Automatic address governance method and device based on intelligent agent and medium

Abstract

The invention relates to the field of geographic information systems, and discloses an agent-based automatic address governance method and device and a medium, and the method comprises the steps: obtaining a to-be-governed address through a data perception and preprocessing agent; inputting the address to be governed into a master control agent; the main control agent analyzes the address to be governed and gives an analysis result; according to the analysis result, selecting a tool execution agent to process the address to be treated; and outputting the governed address, and storing all data by adopting a memory and interaction agent. According to the method, the analysis accuracy, the semantic understanding depth and the robustness of the LLM in processing complex and non-standard addresses containing confusion information are greatly improved, meanwhile, the requirement for manual intervention is greatly reduced, the efficiency of processing mass address data is remarkably improved, and the real-time response speed is remarkably increased.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of geographic information systems, in particular to an agent-based automatic address management method, device and medium. BACKGROUND

[0002] In the field of urban digital management and smart services, address data as a core spatial information asset, its standardization and dynamic management capability directly affects the efficiency of key scenarios such as urban planning, logistics distribution, and emergency response.

[0003] With the expansion of urban space and the diversification of address expression forms, address data management faces three major challenges: 1. Difficulty in semantic analysis and element decomposition of unstructured address text (such as "the third blue sign on the east side of Zhongguancun Entrepreneurship Street"); 2. Address hierarchy heterogeneity caused by regional cultural differences (such as the coexistence of "street-community-doorplate" and "district-road-lane-number" mixed systems); 3. The contradiction between the surge in dynamic management needs and the high cost of manual intervention, especially in the face of address changes and alias derivatives (such as the popular landmark "a coffee wall") in urban renewal, traditional methods are difficult to achieve real-time response.

[0004] Due to the semantic complexity and heterogeneity of the address to be processed, it is very difficult to directly use a model to complete the management of the address. In addition, although existing address management technology has achieved high automation in sub-tasks such as address standardization and geosegmentation, the overall process of address management is still linearly executed according to the set rules. That is, even for very standard addresses to be processed, they still need to be processed through all steps. When processing large-scale address data, the efficiency is extremely low (that is, addresses that only need to be partially processed or do not need to be processed are still processed using all steps). And when the characteristics of the address change, the processing steps may need to be adjusted on a large scale, and the cost of designing a new address management system is high. SUMMARY

[0005] The purpose of the present application is to propose an agent-based automatic address management method, device and medium, to solve the technical problems of element structure difficulty, address analysis difficulty to respond in real time, low efficiency and high cost existing in the current address data management.

[0006] The present application proposes an innovative solution to the above problems, 1. For challenges 1 and 2, address management is divided into multiple sub-tasks, each sub-task designs a professional model or algorithm to solve a specific address problem.

[0007] 2. For challenge 3, introduce an agent dynamic decision mechanism. The agent makes the following two decisions: a. Through the agent's autonomous evaluation of address text complexity (such as the proportion of invalid elements, regional expression characteristics), real-time selection of processing path (such as "invalid element filtering → address standardization → address completion" or "invalid element filtering → address completion"), b. The processing steps are regarded as the calling tools of the agent, so it is very convenient to extend new steps, which can respond to changes in address governance more quickly.

[0008] Specifically, the application provides an agent-based automatic address governance method, device and medium, the method comprising the following steps: S1, acquiring an address to be governed by using a data perception and preprocessing agent; S2, inputting the address to be governed into a master agent; S3, the master agent analyzes the address to be governed and gives an analysis result; S4, selecting a tool execution agent to process the address to be governed according to the analysis result; S5, outputting the governed address, and storing all data in steps S1-S5 by using a memory and interaction agent.

[0009] A storage medium stores instructions and data for implementing an agent-based automatic address governance method.

[0010] An agent-based automatic address governance device comprises a processor and the storage medium; the processor loads and executes the instructions and data in the storage medium to implement an agent-based automatic address governance method.

