CN-121979924-A - Educational policy retrieval method and system integrating policy corpus and large model

CN121979924ACN 121979924 ACN121979924 ACN 121979924ACN-121979924-A

Abstract

The invention provides an education policy retrieval method and system integrating a policy corpus and a large model, and relates to the technical field of artificial intelligence. Firstly, a dynamic knowledge network construction module is used for constructing a dynamically updatable education policy knowledge graph from four dimensions of time, space, effectiveness and main body, and accurately modeling version evolution, region applicability and effectiveness relation among policies to provide a structural basis. And secondly, based on the atlas, the collaborative reasoning engine analyzes the complex intention of the natural language query of the user by using a domain fine-tuned generation type artificial intelligent model, and ensures the accurate compliance of the interpretation answer through a controllable generation technology under knowledge constraint. Finally, the feedback optimization system discovers high-frequency contradiction points in policy execution by collecting multidimensional user behaviors and explicit feedback, drives bidirectional optimization of a knowledge graph and a generation model, enables the system to continuously adapt and evolve, and achieves dynamic reasoning and closed-loop optimization of education policy interpretation.

Inventors

WU LIJUN
WU XINJIE
Zhao Zhanshu
WANG LINHAO
ZHU NING
CAO SHIYONG
PENG SHOUYE
KANG FENGWEI
YANG WUMENG

Assignees

北京尚睿通科技有限公司

Dates

Publication Date: 20260505
Application Date: 20260113

Claims (10)

1. A method for educational policy retrieval that merges a corpus of policies with a large model, comprising: receiving natural language query information input by a user; Analyzing the natural language query information by using a generated artificial intelligent model, and identifying and obtaining user intention characteristics, wherein the user intention characteristics comprise time, place, policy main body, item entity and query intention; traversing a dynamic education policy knowledge graph based on the user intention characteristics, and outputting a search result, wherein the search result comprises current effective policy terms, historical version data and difference information; And performing checksum constraint on accuracy, timeliness and traceability based on the natural language query information, the query intention and the search result, and generating an educational knowledge interpretation answer.
2. The method for educational policy retrieval that merges a corpus of policies with a large model of claim 1, further comprising: Collecting multi-source heterogeneous education policy data; Based on the education policy data, establishing a version evolution relation graph between policy documents through entity identification and relation extraction, wherein the version evolution relation graph comprises revision, revocation and supplement relations; Based on the version evolution relation diagram, storing differences among versions at a clause level by adopting a differential storage technology to obtain a version evolution relation diagram with difference information; based on the education policy data and administrative division codes, carrying out hierarchical geocoding and modeling on the application range of the policy to obtain an education policy geospatial map; based on the education policy data and a predefined law applicable rule base, carrying out conflict detection and priority pre-judgment on policies with inconsistent regulations in the same scope, and generating an education policy effectiveness map; based on the education policy data, labeling the policy main body and the matters related to the policy terms in a fine mode through named entity identification and relation extraction to form multi-label classification, and obtaining an education policy theme label diagram; and generating the dynamic education policy knowledge map in a fusion way based on the version evolution relation map with the difference information, the education policy geospatial map, the education policy effectiveness map and the education policy theme label map.
3. The method for educational policy retrieval that merges a corpus of policies with a large model of claim 1, further comprising: Taking a general large language model as a base model; Acquiring special training data in the field of education policies, wherein the special training data comprises policy text, administrative question and answer on politics answer pairs and policy consultation dialogue data; and performing field adaptability training on the base model through a parameter efficient fine tuning technology by using the special training data to obtain the generated artificial intelligent model, wherein the generated artificial intelligent model is used for analyzing the structured user intention characteristics from the natural language query.
4. The method for educational policy retrieval that merges a corpus of policies with a large model of claim 1, further comprising: collecting explicit scores of the educational knowledge reading answers and implicit interactive behavior data of the user, wherein the implicit interactive behavior data comprises a query retry rate, a result click rate and page residence time; based on the explicit scoring and implicit interaction behavior data, data mining is carried out, and high-frequency contradiction points in the education policy execution process are determined; determining a plurality of clauses related to the high-frequency contradiction points based on the high-frequency contradiction points in the education policy execution process and the dynamic education policy knowledge graph; And updating the dynamic education policy knowledge graph based on the high-frequency contradiction points and a plurality of clauses related to the high-frequency contradiction points to obtain an updated dynamic education policy knowledge graph.
5. The method for searching for educational policies by fusing a corpus of policies and a large model according to claim 4, wherein the data mining based on the explicit scoring and implicit interaction data, determining high frequency contradictory points in the execution process of the educational policies comprises: clustering the negative feedback text by using the BERT-Topic Topic model, extracting high-frequency contradiction keywords, and carrying out matching positioning with policy clauses in the knowledge graph to obtain a clustering result; based on the clustering result, generating a policy execution contradictory thermodynamic diagram, and representing the negative feedback quantity of the clauses in a visual instrument panel by color depth; and determining the high-frequency contradiction point based on the clustering result and the policy execution contradiction thermodynamic diagram.
6. The method for retrieving educational policies integrating a corpus of policies with a large model according to claim 1, wherein said parsing the natural language query information using a generated artificial intelligence model, identifying user intent features, comprises: Inputting the natural language query information into the generated artificial intelligence model; And obtaining a structured analysis result output by the generated artificial intelligence model, wherein the structured analysis result comprises an entity list identified from natural language query information and an inferred query intention list, and the entity list comprises a time entity, a place entity, a policy main entity and a matter entity.
7. The method for retrieving educational policies integrating a corpus of policies with a large model according to claim 1, wherein traversing the dynamic educational policy knowledge graph based on the user intent features, outputting a retrieval result comprises: Taking the identified time, place, policy main body and item entity as query conditions, and performing traversal search in a map database; The retrieval process follows the time effectiveness, the spatial range and the main body association relationship modeled in the map; And outputting the effective policy and clause set, related historical version information and clause level difference information obtained by a difference storage technology, which conform to the current query space-time range.
8. The method for educational policy retrieval that merges a corpus of policies with a large model according to claim 1, wherein the generating an educational knowledge interpretation answer based on the natural language query information, the query intent, and the retrieval result, performing checksum constraint on accuracy, timeliness, and traceability, comprises: constructing a knowledge constraint object based on the search result, wherein the knowledge constraint object comprises a policy and clause original text to be followed, a policy document number to be referred, a revoked policy list and clause level difference information; Inputting the natural language query information, the query intention and the knowledge constraint object into the generated artificial intelligent model together to obtain a plurality of candidate generated contents; Performing constraint verification on the plurality of candidate generated contents to obtain verification results of each candidate generated content, wherein the constraint verification comprises fact consistency verification, illusion inhibition verification and timeliness verification, wherein the fact consistency verification is used for ensuring that key facts in the generated contents are consistent with clause original texts in the knowledge constraint object, the illusion inhibition verification is used for preventing the generated contents which are not in basis in the knowledge constraint object, the timeliness verification is used for ensuring that generated answers clearly indicate timeliness states of effective policies, and the traceability verification is used for automatically attaching corresponding policy clause source indexes after key assertion of the answers; And determining the educational knowledge interpretation answer based on the verification result.
9. The method for educational policy retrieval that merges a corpus of policies with a large model of claim 1, further comprising: Giving credibility weights to different sources; When the same policy file is acquired from a plurality of information sources, the version with the highest credibility weight is selected as a main data source, and other versions are selected as check data sources.
10. An educational policy retrieval system integrating a corpus of policies with a large model, characterized in that the system comprises an electronic device comprising a memory and a processor, the memory storing a computer program, the processor being for invoking and running the computer program stored in the memory to perform the method according to any of claims 1 to 9.

