PaperlayerView exact revision on arXiv

Source-linked AI summary

Spurious Correlations in Reference-Free Evaluation of Text Generation

Esin Durmus, Faisal Ladhak, Tatsunori Hashimoto

arXiv:2204.09890v1cs.CL

TL;DR

Reference-free evaluation could reduce reliance on costly human judgments, but strong correlations may mask dependence on spurious features. The paper tests these metrics across generation tasks and adversarially trains a faithfulness metric; existing metrics struggle on abstractive, faithful systems, while the proposed metric improves that ranking and performs effectively overall.

  • Problem

    High correlations with human judgments do not establish robustness, since reference-free metrics for summarization and dialogue may track word overlap, length, and perplexity.

    Source: §1 · p. 1

  • Method

    The paper compares learned faithfulness metrics with word-overlap measures and adversarially trains a metric to make its predictions less reliant on density.

    Source: §3.1 · p. 3, §5 · p. 7

  • Results

    Existing learned metrics do not outperform density overall and perform worse within abstractive-faithful systems, while the adversarial metric improves that ranking and reaches 85.27% overall pairwise accuracy.

    Source: §4.2 · p. 6, §5 · p. 8

  • Takeaways & Limitations

    Adversarially reducing reliance on a spurious correlate can improve ranking of abstractive, faithful systems while preserving effective overall pairwise ranking.

    Source: §1 · p. 1, §5 · p. 8

Abstract

from arXiv · show

Model-based, reference-free evaluation metrics have been proposed as a fast and cost-effective approach to evaluate Natural Language Generation (NLG) systems. Despite promising recent results, we find evidence that reference-free evaluation metrics of summarization and dialog generation may be relying on spurious correlations with measures such as word overlap, perplexity, and length. We further observe that for text summarization, these metrics have high error rates when ranking current state-of-the-art abstractive summarization systems. We demonstrate that these errors can be mitigated by explicitly designing evaluation metrics to avoid spurious features in reference-free evaluation.

1 Introduction

Reference-free metrics promise faster, less costly NLG evaluation, but high human-score correlations may conceal reliance on spurious features and weak system selection. The paper examines these risks in summarization and dialogue and tests adversarial training as a mitigation.

1 Introduction

  • Reference-free metrics have drawn interest as a potentially viable alternative to costly human evaluation because many report high correlations with human judgments.The cited work describes these metrics as a way to assess generated outputs without human references.Source: §1 · p. 1
  • Across summarization and dialogue, simple correlates such as word overlap, length, and perplexity can achieve human-judgment correlations similar to learned metrics.Learned metrics correlate more strongly with these spurious measures than human scores do, suggesting possible reliance on them.Source: §1 · p. 1
  • For summarization faithfulness, word overlap matches learned metrics in system-level ranking, while learned metrics perform significantly worse among abstractive, faithful state-of-the-art systems.The paper attributes this weakness to similarly low word overlap across those systems.Source: §1 · p. 1
  • Adversarial training yields effective overall pairwise ranking with lower word-overlap correlation and improves ranking between abstractive, faithful systems.These systems are a failure mode for existing reference-free faithfulness metrics.Source: §1 · p. 1

2 Reference-free Evaluation of Text Generation

This section defines reference-free metrics and two ways to evaluate them: example-level correlations and system-level rankings. It also explains why metric performance can depend on the evaluation examples and systems, motivating a distinction between genuine and spurious correlates.

2 Reference-free Evaluation of Text Generation

  • A reference-free metric F(x, y) scores output y given input x, aiming to score desirable output attributes; prior work evaluates it through correlation with human judgments H(x, y).Example-level evaluation computes Pearson or Spearman correlation on sampled evaluation data, while system-level evaluation measures agreement with human rankings across model pairs.Source: §2 · p. 2
  • Example-level human correlations measure Pearson or Spearman agreement between metric scores and human judgments over sampled evaluation examples.This is one procedure used by existing reference-free evaluations.Source: §2 · p. 2
  • System-level evaluation measures whether a metric identifies the better model among a set of models, using agreement with human judgments over model pairs.The passage defines system-level ranking as accuracy in matching human evaluation across all distinct model pairs.Source: §2 · p. 2
  • Metric evaluations may depend strongly on the sampled examples and model set; a grammar checker could score well if dialogue data contains only polished responses or ungrammatical nonsense.The example illustrates how an evaluation distribution can make a simple signal appear effective despite the challenge of predicting response quality.Source: §2 · p. 2
  • A spurious correlate is an attribute associated with human judgments on the evaluation distribution but not on a carefully constructed test distribution.A metric is spuriously correlated when it loses correlation with human judgments on the test distribution while remaining correlated with the attribute there.Source: §2 · p. 2

3 Example-level Analysis of Learned Evaluation Metrics

The section compares reference-free evaluation metrics for summarization and dialog generation with simple spurious correlates. Across both tasks, the correlates can match or outperform learned metrics, and learned metrics often track the correlates more closely than human judgments do.