[0011] The application provides the following beneficial effects: 1. The model performance is greatly improved, breaking through the LLM application bottleneck: through fine-tuning of large language models for address governance tasks (such as LoRA fine-tuning, optimizing learning rate, loss function, etc.), advanced Prompt Engineering (such as thought chain, dynamic tool tip, role playing), and deep bidirectional fusion with knowledge graph (input enhancement of knowledge injection and output verification of knowledge guidance), the application greatly improves the address parsing accuracy, semantic understanding depth and robustness of LLM when dealing with complex, non-standard and ambiguous information.

[0012] 2. A highly automated and collaborative governance process can be constructed: a collaborative architecture of multiple agents (master agent, llm service agent, knowledge retrieval and reasoning agent, tool execution agent, etc.) and model context protocol (MCP) to realize intelligent decomposition of complex address governance tasks, dynamic planning and scheduling on demand, and parallel processing and real-time state sharing between agents. MCP ensures the standardization, efficiency and traceability of information exchange between agents, making the entire governance process highly automated, greatly reducing the need for manual intervention, and the execution process is transparent and can be traced back, thereby significantly improving the efficiency and real-time response speed of processing massive address data. BRIEF DESCRIPTION OF DRAWINGS

[0013] Figure 1 is a simple process flow diagram of the method of the present application; Figure 2 is a hardware device working schematic of an embodiment of the present application. DETAILED DESCRIPTION

[0014] In order to make the purpose, technical scheme and advantages of the present application clearer, the embodiments of the present application will be further described below with reference to the drawings.

[0015] Before formally describing the present application, a general description of the present application will be given first for easy understanding.

[0016] Please refer to Figure 1 The present application provides an automatic address governance method based on agent, comprising the following steps: S1, acquiring the address to be governed by using a data perception and preprocessing agent; It should be noted that the data perception and preprocessing agent (Perception&Preprocessing Agent) is responsible for obtaining raw address data from multiple sources (such as business system database, batch file import, real-time API interface).

[0017] At the same time, the data perception and preprocessing agent also performs preliminary number cleaning operations, such as removing obvious illegal characters, unifying full and half angles, processing common noise methods (such as "telephone: 1456323.."), performing preliminary word segmentation and address type preliminary judgment (such as judging whether it is a structured address or a spoken description), and preparing for the processing of subsequent agents.

[0018] S2, inputting the address to be governed into a master agent; S3, the master agent analyzes the address to be governed and gives an analysis result; It should be noted that the master agent includes a large model service agent, a knowledge retrieval agent and a reasoning agent.

[0019] It should be noted that the master agent (Team Agent) as the "brain" is responsible for receiving the address information to be governed, conducting preliminary evaluation and task planning. It will dynamically decompose tasks according to the complexity of the address and the governance goal, and assign sub-task sequences to other professional agents through the model context protocol mechanism (MCP).

[0020] Among them, MCP is a key technology for efficient collaboration of multiple agents. It is a standardized protocol for standardizing information exchange format, data structure and semantic interface between agents, large language models, knowledge graphs and various tools. Its core goal is to ensure consistency, traceability, understandability and interoperability of information transmission between different modules of the system.

[0021] The implementation is as follows: MCP usually defines a set of message structures based on serialization formats such as json, XML or protobuf. Each message may contain: task id, session id, timestamp, sender agent identifier, receiver agent identifier, task type, input data (such as address to be processed, context information), processing parameters, output results, status code, error information, confidence score, reasoning path record, etc.

[0022] In the present application, the master agent takes the large model service agent as the core, inputs the prompt words constructed by the reasoning agent to the large model service agent for thinking and analysis, obtains the analysis results, and outputs the analysis results after checking by the knowledge retrieval agent.

[0023] Among them, the large model service agent (LLM Service Agent): embed one or more large language models adjusted and optimized for specific parameters. Its core responsibility is to perform deep semantic understanding of address data, structured analysis (convert unstructured text into structured address elements), standardization conversion (such as number format, hierarchical expression unification), key element extraction, address style conversion and other core natural language processing tasks.

[0024] Specifically, the present application not only simply calls the general LLM API, but emphasizes a series of targeted parameter adjustments, optimization strategies and case fusion on LLM.

[0025] Model selection and foundation: select pre-trained large language models that perform well in Chinese understanding and generation as the foundation, such as Alibaba's Qwen series, Zhi Spectrum AI's GLM series, and variants of BERT, GPT architecture.

