CN-121979937-A - Public opinion data structuring method based on Re-Act framework

CN121979937ACN 121979937 ACN121979937 ACN 121979937ACN-121979937-A

Abstract

The invention discloses a public opinion data structuring method based on Re-Act framework, which relates to the technical field of public opinion data processing and intelligent structuring, the invention firstly constructs an Re-Act public opinion structuring system comprising a thinking module, a moving module and an observing module, and finally, based on the full-link data, the data are processed by a structuring tool, and the standardized data and the relation map are output and then are terminated. The scheme optimizes the structured quality by combining multi-module processing, outputs standardized data and relationship graphs, adapts public opinion dynamic characteristics, provides complete and accurate support for deep analysis and decision, and overcomes the limitations of easy breakage and poor adaptability of the traditional method link through multiple rounds of closed loop dynamic complement of public opinion cause, response, action, interaction and ending data under the Re-Act framework.

Inventors

LIU SHAOBING
WANG HAIRONG
LV XIAOBAO

Assignees

曙光天玑数据科技(江苏)有限公司
中科曙光南京研究院有限公司

Dates

Publication Date: 20260505
Application Date: 20260106

Claims (10)

1. The public opinion data structuring method based on Re-Act framework is characterized by comprising the following steps: The method comprises the steps of S1, constructing a public opinion data structuring system based on an Re-Act framework, wherein the system comprises a thinking module, a action module and an observation module, wherein an event element association reasoning rule base is arranged in the thinking module and is integrated with an execution engine of the public opinion structuring system; S2, performing closed loop iteration through multiple rounds of thinking decision, action execution, result observation and strategy, wherein the thinking module generates an action instruction, the action module directionally collects public opinion event core information and subdivision process data according to the instruction, the observation module feeds back data evaluation results, and the process is iterated until the cause, response, action, interaction and ending all links are covered; And S3, based on complete full-link data, multi-dimensional data extraction and main body relation carding are completed through the structuring tool, and the flow is terminated after standardized data and relation maps are output.
2. The public opinion data structuring method based on Re-Act framework of claim 1, wherein the specific process of S1 is as follows: constructing a public opinion data structuring system based on an Re-Act framework, wherein the system consists of a thinking module, a action module and an observation module; the thinking module is internally provided with an event element association reasoning rule base which is integrated with an execution engine of the public opinion structuring system and has the functions of receiving a context state, generating reasoning sentences, driving a tool to call and iteratively updating cognition; The action module integrates a search tool and a structuring tool, the search tool is combined with keyword inverted index through semantic embedded vector retrieval, and is used for butting a multi-source public opinion knowledge base to obtain an original text and metadata, and the structuring tool is used for converting verified information into a standardized data format with preset service dimensions based on a large model technology; The observation module adopts a three-layer mixed architecture of a data preprocessing layer, a rule evaluation layer and an LLM semantic analysis layer, the data preprocessing layer performs preprocessing operation, the rule evaluation layer performs data coverage evaluation according to a preset public opinion event core link template and a quantization index, and the LLM semantic analysis layer completes fuzzy information conversion, implicit deletion identification and data integrity and relationship logic verification.
3. The public opinion data structuring method based on Re-Act framework of claim 2, wherein the specific process integrated in the public opinion structuring system execution engine is as follows: in the execution engine of the public opinion structuring system, the workflow of the thinking module is as follows: accepting a context state comprising original input of a user, historical search results and extracted structured fragments; Generating an inference statement, namely guiding the large language model to output the inference statement conforming to task logic based on a preset system prompt; driving follow-up actions, namely directly determining the conclusion of the reasoning statement to call a tool in the next step; and iteratively updating cognition, namely continuously correcting an reasoning path by a thinking module along with the injection of a return result of a new tool until all business dimension information is complete.
4. The public opinion data structuring method based on Re-Act framework of claim 2, wherein the specific process of integrating the search tool and the structuring tool by the action module is as follows: The action module does not make an autonomous decision, but strictly responds to task demands proposed by the thinking module, and performs two key operations: an information acquisition class operation, namely retrieving an original text related to a current event from a multi-source public opinion database by calling a search tool; the result generation class operation, namely converting the verified information into a standardized data format conforming to a preset service dimension by calling a structuring tool; the action module includes two core sub-tools: (1) Search tool: receiving natural language query words, butting a pre-constructed public opinion knowledge base, and returning original text fragments with high correlation degree and metadata thereof; The calling format is Search query, wherein the query is dynamically generated by a thinking module and can contain entity, time range and domain definition words; the method is realized by combining vector search based on semantic embedding with keyword inverted index to support multi-hop expansion and authoritative weighted sorting; (2) Structuring tool: the function is that after the thinking module judges that all dimension information is complete, the tool is called to integrate the scattered observation results into a uniform structured data object; the calling format is structure [ text ], wherein the text is news meeting the search; the technology is realized by carrying out structural extraction on news information based on the capability of a large model.
5. The public opinion data structuring method based on Re-Act framework of claim 1, wherein the specific process of S2 is as follows: Starting a multi-round thinking decision, action execution, result observation and strategy iteration closed-loop flow, and generating an action instruction containing tool types and acquisition parameters by a thinking module based on a context state and a built-in event element association reasoning rule base; The action module responds to the action instruction, interfaces with the multi-source co-opinion knowledge base through the integrated search tool, and directionally collects core information of public opinion events and subdivision process data; The observation module processes the acquired data through the data preprocessing layer, the rule evaluation layer and the LLM semantic analysis layer, generates a data evaluation result containing link coverage conditions and missing link identifiers and feeds the data evaluation result back to the thinking module; the thinking module iteratively corrects the action instruction according to the feedback result, and repeats the closed loop flow until the acquired data completely covers the public opinion event cause, response, action, interaction and ending all links.