3 Example-level Analysis of Learned Evaluation Metrics

  • The analysis compares learned metrics with word overlap for summarization and with perplexity, length, and their combination for dialog quality.Coverage is the percentage of summary words also present in the article; density is the average length of summary segments extracted from it. Perplexity and length do not account for dialog context.Source: §3 · p. 2, §3.1 · p. 3, §3.2 · p. 4

3.1 Faithfulness Evaluation in Text Summarization

  • Density correlates with human faithfulness judgments similarly to DAE and significantly better than FactCC, while both learned metrics correlate more with density than with human scores.The analysis uses human annotations for 16 summarization models on CNN/DM.Source: Figure 1 · p. 3, §3.1 · p. 3, Table 1 · p. 3

3.2 Learned Metrics for Dialog Generation

  • On TopicalChat, PPL+Len significantly outperforms learned metrics; on PersonaChat and DailyDialog, it performs similarly to the best learned metric.Across dialog datasets, reference-free metrics correlate more with spurious measures than human scores do; USL-H is more consistent than other learned metrics and significantly exceeds MAUDE and DialogRPT on PersonaChat and TopicalChat.Source: Figure 2 · p. 4, Table 2 · p. 4, §3.2 · p. 5
  • For groundedness, coverage and density correlate with human scores significantly more than MNLI+Adv and USL-H, while the learned metrics track spurious correlates more than human scores.Groundedness is evaluated using entailment annotations for responses conditioned on external evidence.Source: Figure 3 · p. 5, §3.2 · p. 5, Table 3 · p. 5
  • The authors recommend comparing reference-free metrics with potential spurious correlates and developing a standard set of such correlates.The recommendation follows their observation that metrics can perform similarly to simple correlates on datasets beyond their original evaluations.Source: §3.2 · p. 5

4 Learned Metrics in System-level Evaluation

System-level rankings of summarization faithfulness can look strong overall while failing within the abstractive-faithful (AF) group. Density matches the best learned metric overall, but adversarial training performs better than existing learned metrics within AF.

4.1 Pairwise Ranking of Systems

  • System-level rankings show the same concern as example-level correlations: density and coverage appear as spurious correlates.This motivates finer-grained analysis of rankings among abstractive, faithful systems.Source: §4.1 · pp. 5–6
  • The AF group comprises systems with faithfulness above 4.5 out of 5 and density below 30.Figure 4 contrasts rankings across all systems with rankings within this group.Source: Figure 4 · p. 6

4.2 Results

  • Density achieves the highest overall accuracy, while learned metrics perform relatively poorly within AF, where systems are close in density.Table 4 reports significantly lower ranking accuracy for all metrics within AF than across all system pairs; density matches DAE in both settings.Source: §4.2 · p. 6, Table 4 · p. 6
  • Adversarial training performs significantly better than previously proposed metrics for pairwise ranking within the abstractive-faithful group.Table 5 compares DAE and the adversarial metric across all pairs and within AF.Source: Table 5 · p. 7

5 Adversarial Model

The paper trains a faithfulness metric to make its representations less predictive of summary density. The resulting metric has lower density correlation than FactCC and DAE and performs better within AF systems while reaching 85.27% overall pairwise accuracy.

Adversarial approach

  • The approach tests whether learning representations less reliant on spurious correlates can improve system-level metric performance.The preceding system-level analysis found learned faithfulness metrics unable to outperform density.Source: §5 · p. 7
  • The model uses FactCC’s synthetic data and is essentially FactCC-Electra with an added adversarial density-prediction head.Electra was chosen to match DAE’s architecture; FactCC-Electra serves as an ablation.Source: §5 · p. 7
  • A density-prediction adversary uses reversed gradients to make the encoder representation less predictive of summary density while predicting faithfulness.The density predictor operates on the model’s internal representation; λ is increased gradually from 0 to 1.Source: §5 · p. 7

Results

  • Density correlation is 27.71% for the adversarial metric, versus 59.10% for FactCC and 76.37% for DAE.The FactCC-Electra model performs worse than original FactCC, consistent with Goyal and Durrett (2021).Source: §5 · pp. 7–8
  • The adversarial metric reaches 85.27% accuracy in overall system pairwise ranking and significantly outperforms existing learned metrics within AF.The AF comparison is reported in Table 5.Source: §5 · p. 8

6 Related Work

Prior evaluation research largely examined reference-based metrics, while newer work assessed reference-free metrics across datasets and robustness dimensions. This paper distinguishes itself by studying spurious correlations in summarization and dialog-generation metrics.

Related work

7 Conclusion

The paper concludes that overall human-score correlation is an incomplete test of reference-free metrics: simple spurious correlates can match learned metrics, which struggle to rank abstractive, faithful systems. Its adversarial faithfulness metric is more robust across distribution shifts and ranks those systems better.

7 Conclusion

  • Learned metrics fail to rank abstractive, relatively faithful summarization systems well despite relatively high overall system-level ranking performance.These systems represent the current state-of-the-art regime discussed in the conclusion.Source: §7 · p. 8
  • A faithfulness metric trained not to rely on spurious overlap is more robust across distribution shifts and better ranks abstractive, faithful systems.Source: §7 · p. 8
  • The paper recommends evaluating metrics across different system distributions, using simple spurious correlates to flag possible overestimates, and collecting large-scale human evaluations across varied systems.The conclusion points to Fabbri et al. (2020) as an example of such a dataset.Source: §7 · pp. 8–9

A Text Summarization Models

Section 4 uses the models listed in Table 6.

A Text Summarization Models

Related papers

Loading 2204.09890v1…