[0026] For specific, mode relatively fixed sub-tasks (such as specific type of address element into the score, digital format conversion), different basic models are used to train lightweight special models (such as BERT-based sequence labeling model), and the LLM service agent is used for scheduling and result fusion to form a hybrid strategy of "LLM + small model".

[0027] Regarding the model fine-tuning strategy: the parameter adjustment and fine-tuning of the application is not the conventional operation of general LLM, but a specialized improvement based on a deep understanding of the inherent characteristics of address data (such as the hierarchical dependence of administrative divisions, the geographical proximity of POIs and roads, the coexistence of fixed address patterns and colloquial expressions, and the density of entity nouns). The purpose is to guide the LLM to learn the special "grammar" and "semantics" of address language, so that it can not only understand the surface meaning of the text, but also capture the structure and logic behind the address. This is the core difference and innovation of the application compared to general LLM applications.

[0028] In the process of lora fine-tuning, the application sets rank to 8, lora_alpha to 64, uses AdamW optimizer, and sets the initial learning rate to 1e-5. In the training, the batch_size is set to 2, and a total of 3000 steps are trained. The innovation lies in that, in the fine-tuning process, in addition to the standard cross-entropy loss, an auxiliary loss term for address structure integrity and administrative division hierarchy consistency is additionally introduced. The auxiliary loss term is as follows:

[0029] wherein, is the auxiliary loss term, is the administrative level set pair set (such as province-city, city-district) that needs to be verified; is the level weight (such as province-city error weight = 0.7, city-district level = 0.3); is the knowledge graph relationship verification function; is the core address element such as province, city, street, and house number; is the element importance weight (such as province = 0.4, city = 0.3, house number = 0.2); is the prediction confidence of the model for the element.

[0030] Taking the address "X province Y city Z district" analysis as an example: The model analyzes that: province = X province (conf = 0.95), city = Y city (conf = 0.90), and district = Z district (conf = 0.85); Knowledge graph verification: if Y city does not contain Z district in the knowledge base, then

[0031] Loss calculation:

[0032] Therefore, the present application is to enforce administrative hierarchy logical correctness; by improve the reliability of key element identification.

[0033] Through the newly designed loss function, those resolved counties and their belonging cities that do not match in the knowledge network are punished. At the same time, the key address elements (such as province, city, detailed doorplate) in the training data are weighted detection, to ensure that the model focuses on learning the extraction of these core information.

[0034] Among them, the reasoning intelligent agent is used to construct the prompt word template, and the semantic network rule reasoning engine is used to execute the pre-defined address related rules.

[0035] Specifically, the definition of the prompt word is to input a certain text into the large model. The prompt word in the present application can be generated by the prompt word template and external input variables. For example: Prompt word template: <target> process the address to be managed into a standard address; <tools that can be used>; <tool information>; <mode> answer as follows: Address to be managed: input address to be managed; Think: next action; Observe: the result of the action; (this process can be repeated for many rounds) <key points> When selecting and calling tools, please strictly output in the following format: <tool name: <<called tool>>, parameter: <<input parameter value>>; Start; Address to be managed: <<address to be managed>>; Think: <<thinking content>>; Observe: <<tool calling result>>; When initializing, replace <<address to be managed>> with the input address to be processed, replace <<tool information>> with the address management tool that can be called. Then replace <<thinking content>> with the content input by the large model, and replace <<tool calling result>> with the output result of the called tool.

[0036] Specifically, the rule reasoning engine provides support for RDF, RDFS, OWL and a rule reasoning machine dedicated to address verification through a powerful semantic network framework.

[0037] Definition and management of rules (rule examples): Rule 1: The county must belong to the city it claims; Rule 2: A contains B, B contains C, then A contains C (transitivity); Rule 3: Complete the street according to the road name and POI name (exemplary, actual rules are more complex); Rule 4: Alias inference (based on alias relationships in the knowledge graph).

[0038] The above reasoning application scenarios mainly include: Address element verification: Check if there are logical conflicts between the parsed address elements (e.g., zip code does not match the county, street does not belong to the claimed county.

