Large language model misleading information detection and correction system

Abstract

The invention provides a system for detecting and correcting misleading information of a large language model, which is characterized in that a state machine processing engine is integrated and configured to be a scheduling processing flow based on a deterministic finite state automaton, the state transition is driven by the attribute of an input problem and a knowledge state signal, and a dynamic knowledge updating unit is designed; knowledge is continuously grabbed from an external authoritative source through a lightweight knowledge crawler, knowledge freshness score is calculated to realize driving, and meanwhile, the timeliness of a large language model knowledge base is maintained through an incremental knowledge injection mechanism; and sequentially scheduling question detection, question error correction and question answering stages based on a state machine processing engine and a dynamic knowledge updating unit, and iteratively updating until a final accurate response is generated. The method has theoretical preciseness and knowledge real-time performance, correctness of model behaviors can be ensured mathematically, newest knowledge can be dynamically integrated, and therefore continuous and reliable services can be provided in complex and changeable real scenes.

Images