Source-linked AI summary
The Factual Inconsistency Problem in Abstractive Text Summarization: A Survey
Yichong Huang, Xiachong Feng, Xiaocheng Feng, Bing Qin
TL;DR
Abstractive summarization models improve fluency and readability but can distort or fabricate source facts, while conventional evaluation may overlook factual correctness. This survey reviews factual-consistency metrics and optimization methods; it reports that evaluation remains open, with semantic-similarity methods correlating with human judgments at no more than 0.5, and reviews correction systems with documented gains and limitations.
Problem
Neural abstractive summaries can distort or fabricate source facts, while token-overlap evaluation may miss factual contradictions.
Source: §1 · p. 1Method
The survey reviews factual-consistency evaluation metrics, meta-evaluations, and factual-consistency optimization methods for summarization systems.
Source: §1 · p. 2, Triple-based · p. 3Results
Semantic-similarity methods reached state-of-the-art factual-consistency evaluation performance on XSum after parameter search, but correlation with human evaluation was no more than 0.5.
Source: §3.3 · p. 5Takeaways & Limitations
Factual-consistency evaluation remains an open issue, while reviewed correction methods show measurable success but do not cover all error types.
Source: §3.3 · p. 5, §4.3 · p. 6Post-editing methods are limited because corrupted reference summaries differ in data distribution from model-generated summaries.
Source: §4.3 · p. 6
Abstract
from arXiv · showhide
Recently, various neural encoder-decoder models pioneered by Seq2Seq framework have been proposed to achieve the goal of generating more abstractive summaries by learning to map input text to output text. At a high level, such neural models can freely generate summaries without any constraint on the words or phrases used. Moreover, their format is closer to human-edited summaries and output is more readable and fluent. However, the neural model's abstraction ability is a double-edged sword. A commonly observed problem with the generated summaries is the distortion or fabrication of factual information in the article. This inconsistency between the original text and the summary has caused various concerns over its applicability, and the previous evaluation methods of text summarization are not suitable for this issue. In response to the above problems, the current research direction is predominantly divided into two categories, one is to design fact-aware evaluation metrics to select outputs without factual inconsistency errors, and the other is to develop new summarization systems towards factual consistency. In this survey, we focus on presenting a comprehensive review of these fact-specific evaluation methods and text summarization models.
1 Introduction
Abstractive summarization produces fluent, human-readable text but can distort or fabricate source facts. This survey focuses on factual consistency evaluation and optimization methods addressing that problem.
1 Introduction
- Text summarization condenses documents while preserving their main information, and approaches are either extractive or abstractive.Extractive systems select and arrange salient source sentences, whereas abstractive systems generate novel sentences through rephrasing or new words.Source: §1 · p. 1
- Neural abstractive models generate fluent summaries, but their unconstrained generation can produce factual inconsistencies.Such errors distort or fabricate facts in the source; one example exaggerates a magnitude-4.8 earthquake into a powerful quake.Source: §1 · p. 1
- Up to 30% of summaries generated by abstractive models contain factual inconsistencies, threatening system credibility and usability.The survey illustrates this problem with source-supported facts distinguished from inconsistent generated claims.Source: §1 · p. 1, Figure 1 · p. 1
- N-gram overlap metrics can miss factual errors because summaries with nearly identical words may express opposite meanings.The examples “I am having vacation in Hawaii” and “I am not having vacation in Hawaii” share nearly all unigrams and bigrams.Source: §1 · pp. 1–2
- The survey reviews automatic factual-consistency metrics, their meta-evaluations, and methods for optimizing factual consistency in summarization systems.It also synthesizes lessons from prior methods and identifies possible future directions.Source: §1 · p. 2, Figure 2 · p. 2
2 Background
The background introduces abstractive summarization and explains how neural generation can produce fluent text without explicitly rewarding faithfulness. It defines factual inconsistency as either contradiction or unsupported content relative to the source.
