Public opinion data structuring method based on Re-Act framework

Abstract

The invention discloses a public opinion data structuring method based on a Re-Act framework, and relates to the technical field of public opinion data processing and intelligent structuralization, and the method comprises the steps: firstly constructing a Re-Act public opinion structuring system comprising a thinking module, an action module and an observation module, then collecting data through multiple rounds of closed-loop iteration until a full link is covered, and finally obtaining the data. And finally, processing based on the full-link data through a structured tool, outputting standardized data and a relation graph, and terminating. According to the scheme, public opinion cause, response, action, interaction and ending full-link data are dynamically complemented through multiple rounds of closed loops of thinking, action and observation under the Re-Act framework, structured quality is optimized by combining multi-module processing, standardized data and a relation graph are output, public opinion dynamic characteristics are adapted, a complete and accurate support is provided for deep analysis and decision making, and the public opinion analysis and decision making efficiency is improved. And the limitation that the link is easy to break and the adaptability is poor in the traditional method is overcome.

Description

Technical Field

[0001] This invention relates to the field of public opinion data processing and intelligent structuring technology, specifically to a public opinion data structuring method based on the Re-Act framework. Background Technology

[0002] With the explosive growth of information from social media, news platforms, short videos, and other channels, public opinion data is characterized by fragmentation, multi-source nature, and dynamism. Enterprises, governments, and other entities have an increasingly urgent need for "structured full-chain public opinion events, clear understanding of multi-party relationships, and precise decision-making basis." To effectively address public opinion, it is necessary to capture the entire process from "cause-response-action-interaction-conclusion" and clarify the relationships among multiple stakeholders in order to achieve public opinion risk assessment and trend prediction.

[0003] The core demands of current public opinion data processing focus on three points: First, the integrity of the event chain, which needs to cover the key links from the initial outbreak to the subsequent conclusion, so as to avoid one-sided public opinion analysis due to missing data; second, the depth of data structuring, which needs to break down the data from multiple dimensions such as "industry, public opinion, and subject relationship", rather than just staying at the level of text extraction; and third, dynamic adaptability, which needs to adjust the collection strategy according to real-time data feedback to cope with sudden changes in public opinion events.

[0004] Existing technologies, such as the invention patent application with publication number CN119829723A, disclose an AI-driven method for structured storage and retrieval of document data. This method includes: analyzing and understanding document content using natural language processing (NLP) technology to extract structured question-and-answer knowledge and construct a structured question-and-answer knowledge base; analyzing and understanding user query requests and historical dialogues using NLP technology and a large-scale pre-trained language model to extract the user's true intent and needs, generating structured demand data; and vectorizing user input using a pre-trained language model based on user input and demand analysis results, extracting relevant information from a vector database through a dynamic retrieval strategy, and generating and outputting the final answer. This invention, through in-depth analysis of documents using NLP technology, extracts structured question-and-answer knowledge and utilizes the semantic understanding capabilities of a large model to improve the performance and user experience of the question-and-answer system.

[0005] As can be seen from the above solutions, traditional public opinion data processing methods have significant limitations: on the one hand, the reliance on a fixed "collection-structuring" model cannot dynamically supplement missing data, easily leading to breaks in the event chain; on the other hand, the lack of an efficient "decision-feedback" mechanism makes it difficult to identify ambiguous information and hidden gaps, and the structured results are insufficient to support in-depth public opinion analysis. Against this backdrop, there is an urgent need for a technical solution that integrates dynamic decision-making and closed-loop feedback to achieve full-chain structured processing of public opinion data. Summary of the Invention

[0006] To address the aforementioned technical shortcomings, the present invention aims to provide a method for structuring public opinion data based on the Re-Act framework.

[0007] To solve the above technical problems, the present invention adopts the following technical solution: The present invention provides a public opinion data structuring method based on the Re-Act framework, including the following steps: S1: Construct a public opinion data structuring system based on the Re-Act framework, which includes a thinking module, an action module, and an observation module; the thinking module has a built-in event element association reasoning rule library, which is integrated into the execution engine of the public opinion structuring system; the action module integrates search tools and structuring tools, the search tools are connected to a multi-source compliant public opinion knowledge base, and the structuring tools are used for information standardization transformation; the observation module adopts a three-layer hybrid architecture of data preprocessing layer, rule evaluation layer, and LLM semantic analysis layer.

