Government affair hotline appeal knowledge graph construction method based on agent technology

Abstract

The invention discloses a government affair hotline appeal knowledge graph construction method based on an agent technology, and the method comprises the following steps: sorting basic data, and building a public entity; regularly arranging result data of entity recognition for performing all-parameter fine tuning so as to enhance the capability of a large language model in an entity recognition task and better correspond an entity recognition result to an existing tag to form an entity recognition large model; compiling an appeal analysis agent program based on the entity identification large model; and self-learning and evolving the knowledge graph. The knowledge graph is constructed for supporting applications such as data analysis and question and answer service in the government affair service field, and has multiple functions such as citizen historical appeal recognition, public opinion early warning and reply text generation. The constructed knowledge graph can be effectively applied to government affair service application based on the knowledge graph, the application capability is effectively enhanced, and the platform operation efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the field of government service technology, and in particular to a method for constructing a government hotline demand knowledge graph based on intelligent agent technology. Background Art

[0002] In recent years, in the process of optimizing government services and improving administrative efficiency, promoting the application of new technologies and improving the level of convenient services have become important value orientations. The government requires innovative application of new technologies such as big data, blockchain, and artificial intelligence, and promotes the transformation of government services from human services to human-computer interaction, and from experience judgment to data analysis; explore the application of technologies such as natural language big models, strengthen intelligent customer service capabilities, and provide intelligent question and answer, intelligent search, intelligent guidance and other services. However, the existing government hotline services have problems such as being unable to cope with more diverse and complex application scenarios. Summary of the invention

[0003] The purpose of the present invention is to provide a method for constructing a knowledge graph of government hotline demands based on intelligent agent technology to address the above-mentioned problems. The knowledge graph is used to support data analysis, question-answering services and other applications in the field of government services, including identification of historical demands of citizens, public opinion warning, and reply text generation. The constructed knowledge graph can be effectively applied to government service applications based on the knowledge graph, effectively enhancing application capabilities and improving platform operation efficiency.

[0004] The technical solution of the present invention is as follows: A method for constructing a knowledge graph of government hotline demands based on agent technology includes the following steps: Organize basic data and establish common entities; Regularly organize entity recognition result data for fine-tuning all parameters in order to enhance the ability of the large language model in entity recognition tasks and better correspond entity recognition results to existing labels to form a large entity recognition model; Write a demand analysis agent program based on the entity recognition big model; Self-learning and evolution of knowledge graphs; The steps to construct an agent program include: Set goals and requirements: clarify the goals and functional requirements of the agent; Choose a platform and tools; Basic settings of the intelligent agent, including base model configuration, security settings and environment construction; Configuring knowledge and skills, including knowledge base construction, skill development, and usage process design; Testing and debugging.

[0005] Furthermore, the agent program identifies relations and entities in the text, extracts triple data entity-attribute-attribute value from the numbered demand as an entity, associates it with the common entity, obtains triple data entity-relationship-entity, and stores it in the knowledge graph; The attributes are: appeal subject, location information, appellant, appeal object, citizen emotions and attitudes, citizen expected solution, evidence and basis, time sensitivity; the appellant uses mobile phone number or email address as identity identifier; corresponding entities are established for the appeal subject, location information, appellant, appeal object, citizen emotions and attitudes, and this appeal entity is associated with them.

[0006] Furthermore, the intelligent agent queries whether the existing entities contain attribute values ​​that do not exist in the public entities through fuzzy query and vectorized text matching, so as to decide whether to prompt the management personnel to create a new appeal topic and other operations.

[0007] Furthermore, the self-learning and evolution of the knowledge graph includes the following steps: Data collection and labeling: continuously collect citizen demand data and label the collected data, including demand theme, location information, demander, complaint object, citizen emotions and attitudes, and expected solutions. Use automated tools for preliminary labeling and then perform manual verification. Model training and optimization, combining data from the government sector to fine-tune all parameters to adapt to specific tasks. The previously annotated data can be used as a training set during the fine-tuning process; The knowledge graph is dynamically updated, using the trained model to identify and link entities in new data, and to associate the identified entities with existing entities in the knowledge graph; User feedback and correction, establish a user feedback mechanism to allow staff to evaluate and provide feedback on the information in the knowledge graph.

