Source-linked AI summary
Annotating and Modeling Fine-grained Factuality in Summarization
Tanya Goyal, Greg Durrett
TL;DR
Factuality detection for summarization lacks a clear account of which errors matter and whether synthetic training data reflects real generation errors. The paper compares synthetic and human-labeled supervision at multiple annotation granularities, finding that fine-grained human annotations better support detection and localization, while localized errors can improve XSum summarization training.
Problem
Existing factuality datasets and models do not clearly reveal which errors summarization systems make or where current detection techniques succeed and fail.
Method
The paper compares synthetic and human-labeled factuality data and models errors at word-, dependency-, and sentence-level granularity.
Results
Fine-grained human annotations more reliably detect and localize errors, while synthetic datasets poorly match errors in challenging abstractive summarization.
Takeaways & Limitations
Using localized factuality errors to modify summarization training produces more factual models on error-prone gold datasets.
Takeaways & Limitations
Factuality judgments about inferences, world knowledge, and implicatures are not standardized and depend on annotators and task.
Abstract
from arXiv · showhide
Recent pre-trained abstractive summarization systems have started to achieve credible performance, but a major barrier to their use in practice is their propensity to output summaries that are not faithful to the input and that contain factual errors. While a number of annotated datasets and statistical models for assessing factuality have been explored, there is no clear picture of what errors are most important to target or where current techniques are succeeding and failing. We explore both synthetic and human-labeled data sources for training models to identify factual errors in summarization, and study factuality at the word-, dependency-, and sentence-level. Our observations are threefold. First, exhibited factual errors differ significantly across datasets, and commonly-used training sets of simple synthetic errors do not reflect errors made on abstractive datasets like XSum. Second, human-labeled data with fine-grained annotations provides a more effective training signal than sentence-level annotations or synthetic data. Finally, we show that our best factuality detection model enables training of more factual XSum summarization models by allowing us to identify non-factual tokens in the training data.
1 Introduction
The paper examines why factuality detection methods do not reliably reflect errors in abstractive summaries and compares synthetic and fine-grained human supervision. It finds that fine-grained annotations support better error detection and can improve summarization training.
- Existing factuality efforts provide limited insight into which errors state-of-the-art summarizers make and how to fix them.
- Synthetic corruption and paraphrasing target error distributions unlike those in generated summaries, leading to poor practical factuality detection.
- CNN/Daily Mail and XSum generated summaries exhibit substantially different factuality-error distributions, so one dataset-creation approach cannot serve both domains.
- Fine-grained human annotations enable more reliable detection and localization of errors than sentence-level supervision in highly abstractive XSum.
- A modified training objective uses detected error spans to train more factual summarization models on noisy gold data.
2 Training Datasets to Compare
The paper compares synthetic factuality datasets built from corruptions or paraphrases with annotation formats derived from spans and dependencies. It evaluates these approaches across XSum and CNN/DailyMail, whose abstraction levels and factuality errors differ.
- The study compares synthetic-error distributions with factuality errors produced by models trained on XSum and CNN/DailyMail.XSum uses highly abstractive first-sentence summaries, whereas CNN/DailyMail references have higher source overlap.
- Entity-centric Synthetic Data (Ent-C): Entity-centric transformations perturb entities and noun phrases through swaps, negation, noise injection, and paraphrasing.The transformations include entity and number swapping, pronoun swapping, sentence negation, arbitrary noise injection, and backtranslation.
- Entity-centric Synthetic Data (Ent-C): The entity-centric dataset uses 50k labeled CNN/DailyMail pairs and a similarly sized XSum dataset, with rule-based corruptions yielding span-level annotations.
- Generation-centric Synthetic Data (Gen-C): Generation-centric data treats 10th-best paraphrases as more error-prone than 1-best generations and labels their new information as non-factual.This approach automatically produces sentence-level and dependency-level annotations for both domains, with 40k training examples per domain.
- Types of supervision: The compared datasets support different annotation granularities, requiring format unification for sentence-level and dependency-level model comparisons.
- Types of supervision: Span highlights can be converted into dependency labels by marking an arc non-factual when either its head or child word is highlighted.
3 Analysis of Error Types
The paper develops an error taxonomy and compares factuality errors in real summarization outputs with synthetic corruptions. Error distributions differ across domains and synthetic data, limiting transfer to realistic generation errors.