[0008] S2: Through multiple rounds of thinking and decision-making, action execution, result observation and strategy iteration closed loop, the thinking module generates action instructions, the action module collects core information and detailed process data of public opinion events according to the instructions, the observation module provides feedback on data evaluation results, and iterates the above process until it covers the entire chain of cause, response, action, interaction and conclusion.

[0009] S3: Based on complete end-to-end data, the process terminates after extracting multi-dimensional data and sorting out the main relationships using structured tools, outputting standardized data and relationship graphs.

[0010] The beneficial effects of this invention are as follows: This invention provides a method for structuring public opinion data based on the Re-Act framework. First, a Re-Act public opinion structuring system containing thinking, action, and observation modules is constructed. Then, data is collected through multiple rounds of closed-loop iterations to cover the entire chain. Finally, based on the full-chain data, it is processed by structuring tools to output standardized data and relationship graphs before termination. This solution dynamically completes the entire chain of public opinion data—from the cause, response, action, interaction, and conclusion—through multiple rounds of closed-loop processing within the Re-Act framework. Combined with multi-module processing to optimize structuring quality, it outputs standardized data and relationship graphs, adapting to the dynamic characteristics of public opinion and providing complete and accurate support for in-depth analysis and decision-making. This overcomes the limitations of traditional methods, such as easily broken chains and poor adaptability. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1This is a schematic diagram of the implementation steps of the method of the present invention.

[0013] Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation

[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] See Figure 1 and Figure 2 As shown, a method for structuring public opinion data based on the Re-Act framework includes the following steps: S1: Constructing a public opinion data structuring system based on the Re-Act framework, which includes a thinking module, an action module, and an observation module; the thinking module has a built-in event element association reasoning rule base, which is integrated into the execution engine of the public opinion structuring system; the action module integrates search tools and structuring tools, the search tools are connected to a multi-source compliant public opinion knowledge base, and the structuring tools are used for information standardization and transformation; the observation module adopts a three-layer hybrid architecture of data preprocessing layer, rule evaluation layer, and LLM semantic analysis layer.

[0016] In a specific embodiment, the specific process of S1 is as follows: constructing a structured system of public opinion data based on the Re-Act framework, which consists of a thinking module, an action module, and an observation module.

[0017] The thinking module has a built-in event element association reasoning rule library, which is integrated into the public opinion structured system execution engine. It has the functions of receiving context status, generating reasoning statements, driving tool calls, and iteratively updating cognition.

[0018] The action module integrates search tools and structured tools. The search tools use semantic embedding vector retrieval combined with keyword inverted indexes to connect to multi-source compliance public opinion knowledge bases to obtain raw text and metadata. The structured tools are based on large model technology to transform verified information into standardized data formats for preset business dimensions.

[0019] The observation module adopts a three-layer hybrid architecture consisting of a data preprocessing layer, a rule evaluation layer, and an LLM semantic analysis layer. The data preprocessing layer performs preprocessing operations, the rule evaluation layer evaluates data coverage based on preset templates of core elements of public opinion events and quantitative indicators, and the LLM semantic analysis layer completes the transformation of fuzzy information, identification of implicit missing information, and verification of data integrity and relational logic.

[0020] It should be noted that the event element association reasoning rule base is the core data component of the thinking module, which stores the element association logic, tool mapping rules and data verification logic of each link of public opinion events.

[0021] The logic of element association in each link: refers to the correspondence and mutual triggering logic of core elements in the entire link of the cause, response, action, interaction and conclusion of the public opinion event, and within each link. It is the core basis for reasoning and instruction generation in the thinking module.

[0022] Tool mapping rules: These refer to the pre-defined calling rules in the rule base for the corresponding calling rules of tools in each link of public opinion events and action modules, which determine the types of tools required for data collection or standardized processing in different links.

[0023] Data verification logic refers to verifying the logical consistency of public opinion data within each link, the coherence of the relationship between different links, and the reasonableness of the matching between subject behavior and corresponding relationship. At the same time, it combines LLM semantic analysis to correct ambiguous data and identify the standardized logical standard of implicit missing data.

[0024] The execution engine of the public opinion structuring system is the operating platform for the thinking module. It is responsible for coordinating the data flow between the thinking module, the action module, and the observation module, and also provides context state storage and retrieval support for the thinking module.

