AI Output Error Highlighting Using Token-Level Confidence Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI systems do not effectively inform users about the likelihood of errors in their output data, making it difficult for users to recognize the degree of possibility that elements in the output include errors.
Innovation Solution
An information processing system that includes a processor to acquire input and output data generated by AI, and provides error possibility information indicating the likelihood of errors in the output data, attaching this information to specific elements within the output, such as text data, using a large language model to analyze the contribution of knowledge neurons in predicting subsequent elements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If AI generates output data in response to input data, then productivity and automation are improved, but the reliability and accuracy of the output data deteriorate due to potential errors
Solution Approach 1:
The system implements feedback by analyzing the generated output data through the processor to determine error possibility, then providing this information back to the user. This allows the system to self-evaluate its output quality and inform users about potential errors, resolving the contradiction between high-speed AI generation and reliable output.
2Measurement precision
If error possibility information is provided for all elements in output data, then measurement precision and reliability are improved, but device complexity and information processing load increase
Solution Approach 1:
The system applies local quality by providing error possibility information selectively rather than uniformly across all elements. The processor determines error possibility for each element based on its specific characteristics and the AI model's confidence, allowing detailed precision where needed while maintaining simplicity where not required.
3Loss of information
If error possibility information is attached to specific elements in output data, then information completeness is improved, but ease of operation and user comprehension become more difficult
Solution Approach 1:
The system segments error possibility information by attaching it to specific elements within the output data rather than providing a single overall error rating. This segmentation allows users to comprehend the output element-by-element, maintaining ease of operation while providing complete error information where relevant.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
An information processing system includes a processor configured to: acquire input data; acquire output data generated by artificial intelligence in response to the input data; and output the output data and error possibility information indicating a degree of possibility that an element of multiple elements in the output data includes an error by the artificial intelligence.