Abstractive Summarization Model Factual Consistency via Virtual Samples
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing abstractive summarization methods face challenges in maintaining factual consistency and adaptiveness, often distorting important information or introducing new factual relationships, and their performance varies significantly with the content of the original text.
Innovation Solution
A method that selects common phrases from existing sample pairs based on prediction probabilities to generate virtual sample pairs that contradict the summary model's knowledge, allowing for improved factual adaptiveness and consistency, and uses contrastive learning to enhance the summary model's performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If abstractive summarization method is used to generate summaries, then summary quality and flexibility are improved, but factual consistency with original text deteriorates
Solution Approach 1:
The patent implements a feedback mechanism where the summary model continuously refines its predictions by comparing generated summaries against the original text and adjusting based on factual consistency evaluation. This iterative feedback loop enables the model to maintain flexibility in summary generation while correcting factual deviations through multiple refinement stages.
Solution Approach 2:
The patent changes the parameter of prediction probability threshold dynamically during the summarization process. By adjusting this threshold based on the confidence levels and factual consistency scores, the model can adaptively balance between creative summary generation and factual accuracy, resolving the contradiction between flexibility and reliability.
2Productivity
If deep learning model changes words to represent key factual relationships, then summary generation capability is improved, but factual consistency deteriorates
Solution Approach 1:
The patent replaces the purely statistical word prediction mechanism with a hybrid system that incorporates factual consistency checking and knowledge base verification. This substitution of the mechanical prediction process with a more sophisticated verification mechanism prevents factual distortion while maintaining efficient summary generation capability.
Solution Approach 2:
The patent introduces an intermediary verification layer between the summary generation model and the final output. This intermediary component acts as a mediator that checks factual consistency and corrects distorted information before the summary is finalized, thus preventing loss of accurate factual information while preserving generation efficiency.
3Adaptability or versatility
If summary model is trained on diverse original text content, then adaptability to different content is improved, but factual consistency varies with content type
Solution Approach 1:
The patent applies local quality by tailoring the summary generation approach to specific content types and domains. The model adjusts its prediction probability thresholds and factual verification strategies based on the specific characteristics of the input text, ensuring consistent factual accuracy across different content types while maintaining overall adaptability.
Solution Approach 2:
The patent implements dynamics by making the model's factual consistency verification mechanism adaptive to different content types. The verification strictness and prediction probability thresholds dynamically adjust based on the detected content characteristics, allowing the model to maintain high factual consistency across diverse content while preserving content-specific adaptability.
Data Source
AI summary
A method for generating a summary performed by at least one computing device is provided. The method may include: acquiring a first sample pair, the first sample pair including an original text and a summary corresponding to the original text, extracting a common phrase that appears simultaneously in the original text and the summary of the first sample pair, selecting a first phrase among common phrases based on a prediction probability of a summary model for the common phrases, generating a second sample pair by modifying the first phrase in the original text and the summary of the first sample pair, and updating the summary model by performing a summary task using the second sample pair.