[0025] Keyword inverted index: By constructing a mapping index table between keywords and data storage locations, the original data containing the target keywords in the knowledge base can be quickly located after receiving query terms, greatly improving the efficiency of data retrieval.

[0026] Multi-source compliant public opinion knowledge base: The data source for search tools integrates public opinion data from multiple platforms such as Weibo and news platforms, and all data complies with relevant data security and privacy protection regulations, ensuring that the collected original text and metadata are legal and usable.

[0027] Pre-set core link template for public opinion events: refers to a standardized framework that is set in advance and covers the entire chain of public opinion events, including the cause, response, action, interaction and conclusion, and is used by the rule evaluation layer to evaluate data coverage.

[0028] Quantitative evaluation indicators include quantifiable standards such as data coverage thresholds and data volume thresholds, which are used to objectively determine the data coverage status and data volume, avoiding subjective evaluation bias.

[0029] Preferably, the specific process of integrating the public opinion structuring system execution engine is as follows: within the execution engine of the public opinion structuring system, the workflow of the thinking module is as follows: accepting context state: including the user's original input, historical search results and extracted structured fragments.

[0030] Generate inference statements: Based on preset system prompts, guide the large language model to output inference statements that conform to the task logic.

[0031] Driving subsequent actions: The conclusion of the reasoning statement directly determines the next step of calling the tool.

[0032] Iterative updates to understanding: As new tools return results, the thinking module continuously corrects the reasoning path until all business dimension information is complete.

[0033] It should be noted that the preset system prompts are pre-configured instruction templates that conform to the structured task logic of public opinion data. The content includes task objectives, reasoning directions, and output specifications, which are used to guide the large language model to output logically coherent reasoning statements that meet the task requirements.

[0034] Preferably, the specific process of integrating search tools and structured tools into the action module is as follows: the action module does not make autonomous decisions, but strictly responds to the task requirements proposed by the thinking module and performs two types of key operations: information acquisition operations: by calling the search tool, it retrieves the original text related to the current event from the multi-source public opinion database.

[0035] Result generation operations: By calling structured tools, verified information is transformed into a standardized data format that conforms to preset business dimensions.

[0036] The action module contains two core sub-tools: (1) Search tool: Function: Receive natural language query terms, connect to the pre-built public opinion knowledge base, and return highly relevant original text fragments and their metadata.

[0037] Call format: Search[query], where query is dynamically generated by the Thinking module and can include entities, time ranges and domain qualifiers.

[0038] Technical implementation: Vector retrieval based on semantic embedding combined with keyword inverted index, supporting multi-hop expansion and authority-weighted sorting.

[0039] (2) Structured tools: Function: After the thinking module determines that all dimensional information is complete, this tool is called to integrate the scattered observation results into a unified structured data object.

[0040] Call format: Structuring[text], where text is the news that matches the search results.

[0041] Technical implementation: Structured extraction of news information based on the capabilities of large models.

[0042] It should be noted that the preset business dimensions refer to the core classification dimensions of the data extracted by structured tools, which are set in advance according to the needs of public opinion analysis. These dimensions include the main characteristics and event attributes of the elements in the link process.

[0043] Standardized data format: The output format of result generation operations refers to the data format that conforms to the preset business dimensions and unified structural specifications.

[0044] Pre-built public opinion knowledge base: The data source carrier of the search tool is a knowledge base formed by integrating compliant public opinion data from multiple channels in advance.

[0045] Multi-hop expansion: The search tool automatically extends its search to retrieve data that is logically related to the initially retrieved data related to the public opinion event, thereby expanding the scope of data collection.

[0046] Authority-weighted ranking: The search tool assigns weights to search results based on the authority of the original text source, and outputs the results in descending order of weight to prioritize the provision of highly credible data.

[0047] S2: Through multiple rounds of thinking and decision-making, action execution, result observation and strategy iteration closed loop, the thinking module generates action instructions, the action module collects core information and detailed process data of public opinion events according to the instructions, the observation module provides feedback on data evaluation results, and iterates the above process until it covers the entire chain of cause, response, action, interaction and conclusion.

[0048] In a specific embodiment, the specific process of S2 is as follows: initiating a closed-loop process of multi-round thinking and decision-making, action execution, result observation and strategy iteration, the thinking module generates action instructions containing tool type and collection parameters based on the context state and the built-in event element association reasoning rule base.

[0049] The action module responds to action commands and connects to a multi-source compliant public opinion knowledge base through integrated search tools to collect core information and detailed process data of public opinion events in a targeted manner.

