PaperlayerView exact revision on arXiv

Source-linked AI summary

TRUE: Re-evaluating Factual Consistency Evaluation

Or Honovich, Roee Aharoni, Jonathan Herzig, Hagai Taitelbaum, Doron Kukliansy, Vered Cohen, Thomas Scialom, Idan Szpektor, Avinatan Hassidim, Yossi Matias

arXiv:2204.04991v3cs.CL

TL;DR

Factual-consistency metrics lack unified cross-task evaluation, while generated text can contain inconsistencies that limit practical use. TRUE standardizes 11 datasets and evaluates diverse metrics with an example-level protocol. Large-scale NLI and QG-QA methods perform strongly and complementarily, though subtle inconsistencies and difficult inputs remain challenging.

  • Problem

    Factual-consistency metrics are developed in isolation, and system-level correlations leave their example-level accuracy unclear.

    Source: §1 · p. 1, §2.3 · p. 5

  • Method

    TRUE standardizes 11 datasets across diverse tasks and evaluates 12 metrics using an example-level factual-consistency protocol.

    Source: §1 · p. 2, §2 · p. 2

  • Results

    Large-scale NLI and QG-QA approaches achieve strong, complementary results across multiple tasks and datasets.

    Source: §1 · p. 2, §8 · p. 9

  • Takeaways & Limitations

    The authors recommend NLI and QG-QA methods for factual-consistency evaluation and encourage unified benchmarking with TRUE.

    Source: §7 · p. 9, §8 · p. 9

    The evaluated metrics struggle with subtle inconsistencies, long inputs, labeling issues, and personal statements, while TRUE focuses on English text-to-text tasks.

    Source: §5 · p. 8, §7 · p. 9

Abstract

from arXiv · show

Grounded text generation systems often generate text that contains factual inconsistencies, hindering their real-world applicability. Automatic factual consistency evaluation may help alleviate this limitation by accelerating evaluation cycles, filtering inconsistent outputs and augmenting training data. While attracting increasing attention, such evaluation metrics are usually developed and evaluated in silo for a single task or dataset, slowing their adoption. Moreover, previous meta-evaluation protocols focused on system-level correlations with human annotations, which leave the example-level accuracy of such metrics unclear. In this work, we introduce TRUE: a comprehensive survey and assessment of factual consistency metrics on a standardized collection of existing texts from diverse tasks, manually annotated for factual consistency. Our standardization enables an example-level meta-evaluation protocol that is more actionable and interpretable than previously reported correlations, yielding clearer quality measures. Across diverse state-of-the-art metrics and 11 datasets we find that large-scale NLI and question generation-and-answering-based approaches achieve strong and complementary results. We recommend those methods as a starting point for model and metric developers, and hope TRUE will foster progress towards even better evaluation methods.

1 Introduction

TRUE addresses fragmented factual-consistency evaluation by standardizing diverse datasets and introducing an example-level comparison protocol. Its assessment finds that large-scale NLI and QG-QA approaches perform strongly and complementarily across tasks, while exposing limitations such as long inputs and personal statements.

Motivation

  • Automatic factual-consistency evaluation can support better evaluation and generation by filtering training data or augmenting it for controlled generation.The motivation is detecting whether generated text is consistent with its grounding text.Source: §1 · p. 1

Benchmark and evaluation framework

  • TRUE unifies factual-consistency evaluation by standardizing 11 published datasets and comparing methods across tasks and domains.The collection covers summarization, knowledge-grounded dialogue, paraphrasing, and fact verification.Source: §1 · p. 2
  • The study replaces less actionable system-level correlations with ROC AUC for detecting inconsistent examples under binary labels.This protocol measures recall and precision-relevant behavior for individual examples rather than only correlations with human judgments.Source: §1 · pp. 1–2

Results

  • Large-scale NLI and QG-QA metrics achieve significantly better results across a wide variety of tasks and datasets.The assessment covers 12 metrics and identifies these approaches as strong performers.Source: §1 · p. 2
  • Combining NLI and QG-QA methods yields even better results, indicating complementary strengths and room for further improvement.The paper recommends these approaches as a starting point for model and metric developers.Source: §1 · p. 2

Limitations and future work

  • All evaluated approaches struggle with long inputs, labeling issues, and personal statements.Source: §1 · p. 2

2 Standardizing Factual Consistency

TRUE standardizes factual consistency across 11 datasets and diverse tasks using binary, grounding-based labels, then evaluates metrics with example-level ROC AUC. This framework addresses inconsistent labeling and the limited usefulness of system-level correlations for binary inconsistency detection.