- Error taxonomy: The taxonomy divides errors into entity-related, event-related, noun phrase-related, and other categories, with the first three split into intrinsic and extrinsic errors.Intrinsic errors misinterpret source information, whereas extrinsic errors hallucinate unsupported information; other errors are not split by this distinction.
- Taxonomy rationale: The paper’s taxonomy targets grammatical roles of errors rather than their effects on readers.This framing is intended to support improving summarization models.
- Data and analysis: The analysis annotates errors in BART-generated XSUM summaries, CNN/DM model summaries, and synthetic Ent-C and Gen-C examples.The XSUM annotations cover 100 examples, while CNN/DM annotations cover 50 summaries from 10 models.
- Observed errors: 60% of errors made by BART models on XSUM are extrinsic, while 40% are intrinsic.The paper links these patterns to noisy, highly abstractive gold summaries and long-range dependency errors.
- Synthetic versus real errors: Synthetic datasets do not reflect the error distributions of actual generation models, and Ent-C covers intrinsic event errors mainly through pronoun swaps.Because CNN/DM and XSUM exhibit different errors, synthetic data modeled on one setting may not transfer effectively to the other.
4 Factuality Models to Compare
The paper compares sentence-level and dependency-arc factuality models for judging whether generated summaries are supported by their source documents. The arc-based formulation localizes errors, while weak supervision adapts it when fine-grained arc labels are unavailable.
- Model formulations: A sentence-factuality model makes one factuality judgment for an entire summary, whereas an arc-factuality model judges dependency arcs independently.Arc-level decisions are combined into a sentence-level decision, enabling localization of factuality errors.
- Arc-Factuality model: The DAE model encodes the document and summary, derives representations for each dependency arc, and predicts factual or non-factual labels.Each arc representation concatenates encoder outputs for its head and child words.
- Arc-level decisions: An arc is non-factual when its dependency relationship is not entailed by the document, so different arcs in one summary can receive different labels.For example, “arrested ← games” is unsupported while “seven ← games” is supported in the illustrated example.
- Training data: DAE training uses arc-labeled examples derived from synthetic or human-labeled data.The training examples pair a document and summary with labels for the summary’s dependency arcs.
- Weak supervision: DAE-Weak trains from sentence-level labels by constraining factual summaries to have all factual arcs and non-factual summaries to contain at least one unsupported arc.The model maximizes the marginal likelihood of arc labelings that satisfy these constraints.
5 Experiments
Experiments show that synthetic factuality datasets poorly represent errors made by abstractive summarizers, especially on XSUM, whereas fine-grained human annotations substantially improve detection.
- 5.1 Evaluation of Synthetic Training Datasets: Around 75% accuracy on human-annotated CNN/DAILYMAIL data contrasted with over 90% on held-out synthetic data, indicating distribution mismatch.Synthetic-trained models fit their training distributions well but generalize poorly to actual generation errors.
- 5.1 Evaluation of Synthetic Training Datasets: On XSUM, synthetic-trained factuality models perform very poorly, close to the majority-label baseline.The result reflects a fundamental difference between XSUM generation errors and artificial corruption mechanisms.
- 5.2 Human Annotated Dataset Evaluation: Small amounts of human-annotated data outperform large auto-generated datasets by a large margin on XSUM.Fine-grained annotations significantly boost performance over sentence-level models.
- 5.2 Human Annotated Dataset Evaluation: Models using fine-grained annotations significantly outperform sentence-level models, while DAE-Weak also improves on the sentence-level model by localizing errors.The comparison supports decomposing factuality errors rather than predicting only sentence-level labels.
6 Localization of errors
The study evaluates factuality error localization at dependency and word levels, then uses localized errors to mask unsupported tokens during summarization training. This approach yields lower factuality error rates and higher human ratings than the compared systems.
- 6.1 Localizing Factuality on XSUM: Error localization is evaluated at dependency-arc and word levels, extending factuality assessment beyond sentence-level performance.The study examines whether models can identify where errors occur within generated summaries.
- 6.1 Localizing Factuality on XSUM: DAE outperforms DAE-Weak at both granularity levels, supporting fine-grained annotations for more reliable localization.DAE-Weak nevertheless achieves comparable dependency-level recall, and both models are recall-oriented.
- 6.1 Localizing Factuality on XSUM: A selected DAE checkpoint achieves 83.9 recall when optimized directly for that metric.This checkpoint is used for the downstream application in Section 6.2.
- 6.2 Downstream Applications: The downstream method identifies unsupported facts in XSUM training data and ignores their corresponding tokens during summarizer training.The modified objective maximizes likelihood only for tokens marked factual by the DAE model.
