Knowledge-Base Mediated LLM Response Processing for Hallucination Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current human-computer interaction systems using large language models (LLM) suffer from hallucination issues, leading to low reliability due to generated content not matching actual situations or being irrelevant to the input messages.
Innovation Solution
An information processing method that outputs a first response message based on a target knowledge base and then inputs it into one or more target large language models, ensuring the matching degree of the final response message with the input message is higher than the initial response.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If large language models are used to process user input messages directly, then the system can provide interactive responses, but the generated content may cause hallucination problems and reduce reliability
Solution Approach 1:
The patent introduces a knowledge base as an intermediary component between the user input and the large language model. The system first retrieves relevant information from the knowledge base based on the user's input message, then feeds this retrieved information to the large language model for response generation. This intermediary structure ensures that the model bases its responses on verified knowledge rather than generating potentially hallucinated content, thereby improving reliability while maintaining interactive capability
Solution Approach 2:
The system performs preliminary information retrieval from the knowledge base before the large language model generates the response. By pre-fetching and verifying relevant information in advance, the system prepares accurate context data that the model can then use to generate reliable responses, preventing hallucination issues before they occur
2Measurement precision
If a knowledge base is used to generate response messages, then the matching degree with input messages improves, but the system complexity increases due to multiple processing components
Solution Approach 1:
The patent divides the response generation system into two distinct functional modules: a knowledge base retrieval module and a large language model generation module. The retrieval module handles information extraction and verification, while the generation module focuses on natural language response formulation. This segmentation allows each component to specialize in its strength, improving overall response accuracy while maintaining clear system architecture
Solution Approach 2:
The system merges the strengths of two different approaches: the factual accuracy of knowledge base retrieval and the natural language generation capability of large language models. By combining these two components in a sequential processing pipeline, the system achieves both high matching degree with input messages and natural, fluent responses
Data Source
AI summary
An information processing method, an electronic device, and a storage medium are provided in the present disclosure. The information processing method includes in response to obtaining of a target input message, outputting a first response message based on a target knowledge base; and inputting the first response message into at least one target large language model, and outputting a second response message, where the at least one target large language model is different from the target knowledge base; and a matching degree between the second response message and the target input message is higher than a matching degree between the first response message and the target input message.