Summary Model Training with Negative Sampling for Factual Consistency

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing abstractive summarization methods using deep learning models struggle with factual consistency, often distorting important information or introducing new, non-existent information in summary sentences.

Innovation Solution

A method involving likelihood and unlikelihood losses is employed to train a summary model, where likelihood loss is calculated using positive samples and unlikelihood loss is calculated using negative samples generated by masking or removing main keywords, ensuring the model maintains the factual integrity of the original text.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Manufacturing precision

If abstractive summarization is used to generate new summary sentences, then summarization quality and conciseness are improved, but factual consistency deteriorates

Engineering Contradiction:

Improvesummarization qualityVSAvoidfactual consistency

Core Design Contradiction:

Manufacturing precisionVSReliability

Solution Approach 1:

The patent applies preliminary anti-action by introducing a negative sampling mechanism that pre-empts factual inconsistencies. During training, the model is exposed to negative samples (incorrect summaries) and learns to avoid generating them by calculating unlikelihood loss. This preventive approach counteracts the tendency of abstractive models to generate factually inconsistent summaries while maintaining their summarization quality.

Inventive Principle:

Principle #9Preliminary anti-action

Solution Approach 2:

The patent implements feedback through a dual-loss training mechanism. The model receives feedback from both likelihood loss (encouraging accurate summarization) and unlikelihood loss (penalizing factual inconsistencies). This feedback loop continuously guides the model to improve factual consistency while maintaining summarization quality, resolving the contradiction between the two objectives.

Inventive Principle:

Principle #23Feedback

2Adaptability or versatility

If deep learning model generates new keywords and sentences, then abstractive summarization capability is improved, but information distortion increases

Engineering Contradiction:

Improveabstractive summarization capabilityVSAvoidinformation distortion

Core Design Contradiction:

Adaptability or versatilityVSLoss of information

Solution Approach 1:

The patent prevents information distortion by introducing negative samples that represent distorted information scenarios. The model is trained to recognize and avoid these distorted patterns through unlikelihood loss calculation, thereby maintaining high abstractive capability while preventing information distortion from occurring in the first place.

Inventive Principle:

Principle #9Preliminary anti-action

Solution Approach 2:

The patent introduces an intermediary mechanism (negative sampling and unlikelihood loss) that mediates between the model's abstractive generation capability and information fidelity. This intermediary structure allows the model to generate creative summaries while the unlikelihood loss acts as a constraint to prevent information distortion, balancing adaptability with information preservation.

Inventive Principle:

Principle #24Intermediary (Mediator)

3Ease of operation

If summary model changes keywords representing main factual relationship, then summarization fluency is improved, but factual accuracy deteriorates

Engineering Contradiction:

Improvesummarization fluencyVSAvoidfactual accuracy

Core Design Contradiction:

Ease of operationVSMeasurement precision

Solution Approach 1:

The patent applies preliminary anti-action by pre-training the model with negative samples that contain altered keywords and factual relationships. This prepares the model to recognize and avoid changing key factual keywords, thereby maintaining both fluency and factual accuracy simultaneously rather than allowing fluency improvements to compromise accuracy.

Inventive Principle:

Principle #9Preliminary anti-action

Solution Approach 2:

The patent changes the training parameters by introducing unlikelihood loss as an additional optimization objective. This parameter change modifies the model's learning behavior to preserve factual keywords while maintaining fluency, resolving the contradiction between ease of operation (fluency) and measurement precision (factual accuracy).

Inventive Principle:

Principle #35Parameter changes

Data Source

PatentUS12619824B2Method for generating summary and system therefor

Publication Date: 2026.05.05 SAMSUNG SDS CO LTD

AI summary

Provided are a method for generating a summary and a system therefor. The method according to some embodiments may include calculating a likelihood loss for a summary model using a first text sample and a first summary sentence corresponding to the first text sample, calculating an unlikelihood loss for the summary model using a second text sample and the first summary sentence, the second text sample being a negative sample generated from the first text sample, and updating the summary model based on the likelihood loss and the unlikelihood loss.