level and feedback level, and improve the efficiency of government processing. The technical solution adopted is: data collection: through the government extranet data capture method, data is collected from the government extranet platform, paper documents and third-party databases in real time or in batches to obtain the original data, and Apache NiFi is used to build a data pipeline to connect to the interface of enterprise-friendly services and grassroots governance; data intelligent analysis and processing: in-depth analysis, semantic understanding and regular processing of the collected original data, extraction of structured information and generation of decision support content, so as to realize the transformation from data to knowledge; optimization of process execution through the Activity 7 process engine; closed-loop interactive feedback.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and smart service technology, and specifically to a method and system for optimizing government function efficiency based on multimodal data fusion. Background Art

[0002] Government efficiency refers to the ratio between administrative outcomes achieved and administrative resources (including human, material, financial, and time) consumed by relevant units in fulfilling their responsibilities. It reflects the quality, speed, and effectiveness of government work and is a key indicator for measuring governance capacity and performance. Government efficiency requires not only that administrative activities reach a certain scale in terms of quantity but also that their quality meet established standards and requirements. It focuses not only on the direct outcomes of administrative activities but also on the actual benefits and value these outcomes generate for society.

[0003] Traditional government systems rely on manual Excel export for data integration and require serial approvals, taking an average of 5-7 days. This indicates the following deficiencies in existing government systems:

[0004] ① Serious data silos: Each layer of the system is built independently, the data format is fragmented (Excel / PDF / scanned copies), and cross-departmental calls require manual export and import, which takes up more than 30% of the time.

[0005] ② Redundant and inefficient processes: Traditional OA systems rely on paper-based approvals and serial signatures. The average number of transfer nodes for a single transaction is ≥5, and the overdue completion rate is as high as 25%.

[0006] ③ Decision-making relies on experience: Indicator analysis is mostly based on manual experience, lacking data modeling and simulation tools, and the decision-making error rate fluctuates between 15% and 20%.

[0007] ④ Lack of feedback mechanism: Task execution status is not visible, exception handling relies on manual tracing, and the problem response cycle exceeds 48 hours.

[0008] Therefore, how to optimize government functions from the data level, process level, decision-making level and feedback level and improve the efficiency of government processing is a technical problem that needs to be solved urgently. Summary of the Invention

[0009] The technical task of the present invention is to provide a method and system for optimizing the efficiency of government functions based on multimodal data fusion to solve the problem of how to optimize government functions from the data level, process level, decision-making level and feedback level and improve the efficiency of government processing.

[0010] The technical task of the present invention is achieved in the following manner: a method for optimizing the efficiency of government functions based on multimodal data fusion, the method being specifically as follows:

[0011] Data collection: Data is collected in real time or in batches from government extranet platforms, paper documents, and third-party databases through government extranet data capture. The system then uses Apache NiFi to build data pipelines, connect to interfaces for enterprise services and grassroots governance, and support real-time subscriptions to JSON and / or XML data.

[0012] Intelligent data analysis and processing: Conduct in-depth analysis, semantic understanding, and regularized processing of collected raw data, extract structured information, and generate decision-support content, realizing the transformation from data to knowledge;

[0013] Optimize process execution through Activity 7 process engine;

[0014] Closed-loop interactive feedback: When the Activity 7 process engine detects that a task has timed out for more than the set threshold number of times, it triggers the RPA robot to automatically send a reminder notification and simultaneously back up the notification to the emergency processing queue. Based on the department's monthly completion rate and public satisfaction indicators, the reinforcement learning algorithm adjusts the process priority weight to achieve flexible resource allocation.

[0015] As a preference, during the data collection process, for unstructured data, OCR recognition is used and then the key fields are extracted through the Hairuo model; among them, the key fields include document number, issuing unit and validity period.

[0016] Preferably, data collection also includes cross-domain crawler enhancement, as follows;

[0017] Establish a dynamic crawler strategy library, preset XPath parsing rules for different website structures, and store the XPath configurations of different websites in the dynamic crawler strategy library in JSON format;

[0018] Interactively debug XPath using Scrapy's Scrapy Shell or third-party tools such as XPath Helper;

[0019] When the crawler starts, it loads the corresponding XPath configuration from Redis or MySQL based on the current domain name. After obtaining the corresponding XPath configuration data, it filters out noise interference and cleans the XPath extraction results using regular expressions to remove irrelevant characters (such as \n and \t).