[0050] The observation module processes the collected data through a data preprocessing layer, a rule evaluation layer, and an LLM semantic analysis layer, generating data evaluation results that include the coverage of links and the identification of missing links, and then feeding them back to the thinking module.

[0051] The thinking module iteratively modifies action instructions based on feedback results, repeating the above closed-loop process until the collected data fully covers the entire chain of public opinion events, including the cause, response, action, interaction, and conclusion.

[0052] Preferably, the specific process of generating action instructions containing tool type and collection parameters is as follows: the thinking module first obtains the current context state of public opinion data collection and calls the built-in event element association reasoning rule library.

[0053] The historical data collection records in the current context are compared with the covered link information and the entire link of the public opinion event to obtain the uncovered links that have not acquired data and the data-insufficient links whose data volume is less than the preset data volume threshold, and the specific type of the two types of links is recorded.

[0054] Based on the mapping relationship between the uncovered links and links with insufficient data in the rule base and the search tool, the identified link types are matched, and it is determined that the current tool type to be called is the search tool integrated into the action module, so as to supplement the missing data of the corresponding link through this tool.

[0055] Based on the event element association logic in the rule base, and combined with the core element requirements of the currently missing link, collection parameters are generated, including data collection dimensions, related element filtering conditions, and knowledge base docking scope.

[0056] Finally, the tool types and collected parameters are integrated to form structured action instructions.

[0057] It should be noted that the rule base quantification requirements refer to the quantifiable standards set for data collection at each public opinion link in the event element association reasoning rule base, such as data volume thresholds.

[0058] The preset data volume threshold is a critical value used to determine whether the data volume meets the standard. It is set by professionals according to the collection needs, and no specific numerical limit is set here.

[0059] Related element filtering conditions: These refer to the filtering rules set by the thinking module based on the correlation logic of event elements and the characteristics of missing links in order to avoid invalid data collection and accurately locate target data.

[0060] Knowledge base integration scope: refers to the sub-scope of the multi-source compliant public opinion knowledge base specified by the thinking module from which the search tool needs to collect data, such as specifying news information platforms when collecting data on the cause stage and specifying social media comment sections when collecting data on the interaction stage.

[0061] Structured action instructions: These are instructions generated by the thinking module based on the context state and the rule base for associating event elements. First, the tool type is determined by the mapping logic between links and tools. Then, the core element requirements of the missing links are combined to generate collection parameters containing data collection dimensions, related element filtering conditions, and knowledge base docking scope. Finally, the two are integrated to generate instructions in a standardized format that can be directly parsed by the action module.

[0062] Preferably, the specific process of targeted collection of core information and detailed process data of public opinion events is as follows: the action module first parses the action instructions issued by the thinking module and determines the tool type and collection parameters specified in the instructions.

[0063] The search tool uses semantic embedding vector retrieval combined with keyword inverted indexing and a multi-source compliant public opinion knowledge base. First, the keyword inverted indexing is used to initially locate the original text and metadata related to the public opinion event in the knowledge base. Then, the semantic embedding vector retrieval is used to filter the semantic relevance of the initial location results, obtaining data with a similarity greater than the preset threshold, and extracting content that matches the core information and subdivided process data of the public opinion event.

[0064] Finally, the extracted content is preprocessed to ensure that the collected content corresponds to the target link, and then the processed core information and subdivided process data are output to the observation module.

[0065] It should be noted that the semantic embedding vector retrieval transforms the original text in the public opinion knowledge base and the collection requirements in the action instructions into high-dimensional semantic vectors. By calculating the similarity between vectors, data with high semantic relevance to the collection requirements is selected.

[0066] Semantic relevance filtering: After the search tool initially locates data through keyword inverted index, it uses semantic embedding vector retrieval technology to calculate the semantic vector similarity between the initial location results and the collection requirements in the action instructions, retains highly relevant data according to a preset similarity threshold, and removes semantically irrelevant data.

[0067] The preset similarity threshold is a critical value used to determine whether the similarity is acceptable. It is set by professionals according to the collection requirements, and no specific numerical limit is imposed here.

[0068] Preferably, the specific process of generating data evaluation results containing the coverage of links and the identification of missing links and feeding them back to the thinking module is as follows: the observation module first receives the core information of the public opinion event and the subdivided process data output by the action module, and transmits them to the data preprocessing layer for preprocessing operations to obtain the preprocessed data to be evaluated.

[0069] Then, the data to be evaluated is input into the rule evaluation layer. The rule evaluation layer calls the preset templates for the entire chain of public opinion events, including the cause, response, action, interaction and conclusion, and the corresponding quantitative evaluation indicators. It verifies the coverage status and data validity of the data to be evaluated in each chain link, and generates the preliminary link coverage status of each link, including those that are covered, partially covered and not covered. At the same time, it marks the links that are not covered or have insufficient data as preliminary missing links.

[0070] The preliminary results output by the rule evaluation layer are then passed to the LLM semantic analysis layer. The LLM semantic analysis layer identifies and completes fuzzy data and implicit missing information through contextual semantic association analysis, verifies and corrects the coverage of the preliminary links and the identification of missing links, and determines the specific type of missing links.

[0071] Finally, observe the coverage of the integrated and corrected links and the identification of missing links, generate data evaluation results in a preset structured format, and transmit them to the thinking module through the feedback interface of the public opinion structured system execution engine.

[0072] It should be noted that the identification and completion judgment of ambiguous data and implicit missing information are as follows: if the data can be deduced from the contextual semantic logic of the existing link data, data completion suggestions can be generated; if it cannot be deduced from the existing data, it is marked as missing information that needs to be collected.

[0073] Verify and correct the initial coverage and missing link identification, and determine the specific type of missing link: If a partially covered link is identified as having a hidden missing element, causing the satisfaction level of the core element item in that link to be less than the preset partial coverage threshold, it is corrected to be uncovered; if a previously uncovered link is corrected as having a core element item satisfaction level greater than the preset partial coverage threshold or the preset covered threshold after the fuzzy data is clarified and key elements are added, it is corrected to the corresponding coverage status; if a previously covered link is identified as having a hidden missing element that was not captured by the rule evaluation layer, and this missing element affects the integrity of the link, it is corrected to be partially covered.

[0074] The preset partial coverage threshold is a critical value used to determine whether partial coverage is acceptable. It is set by professionals according to the correction requirements, and no specific numerical limit is set here.

[0075] The preset coverage threshold is a critical value used to determine whether the coverage is qualified. It is set by professionals according to the correction needs, and no specific numerical limit is set here.

[0076] Preset structured format: The unified format standard for the data evaluation results generated by the observation module is set in advance according to the instructions and correction requirements of the thinking module, including fixed fields such as the coverage status after each link correction and the specific type of missing link.

[0077] S3: Based on complete end-to-end data, the process terminates after extracting multi-dimensional data and sorting out the main relationships using structured tools, outputting standardized data and relationship graphs.

[0078] In a specific embodiment, the specific process of S3 is as follows: based on the acquired complete data of the entire chain of public opinion events, the structured tool first extracts the core element data corresponding to each stage of the public opinion event's cause, response, action, interaction and conclusion according to the preset business dimensions.

[0079] Then, through large-scale modeling technology, we analyze the logical relationships between data, and sort out the types of public opinion participants and related parties and their interactions.

[0080] Then, the data consistency verification mechanism is used to correct the correlation deviation, and finally, standardized data that conforms to the preset format and a relationship graph with the subject or element as nodes and the correlation relationship as edges are output. After completion, the entire structured process is terminated.

[0081] It should be noted that the participants in public opinion events refer to the entities that directly initiate, execute, or lead key actions in public opinion events, such as the entity that triggers the event and the entity that executes the response.

[0083] Stakeholders in public opinion events refer to entities that indirectly participate in public opinion events, influence the development of the events, but do not dominate the core processes, such as third-party evaluation agencies involved in the interactive process and public groups providing feedback.

[0084] Classification of Subjects and Related Parties: This refers to the process of classifying the roles of participating subjects and related parties based on a pre-defined classification system of subjects and related parties and the relationship logic analyzed by the large model. For example, participating subjects can be classified as cause-triggered, response-execution, and action-implementation types, while related parties can be classified as third-party intervention, public interaction, and regulatory supervision types.

[0085] Interactions between subjects and related parties: This refers to the interactions between participating subjects and between participating subjects and related parties, based on the logic of data relationships. Common types include triggering and responding, and execution and monitoring.