Description

Educational policy retrieval method and system integrating policy corpus and large model Technical Field The invention relates to the technical field of artificial intelligence, in particular to an education policy retrieval method and system integrating a policy corpus and a large model. Background The education policy is a core carrier for the national and local education authorities to carry out education treatment, resource allocation and reform promotion. Along with the acceleration of the modernization progress of the education management system, the number of education policies is increased, the education policies are updated frequently, the hierarchical relationship is complex, and the education policies are closely related to regions, audiences and historical versions. The demands of the mass educators, parents of students and the public for timely, accurate and deep understanding of policy contents are becoming urgent. In this context, intelligent interpretation and analysis techniques for educational policies have developed. At present, education policy interpretation is mostly carried out by keyword matching and static semantic library searching, policy texts are collected through crawlers, and indexing is carried out based on manually preset label systems (such as recruitment, subsidy and approval) and keywords. The user can only inquire through a limited drop-down menu or a keyword with a fixed format, and rely on a preset searching intention semantic library to match information. The policy interpretation mode semantic static rigidification can not understand the dynamic evolution of the policy semantic, and the non-preset, emerging or complex combination policy intention can not be identified for shake number admission, double-subtraction non-disciplinary training supervision and the like. Only a document list containing keywords can be returned, and logical relations such as revision, revocation, replacement and the like among policies cannot be revealed, so that a user cannot easily judge the current validity and history of the policies. Only the original text segments can be provided, the policy clauses cannot be interpreted, compared or inferred, and the compound questions requiring deep analysis, such as what changes are made in the past year, what differences are specified in the A place and the B place, and the like, cannot be answered. The current education policy reading mode has the problem of low credibility. Disclosure of Invention The invention provides an education policy retrieval method and system integrating a policy corpus and a large model, which solve the problem of low reliability of the current education policy interpretation mode. The invention provides an education policy retrieval method integrating a policy corpus and a large model, which comprises the steps of receiving natural language query information input by a user, analyzing the natural language query information by utilizing a generated artificial intelligent model, identifying and obtaining user intention characteristics, wherein the user intention characteristics comprise time, place, policy main body, item entity and query intention, traversing a dynamic education policy knowledge graph based on the user intention characteristics, outputting a retrieval result, wherein the retrieval result comprises current effective policy terms, historical version data and difference information, and verifying and restraining accuracy, timeliness and traceability based on the natural language query information, the query intention and the retrieval result to generate an education knowledge interpretation answer. In one possible implementation, the method further comprises the steps of collecting multi-source heterogeneous education policy data, establishing a version evolution relation diagram between policy documents through entity identification and relation extraction based on the education policy data, wherein the version evolution relation diagram comprises revisions, abolishments and supplementary relations, obtaining a version evolution relation diagram with difference information by adopting a difference storage technology and storing differences among versions at a clause level based on the version evolution relation diagram, obtaining a version evolution relation diagram with the difference information based on the education policy data and administrative division codes, carrying out hierarchical geocoding and modeling on application ranges of the policies to obtain an education policy geospatial diagram, carrying out conflict detection and priority pre-judgment on the policies with inconsistent regulations in the same range based on the education policy data and a predefined law application rule base, generating an education policy effect diagram, identifying and relation extraction based on the education policy data through naming entities, forming a multi-label classification for a policy main body and a de