Source-linked AI summary
Evaluating the Factual Consistency of Abstractive Text Summarization
Wojciech Kryściński, Bryan McCann, Caiming Xiong, Richard Socher
TL;DR
Factual inconsistency in abstractive summaries is not captured by standard summarization metrics, motivating source-grounded verification. The paper trains a weakly supervised BERT-based model with transformation-generated data and explanatory span extraction, and reports superior performance to NLI and fact-checking baselines alongside faster, more consistent human verification. Its document-sentence setting does not handle dependencies across summary sentences, and commonsense errors remain difficult to represent with transformations.
Problem
Up to 30% of abstractive summaries contain factual inconsistencies, while standard evaluation does not verify whether statements are entailed by source documents.
Method
A weakly supervised BERT-based model is trained on rule-transformed source sentences to classify consistency and extract supporting or erroneous spans.
Results
The model substantially outperforms models trained on MNLI and FEVER data, while highlights make human annotation 21% faster and increase Fleiss’ κ by 38%.
Takeaways & Limitations
Explanatory span extraction provides useful assistance to humans verifying factual consistency between source documents and generated summaries.
Takeaways & Limitations
Commonsense mistakes are difficult to encode as transformations, and the document-sentence setting does not handle temporal or coreference dependencies across summary sentences.
Abstract
from arXiv · showhide
Currently used metrics for assessing summarization algorithms do not account for whether summaries are factually consistent with source documents. We propose a weakly-supervised, model-based approach for verifying factual consistency and identifying conflicts between source documents and a generated summary. Training data is generated by applying a series of rule-based transformations to the sentences of source documents. The factual consistency model is then trained jointly for three tasks: 1) identify whether sentences remain factually consistent after transformation, 2) extract a span in the source documents to support the consistency prediction, 3) extract a span in the summary sentence that is inconsistent if one exists. Transferring this model to summaries generated by several state-of-the art models reveals that this highly scalable approach substantially outperforms previous models, including those trained with strong supervision using standard datasets for natural language inference and fact checking. Additionally, human evaluation shows that the auxiliary span extraction tasks provide useful assistance in the process of verifying factual consistency.
1 Introduction
Abstractive summarization can produce factually inconsistent statements, exposing a gap in standard evaluation. The paper proposes a weakly supervised model that verifies consistency and explains its decisions.
- Motivation: Up to 30% of summaries generated by abstractive models contain factual inconsistencies.Such inconsistency can make automatically generated summaries virtually useless in practice.
- Problem definition: Factual consistency checking tests whether summary statements are entailed by the source document, unlike fact checking against external knowledge.The task may require incorporating the entire source document rather than short sentence pairs.
- Approach: The proposed weakly supervised BERT-based model verifies factual consistency and highlights pertinent spans in both source documents and generated summaries.Training data comes from rule-based transformations applied to source documents.
- Contribution: The approach substantially improves over models trained with strong supervision from existing NLI and fact-checking datasets.The comparison concerns factual consistency verification for generated summaries.
2 Related Work
Prior work evaluates factual accuracy with extracted facts or NLI-based comparisons, while related training strategies incorporate entailment without directly verifying their effect on factual consistency.
- Fact extraction: Goodrich et al. define factual accuracy as precision between facts extracted from generated summaries and source documents.Their Transformer-based fact extraction approach outperformed non-model metrics such as ROUGE in human evaluation.
- Fact extraction: Fact extraction methods remain challenged by negated relations and relation names expressed through synonyms.These limitations were identified alongside otherwise positive evaluation results.
- NLI-based checking: Falke et al. re-rank candidate summaries using NLI entailment probabilities between source-document and summary sentences.Their models were trained on SNLI and MNLI data.
- Evaluation: ROUGE does not correlate with factual correctness in the cited comparison.This motivates evaluating factual consistency separately from lexical overlap.
- Training strategies: Multi-task and multi-reward approaches added textual entailment to summarization training, but did not verify whether this improved entailment.The cited works used NLI as an auxiliary task or reward.
3 Methods
The method uses document-level context and weakly supervised transformations to train factual consistency models, with optional span explanations and manually annotated evaluation data.
- Methods: Sentence-to-sentence checking is insufficient because summary sentences can paraphrase multiple source fragments and source coreference can bind information across sentences.Common errors include incorrect entities, numbers, pronouns, negations, and commonsense claims.
