CN-122021732-A - Large model illusion eliminating method, device, equipment and storage medium

CN122021732ACN 122021732 ACN122021732 ACN 122021732ACN-122021732-A

Abstract

The invention discloses a large model illusion elimination method, a device, equipment and a storage medium, which relate to the technical field of artificial intelligent natural language processing and comprise the steps of obtaining initial generation content, carrying out atomization disassembly on the initial generation content to obtain an atomic fact set, carrying out internal knowledge base semantic matching and networking retrieval cross-validation on each atomic fact in the atomic fact set to obtain evidence coverage, generating an evidence blank report when the evidence coverage is lower than a preset coverage threshold, carrying out differential rewriting on the initial generation content based on the evidence blank report to obtain target generation content, carrying out evidence coverage quantification check on the generation content through an atomization disassembly and internal and external double-source cross-validation mechanism, blocking illusion cascade in the early reasoning stage and feeding back validation results to carry out iterative rewriting, and obtaining high confidence content subjected to explicit validation.

Inventors

CHEN BIN
Pang Yudie
LU YUYANG

Assignees

超维联构(无锡)智能科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260109

Claims (10)

1. A method of large model illusion cancellation, the method comprising: Acquiring initial generation content; performing atomization disassembly on the initial generation content to obtain an atomic fact set; Performing internal knowledge base semantic matching and networking retrieval cross-validation on each atomic fact in the atomic fact set to obtain evidence coverage; generating an evidence blank report when the evidence coverage rate is lower than a preset coverage rate threshold value; and performing differential rewriting on the initial generated content based on the evidence blank report to obtain target generated content.
2. The method of claim 1, wherein the step of performing an atomization disassembly on the initially generated content to obtain an atomic fact set comprises: performing dependency syntactic analysis on the initial generated content to obtain a clause result and a part-of-speech tagging result; according to the parallel conjunctions and clause structures in the clause results, the clause results are disassembled into propositions units; Performing reference resolution on the proposition unit, and reducing pronouns into entity nouns to obtain self-explanatory propositions; And carrying out triplet standardization on the self-explanatory proposition to obtain an atomic fact set.
3. The method of claim 2, wherein the step of performing dependency syntax analysis on the initially generated content to obtain a clause result and a part-of-speech tagging result comprises: Sentence boundary detection is carried out on the initial generated text, and a sentence dividing result is obtained; And marking the parts of speech of each word in the clause result to obtain a part of speech marking result, and identifying a subject component, a predicate component and an object component according to the part of speech marking result to construct a core logic skeleton.
4. The method of claim 1, wherein the step of performing internal knowledge base semantic matching and networking retrieval cross-validation on each atomic fact in the set of atomic facts to obtain evidence coverage comprises: Converting each atomic fact in the atomic fact set into an atomic fact vector, and carrying out semantic similarity calculation on each atomic fact vector and evidence segments in an internal knowledge base to obtain an internal matching score; When the internal matching score is lower than a preset semantic matching score threshold value, the atomic facts are sent to a networking retrieval interface so as to acquire real-time retrieval fragments through the networking retrieval interface, and the real-time retrieval fragments are converted into external evidence vectors; calculating cosine similarity of the atomic fact vector and the external evidence vector to obtain an external matching score; And determining the coverage state of each atomic fact according to the internal matching score and the external matching score, and counting the proportion of the covered atomic facts in the atomic fact set to obtain evidence coverage rate.
5. The method of claim 4, wherein the step of converting each atomic fact in the set of atomic facts into an atomic fact vector and performing semantic similarity calculation on each of the atomic fact vectors and evidence segments in an internal knowledge base to obtain an internal matching score comprises: mapping each atomic fact into an atomic fact vector in a high-dimensional vector space through a pre-trained word vector model; searching a preset number of target evidence segments in the internal knowledge base based on the atomic fact vector to obtain a candidate evidence set; converting each candidate evidence segment in the candidate evidence set into an internal evidence vector; And calculating similarity values of the atomic fact vector and each internal evidence vector through a cosine similarity function, and determining an internal matching score according to the similarity values.
6. The method of claim 4, wherein the step of obtaining a real-time search segment through the networked search interface, converting the real-time search segment into an external evidence vector, comprises: Extracting a search keyword from the atomic facts, and constructing a search query statement; Sending the search query statement to a search engine application program interface; receiving a search result list returned by the search engine application program interface, and extracting text fragments in the search result list to obtain real-time search fragments; the real-time search segment is converted into an external evidence vector through a pre-trained word vector model.
7. The method of claim 1, wherein the step of generating a evidence blank report when the evidence coverage is below a preset coverage threshold comprises: identifying uncovered target atomic facts in the atomic facts set, and generating an evidence blank report based on the uncovered target atomic facts; The step of differentially rewriting the initially generated content based on the evidence blank report to obtain target generated content comprises the following steps: sending the evidence blank report to a generating end, so that the generating end carries out differential adjustment on the initial generated content according to the evidence blank report to obtain rewritten content, wherein the differential adjustment comprises deleting uncovered atomic facts or converting the uncovered atomic facts into speculative expressions; performing atomization disassembly and cross-validation on the rewritten content again until the evidence coverage rate is greater than or equal to the preset coverage rate threshold value, and judging that the verification is passed; and generating the content by taking the rewritten content passing verification as a target.
8. A large model illusion elimination apparatus, the apparatus comprising: the content acquisition module is used for acquiring the initial generated content; the atomization disassembly module is used for performing atomization disassembly on the initial generation content to obtain an atomic fact set; The cross verification module is used for carrying out semantic matching of an internal knowledge base and cross verification of networking retrieval on each atomic fact in the atomic fact set to obtain evidence coverage rate; And the differential rewriting module is used for generating an evidence vacancy report when the evidence coverage rate is lower than a preset coverage rate threshold value, and performing differential rewriting on the initially generated content based on the evidence vacancy report to obtain target generated content.
9. A large model illusion elimination apparatus, characterized in that the apparatus comprises a memory, a processor and a computer program stored on the memory and executable on the processor, the computer program being configured to implement the steps of the large model illusion elimination method according to any of the claims 1 to 7.
10. A storage medium, characterized in that the storage medium is a computer-readable storage medium, on which a computer program is stored, which computer program, when being executed by a processor, implements the steps of the large model illusion elimination method according to any of the claims 1 to 7.

