Understanding Factuality in Abstractive Summarization with FRANK: A Benchmark for Factuality Metrics

By Artidoro Pagnoni, Vidhisha Balachandran, Yulia Tsvetkov. 2021

396 citations recorded by the source corpus.

Modern summarization models generate highly fluent but often factually unreliable outputs. This motivated a surge of metrics attempting to measure the factuality of automatically generated summaries. Due to the lack of common benchmarks, these metrics cannot be compared. Moreover, all these methods treat factuality as a binary concept and fail to provide deeper insights on the kinds of inconsistencies made by different systems. To address these limitations, we devise a typology of factual errors and use it to collect human annotations of generated summaries from state-of-the-art summarization systems for the CNN/DM and XSum datasets. Through these annotations we identify the proportion of different categories of factual errors and benchmark factuality metrics, showing their correlation with human judgement as well as their specific strengths and weaknesses.

Subfields: Artificial Intelligence, Molecular Biology, Natural Language Processing, Evaluation & Benchmarks

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Molecular Biology

In Natural Language Processing

In Evaluation & Benchmarks

Citations in the corpus

Language Generation Models Can Cause Harm: So What Can We Do About It? An Actionable Survey (2022, background): “Being trained on synthetic data, such detectors typically do not generalize and have low human judgment correlations (Pagnoni et al., 2021).”
Survey of Hallucination in Natural Language Generation (2022, background): “Tolerance towards hallucinations is very low/strict in both summarization [119] and data-to-text tasks [120, 164, 166] because it is essential to provide faithful generation.”
Generating Training Data with Language Models: Towards Zero-Shot Language Understanding (2022, background): “While PLMs have demonstrated remarkable text generation and understanding capability, they can come with potential risks or harms [2, 3, 5] such as generating misinformation [46] or amplifying harmful biases [49].”
Teaching Small Language Models to Reason (2022, background): “The concerns here are that both the teacher and student model may potentially generate non-factual (Ji et al., 2022; Pagnoni et al., 2021; Kreps et al., 2022) or offensive output (Gehman et al., 2020).”
Controlled Text Generation as Continuous Optimization with Multiple Constraints (2021, background): “However, once such models are trained, they are prone to degeneration [66] and biased, non-factual outputs [16, 45] as it is difficult to control the characteristics or attributes of the generated text without architectural modifications [26, 28, 35] and fine-tuning the models on attribute-specific ”
Hallucinated but Factual! Inspecting the Factuality of Hallucinations in Abstractive Summarization (2021, methodology): “Follow-380 ing Pagnoni et al. (2021), we also report the per-381 centage µ of annotators that agree with the majority382 class.”
A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions (2023, neutral): “…such as abstractive summarization (Kryscinski et al., 2020; Wang et al., 2020a; Maynez et al., 2020; Fabbri et al., 2021; Goyal and Durrett, 2021; Pagnoni et al., 2021; Tang et al., 2022), data-to-text(Tian et al., 2019; Parikh et al., 2020), and machine translation (Zhou et al., 2021).”
TRUE: Re-evaluating Factual Consistency Evaluation (2022, background): “Other works on meta-evaluation of factual consistency across datasets include GO-FIGURE (Gabriel et al., 2021) FRANK (Pagnoni et al., 2021) and SummaC (La-ban et al., 2021), however they all focus solely on summarization.”
FactGraph: Evaluating Factuality in Summarization with Semantic Graph Representations (2022, positive): “In this experiment, F ACT CC+ and F ACT G RAPH are trained with the F ACT C OLLECT data without the Pagnoni et al. (2021)’s subset, which is used as dev and test sets, according to its split.”
Investigating Crowdsourcing Protocols for Evaluating the Factual Consistency of Summaries (2021, background): “However, current metrics still do not correlate highly with human judgments on factual consistency (Koto et al., 2020; Pagnoni et al., 2021).”
SQuALITY: Building a Long-Document Summarization Dataset the Hard Way (2022, background): “This mismatch may contribute to the poor correlation of ROUGE with human judgments of quality for these datasets (Pang et al., 2021a; Pagnoni et al., 2021; Scialom et al., 2021, i.a.).”
FActScore: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation (2023, background): “Stable LM mation that are a mixture of true or false, 2 making a binary judgment inadequate (Pagnoni et al., 2021).”
Don’t Say What You Don’t Know: Improving the Consistency of Abstractive Summarization by Constraining Beam Search (2022, result): “5), a result in line with a very recent evaluation of factuality measures (Pagnoni et al., 2021).”
Underreporting of errors in NLG output, and what to do about it (2021, background): “Previously, Popović (2020) asked error annotators to identify issues with comprehensibility and adequacy in machine-translated text.”
News Summarization and Evaluation in the Era of GPT-3 (2022, background): “Annotator rationales often mentioned misleading or incorrect information as the primarily reason for selecting BRIO as the worst summary, confirming the issues that have been observed with XSum -trained models (Maynez et al., 2020; Pagnoni et al., 2021; Goyal and Durrett, 2021).”
Revisiting the Gold Standard: Grounding Summarization Evaluation with Robust Human Evaluation (2022, methodology): “For example, decomposing referencefree evaluation according to the presence of errors or characteristics from a pre-defined taxonomy greatly improves agreement in the FRANK (Pagnoni et al., 2021) factual consistency proto-”
Spurious Correlations in Reference-Free Evaluation of Text Generation (2022, methodology): “In particular we evaluate FactCC (Kryscinski et al., 2020) and DAE (Goyal and Durrett, 2021), which have been shown to achieve higher example-level correlations with human judgements than existing faith-fulness evaluation metrics (Pagnoni et al., 2021).”