[0020] More preferably, the corresponding XPath configuration is loaded from Redis or MySQL based on the current domain name as follows:

[0021] Use the re.search() function to search for the first position that matches the specified regular expression pattern in the string and return a match object;

[0022] Use the group() method to extract the matching string from the match object: if a matching date format is found, the corresponding date string will be returned.

[0023] As a preferred option, the data intelligent analysis and processing is as follows:

[0024] For government documents, tags and summaries are automatically generated through word segmentation analysis;

[0025] For form data, based on the Hainuo big model, trend forecast reports are generated for economic operations and key project data, and timeout risks are automatically marked. SQL aggregation rules are preset, combined with MyBatis dynamic queries, and indicators of each project are calculated in real time and output to the Elasticsearch index library.

[0026] As a preferred method, the process execution is optimized through the Activity 7 process engine as follows:

[0027] Process modeling: First, design a BPMN flowchart using Camunda Modeler, define task nodes and gateway conditions (parallel / exclusive branches), then deploy BPMN using the Activity 7 process engine. Finally, design a flowchart using the process editor, which automatically parses the flowchart and generates process instances. Dynamic routing (for example, redirecting timed tasks to the supervision node) is implemented using the conditionExpression of the Exclusive Gateway. Timeout warning rules are built in, and timer events are embedded in BPMN process nodes. Business system data is automatically captured daily through the Spring Boot scheduler. When the timer expires, the Activity 7 process engine's ExecutionListener interface is used to call the RPA robot to send a reminder message, synchronously updating the task status to the emergency queue and pushing it to the monitoring dashboard via RocketMQ.

[0028] Intelligent dispatch mechanism: Based on the Hainuo big model, a task-personnel matching model trained with historical data was developed. This model was constructed by extracting task attributes (business area, urgency) and personnel attributes (processing time, professional qualifications, and historical success rate). The task-personnel matching model was trained by daily incremental updates of historical data, incorporating the latest task data for optimization and fine-tuning, resulting in a 40% increase in dispatch accuracy.

[0029] Process monitoring dashboard: displays process throughput, average processing time, and backlog task heat map in real time, and supports drilling down to individual task details (handling records, attachment traces).

[0030] Preferably, the pre-training of the historical data training task-personnel matching model is as follows:

[0031] The collected raw data is cleaned and processed: first, web page tags, special characters, and incorrectly encoded noise data are filtered out; then, a deduplication and clustering component is built, supporting LSH features and dense embedding features, to cluster and deduplicate trillions of data within a set time; then, data length is standardized, discarding text that does not reach the set minimum threshold length and segmenting text that exceeds the set maximum threshold length; finally, content compliance is checked and filtered through both technical and manual review to remove corpus with content security risks;

[0032] Knowledge base construction: Based on document parsing and semantic vectorization technology, we vectorize the content of PDF, Word, TXT, and HTML documents. We use a deep neural network model (Sentence-BERT) to convert knowledge into high-dimensional vectors and store them in a dedicated vector database (Pinecone).

[0033] Prompt engineering optimization: Using structured prompt templates and dynamic example injection technology, combined with the Chain-of-Thought to guide the model's step-by-step reasoning, and optimize parameters to generate results;

[0034] Fine-tuning the historical data training task-personnel matching model: Fine-tuning the historical data training task-personnel matching model based on application requirements is performed through supervised fine-tuning, reward models, and low-rank adaptation (LoRA).

[0035] Fusion of retrieval enhancement technologies: Integrates retrieval enhancement generation (RAG) technology to obtain precise knowledge fragments through semantic and keyword hybrid search engines, and combines cross-encoder and reranking to optimize context quality and improve answer accuracy.