[0086] Data consistency verification mechanism: To ensure the logical accuracy of structured public opinion data, a verification dimension is defined based on preset logical rules and subject relationship constraints for the entire chain of public opinion events, including cause, response, action, interaction, and conclusion. The similarity between the basic data and the verification dimension is compared to identify correlation deviations such as data contradictions, broken associations, and misaligned relationships. The cause of the deviation is determined by combining the semantic tracing of the observation module (LLM) and the event element association reasoning rule base of the thinking module. Corrections are performed according to the cause, and the corrected data is re-verified until it meets the preset verification pass threshold.

[0087] Verification dimensions refer to those determined based on the logical rules and subject relationship constraints of the entire link, including the logical consistency of data within the same link link, the consistency of data association between different link links, and the reasonableness of matching subject behavior with corresponding relationships, to determine whether there is any correlation deviation in the basic data.

[0088] Data contradiction: refers to a conflict in basic data within the same logical consistency dimension, such as the response time within the same response chain being recorded as both October 6th and October 7th.

[0089] Disconnection: This refers to the lack of basic data in different links and their coherence. For example, the response link promises to rectify within 3 days, but the action link has no data on any rectification measures.

[0090] Relationship misalignment: refers to an anomaly in the basic data regarding the reasonableness of the match between the subject's behavior and the relationship, such as the release of official event conclusions by public stakeholders.

[0091] The preset verification pass threshold is a critical value used to determine whether the verification passes or fails. It is set by professionals according to the correction requirements, and no specific numerical limit is set here.

[0092] Preset format: refers to a standardized data presentation format that is set in advance according to the needs of public opinion analysis. It includes the core elements of each link in the entire chain of public opinion event, such as the cause, response, action, interaction and conclusion, as well as the types of participating entities and related parties and their interaction relationships.

[0093] Preferably, the specific process of correcting the correlation deviation by combining the data consistency verification mechanism is as follows: obtaining the multi-dimensional extracted data and subject relationship sorting results initially output by the structured tool as the basic data for consistency verification.

[0094] Then, based on the pre-set logical rules and subject relationship constraints of the entire chain of public opinion events, including the cause, response, action, interaction and conclusion.

[0095] Then, by comparing the similarity between the basic data and the verification dimensions through data consistency verification, we can identify correlation deviations such as data contradictions, broken associations, and misaligned relationships.

[0096] Then, the LLM semantic analysis layer of the observation module is called to perform semantic source tracing of the correlation deviation, and the cause of the deviation is determined by combining the event element correlation reasoning rule library built into the thinking module.

[0097] Targeted corrections are made based on the cause of the deviation: if the deviation is due to missing data, the action module is triggered to collect key data from the corresponding link; if the deviation is due to data conflict, data that conforms to the development pattern of the public opinion event is retained.

[0098] Finally, the consistency check is re-executed on the corrected data and subject relationships until the preset check pass threshold is met, thus completing the correlation deviation correction.

[0099] The examples described in this invention are not limited to the specific embodiments listed above. The examples are merely illustrative to facilitate understanding of the invention and do not constitute a limitation on the scope of protection of this invention. Any modifications, equivalent substitutions, etc., made within the spirit and principles of this invention should be included within the scope of protection.

[0100] The above description is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined in this specification, they should all fall within the protection scope of the present invention.

Claims

1. A method for structuring public opinion data based on the Re-Act framework, characterized in that, Includes the following steps: S1: Construct a structured public opinion data system based on the Re-Act framework. This system includes a thinking module, an action module, and an observation module. The thinking module has a built-in event element association reasoning rule library, which is integrated into the execution engine of the structured public opinion system. The action module integrates search tools and structured tools. The search tools are connected to a multi-source compliant public opinion knowledge base, and the structured tools are used for information standardization and transformation. The observation module adopts a three-layer hybrid architecture consisting of a data preprocessing layer, a rule evaluation layer, and an LLM semantic analysis layer. S2: Through multiple rounds of thinking and decision-making, action execution, result observation and strategy iteration closed loop, the thinking module generates action instructions, the action module collects core information and detailed process data of public opinion events according to the instructions, the observation module provides feedback on data evaluation results, and iterates the above process until it covers the entire chain of cause, response, action, interaction and conclusion. S3: Based on complete end-to-end data, the process terminates after extracting multi-dimensional data and sorting out the main relationships using structured tools, outputting standardized data and relationship graphs.

