PaperlayerView exact revision on arXiv

Source-linked AI summary

On Improving Summarization Factual Consistency from Natural Language Feedback

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

arXiv:2212.09968v2cs.CL

TL;DR

Summarization systems can produce fluent outputs that still contain factual errors, a quality failure that can undermine user trust. The paper builds DeFacto, a dataset of human feedback and demonstrations, and studies three feedback-based generation tasks. Human-edited summaries are more factually consistent than reference summaries and initial outputs, while the authors identify limits in dataset scale and scope.

  • Problem

    Summarization systems can produce fluent, relevant summaries that still contain factual errors, which can risk user decisions and erode trust.

    Source: §1 · p. 1

  • Method

    The authors collect DeFacto, pairing human-edited summaries with explanations and corrective instructions, and study summary editing, feedback generation, and joint feedback-and-edit generation.

    Source: §1 · p. 2

  • Results

    Human-edited summaries are more factually consistent than reference summaries and initial system outputs, and the Editor model outperforms a baseline that generates only edited summaries.

    Source: §3.1 · p. 4

  • Takeaways & Limitations

    The authors conclude that natural-language feedback can improve summary factual consistency and may be explored for other text qualities.

    Source: §8 · p. 9

    The annotation process is time-consuming, leaving a relatively small dataset based on one summarization dataset and one base model.

    Source: Limitations · p. 10

Abstract

from arXiv · show

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.

1 Introduction

The introduction frames factual consistency as a user-expected quality that summarization systems still miss, with errors undermining trust. It proposes using human demonstrations and natural-language feedback to align generated summaries with that quality.

1 Introduction

  • Factual errors in fluent, relevant summaries can create risks for users and make systems harder to trust for decision-making.The introduction links model-user mismatch to training datasets that are not collected around predefined user needs.Source: §1 · p. 1
  • The paper studies three feedback-driven tasks: editing summaries from feedback, predicting feedback, and predicting both feedback and an edited summary.Factual consistency is the user-required quality used to study these scenarios.Source: Figure 1 · p. 1, §1 · p. 2
  • DeFacto combines human-edited summaries with corrective instructions and explanations about factual consistency.The introduction presents the dataset as supporting the three tasks and analyses of factual errors and editing operations.Source: §1 · p. 2
  • The Editor model outperforms a baseline that generates only the edited summary, suggesting feedback can benefit training for this task.The introduction also reports that Editing can use human feedback to adjust initial outputs, Critic can generate meaningful feedback, and Editor can correct factuality errors without explicit human intervention.Source: §1 · p. 2

2 The DEFACTO Dataset

DeFacto pairs summaries generated for XSum with human judgments, explanations, evidence, corrective instructions, and edits. Its annotation process asks workers to assess and categorize errors, justify judgments, and make minimal corrections; reported agreement varies across annotation tasks.

2 The DEFACTO Dataset

  • The dataset uses PEGASUS-generated summaries for XSum validation and test examples; 71.1% of annotated summaries contain factual errors.The collection included around 1,000 test examples and 1,500 validation examples, with one annotation per example and two additional annotations for 100 test examples.Source: §2 · p. 3, Table 1 · p. 3, §2.2 · p. 3

2.1 Annotation Process

  • Annotators assess summary consistency, categorize errors, explain their judgments, select supporting source evidence, write corrective instructions, and minimally edit the summary.Instructions use six templates: Remove, Add, Replace, Modify, Rewrite, and Others.Source: §2.1 · p. 3

2.2 Data Collection

  • Agreement was 0.5552 for extrinsic errors, 0.1899 for intrinsic errors, and 0.5260 for any factual errors; evidence matching was 0.4403.For explanations, instructions, and edited summaries, average ROUGE-1 F-scores were 30.52, 50.96, and 71.77, respectively.Source: §2.2 · p. 3

3 DEFACTO Analysis

DEFACTO analysis finds that human-edited summaries are more factually consistent than reference summaries and initial system outputs. It also examines their similarity and intrinsic properties, and the patterns in corrective instructions.

3 DEFACTO Analysis

  • 71.1% of examples contain factual errors, including 58.8% with extrinsic errors, 22.0% with intrinsic errors, and 9.63% with both.The collected XSum validation data were split into training and validation sets for the experiments.Source: §3 · p. 3

3.1 Edited Summary

  • Human-edited summaries are more factually consistent than both reference summaries and initial system outputs.The comparison uses the automatic factuality metrics DAE and QAFactEval.Source: §3.1 · p. 4
  • Human-edited summaries are highly similar to initial outputs, but less similar to reference summaries than the initial outputs are.The authors suggest the reference summaries and initial outputs may share factual errors, increasing their similarity.Source: §3.1 · p. 4
  • Human-edited summaries are less abstractive than initial outputs and reference summaries, consistent with a faithfulness–abstractiveness trade-off.The authors note that removing non-factual information, a common correction for extrinsic errors, may also reduce abstractiveness.Source: §3.1 · p. 4

3.2 Instructions

  • Correcting an initial summary takes an average of 1.52 instructions; removing and replacing information are the most frequent operations.Extrinsic errors are more likely to be corrected by removal, while intrinsic errors are more likely to be corrected by replacement.Source: Figure 3 · p. 4, §3.2 · p. 5

4 Summary Editing for Improving Factual Consistency

This section frames factual-consistency editing as a Seq2Seq task guided by human feedback, then evaluates fine-tuned and zero-shot systems.