- Methods: Each summary sentence is verified against the entire source document in the proposed document-sentence approach.This broader context is intended to handle ambiguities that sentence pairs cannot resolve.
- Training data: Because no supervised factual-consistency dataset exists, the approach creates weakly supervised examples from unannotated in-domain source documents.Large-scale human annotation is described as prohibitively expensive and time consuming.
- Training data: Semantically invariant and variant transformations generate positive and negative claims, respectively.Transformations include paraphrasing, negation, entity and number swaps, pronoun swaps, and token noise.
- Development and test data: Crowdsourced annotation quality and agreement were too low for reliable task labels, and related annotation costs became prohibitively high.The reported limitation concerns expanding the manually annotated evaluation data through crowdsourcing.
- Models: FactCC uses BERT to classify document-claim pairs as CONSISTENT or INCONSISTENT, while FactCCX adds heads that select supporting and erroneous spans.The span-selection heads provide source support and identify possible mistakes in claims.
4 Experiments
Experiments compare FactCC with models transferred from MNLI and FEVER, test its transfer to sentence ranking, and assess the explainable FactCCX variant. FactCC substantially outperforms related-task classifiers and other NLI models, while FactCCX retains strong performance and returns informative spans.
- Training data: 1,003,355 training examples were generated from CNN/DailyMail articles, with 50.2% labeled INCONSISTENT and 49.8% CONSISTENT.The examples were produced using the paper’s rule-based data-generation procedure.
- Transfer from related datasets: FactCC substantially outperforms classifiers trained on MNLI and FEVER despite using weakly supervised data.Models were evaluated using class-balanced accuracy and F1 score; the authors attribute differences partly to a domain gap and summarization-specific errors.
- Sentence ranking: FactCC transfers well from document-sentence training to sentence ranking and outperforms all compared NLI models, including BERT fine-tuned on MNLI.The ranking task pairs an article sentence with positive and negative claims and measures how often the positive claim receives the higher correctness probability.
- Explainable model: FactCCX shows a small performance drop relative to classifier-only FactCC but still substantially outperforms the other two models while returning informative span selections.Because the test summaries lack span-quality annotations, span quality is assessed through human evaluation.
5 Analysis
Human evaluations indicate that FactCCX’s highlighted spans assist factual-consistency assessment, improving both perceived helpfulness and annotation efficiency. However, the approach remains limited by commonsense errors and cross-sentence temporal or coreference dependencies.
- Crowdsourced Experiments: 91.75% of annotators found article highlights at least somewhat helpful, compared with 81.33% for claim highlights.These assessments remained broadly stable after filtering annotations by agreement with author or majority-vote labels.
- Crowdsourced Experiments: FactCCX highlights overlapped human-selected article and claim spans with 65.33% and 65.66% accuracy, respectively.The corresponding token-level F1 scores were 0.6207 for article highlights and 0.6650 for claim highlights; removing noisy annotations increased both metrics.
- Crowdsourced Experiments: Highlights made annotators complete factual-consistency tasks 21% faster and increased Fleiss’ κ inter-annotator agreement by 38%.The comparison used parallel tasks with and without model-generated highlights.
- Crowdsourced Experiments: Crowdsourcing results support span selections as a valuable asset for human-based factual-consistency checking.The experiments evaluated both perceived helpfulness and agreement with independently selected human spans.
- Limitations: Commonsense mistakes are the majority of model errors and are difficult to represent as rule-based transformations for training data.The authors note that humans can spot these errors easily, but defining transformations to generate them is difficult.
- Limitations: Temporal inconsistencies and incorrect coreference across summary sentences are not handled by the document-sentence setting.This limitation arises because certain errors depend on relationships between different summary sentences.
6 Conclusions
The proposed approach verifies factual consistency in abstractive summaries and outperforms models trained on textual-entailment and fact-checking data. Human experiments further indicate that its explainable checking model can assist consistency verification, while broader data augmentation remains future work.
- Document-sentence-level checking handles a broader range of errors than previously proposed sentence-sentence approaches.
- Quantitative studies show the approach outperforms models trained on textual-entailment and fact-checking data.
- Human experiments indicate that the explainable factual consistency model can assist people checking summaries for factual consistency.
- Future work could augment training data with claims spanning multiple source-document sentences or containing commonsense mistakes.