Apeach Eval

apeach-eval

APEACH: Attacking Pejorative Expressions with Analysis on Crowd-Generated Hate Speech Evaluation Datasets — Yang et al. (2022) (arXiv:2202.12459, 2022)

What this evaluates

Evaluates the ability of NLP models to detect hate speech in Korean text. It specifically probes domain-agnostic generalizability and resistance to common inductive biases like text length or topic distribution.

Datasets

• APEACH — total ?; splits: test (-1)

Metrics

• accuracy (primary) — range: [0, 1]
  • Standard binary classification accuracy: the proportion of correctly classified instances (hate speech vs. non-hate speech) out of the total evaluation set.

Input / output format

Input: Korean text sentences, optionally accompanied by a topic category from a predefined list of 10 discrimination/hate speech themes.

Output: Binary label: 'hate speech' (positive) or 'non-hate speech' (negative).

Scoring recipe

def compute_accuracy(predictions, gold_labels):
    correct = sum(1 for p, g in zip(predictions, gold_labels) if p == g)
    return correct / len(gold_labels)

Common pitfalls

• Models may exploit text length bias, as hate speech instances are often longer than neutral ones.
• Topic prompt ordering can cause workers to default to top dropdown options, creating distributional skew.
• Pretraining data overlap can artificially inflate performance on domain-specific hate speech benchmarks.

Evidence (verbatim from paper)

In our study, the hate speech (positive sample) and non-hate speech (negative sample) serve as an element of the balanced dataset for the detection task.

Citation

@misc{yang2022apeach,
  title={APEACH: Attacking Pejorative Expressions with Analysis on Crowd-Generated Hate Speech Evaluation Datasets},
  author={Yang et al. (2022)},
  year={2022},
  note={arXiv:2202.12459}
}

• arXiv: 2202.12459