2. The method for structuring public opinion data based on the Re-Act framework according to claim 1, characterized in that, The specific process of S1 is as follows: Construct a structured system for public opinion data based on the Re-Act framework, which consists of a thinking module, an action module, and an observation module; The thinking module has a built-in event element association reasoning rule library, which is integrated into the execution engine of the public opinion structuring system. It has the functions of receiving context status, generating reasoning statements, driving tool calls, and iteratively updating cognition. The action module integrates search tools and structured tools. The search tools use semantic embedding vector retrieval combined with keyword inverted index to connect to a multi-source compliance public opinion knowledge base to obtain raw text and metadata. The structured tools are based on large model technology to transform the verified information into a standardized data format for preset business dimensions. The observation module adopts a three-layer hybrid architecture consisting of a data preprocessing layer, a rule evaluation layer, and an LLM semantic analysis layer. The data preprocessing layer performs preprocessing operations, the rule evaluation layer evaluates data coverage based on preset templates of core elements of public opinion events and quantitative indicators, and the LLM semantic analysis layer completes the transformation of fuzzy information, identification of implicit missing information, and verification of data integrity and relational logic.

3. The method for structuring public opinion data based on the Re-Act framework according to claim 2, characterized in that, The specific process of integrating the execution engine of the public opinion structuring system is as follows: Within the execution engine of the public opinion structuring system, the workflow of the thinking module is as follows: Accepts contextual state: including original user input, historical search results, and extracted structured fragments; Generate inference statements: Based on preset system prompts, guide the large language model to output inference statements that conform to the task logic; Driving subsequent actions: The conclusion of the reasoning statement directly determines the next step of calling the tool; Iterative updates to understanding: As new tools return results, the thinking module continuously corrects the reasoning path until all business dimension information is complete.

4. The method for structuring public opinion data based on the Re-Act framework according to claim 2, characterized in that, The specific process of integrating search tools and structured tools into the action module is as follows: The Action module does not make autonomous decisions, but strictly responds to the task requirements proposed by the Thinking module, executing two types of key operations: Information retrieval operations: By invoking search tools, retrieve raw text related to the current event from multi-source public opinion databases; Result generation operations: By calling structured tools, verified information is transformed into a standardized data format that conforms to preset business dimensions; The action module contains two core sub-tools: (1) Search tools: Function: Receives natural language query terms, connects to a pre-built public opinion knowledge base, and returns highly relevant raw text fragments and their metadata; Call format: Search[query], where query is dynamically generated by the thinking module and can include entities, time ranges, and domain qualifiers; Technical implementation: Vector retrieval based on semantic embedding combined with keyword inverted index, supporting multi-hop expansion and authority-weighted sorting; (2) Structured tools: Function: After the thinking module determines that all dimensions of information are complete, this tool is called to integrate the scattered observation results into a unified structured data object; Call format: Structuring[text], where text is the news item that matches the search results; Technical implementation: Structured extraction of news information based on the capabilities of large models.

5. A method for structuring public opinion data based on the Re-Act framework according to claim 1, characterized in that, The specific process of S2 is as follows: Initiate a closed-loop process of multi-round thinking and decision-making, action execution, result observation and strategy iteration. The thinking module generates action instructions containing tool type and collection parameters based on the context state and the built-in event element association reasoning rule base. The action module responds to action commands and connects to a multi-source compliant public opinion knowledge base through integrated search tools to collect core information and detailed process data of public opinion events in a targeted manner. The observation module processes the collected data through a data preprocessing layer, a rule evaluation layer, and an LLM semantic analysis layer, generating data evaluation results that include the coverage of links and the identification of missing links, and then feeding them back to the thinking module. The thinking module iteratively modifies action instructions based on feedback results, repeating the above closed-loop process until the collected data fully covers the entire chain of public opinion events, including the cause, response, action, interaction, and conclusion.

6. A method for structuring public opinion data based on the Re-Act framework according to claim 5, characterized in that, The specific process for generating action instructions that include tool type and acquisition parameters is as follows: The thinking module first obtains the current context state of public opinion data collection and then calls the built-in event element association reasoning rule library; By comparing the historical collection records and covered link information in the current context with the entire link of the public opinion event, we can identify the uncovered links for which no data was acquired and the links for which the data volume does not meet the quantitative requirements of the rule base, and record the specific type of the two types of links. Based on the mapping relationship between the uncovered links and links with insufficient data in the rule base and the search tool, the identified link types are matched, and it is determined that the current tool type to be called is the search tool integrated into the action module, so as to supplement the missing data of the corresponding link through this tool; Based on the event element association logic in the rule base, and combined with the core element requirements of the currently missing link, collection parameters are generated, including data collection dimensions, related element filtering conditions, and knowledge base docking scope. Finally, the tool types and collected parameters are integrated to form structured action instructions.

