基于多轮交互意图累积图谱的内容安全评估方法
By constructing an intent accumulation graph and using dynamic threshold adjustment, the problem of identifying progressive inducement attacks in multi-turn dialogues is solved, enabling accurate detection and defense of dialogue risks and significantly improving the accuracy and adaptability of content security assessment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ZHEJIANG CHUANGLIN TECH CO LTD
- Filing Date
- 2026-04-21
- Publication Date
- 2026-07-17
AI Technical Summary
Existing multi-turn dialogue risk detection technologies suffer from problems such as long-distance forgetting, lack of quantitative accumulation, and fixed thresholds when facing progressive inducement attacks. They are difficult to effectively identify cross-turn intent associations and lack interpretability due to their reliance on large-scale labeled data and black-box characteristics.
The method adopts a multi-turn interaction intent accumulation graph approach. By constructing an intent accumulation graph, it calculates a comprehensive risk value using a predefined logical dependency matrix and time decay factor, and dynamically adjusts the interception threshold. By combining intent path accumulation and time decay, it achieves accurate detection and defense of dialogue risks.
It achieves accurate detection of progressively induced attacks, provides interpretability through explicit logical dependency matrix injection, reduces dependence on large-scale labeled data, and improves the system's sensitivity and adaptability through dynamic threshold adjustment, making it suitable for high-frequency interaction scenarios.
Smart Images
Figure CN122087832B_ABST