On Improving Summarization Factual Consistency from Natural Language Feedback

Yixin Liu, Budhaditya Deb, Milagro Teruel, Aaron Halfaker, Dragomir Radev, Ahmed H. Awadallah

TL;DR

This paper tackles the persistent problem of factual inconsistency (hallucination) in text summarization by leveraging natural language feedback. The authors introduce DeFacto, a high-quality dataset containing human demonstrations and detailed feedback (explanations, evidence, and instructions) on system-generated summaries. They demonstrate that fine-tuned language models can successfully use this feedback to correct factual errors, and that training a model to generate feedback before editing acts like a chain-of-thought that significantly boosts performance.

What Problem Does This Paper Address?

While modern abstractive summarization models generate highly fluent text, they frequently suffer from a lack of factual consistency—meaning they hallucinate information or misinterpret the source text.

A major root cause is the training data itself. Popular datasets like XSum are notoriously noisy; prior work shows that up to 76.9% of XSum's reference summaries contain hallucinations. Models trained on this data naturally learn to hallucinate. Gathering entirely new, massive datasets aligned with perfect factual standards is prohibitively expensive. Instead, this paper asks: Can we fix imperfect summaries by training models to follow natural language feedback and human corrections?

Methods

The authors frame summary correction as a sequence-to-sequence (Seq2Seq) problem and explore three distinct natural language generation (NLG) tasks to utilize human feedback:

  1. The Editing Model: Takes the input document, the initial flawed summary, and human feedback (instructions/explanations) to generate a factually corrected summary.
  2. The Critic Model: Takes the input document and initial summary to automatically predict the corrective human feedback.
  3. The Editor Model: Automatically corrects summaries end-to-end by predicting both the feedback and the edited summary given the document and initial summary.
Figure 1: Three NLG tasks studied using our dataset. The Editing model aims to improve the initial system- generated summary given human feedback. The Critic model aims to predict human feedback according to a user-required quality. The Editor model aims to auto- matically correct factual errors by predicting both the human feedback and edited summary.

The authors evaluate these tasks by fine-tuning a family of encoder-decoder models (T5-3B, T0-3B, and T0pp) using teacher forcing, and compare them against zero-shot Large Language Models (LLMs) like GPT-3.5 and GPT-4.

Datasets

The paper introduces DeFacto (Demonstrations and Feedback for FacTual Consistency), a dataset built on top of the XSum validation and test sets. The initial summaries were generated using PEGASUS, a top-performing summarization model.

The dataset contains 2,561 data points, where human annotators performed a multi-step pipeline:

  • Detected and categorized errors (intrinsic vs. extrinsic).
  • Provided a natural language explanation and highlighted source text evidence.
  • Wrote corrective instructions using standardized templates (e.g., Remove, Replace, Modify).
  • Created a final, human-edited, factually consistent summary.
Figure 2: Data example in DEFACTO. The initial summary contains a factual error about the name of the program Planet Earth II. The annotator provided an explanation about why the initial summary is not factually consistent, evidence (i.e., a sentence in the input document) to support their claims, instructions on how to correct the summary, and an edited summary (a demonstration) without the factual error.

Key Insights and Ideas

  • Feedback as a "Chain of Thought": The Editor model (which generates both feedback and the edit) performed better than a baseline that only attempted to output the corrected summary. Forcing the model to explain why something is wrong and how to fix it acts as an intermediate reasoning step that improves final generation quality.
  • Error Type Dictates the Fix: Fine-grained analysis of DeFacto revealed a clear mapping between error types and human edits. Extrinsic errors (hallucinating completely new info) are predominantly fixed by removing information. Intrinsic errors (misinterpreting source info) are more diverse but are most frequently fixed by replacing information.
Figure 3: Distribution of six different types of instruc- tions. removing information and replacing information are the most frequent types. Extrinsic errors are more likely to be corrected by removing while Intrinsic errors are more likely to be corrected by replacing.
  • The LLM Controllability Gap: While zero-shot LLMs (like GPT-3.5) are highly capable of improving factual consistency, they struggle significantly with controllable text generation. They often rewrite the entire summary from scratch rather than making minimal edits, or completely miss errors—GPT-3.5 claimed "No editing instructions needed" 23.9% of the time when given a known factually inconsistent summary.

Key Results

  • Factual Consistency Boost: Human-edited summaries in DeFacto achieved dramatically higher factuality scores compared to both the original system outputs and the original XSum reference summaries (measured via automatic metrics DAE and QAFactEval).
  • Fine-Tuned Model Success: The fine-tuned T0pp model effectively ingested human instructions and explanations. The D+S+I+E variant (Document + Summary + Instruction + Explanation) achieved a DAE factuality score of 0.911 (up from the initial system's 0.704) and closely matched human edits (89.22 ROUGE-1).
  • Competitive End-to-End Correction: When evaluated on the entire test set (including summaries without errors), the end-to-end Editor model achieved competitive or superior performance compared to five state-of-the-art factuality-aware summarization systems (like CLIFF and FactPegasus), despite being trained on significantly less data.

Limitations and Future Work

The authors openly note several limitations:

  • Scale: DeFacto is relatively small (2,561 examples) due to the complex, expensive, and time-consuming nature of human annotation.
  • Diversity: The dataset is restricted to one source dataset (XSum) and one base model (PEGASUS).
  • Subjectivity: Factual consistency is a relatively objective metric. Future work is required to see if natural language feedback can generalize to highly subjective user preferences, such as style, tone, or specific formatting constraints.

Summary updated

Abstract

Despite the recent progress in language generation models, their outputs may not always meet user expectations. In this work, we study whether informational feedback in natural language can be leveraged to improve generation quality and user preference alignment. To this end, we consider factual consistency in summarization, the quality that the summary should only contain information supported by the input documents, as the user-expected preference. We collect a high-quality dataset, DeFacto, containing human demonstrations and informational natural language feedback consisting of corrective instructions, edited summaries, and explanations with respect to the factual consistency of the summary. Using our dataset, we study three natural language generation tasks: (1) editing a summary by following the human feedback, (2) generating human feedback for editing the original summary, and (3) revising the initial summary to correct factual errors by generating both the human feedback and edited summary. We show that DeFacto can provide factually consistent human-edited summaries and further insights into summarization factual consistency thanks to its informational natural language feedback. We further demonstrate that fine-tuned language models can leverage our dataset to improve the summary factual consistency, while large language models lack the zero-shot learning ability in our proposed tasks that require controllable text generation.
BibTeX
@inproceedings{liu2023improving,
  author = {Yixin Liu and Budhaditya Deb and Milagro Teruel and Aaron Halfaker and Dragomir Radev and Ahmed H. Awadallah},
  title = {{On Improving Summarization Factual Consistency from Natural Language Feedback}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2023},
  doi = {10.18653/v1/2023.acl-long.844},
}
Figure