7. A method for structuring public opinion data based on the Re-Act framework according to claim 5, characterized in that, The specific process for targeted collection of core information and detailed process data of public opinion events is as follows: The action module first parses the action instructions issued by the thinking module to determine the tool type and data collection parameters specified in the instructions; The search tool uses semantic embedding vector retrieval combined with keyword inverted indexing and a multi-source compliant public opinion knowledge base. First, the keyword inverted indexing is used to initially locate the original text and metadata related to the public opinion event in the knowledge base. Then, the semantic embedding vector retrieval is used to filter the semantic relevance of the initial location results to obtain data with a similarity greater than the preset threshold. Content that matches the core information and subdivided process data of the public opinion event is then extracted. Finally, the extracted content is preprocessed to ensure that the collected content corresponds to the target link, and then the processed core information and subdivided process data are output to the observation module.

8. A method for structuring public opinion data based on the Re-Act framework according to claim 5, characterized in that, The specific process of generating data evaluation results that include link coverage and missing link identifiers and feeding them back to the thinking module is as follows: The observation module first receives the core information and detailed process data of the public opinion event output by the action module, and then transmits them to the data preprocessing layer for preprocessing to obtain the preprocessed data to be evaluated. Then, the data to be evaluated is input into the rule evaluation layer. The rule evaluation layer calls the preset templates for the entire chain of public opinion events, including the cause, response, action, interaction and conclusion, and the corresponding quantitative evaluation indicators. It verifies the coverage status and data validity of the data to be evaluated in each chain link, and generates the preliminary link coverage status of each link, including those that are covered, partially covered and not covered. At the same time, it marks the links that are not covered or have insufficient data as preliminary missing links. The preliminary results output by the rule evaluation layer are then passed to the LLM semantic analysis layer. The LLM semantic analysis layer identifies and completes fuzzy data and implicit missing information through contextual semantic association analysis, verifies and corrects the coverage of the preliminary links and the identification of missing links, and determines the specific type of missing links. Finally, observe the coverage of the integrated and corrected links and the identification of missing links, generate data evaluation results in a preset structured format, and transmit them to the thinking module through the feedback interface of the public opinion structured system execution engine.

9. A method for structuring public opinion data based on the Re-Act framework according to claim 1, characterized in that, The specific process of S3 is as follows: Based on the complete data of the entire chain of public opinion events, the structured tool first extracts the core element data corresponding to each stage of the public opinion event, including the cause, response, action, interaction and conclusion, according to the preset business dimensions. Then, through large-scale modeling technology, the logical relationships between data are analyzed to sort out the types of public opinion participants and related parties and their interactions. Then, the data consistency verification mechanism is used to correct the correlation deviation, and finally, standardized data that conforms to the preset format and a relationship graph with the subject or element as nodes and the correlation relationship as edges are output. After completion, the entire structured process is terminated.

10. A method for structuring public opinion data based on the Re-Act framework according to claim 9, characterized in that, The specific process of correcting correlation deviations using the data consistency verification mechanism is as follows: Obtain the multi-dimensional extracted data and subject relationship sorting results initially output by the structured tools as the basic data for consistency verification; Furthermore, based on the pre-defined logical rules and subject relationship constraints for the entire chain of public opinion events, including the cause, response, action, interaction, and conclusion; Then, by comparing the similarity between the basic data and the verification dimensions through data consistency verification, we can identify correlation deviations such as data contradictions, broken associations, and misaligned relationships. Then, the LLM semantic analysis layer of the observation module is called to perform semantic source tracing of the correlation deviation, and the cause of the deviation is determined by combining the event element correlation reasoning rule library built into the thinking module. Targeted corrections are performed based on the cause of the deviation: if the deviation is due to missing data, the action module is triggered to collect key data from the corresponding link. In the event of a data conflict, retain the data that aligns with the development pattern of the public opinion event. Finally, the consistency check is re-executed on the corrected data and subject relationships until the preset check pass threshold is met, thus completing the correlation deviation correction.

Images