Description

Large model illusion eliminating method, device, equipment and storage medium Technical Field The invention relates to the technical field of artificial intelligence natural language processing, in particular to a method, a device, equipment and a storage medium for eliminating large model illusions. Background With the wide application of large language models in the field of knowledge question and answer, the generated content has the problem of 'illusion', namely, the text output by the model has logical smoothness on the surface, but contains false information which does not accord with facts or deviates from reference materials, so that the actual illusion and the faithfulness illusion frequently occur, and the credibility and the practicability of the generated content are seriously affected. The prior art mainly relies on retrieval enhancement to generate a static constraint model in a prompt word according to a reference data answer, but the model still jumps out of a reference data logic boundary and introduces false facts when in complex reasoning to form a cascading illusion, and the quantitative verification of the matching degree of each statement and the reference data is lacking. The foregoing is provided merely for the purpose of facilitating understanding of the technical solutions of the present invention and is not intended to represent an admission that the foregoing is prior art. Disclosure of Invention The invention mainly aims to provide a method, a device, equipment and a storage medium for eliminating the illusion of a large model, which aim to solve the technical problem of blocking the illusion cascade phenomenon in the large model generation content through an atomization verification and iteration feedback mechanism. To achieve the above object, the present invention provides a large model illusion elimination method including the steps of: Acquiring initial generation content; performing atomization disassembly on the initial generation content to obtain an atomic fact set; Performing internal knowledge base semantic matching and networking retrieval cross-validation on each atomic fact in the atomic fact set to obtain evidence coverage; generating an evidence blank report when the evidence coverage rate is lower than a preset coverage rate threshold value; and performing differential rewriting on the initial generated content based on the evidence blank report to obtain target generated content. In an embodiment, the step of performing atomization disassembly on the initially generated content to obtain an atomic fact set includes: performing dependency syntactic analysis on the initial generated content to obtain a clause result and a part-of-speech tagging result; according to the parallel conjunctions and clause structures in the clause results, the clause results are disassembled into propositions units; Performing reference resolution on the proposition unit, and reducing pronouns into entity nouns to obtain self-explanatory propositions; And carrying out triplet standardization on the self-explanatory proposition to obtain an atomic fact set. In one embodiment, the step of performing dependency syntax analysis on the initially generated content to obtain a clause result and a part-of-speech tagging result includes: Sentence boundary detection is carried out on the initial generated text, and a sentence dividing result is obtained; And marking the parts of speech of each word in the clause result to obtain a part of speech marking result, and identifying a subject component, a predicate component and an object component according to the part of speech marking result to construct a core logic skeleton. In one embodiment, the step of performing semantic matching of an internal knowledge base and cross-validation of network retrieval on each atomic fact in the atomic fact set to obtain evidence coverage includes: Converting each atomic fact in the atomic fact set into an atomic fact vector, and carrying out semantic similarity calculation on each atomic fact vector and evidence segments in an internal knowledge base to obtain an internal matching score; When the internal matching score is lower than a preset semantic matching score threshold value, the atomic facts are sent to a networking retrieval interface so as to acquire real-time retrieval fragments through the networking retrieval interface, and the real-time retrieval fragments are converted into external evidence vectors; calculating cosine similarity of the atomic fact vector and the external evidence vector to obtain an external matching score; And determining the coverage state of each atomic fact according to the internal matching score and the external matching score, and counting the proportion of the covered atomic facts in the atomic fact set to obtain evidence coverage rate. In an embodiment, the step of converting each atomic fact in the atomic fact set into an atomic fact vector, and performing semantic s