2 Background
- The survey first introduces abstractive summarization methods and factual inconsistency categories to explain why errors arise.These topics provide the background for the survey’s later evaluation and optimization review.Source: §2 · p. 2
2.1 Abstractive Summarization methods
- Neural sequence-to-sequence models encode documents and decode summaries, producing readable text, but word-level likelihood training does not necessarily reward faithfulness.RNNs, CNNs, and Transformers can implement the encoder and decoder.Source: §2.1 · p. 2
2.2 Factual Inconsistency Error
- Intrinsic errors contradict the source, whereas extrinsic errors add facts that are neither supported nor contradicted by it.The survey notes that optimization methods mainly target intrinsic errors, while evaluation metrics do not distinguish the two categories.Source: §2.2 · p. 2
3 Factual Consistency Metrics
The survey organizes factual-consistency metrics by supervision and underlying evaluation tool, then discusses meta-evaluation for comparing them. It distinguishes unsupervised metrics from weakly supervised metrics trained on factual-consistency data.
3 Factual Consistency Metrics
- Factual-consistency metrics are divided into unsupervised and weakly supervised categories.Weakly supervised metrics train on documents, model-generated summaries, and factual-consistency scores.Source: §3 · p. 2
- Unsupervised metrics reuse existing tools and comprise triple-based, textual-entailment-based, QA-based, and other approaches.The survey further organizes these metrics according to the tools on which they rely.Source: §3 · p. 2
- Meta-evaluations compare factual-consistency metrics, and the survey organizes existing metrics in Table 1.The survey introduces two meta-evaluation works alongside its metric taxonomy.Source: §3 · p. 2
3.1 Unsupervised Metrics
Unsupervised factual-consistency metrics compare summaries with source documents using triples, textual entailment, question answering, or similarity-based heuristics. Their effectiveness depends on representation and dataset abstractiveness, while QA-based methods trade interpretability for computational cost.
Triple-based
- Fixed-schema relation extraction makes source and summary facts easier to compare than unspecific-schema triples.Unspecified schemas can produce mismatched triples for equivalent facts, whereas fixed schemas normalize the relation representation.Source: Triple-based · p. 3
- Long-premise NLI training improves factual-consistency evaluation by addressing domain shift and the reasoning demands of summarization.Longer premises require abilities such as coreference resolution and abductive reasoning.Source: Triple-based · p. 3
QA-based
- QA-based metrics generate summary questions, answer them from the source, and score answer similarity, yielding interpretable entity-level checks but high computational cost.QAGS and FEQA use question generation and question answering to compare corresponding answers.Source: QA-based · pp. 3–4, Figure 4 · p. 3
Other Methods
- Word-overlap metrics work better on lowly abstractive CNN/DM, whereas semantic-similarity metrics work better on highly abstractive XSum.Both approaches provide baseline effectiveness, using ROUGE or BLEU for overlap and BERTScore for semantic similarity.Source: Other Methods · p. 4
3.2 Weakly Supervised Metrics
Weakly supervised metrics train factual-consistency models on automatically generated synthetic examples, reducing dependence on manually labeled data. They operate at sentence, quantity-entity, or token granularity, but heuristic examples remain mismatched with real model-generated summaries.
3.2 Weakly Supervised Metrics
- Weakly supervised metrics use synthetic data and target sentence-, entity-, or token-level factual consistency.Their automatic training-data construction sidesteps scarcity of factual-consistency annotations.Source: §3.2 · p. 4
Sentence-level
- FactCC classifies sentence consistency with a BERT-based model trained on paraphrases and altered claims generated from CNN/DM.Negation and entity swaps create negative examples, but rule-based transformations cannot simulate every inconsistency type.Source: Sentence-level · p. 4
Entity-level
- HERMAN labels inconsistent quantity entities such as numbers and dates using sequence labeling trained on modified XSum reference summaries.Positive summaries are references, while negative examples replace quantity entities.Source: Entity-level · p. 4
Token-level
- Token-level evaluation is more fine-grained and explainable and shows higher correlation with human factual-consistency judgments.Its synthetic negatives are created by reconstructing parts of reference summaries.Source: Token-level · p. 4
- Weakly supervised metrics remain limited by expensive human annotations and synthetic examples that differ from summaries produced by real models.Consequently, effectiveness observed during training may be limited in actual factual-consistency evaluation.Source: Token-level · p. 4
3.3 Meta Evaluation
Meta-evaluation compares factual-consistency metrics against human scores from a common annotator group to reduce the comparability problems caused by differing annotation settings. Koto et al. [2020] find that tuned semantic similarity reaches state-of-the-art performance on XSum, but human correlation remains below 0.5.
