Human Feedback Summarization Eval

human-feedback-summarization-eval

Learning to summarize from human feedback — Stiennon et al. (2020) (arXiv:2009.01325, 2020)

What this evaluates

Evaluates abstractive summarization quality by measuring human preference over reference summaries and rating outputs across coverage, accuracy, coherence, and overall quality. It also benchmarks how well learned reward models and automatic metrics correlate with human judgments.

Datasets

• Reddit TL;DR — total ?; splits: test (-1)
• CNN/DailyMail — total ?; splits: test (-1)

Metrics

• preference score (primary) — range: percent
  • Percentage of pairwise comparisons where the model-generated summary is preferred over the reference summary by human annotators.
• Likert scale quality rating — range: [1, 7]
  • 7-point Likert scale rating across four dimensions: coverage, accuracy, coherence, and overall quality.
• reward model validation accuracy — range: percent
  • Percentage of times the reward model's preference prediction agrees with human labeler preferences on a held-out validation set.

Input / output format

Input: Source text (Reddit post or CNN/DM news article)

Output: Generated summary text

Scoring recipe

def compute_preference_score(model_summaries, ref_summaries, human_prefs):
    # human_prefs: list of 1 if model preferred, 0 if ref preferred
    return sum(human_prefs) / len(human_prefs) * 100

def compute_likert_scores(model_summaries, human_ratings):
    # human_ratings: dict of summary_id -> {coverage, accuracy, coherence, overall} in [1,7]
    dims = ['coverage', 'accuracy', 'coherence', 'overall']
    return {d: np.mean([r[d] for r in human_ratings.values()]) for d in dims}

Common pitfalls

• Summary length acts as a confounding factor; longer summaries are often preferred regardless of quality, requiring length-controlled comparisons.
• Automatic metrics like ROUGE and log-probability can fail to track quality improvements and may even show negative correlation with human preference after RL fine-tuning.
• Over-optimizing against the reward model can cause it to become anti-correlated with true human preferences.

Evidence (verbatim from paper)

We measure policy quality as the percentage of summaries generated by that policy that humans prefer over the reference summaries in the dataset. Our policies trained with human feedback significantly outperform our supervised baselines on this metric, with our 1.3B human feedback model significantly outperforming a supervised model 10x its size (61% versus 43% raw preference score against reference summaries).

Citation

@misc{stiennon2020learning,
  title={Learning to summarize from human feedback},
  author={Stiennon et al. (2020)},
  year={2020},
  note={arXiv:2009.01325}
}

• arXiv: 2009.01325