Large Language Model Summarization for Explainable Decisions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Complex decision systems often produce incomprehensible explanations due to their intricate nature, making them functionally useless without additional operations to simplify and clarify their decision-making processes.
Innovation Solution
A summarization model is employed in conjunction with local explainability parameters to generate aggregated summarizations, using a text generation model to provide clear explanations by determining feature importance and generating phrases that highlight changes needed to achieve a target prediction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If complex decision systems are used to make decisions based on multiple variables and factors, then decision accuracy and comprehensiveness are improved, but explanation comprehensibility deteriorates
Solution Approach 1:
The patent segments the complex explanation into multiple components: identified variables, their individual effects on the decision, and prioritized relationships. This segmentation allows the system to present complex decision-making processes in manageable, understandable portions while maintaining complete information about the decision factors.
Solution Approach 2:
The patent introduces an intermediary processing layer that takes the complex output from the decision system and transforms it into simplified explanations. This intermediary generates natural language descriptions that bridge the gap between complex computational results and human-understandable formats, preserving accuracy while improving comprehensibility.
2Loss of information
If comprehensive reports on all variables and factors are provided, then information completeness is improved, but information overload and complexity increase
Solution Approach 1:
The patent applies local quality by providing different levels of detail for different variables based on their importance and relationship to the decision. Critical variables receive more detailed explanations while less important ones are summarized, ensuring information completeness is maintained while reducing overall complexity through selective detail distribution.
Solution Approach 2:
The patent implements partial action by selectively presenting information about variables and factors based on their relevance to the decision outcome. Rather than presenting all possible information equally, the system focuses on the most significant factors while still maintaining completeness of the decision rationale through prioritized presentation.
Data Source
AI summary
A method and related system operations for summarizing information includes determining a plurality of local explainability parameters for a set of inputs using a local explainability model for a prediction model. The method further includes generating a plurality of feature effect indicators for a target prediction and generating a plurality of phrases based on the plurality of feature effect indicators by using a text generation model. The method further includes generating a summarization using a large language model based on the plurality of phrases.