6. The public opinion data structuring method based on Re-Act framework of claim 5, wherein the specific process of generating the action instructions comprising tool types and acquisition parameters is as follows: the thinking module firstly acquires the context state of current public opinion data acquisition and invokes a built-in event element association reasoning rule base; Comparing the historical acquisition record in the current context state with the covered link information and the public opinion event full link to obtain an uncovered link of the unobtained data and a data insufficient link of which the data quantity does not meet the quantitative requirement of the rule base, and recording the specific types of the two types of links; Matching the identified link types according to the mapping relation between the uncovered links and the insufficient links and the search tools in the rule base, and judging the type of the tool to be called currently as the search tool integrated by the action module so as to directionally supplement and collect the missing data of the corresponding links through the tool; Based on event element association logic in a rule base, combining the core element requirements of the current missing link to generate acquisition parameters comprising data acquisition dimensions, association element screening conditions and a knowledge base docking range; finally, integrating the tool type and the acquisition parameters to form a structured action instruction.
7. The public opinion data structuring method based on Re-Act framework of claim 5, wherein the specific process of directionally collecting public opinion event core information and subdivision process data is as follows: The action module firstly analyzes the action instruction issued by the thinking module, and determines the type of the tool designated in the instruction and the acquisition parameters; the method comprises the steps that a search tool starts semantic embedded vector retrieval, combines keyword inverted indexes and a multi-source public opinion knowledge base, initially locates original texts and metadata related to public opinion events in the knowledge base through the keyword inverted indexes, performs semantic relevance screening on a preliminary locating result through semantic embedded vector retrieval to obtain data larger than a preset similarity threshold, and extracts content matched with public opinion event core information and subdivision process data; And finally, preprocessing the extracted content to enable the acquired content to correspond to the target link, and outputting the processed core information and subdivision process data to an observation module.
8. The public opinion data structuring method based on Re-Act framework of claim 5, wherein the specific process of generating the data evaluation result comprising the link coverage condition and the missing link identification and feeding back to the thinking module is as follows: The observation module firstly receives public opinion event core information and subdivision process data output by the action module, and transmits the public opinion event core information and subdivision process data to the data preprocessing layer for preprocessing operation to obtain preprocessed data to be evaluated; Then inputting the data to be evaluated into a rule evaluation layer, calling a preset public opinion event origin, response, action, interaction and ending all-link template and a corresponding quantitative evaluation index by the rule evaluation layer, checking the coverage state and the data validity of the data to be evaluated in each link one by one, generating the preliminary link coverage conditions of covered, partially covered and uncovered links of each link, and marking the links which are not covered or are not enough as missing link preliminary identifications; Then, the preliminary result output by the rule evaluation layer is transmitted to the LLM semantic analysis layer, the LLM semantic analysis layer carries out recognition and completion judgment on fuzzy data and hidden missing information through up-down Wen Yuyi association analysis, the preliminary link coverage condition and missing link identification are verified and corrected, and the specific type of the missing link is determined; And finally, integrating the corrected link coverage condition and the missing link identification by the observation module, generating a data evaluation result according to a preset structural format, and transmitting the data evaluation result to the thinking module through a feedback interface of the public opinion structural system execution engine.
9. The public opinion data structuring method based on Re-Act framework of claim 1, wherein the specific process of S3 is as follows: based on the obtained complete data of the public opinion event full link, the structuring tool firstly extracts core element data corresponding to links of the cause, response, action, interaction and ending of the public opinion event according to preset service dimension; Analyzing the association logic between the data by a large model technology, and combing the type division and interaction relation of the public opinion participation main body and the association party; and then, correcting the association deviation by combining a data consistency checking mechanism, and finally outputting standardized data conforming to a preset format and a relationship graph taking a main body or an element as a node and an association relationship as an edge, and ending the whole structuring flow after finishing.
10. The public opinion data structuring method based on Re-Act framework of claim 9, wherein the specific process of correcting association deviation by combining with a data consistency check mechanism is as follows: Obtaining multi-dimensional extraction data and a main body relation carding result which are preliminarily output by a structuring tool and taking the multi-dimensional extraction data and the main body relation carding result as basic data of consistency verification; Based on preset public opinion event cause, response, action, interaction and ending all-link logic rules and subject relation constraint conditions; The similarity between the basic data and the verification dimension is checked and compared through data consistency, and the association deviation of data contradiction, association fracture and relationship dislocation is identified; then, invoking an LLM semantic analysis layer of the observation module to carry out semantic tracing on the association deviation, and judging a deviation cause by combining an event element association reasoning rule base built in the thinking module; the targeted correction is executed according to the deviation cause, namely, if the data is missing, an action module is triggered to collect the key data of the corresponding link in a supplementary mode, and if the data conflict, the data conforming to the development rule of the public opinion event is reserved; and finally, re-executing consistency verification on the corrected data and the main body relationship until a preset verification passing threshold is met, and finishing association deviation correction.

