LLM Fact-Checking Workflow for Reliable Generated Descriptions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Large language models (LLMs) often provide inaccurate or hallucinated information, making it difficult for entities listing items to verify the truthfulness of additional information, which can mislead users.
Innovation Solution
A system and method using a LLM to generate descriptive text based on a prompt, with fact extraction and comparison to external sources, highlighting differences and providing a user interface to verify accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If LLMs are used to generate additional information about items, then the quantity and variety of information available to users is improved, but the reliability and accuracy of the information deteriorates due to hallucinations
Solution Approach 1:
The patent introduces an intermediary verification system that acts as a mediator between the LLM-generated information and the final output. This system includes: (1) A fact-extraction module that identifies claims in generated text, (2) A search module that queries external knowledge sources to verify facts, and (3) A verification module that compares generated facts with external sources. This intermediary layer filters out hallucinated information while preserving valid generated content, thus maintaining information quantity while improving reliability.
Solution Approach 2:
The patent implements a feedback mechanism where the verification results are fed back into the system to improve future generations. The process includes: (1) Comparing generated facts with verified facts from external sources, (2) Identifying discrepancies and hallucinations, (3) Using this information to refine the LLM's generation process. This feedback loop continuously improves the reliability of generated information while maintaining the beneficial quantity and variety.
2Productivity
If LLMs generate descriptive text with additional information, then the productivity of information creation is improved, but the measurement precision and accuracy of the information deteriorates
Solution Approach 1:
The patent segments the information generation process into distinct modular components: (1) Fact extraction module that identifies specific claims, (2) Search module that queries external sources, (3) Verification module that validates facts, and (4) Output module that presents verified information. This segmentation allows each component to specialize in its function, maintaining high productivity through automation while improving accuracy through dedicated verification steps.
Solution Approach 2:
The patent performs preliminary verification actions before finalizing the generated information. The system extracts facts, searches external sources, and verifies accuracy before presenting the final output. This preliminary action ensures that only verified information is delivered, maintaining measurement precision while the automated process preserves high productivity.
3Ease of operation
If entities rely on LLM-generated information without verification, then the ease of operation is improved, but the object-generated harmful factors increase due to propagation of false information
Solution Approach 1:
The patent implements a self-service verification system where the LLM-generated information automatically verifies itself against external knowledge sources. The system extracts facts from generated text, queries external sources independently, and validates the information without requiring manual intervention. This self-service approach maintains ease of operation for users while eliminating harmful false information through automated verification.
Data Source
AI summary
A system for creating generated descriptive text is provided. A prompt having first facts for an item is received and parsed to extract a first fact in a format. Second facts are generated where the first fact and the second facts are output in the format. A search query is generated that includes the first fact and the second facts and then a search is conducted using the search query. An output is generated based on the results. The output includes a suggested description of the item using at least one first fact of the first facts and the second facts. The output also has a summarization of the plurality of first facts and the second facts along with differences between the first facts and the second facts. A distribution of the plurality of first facts and the second facts in the results is provided in the output.