[0036] A government function efficiency optimization system based on multimodal data fusion, which is used to implement the above-mentioned government function efficiency optimization method based on multimodal data fusion; the system uses the Spring Cloud microservice architecture and combines it with Alibaba Nacos for configuration management and service discovery; the system includes a front-end and a back-end;

[0037] The front-end uses the Uniapp framework to develop mobile apps and web applications, and sends requests to obtain back-end data. The latest dynamic data is displayed on the page, and the component library is used to convert the data of central work and key projects into charts for display.

[0038] The backend uses Spring Boot to build the project framework, integrates MyBatis to operate the database, and creates database tables to store data resources for economic operations and urban management; and writes Mapper interfaces and XML files to implement data addition, deletion, modification and query.

[0039] Preferably, the backend includes:

[0040] The data collection layer is used to collect data in real time or in batches from the government extranet platform, paper documents, and third-party databases through the government extranet data capture method, obtain raw data, and use Apache NiFi to build data pipelines to connect to the interfaces of enterprise services and grassroots governance, supporting real-time subscription of JSON and / or XML data;

[0041] The intelligent processing layer is used to conduct in-depth analysis, semantic understanding, and regular processing of the collected raw data, extract structured information, and generate decision-support content, realizing the transformation from data to knowledge;

[0042] The process execution layer is used to optimize process execution through the Activity 7 process engine;

[0043] The interactive feedback layer is used to trigger the RPA robot to automatically send a reminder notification and synchronize it to the emergency processing queue when the Activity 7 process engine detects that the task timeout exceeds the set threshold number of times. It is then used to adjust the process priority weight through the reinforcement learning algorithm based on the department's monthly completion rate and public satisfaction indicators to achieve flexible resource allocation.

[0044] Preferably, the process execution layer includes:

[0045] The process modeling module is used to design BPMN flowcharts using Camunda Modeler, define task nodes and gateway conditions (parallel / exclusive branches), deploy BPMN using the Activity 7 process engine, and design flowcharts using the process editor. The module automatically parses the flowchart and generates process instances. Dynamic routing is implemented using the conditionExpression of the exclusive gateway. Timeout warning rules are built in, and timer events are embedded in BPMN process nodes. Business system data is automatically captured daily through the Spring Boot scheduler. When the timer expires, the Activity 7 process engine's ExecutionListener interface is used to call the RPA robot to send a reminder message, synchronously updating the task status to the emergency queue and pushing it to the monitoring dashboard via RocketMQ.

[0046] The intelligent dispatch module is used to develop a task-personnel matching model based on historical data training based on the Hainuo big model. This model is constructed by extracting task attributes (business area, urgency) and personnel attributes (processing time, professional qualifications, and historical success rate). The model is trained on historical data and fine-tuned by incorporating the latest task data, improving dispatch accuracy by 40%.

[0047] The process monitoring dashboard module is used to display process throughput, average processing time and backlog task heat map in real time, and supports drilling into individual task details (handling records and attachment traces).

[0048] The method and system for optimizing government function efficiency based on multimodal data fusion of the present invention have the following advantages:

[0049] (1) This invention builds a four-layer system of "data perception - intelligent analysis - process drive - closed-loop feedback". At the data level, it connects multimodal data sources such as government external networks and business systems to build a global data lake, supporting real-time query and correlation analysis. At the process level, it standardizes approval, supervision, and joint investigation processes through the Activity 7 engine, reducing manual intervention by ≥50%. At the decision-making level, it uses large models to generate interpretation summaries, risk warning reports, and optimization suggestions, increasing decision-making efficiency by more than 3 times. At the feedback level, it establishes an adaptive optimization mechanism based on KPI weights and timeliness scores to achieve self-healing of process anomalies and dynamic optimization of resource allocation.

[0050] (2) This invention uses API interfaces, RPA robots, and targeted crawler technology to capture structured / unstructured data from government extranet platforms at all levels in real time, and combines NLP algorithms to achieve automatic classification, labeled storage, and trend prediction, thereby realizing automatic classification and analysis of government extranet data;

[0051] (III) The present invention is compared with the traditional government affairs system as shown in Table 1 below:

[0052] Table 1 Comparison of advantages and disadvantages between the present invention and traditional government affairs systems

[0053]

[0054] (4) This invention improves efficiency: the time required for cross-departmental data transfer is reduced by 90%, and the efficiency of document circulation is increased by 300%;

[0055] (5) The present invention can save costs: reduce repetitive manpower input by more than 50%, saving expenses;

[0056] (6) Invention governance upgrade: Supporting the implementation of innovative scenarios such as "one network for unified management" and "instant approval and instant processing", the public satisfaction rate has increased to more than 95%. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] The present invention will be further described below with reference to the accompanying drawings.