2.1 Definitions and Terminology

  • TRUE defines factual consistency as all factual information in a target text being supported by its grounding text, excluding external-world truth and personal or social statements.Source: §2.1 · pp. 2–3

2.2 Standardization Process

  • The benchmark standardizes 11 datasets from summarization, dialogue, fact verification, and paraphrase detection into binary labels for whole-text consistency.Binary labeling supports cross-task standardization and practical filtering of unfaithful predictions.Source: §2.2 · p. 3

2.2.1 Abstractive Summarization

  • Summarization datasets convert sentence- or score-level annotations into consistent labels only when every summary sentence satisfies the relevant consistency criterion.This includes majority-vote sentence labels in FRANK and QAGS, no-hallucination judgments in MNBM, and unanimous expert scores of 5 in SummEval.Source: §2.2.1 · pp. 3–4

2.2.2 Dialogue Generation

  • Dialogue datasets map groundedness annotations to binary consistency, treating entailed or supported responses as consistent and hallucinated, unsupported, or otherwise non-entailed responses as inconsistent.The standardized set includes BEGIN, Q2, and DialFact, with evidence concatenated when multiple pieces are required.Source: §2.2.2 · p. 4

2.2.3 Fact Verification

  • Fact-verification datasets treat supported claims as consistent and refuted, neutral, or otherwise unsupported claims as inconsistent against their evidence.TRUE applies this mapping to FEVER and VitaminC.Source: §2.2.3 · p. 4

2.2.4 Paraphrase Detection

  • PAWS supplies paraphrase labels as consistency labels for Wikipedia-source examples, although paraphrase is not equivalent to factual consistency.The dataset remains relevant because its non-paraphrases usually have contradicting meanings.Source: §2.2.4 · p. 4

2.3 Meta-Evaluation

  • TRUE evaluates example-level inconsistency detection with ROC AUC, plotting recall against fallout across thresholds without requiring a fixed decision threshold.For datasets with splits, it also tunes a development-set threshold and reports test accuracy.Source: §2.3 · p. 5

3 Evaluation Metrics

TRUE compares standard and state-of-the-art factual consistency metrics across tasks and datasets, normalizing scores where needed and using grounding texts as references. The evaluated families include n-gram, model-based, NLI, and QG-QA approaches.

3 Evaluation Metrics

  • The evaluation compares standard and state-of-the-art metrics across varied tasks and datasets, using grounding texts as references for reference-based methods and normalizing scores to [0,1] when necessary.Source: §3 · p. 5

3.1 N-Gram Based Metrics

  • N-gram baselines including BLEU, ROUGE, and token-level F1 are included because prior work found weak correlation with factual consistency.Source: §3.1 · p. 5

3.2 Model-Based Metrics

  • Model-based metrics cover embedding similarity, learned human-judgment scoring, synthetic-data factual verification, force-decoding probabilities, and token alignment for hallucination detection.The evaluated examples are BERTScore, BLEURT, FactCC, BARTScore, and CTC.Source: §3.2 · p. 5

3.3 Natural Language Inference Metrics

  • ANLI scores entailment probability from the grounding text to the generated text, while SUMMAC aggregates sentence-pair NLI scores with either maximum-per-summary-sentence averaging or a convolutional network.Source: §3.3 · p. 6

3.4 QG-QA Based Metrics

  • QG-QA methods generate questions from spans in generated text, answer them using the grounding text, compare answer spans, and aggregate the resulting scores.Q2 uses NLI to compare answers, while QuestEval also measures relevance by reversing generated and grounding text roles.Source: §3.4 · p. 6

4 Results

Across 11 datasets, NLI-based and QG-QA metrics were the strongest approaches, with complementary errors that made their combination more effective.

Evaluation setup

  • The evaluation used zero-shot metrics on development sets, excluding VitaminC and FEVER from averages because SCZS was trained on VitaminC data containing FEVER examples.Source: §4 · p. 6

Overall results

  • ANLI and SCZS outperformed other approaches on 6 datasets, while Q2 led on 4, with average AUCs of 81.5, 81.4, and 80.7, respectively.The next-best method, BARTScore, had a lower average AUC of 72.2; all other approaches scored 72 or lower on average.Source: §4 · pp. 6–7, Table 3 · p. 7
  • Significance testing found the best method significantly outperformed the second-best on 6 datasets: Q2 led on 3 and the ANLI-based model on 3.Source: §4 · p. 7

Ensemble analysis

  • Averaging Q2, ANLI, and SCZS improved results on most datasets and raised average ROC AUC by 4.5 over the best single metric.All three best metrics scored above 80 on average across the 11 datasets.Source: §4 · p. 7

5 Analysis

Analysis shows that metric performance declines on long groundings, improves with larger models, and remains limited by annotation errors and subtle or domain-specific inconsistencies.