[0008] Furthermore, the specific work of the intelligent agent includes: Continuously process the appeal text: The agent continuously runs to automatically receive and process the appeal text, including text classification, sentiment analysis, and keyword extraction; Data storage: store all original texts and processing results; Labeling data that requires human intervention: When the agent encounters a difficult situation, it should automatically label the data and forward it to the relevant staff.

[0009] Furthermore, the periodic organization of entity recognition result data includes: Data collection: Collect result data from the output of the large language model on the entity recognition task. The data includes the entities recognized by the model and their corresponding text context, location information, and confidence scores. Data cleaning: Clean the collected data to remove noise and irrelevant information; Data labeling: Manual or semi-automatic labeling is performed on the cleaned data to ensure that the entity recognition results match the existing labeling system; Data alignment: Align the sorted entity recognition results with the existing labels to ensure that each entity is correctly mapped to the corresponding label; Data augmentation: Use data augmentation techniques such as text replacement, synonym replacement, and context expansion to generate more training samples; Quality assessment: Conduct quality assessment on the collated data; Data storage: Store the organized data in an appropriate data structure, such as a database or file system, to facilitate the subsequent full parameter fine-tuning process.

[0010] Furthermore, the said collating basic data and establishing a public entity specifically comprises the following steps: Sorting out geographic information, starting from the highest level of administrative divisions, i.e., provincial units, and gradually working down to the most basic administrative divisions; Set up the object entity of the appeal. The object entity of the appeal refers to various organizations that are specifically responsible for handling citizens' requests and affairs, including government departments, public utilities and corporate units that undertake public service functions; Set the appeal subject entity and set the initial appeal subject entity according to the existing government service classification; Set up citizen emotion and attitude entities, set up initial citizen emotion and attitude entities, such as "happy", "sad", "angry", "surprised" emotion labels, and further refine emotion dimensions, such as "anxiety", "dissatisfaction", and "help request".

[0011] Furthermore, the testing and debugging include: Unit testing: Test each functional module individually to ensure that it works as expected; Integration testing: testing the interactions between components to ensure the stability of the overall system; Performance testing: Testing the agent’s performance under high load to ensure it can handle a large number of requests; Security testing: Conduct security audits to find and fix possible security vulnerabilities.

[0012] Furthermore, the basic configuration of the intelligent agent specifically includes: Base model configuration: Select the base model to be used to build the intelligent agent, including determining its deployment and calling methods; Security settings: Set up necessary security measures, such as API key protection and data encryption; Environment setup: configure the development environment and install the required software and tools.

[0013] Compared with the prior art, the present invention has the following beneficial effects: Expanded application paradigm of large language models: The application scope of AI technology represented by large language models has expanded. It is no longer limited to simple tasks, but can cope with more diverse and complex application scenarios while reducing the consumption of human resources.

[0014] Data-driven continuous capability optimization and upgrading: This invention proposes a specific operational method for continuously improving the basic capabilities of AI during operations, and points out the development direction for dealing with complex problems and improving business efficiency and effectiveness.

[0015] The application direction of knowledge graphs in the field of government affairs has been expanded: compared with the organization and induction of knowledge, documents, excellent questions and answers and other materials in the field of government affairs, the application of knowledge graph technology to government affairs demands has strengthened the care for individual citizens, and helped to provide citizens with warm, accurate and personalized services, highlighting the direction of transformation to a service-oriented government.

[0016] Improve the depth of understanding of demands: By associating the key issues of government demands with public entities, we can have a deeper understanding of the background and related factors of the demands. For example, associating citizens' demands for traffic congestion with specific entities such as roads, transportation facilities, and historical events can help relevant departments analyze the problem more comprehensively.

[0017] Optimize resource allocation and improve decision support: By linking public entities, high-frequency demands and hot issues can be identified and predicted, thereby optimizing resource allocation, ensuring that key issues are given priority, and helping decision makers make more scientific and reasonable decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is the agent workflow diagram for this application.

[0019] Figure 2 Work on large-scale entity recognition models and fine-tune optimization processes for the government affairs sector for this application.

[0020] Figure 3 Analyze the internal structure and functions of the intelligent agent for the purpose of this application. DETAILED DESCRIPTION

[0021] It should be noted that relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0022] The features and performance of the present invention are further described in detail below in conjunction with the embodiments.

[0023] See also Figure 1-3 , a method for constructing a knowledge graph of government hotline demands based on agent technology, comprising: S1. Organize basic data and establish public entities.

[0024] Specifically include: (1) Geographic information sorting: Starting from the highest level of administrative divisions, namely provincial units, and gradually sorting down to the most basic administrative divisions, such as streets or villages and towns. This process requires ensuring that each geographic entity is accurately coded and recorded so that when citizens make demands, the relevant geographic information can be accurately mapped to the corresponding administrative division entity.

[0025] (2) Setting of the entity to which the appeal is addressed: The entity to which the appeal is addressed refers to various organizations that are specifically responsible for handling citizen requests and affairs, including but not limited to government departments, public utilities, and corporate units that undertake public service functions. These entities shall clarify their names and their hierarchical relationships in accordance with the records of the business system, so as to clearly define the scope of responsibility and service process. The information of the entity to which the appeal is addressed shall be updated regularly according to the requirements of the competent department to ensure the accuracy and timeliness of the information. At the same time, in order to maintain the stability and continuity of the management system, the information of these entities shall not be changed without permission.

[0026] (3) Setting of appeal subject entities: Based on the existing classification of government services, set the initial appeal subject entities, such as "urban construction", "construction noise", "water transport", etc. The setting of these subject entities is intended to be consistent with the existing classification of government service appeals, ensuring that citizens' appeals can be correctly classified and processed. During the continuous operation of the system, these appeal subject entities will be dynamically managed. The intelligent agent can prompt the addition of new appeal subjects based on the actual situation of citizens' appeals, and the management personnel also have the right to consider removing those appeal subjects that are no longer appropriate or commonly used.

[0027] (4) Citizen emotion and attitude entity setting: Set the initial citizen emotion and attitude entities, such as "happy", "sad", "angry", "surprised" and other emotion labels, and further refine the emotion dimensions, such as "anxiety", "dissatisfaction", "help", etc. These labels help capture the emotional state of citizens when they make demands, thereby providing a more humane response to government services. Similar to the demand subject entity, citizen emotion and attitude labels will also be dynamically managed during the continuous operation of the system. The intelligent agent can prompt the addition of new emotion and attitude entities based on citizen feedback and behavior patterns, and managers can also consider removing labels that are inappropriate or no longer reflect the true emotions of citizens.

[0028] Through the above process, collating basic data and establishing public entities can not only provide support for the accurate classification and efficient processing of citizen demands, but also promote the continuous optimization of government services and the improvement of citizen satisfaction. This step is the cornerstone of building an intelligent and responsive government management and service system.

[0029] S2. Regularly organize the result data of entity recognition for fine-tuning all parameters in order to enhance the ability of the large language model in entity recognition tasks and better correspond entity recognition results to existing labels to form a domain-specific model. This involves systematically reviewing, classifying and optimizing the entity recognition results output by the model to ensure the quality and consistency of the data, thereby providing a solid foundation for fine-tuning all parameters. The specific process of organizing the result data: (1) Data collection: First, we collect result data from the output of the large language model on the entity recognition task. This data includes the entities recognized by the model and their corresponding text context, location information, and confidence scores.

[0030] (2) Data cleaning: Clean the collected data to remove noise and irrelevant information. This may include deleting duplicate entity recognitions, correcting incorrect entity labels, and filtering out recognition results with low confidence.

[0031] (3) Data labeling: Manual or semi-automatic labeling is performed on the cleaned data to ensure that the entity recognition results match the existing labeling system. This may involve classifying entities into predefined categories, such as names of people, places, and organizations.

[0032] (4) Data alignment: Align the entity recognition results with the existing labels to ensure that each entity is correctly mapped to the corresponding label. This step may require manual intervention to resolve inconsistent or ambiguous labels.

[0033] (5) Data augmentation: In order to improve the diversity and coverage of data, data augmentation techniques such as text replacement, synonym replacement, context expansion, etc. can be used to generate more training samples.

[0034] (6) Quality assessment: Perform a quality assessment on the collated data to ensure the accuracy and reliability of the dataset. This may include calculating metrics such as entity recognition precision, recall, and F1 score.

[0035] (7) Data storage: Store the organized data in an appropriate data structure, such as a database or file system, to facilitate the subsequent full parameter fine-tuning process.

[0036] Through the above process, regular organization of entity recognition result data can not only improve the performance of the large language model in entity recognition tasks, but also ensure the efficient correspondence between entity recognition results and existing labels, thereby building a more accurate and reliable large entity recognition model dedicated to the government field.

[0037] S3. Write a demand analysis agent program based on the large model of entity recognition in the government affairs field. The steps to construct the agent program are as follows: (1)Define goals and requirements Clear goals: Determine what the main goal of the agent is, such as providing 24 / 7 customer service, automatically answering common questions, assisting teaching, etc. The goal of the demand analysis agent program is to organize and output a knowledge graph in a specific format from the user's demand text information.

[0038] Functional requirements: List the functions that the agent must implement, including basic functions and advanced functions.

[0039] User experience: Consider user experience factors such as response speed, interaction friendliness, etc.

[0040] Technical specifications: Define the technical requirements of the agent, such as supported languages, operating system compatibility, etc.

[0041] Business objectives: Ensure that the design of the agent is aligned with the long-term strategic goals of the enterprise.

[0042] (2) Selecting platforms and tools Technical evaluation: Research existing development platforms and tools on the market and compare their respective advantages and disadvantages. Evaluate the technical support capabilities of each platform, such as the richness of API interfaces and the completeness of documentation.

[0043] Compatibility check: Confirm the compatibility of the selected platform with existing systems.

[0044] (3) Basic settings of the agent Base large model configuration: Select the base large model to be used to build the intelligent agent, including determining its deployment and calling methods. It can be a single large model or a combination of multiple large models.

[0045] Security settings: Set necessary security measures, such as API key protection, data encryption, etc.

[0046] Environment setup: configure the development environment and install the required software and tools.

[0047] (4) Configuring knowledge and skills Knowledge base construction: Organize and import knowledge bases in related fields, including common questions and their answers.

[0048] Skill development: Develop or integrate specific capabilities, such as search engines, computer skills, etc.

[0049] Use process design: Design smooth agent workflows to ensure that the agent can complete its work as designed.

[0050] (5) Testing and debugging Unit testing: Each functional module is tested individually to ensure that it works as expected.

[0051] Integration testing: Testing the interactions between components to ensure the stability of the overall system.

[0052] Performance testing: Testing the agent’s performance under high load to ensure it can handle a large number of requests.

[0053] Security testing: Conduct security audits to find and fix possible security vulnerabilities.

[0054] (6) Release and deployment Internal Review: Organize your internal team to conduct a comprehensive review of the upcoming release.

[0055] Documentation preparation: Write detailed user manuals and technical documentation.

[0056] Release plan: Develop a detailed release plan, including release schedule, promotion strategy, etc.

[0057] Environment Migration: Migrate an agent from a development environment to a production environment.

[0058] Monitoring and support: Set up a monitoring system to continuously track the performance of the agent.

[0059] Through the above detailed steps, we can ensure that the construction process of the intelligent agent is orderly and efficient, and finally deliver a complete and high-quality product.

[0060] The main function of the intelligent agent program is to identify the relations and entities in the text, extract the triple data entity-attribute-attribute value from the numbered demand as an entity, and then associate it with the aforementioned common entity to obtain the triple data entity-relationship-entity, which is stored in the knowledge graph.

[0061] Specifically, the following contents are extracted as attributes from the text records of citizens' demands: demand subject, location information, demand person (using mobile phone number, email address, etc. as identity identifier), demand object, citizens' emotions and attitudes, solutions expected by citizens, evidence and basis, time sensitivity, and other relevant information. Among them, the demand subject, location information, demand person (using mobile phone number, email address, etc. as identity identifier), demand object, citizens' emotions and attitudes should establish corresponding entities, and associate this demand entity with them.

[0062] In order to identify the above attributes, the agent needs to have tools such as querying geographic coordinates, querying knowledge graphs or other databases. For processing location information, the agent first extracts the administrative division information in the appeal text. If there is administrative division information, it directly establishes a relationship with the administrative division information; if the appeal does not contain administrative division information, but contains location information such as place names and landmark buildings, the agent queries the corresponding geographic coordinates, compares it with the administrative division geographic information of the public entity, and determines the administrative division of the appeal. If the appeal does not contain location information, the agent needs to determine whether the appeal requires knowledge of the administrative division. If it does, it will be submitted to the management personnel for processing. If it does not require knowledge of the administrative division, it will be recorded at the level of the prefecture-level city / autonomous prefecture.