[0058] Attachment Figure 1 A flowchart of the method for optimizing the efficiency of government functions based on multimodal data fusion;

[0059] Attachment Figure 2 This is a structural diagram of the government function efficiency optimization system based on multimodal data fusion. DETAILED DESCRIPTION

[0060] The method and system for optimizing the efficiency of government functions based on multimodal data fusion of the present invention are described in detail below with reference to the accompanying drawings and specific embodiments of the specification.

[0061] Example 1: As shown in the attached Figure 1 As shown, this embodiment provides a method for optimizing government function efficiency based on multimodal data fusion, and the method is specifically as follows:

[0062] S1. Data Collection: Collect data from the government extranet platform, paper documents, and third-party databases in real time or in batches through government extranet data capture to obtain raw data. Build data pipelines using Apache NiFi to connect to interfaces for enterprise services and grassroots governance, supporting real-time subscriptions to JSON and / or XML data.

[0063] S2. Intelligent Data Analysis and Processing: Conduct in-depth analysis, semantic understanding, and regularized processing of collected raw data to extract structured information and generate decision-support content, achieving the transformation from data to knowledge.

[0064] S3, optimize process execution through Activity 7 process engine;

[0065] S4. Closed-loop interactive feedback: When the Activity 7 process engine detects that a task has timed out for more than the set threshold number of times, it triggers the RPA robot to automatically send a reminder notification and synchronize the backup to the emergency processing queue; then, based on the department's monthly completion rate and public satisfaction indicators, the reinforcement learning algorithm adjusts the process priority weight to achieve flexible resource allocation.

[0066] During the data collection process in step S1 of this embodiment, for unstructured data (such as scanned copies of official documents), OCR recognition is performed and then key fields are extracted through the Hairuo model; among them, the key fields include document number, issuing unit and validity period.

[0067] The data collection in step S1 of this embodiment also includes cross-domain crawler enhancement, which is as follows:

[0068] S101. Establish a dynamic crawler strategy library, preset XPath parsing rules for different website structures, and store the XPath configurations of different websites in the dynamic crawler strategy library in JSON format;

[0069] S102. Use Scrapy's Scrapy Shell or third-party tools (such as XPath Helper) to interactively debug XPath; the key code is as follows:

[0070] {

[0071] "XXXXXX.XXXX.cn": {

[0072] "title":" / / h1[@class'policy-title'] / text(),

[0073] "date":" / / span[@id='publish-date'] / text()".

[0074] "content":" / / div[contains(eclass,'main-content")] / / p / text()"

[0075] }

[0076] "XXXX.XX.cn":{

[0077] "title": / / div[eclass.'header'] / h2 / text()".

[0078] "date":"( / / div[eclass='info" ] / span)[1] / text()",

[0079] content":" / / article[eid.'content'] / p / text()"

[0080] }

[0081] };

[0082] The above code is a JSON object that defines rules for extracting webpage data from two different websites. Each website object contains three key-value pairs, one for extracting the webpage title, the publication date, and the body content.

[0083] S103. When the crawler starts, the corresponding XPath configuration is loaded from Redis or MySQL according to the current domain name. After obtaining the corresponding XPath configuration data, noise interference is filtered out. The XPath extraction results are cleaned using regular expressions to remove irrelevant characters (such as \n and \t).

[0084] In step S103 of this embodiment, the corresponding XPath configuration is loaded from Redis or MySQL according to the current domain name as follows:

[0085] S10301. Use the re.search() function to search for the first position in the string that matches the specified regular expression pattern and return a match object.

[0086] S10302. Use the group() method to extract the matching string from the matching object: if a matching date format is found, the corresponding date string will be returned.