[0039] Address element completion: According to the known address elements and the hierarchical relationship and adjacency relationship in the knowledge graph, automatically complete the missing administrative divisions (such as provinces, cities, streets), postal codes, etc.

[0040] Address hierarchy relationship inference: When a place name corresponds to multiple entities, use context and rules to disambiguate. For example, by analyzing other elements in the address (such as county names, POI names) to determine which city's Zhongshan Road the "Zhongshan Road" refers to.

[0041] Address similarity calculation and name duplication: Based on the standardized address elements and relationships obtained through reasoning, more accurately determine the similarity between addresses.

[0042] Among them, the knowledge retrieval agent performs efficient knowledge retrieval (such as querying whether a specific administrative division exists, the standard name and location information of a certain place) according to the request of the master agent, and the output content is used to verify the recall of the parsed results of the relevant standard address elements, complete the missing information, eliminate ambiguity, and perform logical verification.

[0043] The output content is used to verify the recall of the parsed results of the relevant standard address elements, complete the missing information, eliminate ambiguity, and perform logical verification.

[0044] The knowledge retrieval agent performs the following checks: format specification review and toxicity detection.

[0045] The format specification review includes: 1. Whether the json data can be extracted from the output results of the large model service agent; 2. Whether the tool name in the extracted json data belongs to the provided tools; 3. Whether the parameters in the extracted json data meet the input requirements set by the tool.

[0046] The toxicity detection specifically matches the analysis results with a sensitive dictionary. If the matching result is empty, the toxicity detection passes.

[0047] Specifically, Large Model Service Agent (LLM Service Agent): Embed one or more large language models that have been fine-tuned and optimized for specific tasks.

[0048] The core responsibilities include performing deep semantic understanding of address data, structured parsing (converting unstructured text into structured address elements), standardized conversion (such as number format, hierarchical expression unification), key element extraction, address style conversion, and other core natural language processing tasks.

[0049] Accept instructions and context information from the master agent, and output parsing results, confidence levels, or next action recommendations.

[0050] S4, according to the analysis results, select tool execution agent to process the address to be governed; It is worth noting that the tool execution agent (Tool Execution Agent): manages and calls a modular "address governance toolbox".

[0051] This toolbox contains a series of atomic address processing tools or algorithm modules, including: invalid element processing module (based on rules and models), address segmentation module (such as CRF, BiLSTM-CRF model), address adjustment module (based on rule sorting), address completion sub-module (for specific missing types), geocoding / decoding tools, etc. The agent can accurately call specific tools according to the instructions of the master agent and return the execution results.

[0052] Specifically, the invalid element processing module is as follows: Function: Remove POI rules, invalid special characters, and duplicate city information from addresses.

[0053] For example: input: A province B city B city C road D number (MM building) 1 building -1 unit Output: A province B city C road D number 1 building 1 unit Processing flow: input the address to be processed, use a specially trained component extraction model to label each character in the address to be processed, and delete the characters labeled as "invalid".

[0054] Specifically, the address segmentation module is as follows: Function: Segment the input address and label the category of each address element Input: A province B city C road D number 1 building 1 unit 101 room Output: A province / B city / C road / D number / 1 building / 1 unit / 101 room Processing process: input the address to be processed, use a specially trained address segmentation model to process, and output the segmentation result.

[0055] Specifically, the address section adjustment module is as follows: Function: Re-adjust the address elements in the input address according to the predetermined order For example: Input: A province C road D number B city 1 building 1 unit 101 room

city address element is not correct

[0056] Specifically, the address completion module is as follows: Function: Complete the missing elements in the input address For example: Input: B city C road D number 1 building 1 unit 101 room

missing province

[0057] Specifically, it can also include an address specification processing module, which is as follows: Function: Re-describe the input address according to the standard address For example: Input: A province B city C road D number 1-1-101 Output: A province B city C road D number 1 building 1 unit 101 room Process: Input the address to be processed and the standard address (which can be defaulted or input), and use a specially trained domain model to process the address to be processed according to the standard description style and output.

[0058] S5, output the address after governance, and use the memory and interactive agent to store all data in steps S1-S5.