[0063] For the attributes such as the subject of the appeal, the person making the appeal (using mobile phone number, email address, etc. as identity identifier), the object of the appeal, and the emotions and attitudes of citizens, the intelligent agent needs to query whether the existing entity contains attribute values ​​that do not exist in the public entity through fuzzy query and vectorized text matching, so as to decide whether to prompt the management personnel to create a new appeal subject, etc. For example, when extracting the appeal subject, the attribute value of "river channel management in Yunnan" is obtained, and there is no similar entity in the existing appeal subject entity, the management personnel need to be prompted to create a new appeal subject and record the appeal so that the management personnel can re-analyze the appeal after processing.

[0064] The above content is a description of the steps of the present invention for intelligent agent orchestration. Through specific code writing and debugging, a special intelligent agent for constructing knowledge graphs in the government affairs field is obtained.

[0065] S4. Self-learning and evolution of knowledge graphs. The self-learning and evolution of knowledge graphs is a complex process, involving multiple links such as data collection, model training, and knowledge updating. The key technical points and steps to achieve self-learning and evolution of knowledge graphs are: (1) Data collection and labeling. Continuously collect citizen appeal data and label the collected data, including but not limited to the appeal subject, location information, appealer, complaint object, citizen emotions and attitudes, and expected solutions. Use automated tools for preliminary labeling and then conduct manual verification.

[0066] (2) Model training and optimization. Fine-tune all parameters based on data from the government sector to adapt to specific tasks. Previously labeled data can be used as a training set during fine-tuning. As new data is continuously added, the model needs to perform incremental learning, that is, continue training based on the existing model instead of retraining the entire model. When processing data, the aforementioned intelligent agent can submit samples with high uncertainty for manual labeling, and then use these labeled data to further optimize the model.

[0067] (3) Dynamic update of the knowledge graph. Use the trained model to identify and link entities in new data, and associate the identified entities with existing entities in the knowledge graph. If new entities or relationships are found, managers can be automatically prompted to confirm and add them. At the same time, managers are required to review the knowledge graph regularly and delete entities and relationships that are no longer applicable to ensure the accuracy and timeliness of the knowledge graph.

[0068] (4) User feedback and correction. Establish a user feedback mechanism to allow staff to evaluate and provide feedback on the information in the knowledge graph. This feedback can be used to evaluate the quality of the knowledge graph and guide future optimization directions.

[0069] Through the above steps, the knowledge graph can be self-learning and evolving, enabling it to continuously adapt to new data and needs and provide more accurate and efficient services.

[0070] In the process of continuous work and improvement after the launch of the appeal analysis agent, it is necessary to ensure that it can efficiently process citizen appeal texts and mark data that requires human intervention when necessary. At the same time, the upgrade and optimization of the agent also needs to be properly arranged. The specific work of the agent includes: (1) Continuously process citizen appeal texts. The intelligent agent should be designed to run 24 / 7 to automatically receive and process appeal texts submitted by citizens. The processing process includes but is not limited to text classification, sentiment analysis, keyword extraction, etc., so as to quickly understand the core content of the appeal.

[0071] (2) Data storage. All original texts and their processing results should be stored for subsequent analysis and auditing. The storage system must ensure high availability and data consistency to prevent data loss due to system failures.

[0072] (3) Marking data that requires human intervention. When the agent encounters difficult situations (such as ambiguous expressions and special cases), it should automatically mark these data and forward them to relevant staff. The marking mechanism needs to be flexibly configured, and the trigger conditions can be adjusted according to the actual situation.

[0073] The upgrade and optimization arrangements are as follows: (1) Upgrading outside normal working hours. To avoid affecting daily business, the upgrade and optimization of the intelligent body should be carried out during non-peak hours. It is necessary to plan the upgrade window in advance and make adequate testing and backup measures to reduce the impact on the service.

[0074] (2) Optimize and upgrade large entity recognition models for the government sector. Regularly collect data on the performance of agents in actual applications to fine-tune and optimize existing entity recognition models. Use the latest training data and technological advances to improve the accuracy and robustness of the models in specific government scenarios. More advanced algorithms or model architectures can be introduced to improve recognition efficiency and accuracy.

[0075] (3) Entity object update. With policy changes and social development, the intelligent agent needs to update the entity objects in its knowledge base in a timely manner. Human intervention is required to regularly review and update the entity list to ensure that the intelligent agent can correctly identify and process new terms, regulations, etc. Version control should be maintained during the update process to ensure a smooth transition of updates without affecting existing business processes.

