CATs are Fuzzy PETs: A Corpus and Analysis of Potentially Euphemistic Terms

Martha Gavidia, Patrick P. C. Lee, Anna Feldman, Jing Peng

Abstract

Euphemisms have not received much attention in natural language processing, despite being an important element of polite and figurative language. Euphemisms prove to be a difficult topic, not only because they are subject to language change, but also because humans may not agree on what is a euphemism and what is not. Nevertheless, the first step to tackling the issue is to collect and analyze examples of euphemisms. We present a corpus of potentially euphemistic terms (PETs) along with example texts from the GloWbE corpus. Additionally, we present a subcorpus of texts where these PETs are not being used euphemistically, which may be useful for future applications. We also discuss the results of multiple analyses run on the corpus. Firstly, we find that sentiment analysis on the euphemistic texts supports that PETs generally decrease negative and offensive sentiment. Secondly, we observe cases of disagreement in an annotation task, where humans are asked to label PETs as euphemistic or not in a subset of our corpus text examples. We attribute the disagreement to a variety of potential reasons, including if the PET was a commonly accepted term (CAT).

Introduction

In Natural Language Processing (NLP), models are increasingly adept at detecting overt hate speech, yet they often struggle with the nuanced, "slippery" nature of euphemisms. Euphemisms—indirect expressions used to soften the blow of taboo or sensitive topics—present a unique challenge for machine learning. They are highly context-dependent, subject to the "euphemism treadmill" where terms eventually lose their polite status, and inherently subjective. To address this gap, we developed a robust corpus of Potentially Euphemistic Terms (PETs) to help models disambiguate these terms in real-world communication.

The PETs Corpus

We constructed the PETs Corpus by compiling 184 candidate terms from linguistic literature and dictionaries. Using spaCy’s PhraseMatcher, we extracted over 5,500 candidate sentences from the US dialect portion of the GloWbE (Corpus of Global Web-Based English). The final corpus consists of 1,965 sentences, categorized into 1,382 euphemistic usages and 583 literal usages across 129 distinct PETs.

A critical aspect of this dataset is the distinction between euphemistic and literal contexts. For instance, the term "between jobs" acts as a euphemism for unemployment in a personal biography but can be literal in other contexts, such as choosing between multiple job offers.

Examples of PETs used in euphemistic versus literal contexts.

Sentiment Analysis: The "Sentiment Shield"

To quantify the "softening" effect of euphemisms, we conducted a substitution test. We used a roBERTa-base model to score the sentiment of original sentences and then replaced the PETs with their literal counterparts (e.g., replacing "low-income" with "poor").

The results were striking: substituting euphemisms with literal terms caused a 54.6% spike in negative sentiment scores and a 30.0% increase in offensive ratings. This suggests that euphemisms act as a significant buffer, effectively "masking" underlying negativity from current sentiment analysis models.

Percentage change in sentiment and offensiveness labels after substituting PETs with literal terms.

Human Subjectivity and "CATs"

Our analysis also highlighted the inherent ambiguity of these terms. In an inter-rater reliability task involving five NLP graduate students, the Krippendorff’s alpha was only 0.415. This "fair" agreement level reflects the difficulty humans have in distinguishing between a euphemism and a "Commonly Accepted Term" (CAT). CATs are terms that may have originated as euphemisms but have since become the standard, objective terminology in specific fields, such as "venereal disease" in medicine. Annotators often struggle to decide if these are "polite" or just "literal," leading to "fuzzy" data.

Conclusion

The PETs Corpus provides a foundational resource for improving euphemism detection and generation in NLP. Our findings emphasize that euphemisms are not just linguistic curiosities but functional tools that significantly alter the perceived sentiment and offensiveness of text. Future work will expand this research to other English dialects and explore consensus coding methods to better navigate the boundaries of polite language.

Summary updated

Abstract

Euphemisms have not received much attention in natural language processing, despite being an important element of polite and figurative language. Euphemisms prove to be a difficult topic, not only because they are subject to language change, but also because humans may not agree on what is a euphemism and what is not. Nevertheless, the first step to tackling the issue is to collect and analyze examples of euphemisms. We present a corpus of potentially euphemistic terms (PETs) along with example texts from the GloWbE corpus. Additionally, we present a subcorpus of texts where these PETs are not being used euphemistically, which may be useful for future applications. We also discuss the results of multiple analyses run on the corpus. Firstly, we find that sentiment analysis on the euphemistic texts supports that PETs generally decrease negative and offensive sentiment. Secondly, we observe cases of disagreement in an annotation task, where humans are asked to label PETs as euphemistic or not in a subset of our corpus text examples. We attribute the disagreement to a variety of potential reasons, including if the PET was a commonly accepted term (CAT).
BibTeX
@misc{gavidia2022cats,
  author = {Martha Gavidia and Patrick P. C. Lee and Anna Feldman and Jing Peng},
  title = {{CATs are Fuzzy PETs: A Corpus and Analysis of Potentially Euphemistic Terms}},
  year = {2022},
  eprint = {2205.02728},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2205.02728},
}
Figure