[0087] The key codes are as follows:

[0088] raw_date=response.xpath(' / / span[eclass="date" ] / text()').get();

[0089] clean_date =re.search(r'\d{4}-\d{2}-\d{2}', raw_date).group();

[0090] The function of the above code is to extract the first date string that conforms to the YYYY-MM-DD format from the raw_date string and assign it to the variable clean_date.

[0091] The data intelligent analysis and processing in step S2 of this embodiment is as follows:

[0092] For government documents, tags and summaries are automatically generated through word segmentation analysis;

[0093] For form data, based on the Hainuo big model, trend forecast reports are generated for economic operations and key project data, and timeout risks are automatically marked. SQL aggregation rules are preset (such as querying tasks that have timed out for 48 hours and the completion rate of key tasks). Combined with MyBatis dynamic query, various project indicators (such as the proportion of overdue tasks and the completion rate of key items) are calculated in real time and output to the Elasticsearch index library.

[0094] The process execution optimization by the Activity 7 process engine in step S3 of this embodiment is specifically as follows:

[0095] S301, Process Modeling: First, design a BPMN flowchart using Camunda Modeler, define task nodes and gateway conditions (parallel / exclusive branches), then deploy BPMN using the Activity 7 process engine. Finally, design a flowchart using the process editor, which automatically parses the flowchart and generates a process instance. Finally, implement dynamic routing (e.g., redirecting a timed-out task to the supervisor node) using the conditionExpression of the exclusive gateway. The key code is as follows:

[0096] <sequenceflow id="node_1722329536345" name="通过" sourceref="node_1722320397824" targetref="node_1722329536344">

[0097] <conditionexpression xsi:type="tFormalExpression"><![CDATA[${!isRejected}]]>< / conditionexpression>

[0098] < / sequenceflow>

[0099] <sequenceflow id="node_1722566736121" name="驳回" sourceref="node_1722320397824" targetref="node_1722566726388">

[0100] <extensionelements>

[0101] <activiti:executionListener

[0102] event="take" delegateExpression="${rejectListener}" / >

[0103] < / extensionelements>

[0104] <conditionexpression xsi:type="tFormalExpression"><![CDATA[${isRejected}]]>< / conditionexpression>

[0105] < / sequenceflow> ;

[0106] In the above code, the first sequenceFlow is activated when isRejected is false, indicating that the process flows from one node to another. The second sequenceFlow is activated when isRejected is true and triggers an execution listener, rejectListener, when the path is activated. These definitions are used to control the decision points and path selection in the business process, ensuring that the process flows correctly according to different conditions.

[0107] It also includes built-in timeout warning rules and embeds timer events in BPMN process nodes (e.g., "Automatically escalate to a higher-level supervisor if not processed within 48 hours"). Business system data is automatically captured daily through the Spring Boot scheduler. When the timer expires, the Activity 7 process engine's ExecutionListener interface is used to call the RPA robot to send a reminder message, synchronously updating the task status to the emergency queue and pushing it to the monitoring dashboard via RocketMQ. The key code is as follows:

[0108] <boundaryevent id="timeout-event" attachedtoref="task-node">

[0109] <timereventdefinition>

[0110] <timeduration> PT48H< / timeduration> <!-- Trigger after 48 hours -->

[0111] < / timereventdefinition>

[0112] < / boundaryevent> ;

[0113] S302, Intelligent Dispatch Mechanism: Develop a task-personnel matching model trained on historical data based on the Hainuo Big Model. This model is constructed by extracting task attributes (business area, urgency) and personnel attributes (processing time, professional qualifications, and historical success rate). This model is then trained on historical data and fine-tuned by incorporating the latest task data, resulting in a 40% improvement in dispatch accuracy.

[0114] S303, Process Monitoring Dashboard: Displays process throughput, average processing time, and backlog task heat map in real time, and supports drilling down to individual task details (handling records, attachment traces).