[0076] Through the above measures, the intelligent body can not only maintain an efficient daily working state, but also gradually improve its processing capabilities and intelligence level through continuous optimization and upgrading, so as to better serve citizens and government departments. The specific intelligent body update content includes two parts: optimization and upgrading of the large model of entity recognition dedicated to the government affairs field and updating of entity objects.

[0077] In another specific embodiment, Figure 1The figure shows the workflow diagram of the intelligent agent in the knowledge graph construction method of the government hotline appeal. The citizen appeal is in the form of "Our family is registered in Wanghu Community, Jinfang Sub-district Office, Anning City. My brother XXX bought the 2020 medical insurance in 2019 and died of illness a few days later. The medical insurance cost was 250 yuan. In November 2019, we submitted all the information to the Medical Insurance Center to apply for a medical insurance refund. I have visited the Anning City Medical Insurance Center many times to inquire about the handling situation. The staff told me that I could get a refund and it would be postponed during the period, but there has been no news so far. On October 15, 2020, I went to the Anning City Medical Insurance Center again to inquire about the handling situation. The staff told me that I could no longer get a refund for the medical insurance, which I didn’t understand. I hope that the relevant government departments will investigate and handle it and implement our family’s medical insurance refund as soon as possible." Other information of the work order includes the work order number YNKMXXXX and the location of the incident, Anning City.

[0078] S101: Organize this information into JSON format data and input it into the intelligent agent.

[0079] S102: The agent outputs the entity extraction results: The subject of the appeal: medical insurance refund; Location information: Anning City; Applicant: Telephone number; Object of appeal: Anning City Medical Insurance Center; Citizens’ emotions and attitudes: anxiety, confusion, dissatisfaction, seeking help, and expectation; Solutions that citizens expect: Implementing medical insurance refunds; Evidence and basis: The citizen's brother's medical insurance purchase record, the citizen's brother's death certificate, and the materials submitted to the medical insurance center. Consulting staff records: Time Sensitivity: Not sensitive; Other relevant information: None; The agent uses the work order number YNKMXXXX as the entity label and the above items as the attribute value to establish a work order entity node in the knowledge graph. Then, the subject of the appeal is the existing public entity medical insurance refund, so the work order node is connected to the medical insurance refund entity; the location information is Anning City, so the work order node is connected to the public entity of Anning City; the object of the appeal is Anning City Medical Insurance Center, so the work order node is connected to the public entity of Anning City Medical Insurance Center.

[0080] S103: In the matters extracted by the agent, the values ​​of the subject of the appeal, location information, and object of the appeal can be associated with the existing public entity. If the appellant is not in the focus list, there is no need to add a new public entity. If the subject of the appeal in the extracted matters is drone injury, and there is no drone injury node in the public entity, the agent will write the original appeal information and the extracted matters into the auxiliary database and wait for other programs to process.

[0081] Figure 1 The work and fine-tuning optimization process of the large model for entity recognition in the government affairs field are shown. The large model is the base of the demand analysis agent, and its functions are packaged in the agent, and it only receives the requirements of the agent to run.

[0082] S201: The large language model needs prompt words to guide the generation of content, improve the generation quality, control the generation direction, reduce ambiguity, etc. The prompt words used by the appeal analysis agent are as follows: "You are an expert in analyzing event entities. You extract specific content items from people's words and return them in JSON format. You observe {{content}} and propose the appeal theme, location information, appeal person, appeal object, citizen emotions and attitudes, citizen expected solutions, evidence and basis, time sensitivity, and other relevant information." The above is an example of a prompt word, not an actual prompt word. Combining information content means filling the citizen appeal text into the position specified by the prompt word, such as {{content}}, to generate a longer text with instruction tasks that adapts to the working requirements of the large language model.

[0083] S202: Input the above combined information text into the large language model.

[0084] S203: The large language model adopts the reasoning working model according to the input text and generates output. Its output is a string in the format of JSON. The conversion of the string into a JSON object and the subsequent operations are completed by the intelligent agent. The output is in the form of '{"appeal subject":"medical insurance refund","location information":"Anning City","appeal person":"telephone number","appeal object":"Anning City Medical Insurance Center","citizens' emotions and attitudes":["anxiety","confusion","dissatisfaction","help","expectation"],"citizens' expected solution":"implementation of medical insurance refund","evidence and basis":["citizens' brother's medical insurance purchase record","citizens' brother's death certificate","materials submitted to the medical insurance center","records of consulting staff"],"time sensitivity":"insensitive","other relevant information":null}'.