3.3 Meta Evaluation
- Different annotation settings and annotator disagreement make direct comparison of reported metric correlations difficult.Meta-evaluation addresses this by using scores from the same annotator group.Source: §3.3 · pp. 4–5
- Meta-evaluation measures metric quality by correlating metric scores with factual-consistency scores from the same annotators.This provides a common basis for comparing several metrics.Source: §3.3 · p. 5
- On highly abstractive XSum, tuned semantic similarity reaches state-of-the-art factual-consistency evaluation performance.Koto et al. [2020] search optimal pretrained-language-model layers for BERTScore.Source: §3.3 · p. 5
- Human correlation remains below 0.5, leaving factual-consistency evaluation an open research problem.The reported ceiling applies to the meta-evaluated factual-consistency setting described here.Source: §3.3 · p. 5
4 Factual Consistency Optimization
The survey groups factual-consistency optimization methods into fact encoding, textual entailment, post-editing, and other approaches. These methods modify representations, training objectives, generated summaries, data, or inference constraints to reduce factual errors.
4 Factual Consistency Optimization
- Fact-consistency optimization methods are divided into fact encode-based, textual entailment-based, post-editing-based, and other methods.The survey organizes these methods in Figure 5 and Table 2.Source: §4 · p. 5
4.1 Fact-Encode-based
- Fact-encode methods extract relation triples and encode them into summarization models through sequential or graph-based representations.Sequential models concatenate fact descriptions, while graph-based models represent entities and relations with graph neural networks or commonsense knowledge graphs.Source: §4.1 · p. 5, sequential fact encode · p. 5, graph-based fact encode · p. 5
4.2 Textual-Entailment-based
- Textual-entailment-based models incorporate entailment knowledge into encoding, decoding, and reward design for summary generation.The entailment-aware decoder uses RAML, with the reward defined as the entailment score of the generated summary.Source: §4.2 · pp. 5–6
4.3 Post-Editing-based
- Post-editing methods treat generated summaries as drafts and use fact correctors that take the source and draft as input to produce corrected summaries.SpanFact masks entities and selects replacement spans from the source, while an end-to-end BART corrector is designed to handle more error types.Source: §4.3 · p. 6
- 26% of factually inconsistent summaries were corrected by SpanFact, while fewer than 1% of factually consistent summaries were wrongly corrupted.SpanFact is limited to correcting entity errors, and the end-to-end corrector did not outperform it in human evaluation.Source: §4.3 · p. 6
- Post-editing performance is limited by the distribution gap between automatically corrupted reference summaries used for training and model-generated summaries encountered during testing.Both post-editing approaches construct artificial training data automatically rather than using expensive human annotation.Source: §4.3 · p. 6
4.4 Other
- Other approaches filter unfaithful training pairs, impose entity and keyphrase constraints during beam search, or target domain-specific summarization settings.Inference-time constraints can be plugged into abstractive models without modifying their internal structures, but their factual-consistency gains and optimal constraints remain open questions.Source: §4.4 · p. 6
5 Conclusion and Future Directions
The survey concludes that factual-consistency research should expand beyond its current focus on intrinsic errors. It highlights paragraph-level evaluation, extrinsic-error optimization, and factual consistency in other conditional text-generation tasks as future directions.
5 Conclusion and Future Directions
- Future optimization should address extrinsic errors, because existing methods mainly focus on intrinsic errors and maximum likelihood training does not explicitly model document-summary constraints.Entities and quotations are examples of constraints that should appear in the document, and reinforcement learning may help model them during generation.Source: §5 · p. 7
- Paragraph-level factual-consistency metrics should account for relationships between sentences rather than evaluating each sentence independently.The survey characterizes paragraph-level evaluation as more challenging and valuable.Source: §5 · p. 7
- Factual-consistency research should also extend to image captioning, visual storytelling, and other conditional text-generation tasks with cross-modal consistency challenges.The survey presents these directions as guidelines rather than an exhaustive list.Source: §5 · p. 7