[0115] The historical data training task in step S302 of this embodiment - pre-training of the personnel matching model is specifically as follows:

[0116] S30201. Clean and process the collected raw data: First, filter out web page tags, special characters, and incorrectly encoded noise data; then build a deduplication and clustering component that supports LSH features and dense embedding features to cluster and deduplicate trillions of data within a set time; then standardize the data length, discarding text that does not reach the set minimum threshold length and segmenting text that exceeds the set maximum threshold length; finally, perform content compliance detection and filtering through both technical and manual review to remove corpus with content security risks;

[0117] S30202, Knowledge Base Construction: We vectorize the content of PDF, Word, TXT, and HTML documents using document parsing and semantic vectorization technologies. We convert knowledge into high-dimensional vectors using a deep neural network model (Sentence-BERT) and store them in a dedicated vector database (Pinecone).

[0118] S30303, prompt engineering optimization: Using structured prompt templates and dynamic example injection technology, combined with the chain of thought to guide the model's step-by-step reasoning, and optimize parameters to generate results;

[0119] S30304, Fine-tuning the Historical Data Training Task-Personnel Matching Model: Fine-tune the Historical Data Training Task-Personnel Matching Model based on application requirements through supervised fine-tuning, reward models, and low-rank adaptation (LoRA).

[0120] S30305. Retrieval enhancement technology integration: Integrate Retrieval Enhancement Generation (RAG) technology to obtain accurate knowledge fragments through semantic and keyword hybrid search engines, combine cross-encoder and reranking to optimize context quality and improve answer accuracy.

[0121] Example 2: This example provides a system for optimizing the efficiency of government functions based on multimodal data fusion. The system is used to implement the method for optimizing the efficiency of government functions based on multimodal data fusion as described in Example 1. The system uses the Spring Cloud microservice architecture and combines it with Alibaba Nacos for configuration management and service discovery. The system includes a front-end and a back-end.

[0122] The front-end uses the Uniapp framework to develop mobile apps and web applications, and sends requests to obtain back-end data. The latest dynamic data is displayed on the page, and the component library is used to convert the data of central work and key projects into charts for display.

[0123] The backend uses Spring Boot to build the project framework, integrates MyBatis to operate the database, and creates database tables to store data resources for economic operations and urban management; and writes Mapper interfaces and XML files to implement data addition, deletion, modification and query.

[0124] As attached Figure 2 As shown, the backend in this embodiment includes:

[0125] The data collection layer is used to collect data in real time or in batches from the government extranet platform, paper documents, and third-party databases through the government extranet data capture method, obtain raw data, and use Apache NiFi to build data pipelines to connect to the interfaces of enterprise services and grassroots governance, supporting real-time subscription of JSON and / or XML data;

[0126] The intelligent processing layer is used to conduct in-depth analysis, semantic understanding, and regular processing of the collected raw data, extract structured information, and generate decision-support content, realizing the transformation from data to knowledge;

[0127] The process execution layer is used to optimize process execution through the Activity 7 process engine;

[0128] The interactive feedback layer is used to trigger the RPA robot to automatically send a reminder notification and synchronize it to the emergency processing queue when the Activity 7 process engine detects that the task timeout exceeds the set threshold number of times. It is then used to adjust the process priority weight through the reinforcement learning algorithm based on the department's monthly completion rate and public satisfaction indicators to achieve flexible resource allocation.

[0129] The process execution layer in this embodiment includes:

[0130] The process modeling module is used to design BPMN flowcharts using Camunda Modeler, define task nodes and gateway conditions (parallel / exclusive branches), deploy BPMN using the Activity 7 process engine, and design flowcharts using the process editor. The module automatically parses the flowchart and generates process instances. Dynamic routing (for example, redirecting timed tasks to the supervision node) is implemented through the conditionExpression of the Exclusive Gateway. Timeout warning rules are built in, and timer events are embedded in BPMN process nodes (for example, "automatically escalate to a higher-level supervision task if not processed within 48 hours"). Business system data is automatically captured daily through the Spring Boot scheduler. When the timer expires, the Activity 7 process engine's ExecutionListener interface is used to call the RPA robot to send a reminder message, synchronously updating the task status to the emergency queue and pushing it to the monitoring dashboard via RocketMQ.

[0131] The intelligent dispatch module is used to develop a task-personnel matching model based on historical data training based on the Hainuo big model. This model is constructed by extracting task attributes (business area, urgency) and personnel attributes (processing time, professional qualifications, and historical success rate). The model is trained on historical data and fine-tuned by incorporating the latest task data, improving dispatch accuracy by 40%.