[0059] It should be noted that the memory and interactive agent is responsible for persistently storing important data in the governance process, including the original address, the intermediate processing state, the decision and output of each agent, the final governance result, and the complete governance path and log. At the same time, it is also responsible for providing the final, structured, standardized address data to users or other business systems through API interfaces, message queues or user interfaces, etc. Support historical traceability and audit.

[0060] The present application provides a kind of implementation as follows: Enter the address to be processed: No. D, Road C, City B, Province A, Unit 1-1-101 Round 1: Large model prompt input content: <Objective> To process the addresses to be addressed into standard addresses; <Tools Available> Tool 1: Invalid Element Handling Tool Description: Omitted; Tool 2: Standardization Processing Tool Description: Omitted; … Tool 5: Address Completion Tool Description: Omitted; <Method> Answer as follows: Address to be addressed: The address to be addressed that you input; Consider: What is the next step? Observation: The result of the action; (This process can be repeated multiple times) <Important Note> When selecting a tool to invoke, please strictly adhere to the following output format: <Tool name: <<The tool to be called>>, Parameters: <<The parameter values ​​to be entered>>>; start: Address to be addressed: No. D, C Road, City B, Province A, 1-1-101; think: Output content of the large model: The calling tool standardizes the processing.

[0061] Tool Name: Normalization Processing. Parameters: Address to be processed: No. 1-1-101, Building 2, Unit 3, No. 101, Road D, C, City B, Province A; Style address: Room 302, Unit 3, Building 2, No. 101, Road D, C, City B, Province A>> Tool output: Room 101, Unit 1, Building 1, No. D, C Road, City B, Province A Round Two: Large model prompt input content: <Objective> To process the addresses to be addressed into standard addresses. <Tools Available> Tool 1: Invalid Element Handling Tool Description: Omitted Tool 2: Standardization Processing Tool Description: Omitted … Tool 5: Address Completion Tool Description: Omitted <Method> Answer as follows: Address to be addressed: The address to be addressed that you input; Consider: What is the next step? Observation: The result of the action; (This process can be repeated multiple times) <Important Note> When selecting a tool to invoke, please strictly follow the following format when outputting: <Tool Name: <<Invoked Tool>>, Parameters: <<Input Parameter Values>>> start! Address to be addressed: No. D, C Road, City B, Province A, Unit 1-1-101 Thinking: <Tool name: Normalization. Parameters: <Address to be processed: No. D, C Road, City B, Province A, 1-1-101; Style address: Room 302, Unit 3, Building 2, No. D, C Road, City B, Province A>> Observation: The standardized processing result is Room 101, Unit 1, Building 1, Road D, City C, Province A. think: Output content of the large model: Complete the address of the tool you are calling.

[0062] Tool Name: Address Completion. Parameters: Address to be processed: Room 101, Unit 1, Building 1, No. D, Road C, City B, Province A; Tool output: Room 101, Unit 1, Building 1, No. D, Road C, District X, City B, Province A; Third round: Large model prompt input content: <Objective> To process the addresses to be addressed into standard addresses; <Tools Available> Tool 1: Invalid Element Handling Tool Description: Omitted Tool 2: Standardization Processing Tool Description: Omitted … Tool 5: Address Completion Tool Description: Omitted <Method> Answer as follows: Address to be governed: The entered address to be governed Reflection: Next Steps Observation: The results of the action (This process can be repeated multiple times) <Important Note> When selecting a tool to invoke, please strictly adhere to the following output format: <Tool name: <<The tool to be called>>, Parameters: <<The parameter values ​​to be entered>>>; start: Address to be addressed: No. D, C Road, City B, Province A, 1-1-101; Think about: Address completion for calling tools.

[0063] Tool Name: Normalization Processing. Parameters: Address to be processed: No. 1-1-101, Building 2, Unit 3, No. 101, Road D, C, City B, Province A; Style address: Room 302, Unit 3, Building 2, No. 101, Road D, C, City B, Province A>> Observation: The result of the normalization is 1 Unit 101 Room, Building 1, No. 1, C Road, B City, A Province Thought: Call the tool address completion.

