Semantic analysis-based large model security detection method and device, and electronic equipment
By performing overall intent recognition, fine-grained parsing, and semantic aggregation on the input text, the problem of security policy misjudgment in highly sensitive domains by large models is solved, achieving a balance between security and usability, and providing assurance for the reliability and in-depth application of large models in highly sensitive scenarios.
CN122197883APending Publication Date: 2026-06-12SHANGHAI DOUXIANG INFORMATION TECH CO LTD
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANGHAI DOUXIANG INFORMATION TECH CO LTD
- Filing Date
- 2026-01-30
- Publication Date
- 2026-06-12
Smart Images
Figure CN122197883A_ABST
Abstract
The application relates to the technical field of data processing, and provides a large model safety detection method and device based on semantic analysis and electronic equipment, which comprises the following steps: performing overall intention recognition on input text to determine an overall intention category; performing fine-grained semantic analysis and intention judgment on the input text to determine at least one sub-sentence-level semantic target and corresponding intention information; performing semantic normalization aggregation processing on the at least one sub-sentence-level semantic target to generate an aggregated semantic target list; performing comprehensive safety risk assessment based on the overall intention category and the aggregated semantic target list, and outputting a safety decision result. The method balances safety and usability under the premise of ensuring safety compliance, and provides guarantee for the reliability and deep application of a large model in a high-sensitivity scene.
Need to check novelty before this filing date? Find Prior Art