4 Summary Editing for Improving Factual Consistency

  • The section introduces editing initial summaries according to human feedback as a task for improving factual consistency.Source: §4 · p. 5

4.1 Methods

  • The task maps a document, initial system summary, and human instructions and explanation to a human-edited summary; the models are fine-tuned T5 or T0 variants.Source: §4.1 · p. 5

4.2 Experiments

  • Both instructions and explanations guide editing, and their combination performs best; adding the source document further improves performance, especially factual consistency.Source: §4.2 · pp. 5–6
  • Findings from T0pp generalized to T0-3B and T5-3B with few exceptions; T0pp generally outperformed T0-3B except on QAFactEval, while T0-3B generally outperformed T5-3B.Source: §4.2 · p. 6
  • Among summaries generated by the trained T0pp checkpoint from input documents, initial system summaries, and human instructions, 97% were more factually consistent than originals and 91% followed instructions.Three MTurk annotators evaluated the generated summaries; results use majority voting, with ties ignored.Source: §4.2 · p. 6

4.3 Case Study of LLM Summary Editing

  • In zero-shot editing, GPT-3.5 achieved higher factual consistency than fine-tuned T0pp but lower content similarity with the original summary.Source: §4.3 · p. 6

5 Generating Feedback for Improving Factual Consistency

The section trains a Critic to generate editing instructions and evaluates whether those instructions help summary editing, including in zero-shot LLM trials.

5.1 Methods

  • The Critic takes the document and initial system summary as input and generates human-written instructions, fine-tuned with T0 on examples whose initial summaries contain factual errors.Source: §5.1 · pp. 6–7

5.2 Experiments

  • Critic-generated instructions help improve factual consistency over editing from the document and summary alone, though outputs are less similar to human-edited summaries and instructions work less well than human-written ones.Source: §5.2 · p. 7
  • 39% of generated instructions were judged useful for correcting errors and 31% partly useful, while 24% were exactly equivalent to human instructions and 45% partly equivalent.Annotators evaluated 100 test examples; the authors attribute the gap to multiple acceptable ways to correct errors.Source: §5.2 · p. 7

5.3 Case Study of LLM Critic

  • Instructions from zero-shot GPT-3.5 and GPT-4 led to more factual but less similar summaries than fine-tuned models; GPT-3.5 returned “No editing instructions needed” 23.9% of the time despite being told to edit.Source: §5.3 · pp. 7–8

6 Summary Editor with Feedback Generation and Editing

The Editor predicts feedback and an edited summary together, first for known errors and then for arbitrary system outputs that may need no correction.

6 Summary Editor with Feedback Generation and Editing

  • The Editor predicts both human feedback and an edited summary from the document and initial summary, using T0 or T5 with textual tags to mark the output parts.Source: §6 · p. 8, §6.1 · p. 8

6.1 Correcting Known Factual Errors

  • For known factual errors, Editor models improved textual similarity over the document-and-summary-only Editing model, and generating explanations also improved factual consistency.The authors suggest predicting related information alongside a target generation can benefit language-generation performance.Source: §6.1 · p. 8

6.2 Detecting and Correcting Factual Errors

  • For arbitrary outputs, training uses all of DeFacto, preserves summaries without factual errors, substitutes human edits for erroneous ones, and targets explanations as feedback.Source: §6.2 · p. 8
  • On arbitrary outputs, the Editor achieved competitive performance against factuality-focused systems and balanced content similarity with factual consistency despite using much less training data.Source: §6.2 · pp. 8–9

7 Related Work

Related work covers factuality methods and evaluation datasets, then contrasts DeFacto’s instruction-following edits and natural-language feedback with prior work.

7 Related Work

  • Prior factuality approaches include metric-based training, new objectives and multitask learning, post-editing or reranking, and factuality-aware pretraining.Source: §7 · p. 9
  • Prior factuality benchmarks use human scores or binary labels, whereas DeFacto provides natural-language feedback and error-free human-edited summaries.Source: §7 · p. 9
  • Unlike related text-editing work, this paper targets summary factual consistency with instructions; prior summarization feedback studies used numerical quality scores rather than natural-language feedback.Source: §7 · p. 9

8 Conclusions

The paper studies summary factual consistency as a user-expected quality and uses DEFACTO to examine how human demonstrations and feedback can improve text generation. It concludes that feedback can improve factual consistency and may extend to other text qualities.