[0064] <Tool name: address completion. Parameters: <address to be processed: 1 Unit 101 Room, Building 1, No. 1, C Road, B City, A Province>>; Observation: The result of the address completion is 1 Unit 101 Room, Building 1, No. 1, C Road, X District, B City, A Province Thought: Large model output: The final result is: 1 Unit 101 Room, Building 1, No. 1, C Road, X District, B City, A Province.

[0065] End.

[0066] See Figure 2 , Figure 2 is a hardware device working schematic diagram of an embodiment of the present application, which specifically comprises: an automatic address management device based on an agent 401, a processor 402, and a storage medium 403.

[0067] An automatic address management device based on an agent 401: the automatic address management device based on an agent 401 realizes the automatic address management method based on an agent.

[0068] Processor 402: the processor 402 loads and executes instructions and data in the storage medium 403 to realize the automatic address management method based on an agent.

[0069] Storage medium 403: the storage medium 403 stores instructions and data; the storage medium 403 is used to realize the automatic address management method based on an agent.

[0070] In general, the beneficial effects of the present application are: 1. Model performance is greatly improved, breaking through the application bottleneck of LLM: through fine-tuning of large language models for address management tasks (such as LoRA fine-tuning, optimizing learning rate, loss function, etc.), advanced Prompt Engineering (such as thought chain, dynamic tool tips, role-playing), and deep bidirectional fusion with knowledge graph (input enhancement of knowledge injection and output verification of knowledge guidance), the present application greatly improves the parsing accuracy, semantic understanding depth and robustness of LLM when dealing with complex, non-standard, and ambiguous information addresses.

[0071] 2. A highly automated and collaborative governance process can be constructed: with the collaborative architecture of multi-agent (main control agent, llm service agent, knowledge retrieval and reasoning agent, tool execution agent, etc.) and model context protocol (MCP), the complex address governance task is intelligently decomposed, dynamically planned and scheduled on demand, and the parallel processing and real-time state sharing among the agents are realized. MCP ensures the standardization, efficiency and traceability of information exchange among agents, making the entire governance process highly automated, greatly reducing the need for manual intervention, and the execution process is transparent and can be traced back, thereby significantly improving the efficiency and real-time response speed of processing massive address data.

[0072] The above merely describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. An agent-based automated address governance method, characterized in that: The method comprises the following steps: S1, acquiring an address to be governed by a data perception and preprocessing agent; S2, inputting the address to be governed into a master agent; S3, the master agent analyzes the address to be governed and gives an analysis result; S4, according to the analysis result, a tool execution agent is selected to process the address to be governed; S5, outputting the governed address, and storing all data in steps S1-S5 by a memory and interaction agent.

2. The agent-based automated address governance method of claim 1, wherein: The master agent comprises a large model service agent, a knowledge retrieval agent and an inference agent.

3. The agent-based automated address governance method of claim 2, wherein: In step S3, the master agent takes the large model service agent as the core, inputs the prompt words constructed by the inference agent into the large model service agent for thinking and analysis, obtains the analysis result, and outputs the analysis result after checking by the knowledge retrieval agent.

4. The agent-based automated address governance method of claim 2, wherein: The items checked by the knowledge retrieval agent include format specification audit and toxicity detection.

5. The agent-based automated address governance method of claim 4, wherein: The format specification audit includes:

1. whether the json data can be extracted from the output result of the large model service agent; 2. whether the tool name in the extracted json data belongs to the provided tools; 3. whether the parameters in the extracted json data meet the input requirements set by the tools.

6. The agent-based automated address governance method of claim 4, wherein: The toxicity detection specifically refers to matching the analysis result with a sensitive word dictionary, and if the matching result is empty, the toxicity detection is passed.

7. The agent-based automated address governance method of claim 1, wherein: The tool execution agent in step S4 includes an invalid element processing module, a planning processing module, an address section module, an address section adjustment module and an address completion module.

8. A storage medium characterized by: The storage medium stores instructions and data for implementing the automatic address governance method based on the agent in any one of claims 1-7.

9. An agent-based automated address governance device, characterized by: It comprises: A processor and a storage medium; the processor loads and executes the instructions and data in the storage medium to implement the automatic address governance method based on the agent in any one of claims 1-7.

Images