Factual Consistency in Abstractive Summarization

875 concepts570 papers6 proposals

Abstractive summarization models excel at generating fluent, highly coherent text. However, this linguistic prowess often masks a deeper flaw: the tendency to generate plausible-sounding claims, entity swaps, or causal relationships completely unsupported by the source text.

A summary can be perfectly grammatical yet factually unfaithful. This gap between stylistic fluency and factual reliability poses severe risks, particularly in high-stakes domains like medicine, where confusing the direction of clinical evidence can be catastrophic (Jay DeYoung et al., MS^2: Multi-Document Summarization of Medical Studies), or finance, where minor alterations to numerical data render a summary useless (ChaeHun Park et al., Optimizing Domain Specificity of Transformer-based Language Models for Extractive Summarization of Financial News Articles in Korean).

The Root Causes of Factual Inconsistency

To understand why models invent information, we must look at both the data they are trained on and the evolution of their internal capabilities.

Noisy Reference Datasets

Standard summarization datasets frequently contain "gold standard" human summaries that include extrinsic information—background facts or context not present in the source text. As Yanzhu Guo argues in Questioning the Validity of Summarization Datasets and Improving Their Factual Consistency, training on these datasets actively incentivizes models to hallucinate, as they are penalized during training for being strictly faithful when the reference contains outside details.

This mismatch complicates evaluation; Shashi Narayan et al. demonstrate in A Well-Composed Text is Half Done! Composition Sampling for Diverse Conditional Generation that models attempting to maximize reference similarity are often forced to compromise on factual grounding.

The Evolution of Errors

Early sequence-to-sequence models primarily struggled with simple entity misrepresentations, such as swapping names, dates, or dimensions (Yizhu Liu et al., Controlling Length in Abstractive Summarization Using a Convolutional Neural Network).

However, as large language models (LLMs) have grown more capable, their errors have become significantly more sophisticated. As explored in Improving Factual Consistency of News Summarization by Contrastive Preference Optimization, modern models maintain a high degree of logical coherence while generating complex, fluent hallucinations, such as fabricating non-existent causal links or injecting plausible but unsupported background details from their parametric memory.

Why Standard Metrics Fail

Traditional evaluation metrics like ROUGE or BLEU rely on n-gram overlap and semantic similarity, which correlate poorly with factual truth. For example, a summary that states "The drug cured the patient" and one that states "The drug killed the patient" share a high degree of n-gram overlap but possess opposite truth values.

As Asma Ben Abacha et al. show in An Investigation of Evaluation Methods in Automatic Medical Note Generation, these automated metrics fail to distinguish between a harmless stylistic variation and a dangerous factual error. Unsupervised pipelines face a similar hurdle; Jiaxin Ju et al. point out in Monash-Summ@LongSumm 20 SciSummPip: An Unsupervised Scientific Paper Summarization Pipeline that prioritizing word overlap fails to capture whether a summary remains genuinely faithful to the underlying source material.

Paradigms for Detection and Mitigation

To bridge the gap between fluency and faithfulness, researchers have developed several distinct algorithmic frameworks to detect and prevent factual drift.

Granular Natural Language Inference

Natural Language Inference (NLI) models evaluate whether a premise (the source text) logically entails a hypothesis (the generated summary). Historically, applying NLI to document-level summarization failed due to a "granularity mismatch"—a model cannot easily determine if a 50-word summary is entailed by a 1,000-word document all at once.

In SummaC: Re-Visiting NLI-based Models for Inconsistency Detection in Summarization, Philippe Laban et al. resolve this by breaking the evaluation down into sentence-level pairs. By aggregating entailment scores across individual sentence matrices, they significantly improve inconsistency detection accuracy and speed.

Source Document: [S1] [S2] [S3] ... [Sn]
                               │
                               ▼
Generated Summary: [Sum1] [Sum2] [Sum3]
                               │
                               ▼
         Sentence-Level NLI Pairwise Matrix
             (Entailment / Contradiction)

Decompose-and-Aggregate for Long Contexts

When evaluating long-form documents, the computational cost of using massive LLMs to detect hallucinations becomes prohibitive. To address this, Siyi Liu et al. introduce a "decompose-and-aggregate" architecture in Towards Long Context Hallucination Detection. By breaking long contexts into smaller, manageable chunks and processing them with optimized encoder models like BERT, they achieve superior accuracy and run up to 20x faster than larger generative models.

Alternative Reward Functions and Distillation

Instead of optimizing for reference similarity, models can be guided by factual rewards during training. Kristjan Arumae et al. propose a reinforcement learning framework in Guiding Extractive Summarization with Question-Answering Rewards that rewards models based on how well their summaries can answer questions automatically generated from the source text, ensuring the output remains informative and grounded.

For scenarios requiring smaller, deployable models, Melanie Sclar et al. introduce Referee: Reference-Free Sentence Summarization with Sharper Controllability through Symbolic Knowledge Distillation. This framework distills knowledge from massive models into compact ones while applying rigorous symbolic quality filters, producing student models that ultimately surpass their teachers in factual precision.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 630 papers
  2. Version 1 · 2026-10-04 · 782 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_abstractive_hallucination_0cd3997c,
  title = {abstractive hallucination},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/abstractive-hallucination/rdr_0cd3997c46864f33804c5507c7dd0e4a},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 2, frozen text):

@misc{lacuna_abstractive_hallucination_0cd3997c_v2,
  title = {abstractive hallucination},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/abstractive-hallucination/rdr_0cd3997c46864f33804c5507c7dd0e4a/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake