Evaluating the Factual Consistency of Abstractive Text Summarization

By Wojciech Kryscinski, Bryan McCann, Caiming Xiong, Richard Socher. 2019

1,001 citations recorded by the source corpus.

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.

Subfields: Artificial Intelligence, 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 Natural Language Processing

In Evaluation & Benchmarks

Citations in the corpus

Repairing the Cracked Foundation: A Survey of Obstacles in Evaluation Practices for Generated Text (2022, background): “analyze mispredictions on a set of previously annotated summarization corpora (Kryscinski et al., 2020; Wang et al., 2020; Falke et al., 2019; Maynez et al., 2020).”
OpenAssistant Conversations - Democratizing Large Language Model Alignment (2023, background): “These models are often prone to generating inaccurate information about people, places, or facts, a phenomenon commonly known as ‘hallucinations’ [22, 23].”
Language Generation Models Can Cause Harm: So What Can We Do About It? An Actionable Survey (2022, methodology): “ration process (Dixon et al., 2018; Wiegand et al., 2019; Geva et al., 2019; Sap et al., 2021; Kryscinski et al., 2020) and can have unreliable annota-”
Discriminative Adversarial Search for Abstractive Summarization (2020, methodology): “Kryscinski et al. (2019) used them to detect factual consistency in the context of abstractive summarization; Zellers et al. (2019) applied discriminators to detect fake news, in a news generation scenario, reporting high accuracy (over 90%).”
On Faithfulness and Factuality in Abstractive Summarization (2020, background): “More-over, models are usually agnostic to the noises or artifacts of the training data, such as reference divergence, making them vulnerable to hallucinations (Kryscinski et al., 2019a; Wiseman et al., 2017; Dhingra et al., 2019).”
Survey of Hallucination in Natural Language Generation (2022, background): “[89] propose a weakly supervised model named FactCC for evaluating factual consistency.”
Measuring and Improving Consistency in Pretrained Language Models (2021, background): “component of many NLP system (Du et al., 2019; Asai and Hajishirzi, 2020; Denis and Baldridge, 2009; Kryscinski et al., 2020) and obtaining this skill from a PLM would be extremely beneficial and have the potential to make specialized consistency solutions in downstream tasks redundant.”
Length-Controlled AlpacaEval: A Simple Way to Debias Automatic Evaluators (2024, methodology): “Reference-free evaluation metrics Reference-free evaluation metrics have a long history, including classic methods (Louis & Nenkova, 2013) and more recent neural supervised learning methods (Kryscinski et al., 2020; Sinha et al., 2020; Goyal & Durrett, 2020).”
Asking and Answering Questions to Evaluate the Factual Consistency of Summaries (2020, methodology): “This inadequacy leaves human evaluation as the primary method for evaluating the factual consistencies, which has been noted to be challenging even for humans (Daume III and Marcu, 2005; Kryscinski et al., 2019b), in addition to being slow and costly.”
QuestEval: Summarization Asks for Fact-based Evaluation (2021, methodology): “…while recent state-of-the-art generative models (Lewis et al., 2019; Zhang et al., 2019a) produce ﬂuent summaries, they frequently contain false or unsupported information (Kry´sci´nski et al., 2019), a phenomenon also known as neural hallucination (Rohrbach et al., 2018; Zhao et al., 2020).”
Evaluation of Text Generation: A Survey (2020, methodology): “Table 5.1 shows a sample summarization model output, in which the claims made are not consistent with the source document (Kryściński et al., 2019).”
Get Your Vitamin C! Robust Fact Verification with Contrastive Evidence (2021, background): “Generating texts that match given facts is a known challenge (Fan et al., 2020; Kryscinski et al., 2020; Lewis et al., 2020b; Parikh et al., 2020; Shah et al., 2020; Tian et al., 2020) as language models tend to degenerate and hallucinate (Holtzman et al., 2020; Schuster et al., 2020; Zhou et al.,…”
Q^{2}: Evaluating Factual Consistency in Knowledge-Grounded Dialogues via Question Generation and Question Answering (2021, background): “, 2018) and in evaluating the factual consistency of generated summaries (Goodrich et al., 2019; Kryściński et al., 2019; Xu et al., 2020).”
Understanding Factuality in Abstractive Summarization with FRANK: A Benchmark for Factuality Metrics (2021, methodology): “Entailment Classification Goodrich et al. (2019); Kryscinski et al. (2020); Maynez et al. (2020); Goyal and Durrett (2020) model factuality as entailment classification breaking down the summary into smaller units, such as sentences, which are verified against the original article.”
Improving Zero and Few-Shot Abstractive Summarization with Intermediate Fine-tuning and Data Augmentation (2020, background): “size of the desired summaries, and the extent to which the summary is abstractive, all of that can be treated as prior knowledge if the task is to be well-defined (Kryscinski et al., 2020).”
BARTScore: Evaluating Generated Text as Text Generation (2021, methodology): “We break down each test set into four buckets based on the reference length, which are [15, 25), [25, 35), [35, 45), [45, 54] and compute the Kendall’s Tau average correlation of different metrics across all language pairs within each bucket.”
Factual Error Correction for Abstractive Summarization Models (2020, background): “, 2018), and evaluation models for factual consistency in abstractive summarization (Goodrich et al., 2019; Falke et al., 2019; Kryściński et al., 2019; Wang et al., 2020).”
GO FIGURE: A Meta Evaluation of Factuality in Summarization (2020, background): “Especially in document summarization tasks, models that abstract away salient aspects, have been shown to generate text ∗ Work done while ﬁrst author was interning at MSR. with up to 30% factual inconsistencies (Kryscinski et al., 2019; Falke et al., 2019a; Zhu et al., 2020).”