Abstract
Abstractive summarization models often generate factually inconsistent content particularly when the parametric knowledge of the model conflicts with the knowledge in the input document. In this paper, we analyze the robustness of fine-tuning based summarization models to the knowledge conflict, which we call factual adaptiveness. We utilize pre-trained language models to construct evaluation sets and find that factual adaptiveness is not strongly correlated with factual consistency on original datasets. Furthermore, we introduce a controllable counterfactual data augmentation method where the degree of knowledge conflict within the augmented data can be adjustable. Our experimental results on two pre-trained language models (PEGASUS and BART) and two fine-tuning datasets (XSum and CNN/DailyMail) demonstrate that our method enhances factual adaptiveness while achieving factual consistency on original datasets on par with the contrastive learning baseline.
Introduction
Abstractive summarization models have achieved remarkable performance, yet they remain prone to generating factually inconsistent content. A significant driver of these hallucinations is the conflict between a model's internal parametric knowledge—learned during pre-training—and the specific facts presented in a source document. This paper introduces the concept of factual adaptiveness, which measures a model's robustness to such knowledge conflicts. When a document contains information that conflicts with what the model "knows," a non-adaptive model often defaults to its internal knowledge, leading to hallucinations.
Methodology
The authors propose a framework to evaluate and enhance factual adaptiveness through controllable counterfactual data augmentation.
Counterfactual Generation
The process involves identifying named entities in a document and replacing them with alternative entities of the same type. To ensure a genuine knowledge conflict, a Pre-trained Language Model (PLM) is used to verify the model's prior confidence in the original entity. This allows for the creation of evaluation sets that specifically target the model's stubbornness regarding its internal knowledge.
Controllable Augmentation
By mixing a tunable ratio of these counterfactual samples into the fine-tuning dataset, the model is incentivized to rely strictly on the input document. This process is evaluated using two new metrics: $M_{CL}$ (Conditional Likelihood difference), which measures the drop in model confidence when switching from a "known" entity to a counterfactual one, and $M_{FC}$ (Factual Consistency difference), which measures the impact on actual summary quality when a knowledge conflict is introduced.
Results and Discussion
Experiments conducted on PEGASUS and BART using the XSum and CNN/DailyMail datasets reveal that standard factual consistency does not imply factual adaptiveness. Models that perform well on standard benchmarks may still fail when document facts change.
As shown in the results, the proposed method significantly outperforms baselines, including NLL, Filtering, and CLIFF, particularly on the challenging XSum dataset. Qualitative analysis confirms that while baseline models often revert to pre-trained biases (e.g., hallucinating a common country name when an obscure one is provided in the text), the adaptive model correctly identifies and utilizes the document-specific entities. Crucially, this improvement in adaptiveness does not come at the cost of standard performance, as ROUGE scores and original factual consistency remain competitive.
Conclusion
Factual adaptiveness is a critical yet overlooked dimension of summarization quality. The proposed augmentation technique provides a controllable way to ensure models remain faithful to the source text even when it contradicts their internal knowledge. Future research will explore extending this approach to more complex, non-entity-based knowledge conflicts, such as conflicting events or causal relationships.
Summary updated