Input Length

  • All metrics degraded consistently beyond 200-token groundings, although ANLI and Q2 retained AUC above 0.825 on the longest bin.The longest bin contained texts exceeding 500 tokens, despite Q2 and ANLI performing end-to-end QA and NLI.Source: §5 · p. 7, Figure 1 · p. 8

Model Size

  • Larger ANLI, BLEURT, and BERTScore variants improved average ROC AUC by 4.7, 3.7, and 1.3, respectively.The ablation therefore supports a performance advantage for larger model variants.Source: §5 · p. 8, Table 4 · p. 8

Qualitative Analysis

  • Hard examples exposed substantial annotation noise: 35 of 80 cases where all best metrics failed and 27 of 100 cases where one or two failed contained annotation errors, versus 10 of 100 uniformly sampled cases.The high error rate characterized difficult examples rather than the datasets generally.Source: §5 · p. 8
  • Best-performing metrics missed subtle inconsistencies in 21 of 180 examples, partly because score aggregation can hide small contradictory segments.Using a minimum rather than an average could reduce this issue but would increase false negatives.Source: §5 · p. 8
  • Personal statements in dialogue caused false ungrounded classifications in 10 of 62 analyzed responses, motivating exclusion of non-factual content from evaluation.Source: §5 · p. 8

Ensemble Analysis

  • The ensemble succeeded in 25,761 cases where at least one best metric failed, usually because two metrics succeeded while one failed.Among sampled cases, 47% involved a borderline error from one metric and 36% involved a non-borderline error.Source: §5 · p. 8

6 Related Work

TRUE extends factual-consistency meta-evaluation across tasks and datasets, addressing limitations in prior task-specific or correlation-based approaches.

Scope of prior work

  • Prior unified evaluation efforts addressed system outputs or general text-generation metrics, whereas TRUE evaluates factual-consistency methods themselves across tasks.Source: §6 · p. 9
  • Earlier factual-consistency meta-evaluations largely focused on summarization, dialogue, or correlations, while TRUE generalizes evaluation across tasks and datasets.Source: §6 · p. 9

TRUE’s contribution

  • TRUE claims the first cross-task demonstration that large-scale QG-QA and NLI methods are strong and highly complementary.The paper presents this as a basis for stronger baselines and meta-evaluation methodology.Source: §6 · p. 9

7 Discussion and Future Work

TRUE recommends NLI and QG-QA methods for factual consistency evaluation and identifies complementary strengths, while highlighting unresolved challenges for future work.

7 Discussion and Future Work

  • QG-QA and NLI methods perform better than other approaches, especially when combined, motivating their use for factual consistency evaluation.The authors recommend these methods for model developers and for evaluating new metrics.Source: §7 · p. 9
  • ROC AUC is recommended over correlations because it provides more interpretable and actionable evaluation measures.Source: §7 · p. 9
  • The binary annotation scheme supports testing new metrics across tasks and datasets and can unify more detailed labels through conversion protocols.Source: §7 · p. 9
  • Future work should address challenges involving long inputs and personal statements in dialogue.Source: §7 · p. 9

8 Conclusions

TRUE standardizes factual consistency datasets from diverse tasks to enable unified comparison of automatic evaluation methods. Large-scale NLI and QG-QA approaches perform well across tasks and datasets and are complementary, leaving room for improvement.

8 Conclusions

  • TRUE standardizes datasets from diverse tasks into a unified labeling scheme for thorough comparison of automatic evaluation methods.Source: §8 · p. 9
  • Large-scale NLI and QG-QA approaches perform well across multiple tasks and datasets and show complementary strengths.Their complementarity suggests additional headroom for improving factual consistency evaluation.Source: §8 · p. 9

A Additional Data Statistics

The appendix reports grounding-text and generated-text length statistics for TRUE’s datasets.

A Additional Data Statistics

B Implementation Details

The implementation details describe training infrastructure and configurations for the QG-QA and NLI models.

B Implementation Details

  • All models are trained using the t5x library.Source: B Implementation Details · p. 14
  • The QG-QA reimplementation uses T5-11B, 2048-token inputs, beam search, and a tuned F1 overlap threshold.It generates multiple questions and selects the first passing the validation threshold.Source: B Implementation Details · p. 14
  • The NLI model fine-tunes T5-11B on ANLI for 25K steps with learning rate 10^-4 and batch size 32.Inference uses a maximum input length of 2048 tokens.Source: B Implementation Details · p. 14

C ROC Curves

TRUE reports ROC curves and ROC AUC results across datasets and factual-consistency metrics, alongside accuracy and overlap analyses. The evaluation distinguishes grounded from ungrounded text and documents dataset exclusions and threshold-tuning procedures.

C ROC Curves

  • Figure 2 compares ROC curves across the TRUE datasets using the best-performing metrics.The supplied passages do not state the numerical curve values or identify a single overall winner.Source: C ROC Curves · p. 14, Figure 2 · p. 15
  • Table 8 compares ROUGE-L and simple F1 token overlap between generated text and grounding for grounded and ungrounded examples.The Pos columns represent grounded text, while the Neg columns represent ungrounded text.Source: Table 8 · p. 15
  • Table 9 provides ROC AUC results for metrics not reported in Table 3.Source: Table 9 · p. 16
  • Table 10 reports ROC AUC results for different metrics on the TRUE test set, excluding VitaminC from averages because SCZS was trained on it.Highest non-Ensemble scores are highlighted, while the specified SC results are struck through.Source: Table 10 · p. 16
  • Tables 11 and 12 report development- and test-set accuracy, with thresholds tuned on the development or corresponding development sets.VitaminC and, for development averages, FEVER are excluded when training overlap could affect evaluation.Source: Table 11 · p. 16, Table 12 · p. 16

Related papers

Loading 2204.04991v3…