LLM Summary Attribution Using Segment Matching and Source Modification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Large language models (LLMs) do not attribute inadvertently included content to their underlying data sources, necessitating a system to accurately attribute such content in generated natural language summaries.
Innovation Solution
A system processes NL-based summaries using LLMs to identify matching segments with training or fine-tuning datasets, enabling attribution by modifying summaries with source information based on predefined rules.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If LLMs generate NL-based summaries using training data, then summary generation capability is improved, but content attribution to sources is lost
Solution Approach 1:
The system performs preliminary actions by segmenting and indexing the training dataset before LLM generation, storing metadata about data sources. This enables subsequent attribution by comparing generated summary segments against the pre-segmented indexed data, allowing the system to identify and attribute source material without compromising the generative capability of the LLM.
Solution Approach 2:
The patent introduces an intermediary mechanism consisting of a segmentation module and index structure that mediates between the LLM's generative output and the training data sources. This intermediary layer enables the system to trace generated content back to sources by comparing summary segments against the indexed training segments, thereby restoring attribution capability while maintaining generation functionality.
2Reliability
If LLMs are trained on diverse datasets, then language generation quality is improved, but identification of source content becomes more difficult
Solution Approach 1:
The system applies segmentation by dividing the training dataset into discrete segments with associated metadata, and similarly segmenting the generated summary into comparable segments. This segmentation enables systematic comparison between generated content and training data, making source identification feasible even in diverse datasets by breaking down the complex task into manageable segment-level comparisons.
Solution Approach 2:
The patent replaces manual or complex analysis mechanisms with an automated computational approach using the LLM itself to compare summary segments against the indexed training segments. This substitution of mechanical analysis with intelligent comparison enables efficient source identification in diverse datasets by leveraging the LLM's language understanding capabilities rather than relying on simple keyword matching or manual review.
3Measurement precision
If LLMs inadvertently include training data content, then response relevance is improved, but attribution accuracy deteriorates
Solution Approach 1:
The system implements feedback by comparing the generated summary segments against the indexed training segments and using this comparison information to generate attributions. The feedback loop enables the system to recognize when training data content has been inadvertently included in the summary and to provide accurate attributions, thereby improving both relevance measurement and attribution precision simultaneously.
Data Source
AI summary
Implementations described herein relate to attribution of a natural language (NL) based summary generated using a large language model (LLM). Processor(s) of a system can: receive NL based input associated with a client device, generate the NL based summary using the LLM, and process the NL based summary to determine whether a NL based summary segment of the NL based summary matches a dataset segment of a dataset that was utilized to initially train the LLM and/or to fine-tune the LLM. Further, the processor(s) can, in response to determining that the NL based summary segment matches the dataset segment, modify the NL based summary segment of the NL based summary to generate a modified NL based summary. Moreover, the processor(s) can cause the modified NL based summary to be rendered at the client device. The attribution of the NL based summary can be provided as a service to various third-parties.