[0085] S204: Return the above character string to the agent.

[0086] S205: The intelligent agent updates and modifies the knowledge graph as in S103.

[0087] S206: The staff regularly observes the newly added work order nodes in the knowledge graph, checks whether the attributes of the nodes are clearly described and comply with the platform regulations, and checks whether the original citizen appeal text is completely described. The staff regularly selects some high-quality and typical work order nodes and exports them into a suitable format for fine-tuning the large model.

[0088] S207: In order to make the large language model better adapt to a specific task or field, fine-tuning is required. Fine-tuning selects the model and data set to be fine-tuned, and completes the fine-tuning through a specific procedure. The fine-tuning process does not belong to the content of the application for protection of the present invention, and will not be described in detail here.

[0089] Figure 3 The figure shows the internal structure and functions of the demand analysis agent in the related art. In the figure, 300 is the scope of the entire agent, which is a set of computer software programs and logically has three modules: 310 entity recognition module, 320 geographic information processing module, and 330 content item processing module.

[0090] 310 Entity recognition module, used to call the government affairs domain dedicated entity recognition big model, and organize the data returned by the big model. Calling the big model is as described in S201, organizing the data returned by the big model is as described in S203, processing the string in JSON format into a JSON object, and providing it to other modules of the intelligent agent for use.

[0091] The 320 geographic information processing module is used to observe whether the location information is correctly extracted to ensure that the location information is in the correct format. First, observe whether the location information extracted by the 310 module conforms to the location public entity in the knowledge graph. If it conforms, the work of the 320 module ends; if the location information extracted by the 310 module does not conform to the location public entity in the knowledge graph, but is a landmark building or place name, then the 320 module is responsible for calling the map search engine to query the geographic coordinates of the location information, and compare it with the geographic coordinates of the location public entity in the knowledge graph to determine the closest location public entity as the location information analyzed by this intelligent agent. If the location information extracted by the 310 module is empty, the 320 module will find which prefecture or city hotline the work order entered, and use the prefecture or city as the location information analyzed by this intelligent agent, such as Kunming City. The accuracy of such location information is at the prefecture or city level.

[0092] The 330 content item processing module processes four items: the subject of the appeal, location information, the object of the appeal, and the appellant. The final output of the 330 module is the query statement to be executed in the knowledge graph and the auxiliary database. Each item is compared with the public entity in turn. For the three items of the subject of the appeal, location information, and the object of the appeal, determine whether the item value already exists in the existing public entity. If it does, skip the processing and prepare to establish the corresponding connection after adding a work order node to the knowledge graph; if it does not exist, write the corresponding data in the auxiliary database and wait for other programs to process. For the appellant item, the 330 module must determine whether the value of the item belongs to the key focus list. The key focus list is a public entity. If the value of the item belongs to the key focus list, prepare to establish a connection with the corresponding public entity; if it does not belong to the key focus list, no processing will be performed.

[0093] The above-mentioned embodiments only express the specific implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the protection scope of the present application. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the technical solution concept of the present application, and these all belong to the protection scope of the present application.

Claims

1. A method for constructing a knowledge graph of government hotline demands based on agent technology, characterized in that: The following steps are involved: Organize basic data and establish common entities; Regularly organize entity recognition result data for fine-tuning all parameters in order to enhance the ability of the large language model in entity recognition tasks and better correspond entity recognition results to existing labels to form a large entity recognition model; Write a demand analysis agent program based on the entity recognition big model; Self-learning and evolution of knowledge graphs; The steps to construct an agent program include: Set goals and requirements: clarify the goals and functional requirements of the agent; Choose a platform and tools; Basic settings of the intelligent agent, including base model configuration, security settings and environment construction; Configure knowledge and skills, including knowledge base construction, skill development and usage process design; Testing and debugging.

2. According to the method for constructing a knowledge graph of government hotline demands based on agent technology in claim 1, it is characterized in that: The agent program identifies the relations and entities in the text, extracts the triple data entity-attribute-attribute value from the numbered demand as an entity, associates it with the common entity, obtains the triple data entity-relationship-entity, and stores it in the knowledge graph; The attributes are: appeal subject, location information, appellant, appeal object, citizen emotions and attitudes, citizen expected solution, evidence and basis, time sensitivity; the appellant uses mobile phone number or email address as identity identifier; corresponding entities are established for the appeal subject, location information, appellant, appeal object, citizen emotions and attitudes, and this appeal entity is associated with them.