- 6.2 Downstream Applications: DAE-masked training produces lower word and sentence error rates and higher human factuality ratings than the compared systems.The evaluation uses both DAE-based error measures and human judgments on generated summaries.
7 Related Work
Related work has assessed factuality through evaluation models, model modifications, correction, constrained decoding, and adjacent verification tasks. These approaches face the difficulty that generated factual errors are varied and not easily captured by discrete rules.
- Prior Summarization Evaluation: Earlier summarization evaluation emphasized content selection and grammaticality, with comparatively little attention to factuality.The related literature includes human evaluation and methods such as Pyramid that could potentially assess content.
- Methods for Enforcing Factuality: Factuality enforcement has included encoding semantic structures, post-hoc correction, and constrained decoding.These methods modify generation or attempt to repair outputs after generation.
- Methods for Enforcing Factuality: Such techniques struggle to cover the full range of factual errors because factuality is fuzzy and cannot be easily reduced to discrete rules.This motivates modeling factuality from annotated examples rather than relying only on fixed constraints.
- Related Tasks: Related tasks include factuality in radiology summarization, data-to-text generation, and fact verification.Fact-verification claims are usually human-authored straightforward facts, unlike claims embedded in generated summary structures.
8 Conclusion
The paper concludes that synthetic datasets are poorly suited to evaluating factuality in challenging abstractive summarization, while fine-grained human annotations support more factual models.
- 8 Conclusion: Existing synthetic datasets are not well-suited to factuality evaluation of recent summarization models such as BART on challenging domains such as XSUM.The conclusion emphasizes mismatch between synthetic data and errors from modern abstractive systems.
- 8 Conclusion: Human-annotated data, especially fine-grained annotations, can enable training of more factual summarization models.The authors identify improved modeling and data creation as directions for future work.
A Manual Annotation of Errors
Two authors manually characterized factuality errors in a shared set of 50 summaries, finding substantial but imperfect agreement on both erroneous spans and their categories.
- A Manual Annotation of Errors: The authors independently annotated factuality errors in a common set of 50 summaries.The error distributions were derived from manual annotations by two paper authors.
- A Manual Annotation of Errors: 74% of the time, both authors agreed which spans or hallucinations constituted an error.
- A Manual Annotation of Errors: When both authors marked a common span as erroneous, they agreed on its error category 84% of the time.
B Synthetic Dataset Performance on held-out samples
Synthetic factuality models fit held-out synthetic data well, but their performance on human-annotated data is variable because synthetic errors target a different distribution. For XSUM, annotations from one generation model can evaluate other models trained on the same summarization data.
- Synthetic Dataset Performance on held-out samples: Synthetic-data models report very high performance on held-out tests constructed from the same synthetic distributions.This indicates that the models can fit the synthetic domain.
- Synthetic Dataset Performance on held-out samples: Human-annotated CNN/DM performance shows high variance during training, unlike the steadily increasing held-out synthetic accuracy.This behavior also occurs for ENT-C and GEN-C, with greater variance for ENT-C.
- Synthetic Dataset Performance on held-out samples: Optimizing for synthetic error distributions does not necessarily improve performance on errors made by actual generation systems.
- Synthetic Dataset Performance on held-out samples: For highly abstractive XSUM, the study investigates whether annotations from one summarization model can identify errors in outputs from other models.The evaluation uses factuality annotations collected for five XSUM generation models.
- Synthetic Dataset Performance on held-out samples: Performance is similar when training with all models or only other models, indicating cross-model reuse of human factuality annotations.The evaluated models share the same summarization training dataset.
D Implementation Details
The implementation uses ELECTRA-based factuality models and BART-based summarization models, with specified hyperparameters and repeated runs for high-variance results. Human factuality judgments use binary labels and erroneous-span highlighting.
- D Implementation Details: Factuality models are fine-tuned from a 110M-parameter pre-trained ELECTRA discriminator.Five learning-rate hyperparameter trials are used to select the final settings.
- D Implementation Details: Results for the high-variance sentence-factuality model are averaged over three random-seed runs.
- D Implementation Details: BART summarization experiments specify both fine-tuning hyperparameters and decoding parameters for test-set generation.
- D Implementation Details: Mechanical Turk workers judged three generated summaries per article as Factual or Non-Factual and highlighted erroneous spans.The task presented workers with one source article and three corresponding summaries.