[0132] The process monitoring dashboard module is used to display process throughput, average processing time and backlog task heat map in real time, and supports drilling into individual task details (handling records and attachment traces).

[0133] Finally, it should be noted that the above 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 above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for optimizing government function efficiency based on multimodal data fusion, characterized in that: The method is as follows: Data collection: Data is collected in real time or in batches from government extranet platforms, paper documents, and third-party databases through government extranet data capture. The system then uses Apache NiFi to build data pipelines, connect to interfaces for enterprise services and grassroots governance, and support real-time subscriptions to JSON and / or XML data. Intelligent data analysis and processing: Conduct in-depth analysis, semantic understanding, and regularized processing of collected raw data, extract structured information, and generate decision-support content, realizing the transformation from data to knowledge; Optimize process execution through Activity 7 process engine; Closed-loop interactive feedback: When the Activity 7 process engine detects that a task has timed out for more than a set threshold, it triggers the RPA robot to automatically send a reminder and simultaneously back up the notification to the emergency processing queue. Based on the department's monthly completion rate and customer satisfaction indicators, the reinforcement learning algorithm adjusts the process priority weights to achieve flexible resource allocation. The specific steps for optimizing process execution using the Activity 7 process engine are as follows: Process modeling: First, design a BPMN flowchart using Camunda Modeler, define task nodes and gateway conditions, then deploy BPMN using the Activity 7 process engine. Finally, design the flowchart using the process editor, which automatically parses the flowchart and generates process instances. Dynamic routing is then implemented using the conditionExpression of the exclusive gateway. Built-in timeout warning rules are implemented, and timer events are embedded in BPMN process nodes. Business system data is automatically captured daily through the Spring Boot scheduler. When the timer expires, the Activity 7 process engine's ExecutionListener interface is used to call the RPA robot to send a reminder message, synchronously updating the task status to the emergency queue and pushing it to the monitoring dashboard via RocketMQ. Intelligent dispatch mechanism: Based on the Hainuo big model, a historical data training task-personnel matching model is developed. Data is constructed by extracting task and person attributes. The historical data training task-personnel matching model is updated daily incrementally, incorporating the latest task data for optimization and fine-tuning. Process monitoring dashboard: displays process throughput, average processing time, and backlog task heat map in real time, and supports drilling down to individual task details.

2. The method for optimizing government function efficiency based on multimodal data fusion according to claim 1 is characterized in that: During the data collection process, for unstructured data, OCR recognition is used and then the key fields are extracted through the Hairuo model; among them, the key fields include document number, issuing unit and validity period.

3. The method for optimizing government function efficiency based on multimodal data fusion according to claim 1 or 2, characterized in that: Data collection also includes cross-domain crawler enhancements, as follows; Establish a dynamic crawler strategy library, preset XPath parsing rules for different website structures, and store the XPath configurations of different websites in the dynamic crawler strategy library in JSON format; Use Scrapy's Scrapy Shell or third-party tools to interactively debug XPath; When the crawler starts, it loads the corresponding XPath configuration from Redis or MySQL according to the current domain name. After obtaining the corresponding XPath configuration data, it performs noise interference filtering and cleans the XPath extraction results through regular expressions to remove irrelevant characters.

4. The method for optimizing government function efficiency based on multimodal data fusion according to claim 3 is characterized in that: Load the corresponding XPath configuration from Redis or MySQL based on the current domain name as follows: Use the re.search() function to search for the first position that matches the specified regular expression pattern in the string and return a match object; Use the group() method to extract the matching string from the match object: if a matching date format is found, the corresponding date string will be returned.

5. The method for optimizing government function efficiency based on multimodal data fusion according to claim 1 is characterized in that: The data intelligent analysis and processing are as follows: For government documents, tags and summaries are automatically generated through word segmentation analysis; For form data, based on the Hainuo big model, trend forecast reports are generated for economic operations and key project data, and timeout risks are automatically marked. SQL aggregation rules are preset, combined with MyBatis dynamic queries, and indicators of each project are calculated in real time and output to the Elasticsearch index library.