Description

Public opinion data structuring method based on Re-Act framework Technical Field The invention relates to the technical field of public opinion data processing and intelligent structuring, in particular to a public opinion data structuring method based on an Re-Act framework. Background Along with the explosive growth of multi-channel information such as social media, news platforms, short videos and the like, public opinion data presents the characteristics of fragmentation, multisourcing and dynamics, and the demands of bodies such as enterprises, government and the like on the public opinion event full-link structuring, multi-party relation definition and decision basis precision are increasingly urgent. Aiming at public opinion, the public opinion risk assessment and trend prejudgement can be realized by completely capturing the whole flow of 'cause-response-action-interaction-ending', and combing the main relations of multiple parties. The core requirements of current public opinion data processing are concentrated on three points, namely firstly, event link integrity needs to be covered from initial burst to subsequent ending, the situation that public opinion analysis is unilateral due to data loss is avoided, secondly, data structuring depth needs to be achieved by multi-dimensional data disassembly from industry, wind assessment, subject relation and the like instead of only staying on a text extraction level, thirdly, dynamic adaptability needs to be achieved, and the collection strategy needs to be adjusted according to real-time data feedback, so that sudden changes of public opinion events are dealt with. The AI-driven document data structured storage and retrieval method disclosed in the patent application of the invention with the publication number of CN119829723A in the prior art comprises the steps of analyzing and understanding document contents by utilizing a natural language processing technology, extracting structured question-answer knowledge, constructing a structured question-answer knowledge base, analyzing and understanding a query request and a historical dialogue of a user by utilizing the natural language processing technology and a large-scale pre-training language model, extracting real intention and demand of the user, generating structured demand data, vectorizing user input by utilizing the pre-training language model according to user input and demand analysis results, extracting relevant information from a vector database by utilizing a dynamic retrieval strategy, and generating and outputting a final answer. According to the invention, the document is deeply analyzed by the natural language processing technology, the structured question-answering knowledge is extracted, and the semantic understanding capability of a large model is utilized, so that the performance and user experience of the question-answering system are improved. As can be seen from the scheme, the traditional public opinion data processing method has obvious limitations that on one hand, the method depends on an acquisition-structuring mode of a fixed flow, cannot dynamically supplement missing link data and is easy to break an event link, and on the other hand, the method lacks an efficient decision-feedback mechanism, so that fuzzy information and implicit missing are difficult to identify, and a structuring result is difficult to support deep public opinion analysis. Under the background, a technical scheme for integrating dynamic decision and closed loop feedback is needed to realize the full-link structured processing of public opinion data. Disclosure of Invention Aiming at the technical defects, the invention aims to provide a public opinion data structuring method based on an Re-Act framework. The invention provides a public opinion data structuring method based on an Re-Act framework, which comprises the following steps of S1, constructing a public opinion data structuring system based on the Re-Act framework, wherein the system comprises a thinking module, a action module and an observation module, wherein the thinking module is internally provided with an event element association reasoning rule base and is integrated with an execution engine of a public opinion structuring system, the action module integrates a search tool and a structuring tool, the search tool is in butt joint with a multi-source public opinion knowledge base, the structuring tool is used for information standardization conversion, and the observation module adopts a three-layer hybrid structure of a data preprocessing layer, a rule evaluation layer and an LLM semantic analysis layer. S2, through multi-round thinking decision making, action execution, result observation and strategy iteration closed loop, the thinking module generates action instructions, the action module directionally collects public opinion event core information and subdivision process data according to the instructions, the observation module feeds b