8 Conclusions

  • DEFACTO supports three proposed NLG tasks for using human feedback to improve summary factual consistency.The paper also identifies natural-language feedback for other text qualities as a promising direction for future work.Source: §8 · p. 9

Limitations

The annotation process is complex and time-consuming, limiting dataset scale and experimental scope. Extending the methods to subjective qualities is challenging because preferences can be less objective, complicating annotation and evaluation.

Limitations

  • The complex, time-consuming annotation task constrained dataset scale and the study to XSum and one base summarization model, Pegasus.The authors cite time and budget limitations and a relatively low proportion of annotators qualifying after the test.Source: Limitations · p. 10
  • Applying the methods to subjective user preferences is challenging because such qualities can be less well-defined and objective, increasing annotation and evaluation ambiguity.The authors leave generalization to more diverse expectations and preferences for future work.Source: Limitations · p. 10

A Data Collection Details

The data collection details describe XSum-based annotation, instruction templates, and examples of correcting factual errors in summaries.

A Data Collection Details

A.1 Instruction Templates

  • Annotators received six instruction templates for removing, adding, replacing, modifying, rewriting, or giving other instructions about summary content.Source: §A.1 · p. 13
  • A summary edit can require more than one instruction.Source: §A.1 · p. 13

Example 1

  • Another edit replaces Wirral and a job-loss claim with Moreton and the claim that 342 families would be affected.Source: Example 1 · p. 14
  • A court-related example replaces unsupported references to Teesside Crown Court and Middlesbrough with Teesside Youth Court and Stephen Street.Source: Example 1 · p. 14

A.2 Annotated Examples

  • One example removes unsupported first names and a location claim, yielding a summary with those details omitted.Source: Example 1 · p. 14

Example 1

  • Another example replaces an unsupported concert-hall name with Aberdeen Performing Arts.Source: Example 1 · p. 14
  • An example distinguishes a catastrophic composting program from a decision to stop it, revising the summary to clarify which was described as a failure.Source: Example 1 · p. 14

B Factuality Metrics Setting

The factuality metrics setting uses DAE dependency-level predictions to produce summary-level scores and uses QAFactEval's summary-level prediction scores directly.

B Factuality Metrics Setting

  • DAE predicts whether each summary dependency arc is entailed by the input document, and the predictions are used to define a summary-level factuality score.The paper transfers DAE's dependency-level predictions into summary-level scores.Source: B Factuality Metrics Setting · p. 14
  • Figure 4 concerns the number of instructions needed to correct a factually inconsistent system-generated summary.Source: Figure 4 · p. 15
  • QAFactEval prediction scores are used directly because they are already summary-level scores.Source: B Factuality Metrics Setting · p. 15

C.1 Intrinsic Summary Evaluation

This material defines intrinsic summary metrics, describes the distribution of human editing instructions, and documents experimental baselines and language-model prompts. GPT-3.5 sometimes declined to generate instructions for factually inconsistent summaries, unlike GPT-4.

C.1 Intrinsic Summary Evaluation

  • Intrinsic evaluation uses compression rate, Extractive Fragment Coverage over shared word sequences, and a word-level novelty measure.The passages define compression as a ratio of input-document and summary word counts and describe shared word sequences as extractive fragments.Source: §C.1 · p. 15

C.2 Human-written Instructions

  • Human editing can take more than one instruction, and Fig. 4 shows the distribution of instruction counts.Source: §C.2 · p. 15

D Experimental Details

  • Experiments use T5-3B, T0-3B, and T0pp (about 3B, 3B, and 11B parameters); training took one GPU for 3B models and eight GPUs for 11B models.Reported training times were around 8 hours for 3B models and 20 hours for 11B models.Source: D Experimental Details · p. 15

D.1 Baseline Summarization Systems

  • Editor comparisons include CCGS, CLIFF, ReDRESS, FactPegasus, and CompEdit, spanning contrastive, factuality-trained, and compression-based systems.Source: §D.1 · p. 15

D.2 Setting of LLM Case Study for Summary Editing

  • GPT-3.5 editing prompts use temperature zero and request necessary changes while preserving most of the summary; prompts without instructions permit a full rewrite if all information is incorrect.Source: §D.2 · pp. 15–16

D.3 Setting of LLM Case Study for Instruction Generation

  • Instruction-generation prompts define six edit operations, prioritize removal and replacement before addition, and request edits of similar length and style.Source: §D.3 · p. 16
  • GPT-3.5 responded “No editing instructions needed” for 23.9% of examples, compared with 1.8% for GPT-4.The authors describe this as evidence that LLMs still lack controllable instruction-generation ability.Source: §D.3 · p. 16
Loading 2212.09968v2…