6. The method for optimizing government function efficiency based on multimodal data fusion according to claim 1 is characterized in that: The historical data training task - pre-training of the personnel matching model is as follows: The collected raw data is cleaned and processed: first, web page tags, special characters, and incorrectly encoded noise data are filtered out; then, a deduplication and clustering component is built, supporting LSH features and dense embedding features, to cluster and deduplicate trillions of data within a set time; then, data length is standardized, discarding text that does not reach the set minimum threshold length and segmenting text that exceeds the set maximum threshold length; finally, content compliance detection and filtering are performed to remove corpus with content security risks; Knowledge base construction: Based on document parsing and semantic vectorization technology, we vectorize the content of PDF, Word, TXT, and HTML documents. We convert knowledge into high-dimensional vectors through a deep neural network model and store them in a dedicated vector database. Prompt engineering optimization: Using structured prompt templates and dynamic example injection technology, combined with thought chain guidance model step-by-step reasoning, parameter optimization generation effect; Fine-tuning the historical data training task-personnel matching model: Fine-tuning the historical data training task-personnel matching model based on application requirements through supervised fine-tuning, reward models, and low-rank adaptation. Fusion of search enhancement technology: Integrates search enhancement generation technology to obtain precise knowledge fragments through semantic and keyword hybrid search engines, and optimizes context quality by combining cross-coding and re-ranking.

7. A government function efficiency optimization system based on multimodal data fusion, characterized by: The system is used to implement the government function efficiency optimization method based on multimodal data fusion as described in any one of claims 1 to 6; The system uses Spring Cloud microservice architecture and combines Alibaba Nacos for configuration management and service discovery. The system includes front-end and back-end. The front-end uses the Uniapp framework to develop mobile apps and web applications, and sends requests to obtain back-end data. The latest dynamic data is displayed on the page, and the component library is used to convert the data of central work and key projects into charts for display. The backend uses Spring Boot to build the project framework, integrates MyBatis to operate the database, and creates database tables to store data resources for economic operations and urban management; and writes Mapper interfaces and XML files to implement data addition, deletion, modification, and query; The backend includes: The data collection layer is used to collect data in real time or in batches from the government extranet platform, paper documents, and third-party databases through the government extranet data capture method, obtain raw data, and use Apache NiFi to build data pipelines to connect to the interfaces of enterprise services and grassroots governance, supporting real-time subscription of JSON and / or XML data; The intelligent processing layer is used to conduct in-depth analysis, semantic understanding, and regular processing of the collected raw data, extract structured information, and generate decision-support content, realizing the transformation from data to knowledge; The process execution layer is used to optimize process execution through the Activity 7 process engine; The interactive feedback layer is used to trigger the RPA robot to automatically send a reminder notification when the Activity 7 process engine detects that the task has timed out for more than a set threshold, and simultaneously back it up to the emergency processing queue. Based on the department's monthly completion rate and customer satisfaction indicators, the reinforcement learning algorithm adjusts the process priority weight to achieve flexible resource allocation. The process execution layer includes: The process modeling module is used to design BPMN flowcharts using Camunda Modeler, define task nodes and gateway conditions, deploy BPMN using the Activity 7 process engine, and design flowcharts using the process editor. The module automatically parses the flowchart and generates process instances. Dynamic routing is implemented using the conditionExpression of the exclusive gateway. Timeout warning rules are built in, and timer events are embedded in BPMN process nodes. Business system data is automatically captured daily through the Spring Boot scheduler. When the timer expires, the Activity 7 process engine's ExecutionListener interface is used to call the RPA robot to send a reminder message, synchronously updating the task status to the emergency queue and pushing it to the monitoring dashboard via RocketMQ. The intelligent dispatching module is used to develop a historical data training task-personnel matching model based on the Hainuo big model, constructing data by extracting task attributes and personnel attributes; and updating the historical data training task-personnel matching model through daily incremental updates, incorporating the latest task data for optimization and fine-tuning; The process monitoring dashboard module is used to display process throughput, average processing time, and backlog task heat map in real time, and supports drilling down to individual task details.

Images