3. According to claim 2, a method for constructing a knowledge graph of government hotline demands based on agent technology is characterized in that: The intelligent agent queries whether the existing entities contain attribute values ​​that do not exist in the public entities through fuzzy query and vectorized text matching, and decides whether to prompt the administrator to create a new demand topic operation.

4. According to the method for constructing a knowledge graph of government hotline demands based on agent technology in claim 1, it is characterized in that: The self-learning and evolution of the knowledge graph includes the following steps: Data collection and labeling: continuously collect citizen demand data and label the collected data, including demand theme, location information, demander, complaint object, citizen emotions and attitudes, and expected solutions. Use automated tools for preliminary labeling and then perform manual verification. Model training and optimization, combining data from the government sector to fine-tune all parameters to adapt to the task. During the fine-tuning process, the previously annotated data can be used as a training set; The knowledge graph is dynamically updated, using the trained model to identify and link entities in new data, and to associate the identified entities with existing entities in the knowledge graph; User feedback and correction, establish a user feedback mechanism to allow staff to evaluate and provide feedback on the information in the knowledge graph.

5. A method for constructing a knowledge graph of government hotline demands based on agent technology according to claim 2 or 3, characterized in that: The specific work of the agent includes: Continuously process the appeal text: The agent continuously runs to automatically receive and process the appeal text, including text classification, sentiment analysis, and keyword extraction; Data storage: store all original texts and processing results; Labeling data that requires human intervention: When the agent encounters a difficult situation, it should automatically label the data and forward it to the relevant staff.

6. According to the method for constructing a knowledge graph of government hotline demands based on agent technology in claim 1, it is characterized in that: The periodic arrangement of entity recognition result data includes: Data collection: Collect result data from the output of the large language model on the entity recognition task. The data includes the entities recognized by the model and their corresponding text context, location information, and confidence scores. Data cleaning: Clean the collected data to remove noise and irrelevant information; Data labeling: Manual or semi-automatic labeling is performed on the cleaned data to ensure that the entity recognition results match the existing labeling system; Data alignment: Align the sorted entity recognition results with the existing labels to ensure that each entity is correctly mapped to the corresponding label; Data augmentation: Use data augmentation techniques such as text replacement, synonym replacement, and context expansion to generate more training samples; Quality assessment: Conduct quality assessment on the collated data; Data storage: Store the organized data in an appropriate data structure, such as a database or file system, to facilitate the subsequent full parameter fine-tuning process.

7. According to claim 1, a method for constructing a knowledge graph of government hotline demands based on agent technology is characterized in that: The said collating basic data and establishing a public entity specifically comprises the following steps: Sorting out geographic information, starting from the highest level of administrative divisions, i.e., provincial units, and gradually working down to the most basic administrative divisions; Set up the object entity of the appeal. The object entity of the appeal refers to various organizations that are specifically responsible for handling citizens' requests and affairs, including government departments, public utilities and corporate units that undertake public service functions; Set the appeal subject entity and set the initial appeal subject entity according to the existing government service classification; Set up citizen emotion and attitude entities, set up initial citizen emotion and attitude entities, such as "happy", "sad", "angry", "surprised" emotion labels, and further refine emotion dimensions, such as "anxiety", "dissatisfaction", and "help request".

8. According to the method for constructing a knowledge graph of government hotline demands based on agent technology in claim 1, it is characterized in that: The testing and debugging include: Unit testing: Test each functional module individually to ensure that it works as expected; Integration testing: testing the interactions between components to ensure the stability of the overall system; Performance testing: Testing the agent’s performance under high load to ensure it can handle a large number of requests; Security testing: Conduct security audits to find and fix possible security vulnerabilities.

9. According to the method for constructing a knowledge graph of government hotline demands based on agent technology in claim 1, it is characterized in that: The basic settings of the intelligent agent specifically include: Base model configuration: Select the base model to be used to build the intelligent agent, including determining its deployment and calling methods; Security settings: Set up necessary security measures, such as API key protection and data encryption; Environment setup: configure the development environment and install the required software and tools.

Images