Abstractive summarization models are designed to rewrite source documents in their own words, offering a level of conciseness and flow that simple "cut-and-paste" extractive methods cannot match. However, these models suffer from a persistent flaw: hallucinations. They frequently generate fluent, confident sentences that contain names, places, or relationships never mentioned in the source text.
For years, researchers feared a "factuality tax"—the idea that forcing a model to be more truthful would inevitably make it more rigid, less creative, and more likely to simply copy-paste from the source. Recent evidence suggests the opposite. By grounding models in factual structures and teaching them to reject noisy data, we can achieve summaries that are both more accurate and more abstractive.
The Myth of the Factuality Tax
The central tension in summarization has long been the trade-off between "faithfulness" (staying true to the source) and "abstractiveness" (using novel vocabulary and phrasing). Conventional wisdom suggested that if you penalized a model for hallucinating, it would retreat into "extractive" behavior, becoming a glorified copy-machine to avoid making mistakes.
However, work by Tanay Dixit and Muhao Chen with the EFACTSUM framework demonstrates that this tax is not mandatory. They found that models could achieve double-digit improvements in factuality scores while actually increasing their lexical diversity. Similarly, I-Chun Chern and Graham Neubig showed that Contrastive Reward Learning (CRL)—which trains models to prefer accurate summaries over hallucinated ones—improves coherence and relevance alongside truthfulness. These findings suggest that factual accuracy and linguistic quality are complementary objectives, not competing ones.
Why Models Hallucinate: Noise and Over-Generalization
To fix hallucinations, we must understand why they happen. One primary cause is "noisy" training data. Standard datasets often contain human-written summaries that include external information not found in the source text. When models are forced to mimic these summaries, they learn that "making things up" is part of the job.
Meng Cao and Jackie CK Cheung addressed this through rejection learning. Instead of forcing a model to memorize every word of a noisy reference summary, their framework teaches the model to recognize and "reject" unsupported information during training. Interestingly, this doesn't just make the model more truthful; it makes it more abstractive. By not being tethered to specific noisy phrases, the system becomes better at paraphrasing and generating novel n-grams that remain grounded in the actual source facts.
Another cause is "prior bias"—the model’s tendency to generate words that are common in its general training data rather than specific to the document at hand. Kyubyung Chae and Taesup Kim introduced a decoding strategy called Domain-Conditional Pointwise Mutual Information ($\text{PMI}_{DC}$). This method penalizes words that are highly probable given only the general topic of a document, forcing the model to rely on the specific evidence provided in the source text.
Grounding Summaries in Factual Structures
If the goal is to prevent the model from drifting into fiction, we can provide it with a "factual anchor." Chenguang Zhu and Michael Zeng pioneered this with FASUM, a model that extracts a knowledge graph of facts from the source text and uses graph attention to ensure the summary respects those relationships.
This structural approach is often paired with more granular evaluation metrics. Nan Feng and Ramesh Nallapati focused on entity-level consistency, specifically tracking whether names and places in the summary appeared in the source. By cleaning training data of entity errors and using multi-task learning to generate entities and summaries jointly, they showed that truthfulness can be baked into the model's architecture without sacrificing fluency.
Optimization for Truthfulness
The final piece of the puzzle is how we fine-tune these models. Standard training (Maximum Likelihood Estimation) treats every word in a reference summary as equally important. Modern approaches use "preference-based" or "contrastive" objectives to give the model a sense of direction.
- FactPEGASUS, developed by David Wan and Mohit Bansal, integrates factuality into pre-training by selecting factual sentences as targets, leading to massive gains in faithfulness.
- FactGen, by ZhiBin Lan and Jinsong Su, uses Natural Language Inference (NLI) to rank candidate summaries during training, ensuring the model learns to distinguish between a "plausible-sounding" hallucination and a "factually-entailed" statement.
- Even in complex tasks like "perspective summarization," where a model must capture a specific political viewpoint, Narutatsu Ri and Kathleen McKeown found that using Direct Preference Optimization (DPO) allowed models to use more novel vocabulary while simultaneously increasing their adherence to the required perspective.
The collective evidence from these researchers suggests that the "hallucination problem" is not an inherent flaw of abstractive AI, but a consequence of how we train and decode. When we prioritize factual grounding, we don't just get more accurate summaries—we get better ones.