A fact detection algorithm based on sparse key samples

By employing a fact detection algorithm based on sparse key samples, and utilizing lightweight pronoun elimination, dual-path retrieval, and external knowledge base verification, the inefficiency and illusion problems of fact detection in large language models are solved, achieving efficient and reliable fact judgment.

CN122242513APending Publication Date: 2026-06-19GUANGZHOU UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
GUANGZHOU UNIVERSITY
Filing Date
2026-02-04
Publication Date
2026-06-19

AI Technical Summary

Technical Problem

Existing fact detection methods for large language models suffer from low retrieval performance, limited accuracy, and long processing time, making it difficult to meet real-time requirements. Furthermore, they are susceptible to knowledge conflicts and biases within the model, which can lead to illusions.

Method used

A fact detection algorithm based on sparse key samples is adopted. By constructing a relational knowledge base and a vector database, combined with lightweight pronoun elimination and fact entity extraction, dual-path retrieval is performed using sparse keywords and dense semantic retrieval, and semantic consistency judgment is performed using a lightweight natural language reasoning model to output fact detection results.

Benefits of technology

Significantly reduce computational overhead, improve response speed, maintain high detection accuracy, significantly enhance the objectivity and reliability of fact-finding, effectively avoid model illusions, enhance the comprehensiveness and robustness of evidence recall, and build a highly reliable and traceable knowledge infrastructure.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122242513A_ABST
    Figure CN122242513A_ABST
Patent Text Reader

Abstract

This invention discloses a fact detection algorithm based on sparse key samples, belonging to the field of network data security. The method includes: constructing a relational knowledge base and a vector database; performing lightweight pronoun elimination and fact entity extraction on the model's response text, calculating the fact density score of each sentence in the model's response text, and screening high-risk key samples; based on the relational knowledge base and vector database, combining sparse keyword retrieval and dense semantic retrieval, performing dual-path retrieval on the high-risk key samples, and using inverse sorting to fuse and integrate the results to obtain the retrieval text; using a lightweight natural language inference model, performing semantic consistency judgment on the high-risk key samples and the retrieval text, and outputting the fact detection result of the model's response text. This scheme can significantly reduce computational overhead and improve response speed, while significantly improving the objectivity and reliability of fact judgment.
Need to check novelty before this filing date? Find Prior Art