Euphemistic Abuse - A New Dataset and Classification Experiments for Implicitly Abusive Language

Michael Wiegand, Jana Kampfmeier, Elisabeth Eder, Josef RuppenhoferEMNLP 2023 · DOI

Abstract

While NLP models have become adept at catching explicit abuse involving slurs or profanity, they frequently fail to detect "euphemistic abuse"—insults masked as polite or neutral paraphrases. This paper introduces a novel dataset and benchmark for this phenomenon. The study reveals that current state-of-the-art toxicity classifiers are largely blind to these barbs, but performance can be significantly improved through targeted GPT-3 data augmentation and manual linguistic feature engineering, with the latter achieving an F1 score of 75.6, nearing the human baseline of 78.3.

The Challenge of Euphemistic Abuse

Most existing research on implicit toxicity focuses on identity-based hate speech (e.g., antisemitism or misogyny). This paper tackles a "long-tail" phenomenon called euphemistic abuse: utterances that are perceived as abusive even without context, yet contain no "bad" words. Examples include "You inspire me to fall asleep" instead of "You are boring." Because these sentences often use sophisticated vocabulary or creative imagery, they evade standard safety filters. To truly secure online spaces, models must understand the underlying "sting" of a paraphrase, not just a list of prohibited keywords.

Methodology

The authors approached the problem through three main avenues: dataset creation, GPT-3 augmentation, and feature-based classification.

Dataset Creation

The researchers used a multi-stage crowdsourcing pipeline to generate and validate euphemistic insults based on 97 "cue phrases" (explicit insults). A critical component of this process was the generation of contrast sets—non-abusive sentences that are syntactically and semantically similar to the abusive ones—to ensure models could not rely on simple surface patterns.

Figure 1: The multi-stage crowdsourcing pipeline used to generate and validate the euphemistic abuse dataset and its contrast sets.

Classification Strategies

  1. GPT-3 Augmentation: The authors explored whether LLMs could "unmask" euphemisms. They tested an Inclusive variant (training on original sentences concatenated with GPT-3 completions) and an Exclusive variant (using a classifier to label GPT-3 completions and aggregating those labels).
  2. Feature-Based Classification: A logistic regression classifier was built using six manually extracted linguistic features: negated antonyms, opposing sentiment, taboo topics, extremes (hyperbole), lexicalization (idioms), and unusual properties.

Results and Analysis

The evaluation compared commercial tools, research transformers, and the proposed feature-based approach across both within-dataset and cross-dataset scenarios.

Figure 4: Performance comparison of various classifiers in cross-dataset and within-dataset evaluations.

Key Insights

  • The "Blind Spot" of SOTA: Commercial tools like PerspectiveAPI and research models like ToxiGen perform poorly on euphemistic abuse (F1 scores in the low 60s). They are tuned for broader toxicity and miss the nuance of a polite-sounding insult.
  • Linguistics Still Matters: Surprisingly, the manually engineered feature-based classifier (75.6 F1) outperformed the standard RoBERTa transformer (64.5 F1). This suggests that euphemistic abuse relies on specific rhetorical devices that current transformers do not weigh heavily enough by default.
  • LLMs as "De-maskers": GPT-3 is remarkably good at unmasking euphemisms. When asked to complete a euphemistic sentence, it often generates explicit insults, making the underlying toxicity much easier for a secondary classifier to detect.

Conclusion

Euphemistic abuse represents a distinct linguistic category that current toxicity filters largely ignore. The study highlights that the quality of negative data (contrast sets) is vital for robust evaluation; without them, models appear to solve the problem while only learning surface patterns. Future work will focus on automating the detection of high-level linguistic features and incorporating conversational context to better identify these subtle barbs in real-world settings.

Summary updated

BibTeX
@inproceedings{wiegand2023euphemistic,
  author = {Michael Wiegand and Jana Kampfmeier and Elisabeth Eder and Josef Ruppenhofer},
  title = {{Euphemistic Abuse - A New Dataset and Classification Experiments for Implicitly Abusive Language}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2023},
  doi = {10.18653/v1/2023.emnlp-main.1012},
}
Figure