Searching for PETs: Using Distributional and Sentiment-Based Methods to Find Potentially Euphemistic Terms

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

Abstract

This paper presents a linguistically driven proof of concept for finding potentially euphemistic terms, or PETs. Acknowledging that PETs tend to be commonly used expressions for a certain range of sensitive topics, we make use of distributional similarities to select and filter phrase candidates from a sentence and rank them using a set of simple sentiment-based metrics. We present the results of our approach tested on a corpus of sentences containing euphemisms, demonstrating its efficacy for detecting single and multi-word PETs from a broad range of topics. We also discuss future potential for sentiment-based methods on this task.

Introduction

Euphemisms—such as "enhanced interrogation" for "torture" or "between jobs" for "unemployed"—serve as linguistic camouflage, allowing speakers to navigate sensitive or taboo topics with politeness or strategic ambiguity. For Natural Language Processing (NLP), these terms present a significant challenge due to their context-dependency and rapid evolution. This research introduces a proof-of-concept pipeline designed to automatically detect Potentially Euphemistic Terms (PETs) by quantifying the "sentiment shift" that occurs when a phrase is replaced by its literal counterpart.

Motivation

Traditional NLP approaches to euphemism detection often rely on static, pre-defined keyword lists. However, euphemisms are dynamic; they frequently have valid literal meanings (e.g., "passed away" vs. "passed a ball") and change over time. There is a clear need for a more generalized, linguistically-driven method that can identify these terms in the wild without relying solely on fixed dictionaries.

Methodology

The proposed system identifies PETs through a four-stage pipeline combining distributional semantics with transformer-based sentiment analysis:

  1. Phrase Extraction: The system identifies single and multi-word expressions (up to trigrams) using collocation detection.
  2. Phrase Filtering: Candidates are filtered based on topical relevance by calculating the cosine similarity between the candidate's Word2Vec embedding and a list of sensitive topics (e.g., death, politics). Only phrases exceeding a specific similarity threshold are retained.
  3. Phrase Paraphrasing: For each candidate, the system generates "pseudo-literal" interpretations by finding the top 25 most distributionally similar words using Word2Vec.
  4. Phrase Ranking: Using a RoBERTa model, the system scores the sentiment and offensiveness of the original sentence versus the sentence with the candidate replaced by its paraphrases.
Table 1: Pipeline attrition showing the number of candidates and targets retained at each stage.

The core hypothesis is that euphemisms are conscious lexical choices made to reduce negativity. Therefore, replacing a true PET with a literal alternative should trigger a measurable spike in offensiveness or negativity scores.

Data

The system was evaluated using two primary resources:

  • The Euphemism Corpus: A dataset of 1,382 sentences, each annotated with a single gold-standard PET.
  • GloWbE (Corpus of Global Web-Based English): Specifically the US dialect portion, used to train the Word2Vec and Phrase models to ensure distributional semantics aligned with the test data.

Results and Analysis

The algorithm successfully identified the target PET as one of the top two candidates in 52.5% of the test sentences (725 out of 1,382). This significantly outperformed a random selection baseline of approximately 26.3%.

Figure 1: Sample output of the ranking algorithm showing identified PETs and their associated sentiment-shift scores.

Qualitative analysis revealed that the model often discovered valid euphemisms not marked in the original gold-standard dataset. For instance, in a sentence where only "mentally disabled" was tagged, the model also identified "intoxicated person" as a high-ranking PET. Furthermore, the model demonstrated contextual sensitivity, distinguishing between euphemistic and literal uses of the same word (e.g., "dismissed" meaning "fired" vs. "class is over").

Discussion and Future Work

A key insight from this study is that "offensiveness" scores are more predictive than general sentiment scores, confirming that euphemisms are specifically designed to be less offensive rather than just less negative.

However, limitations remain. The system's reliance on early-stage collocation detection means that if a multi-word expression is missed initially, it cannot be recovered. Additionally, highly polarized but non-euphemistic terms (e.g., "sex slaves") can produce false positives due to large sentiment shifts. Future work will explore using contextualized embeddings (like BERT) for the paraphrasing stage and more open-ended topic modeling to reduce reliance on pre-defined sensitive categories.

Summary updated

Abstract

This paper presents a linguistically driven proof of concept for finding potentially euphemistic terms, or PETs. Acknowledging that PETs tend to be commonly used expressions for a certain range of sensitive topics, we make use of distributional similarities to select and filter phrase candidates from a sentence and rank them using a set of simple sentiment-based metrics. We present the results of our approach tested on a corpus of sentences containing euphemisms, demonstrating its efficacy for detecting single and multi-word PETs from a broad range of topics. We also discuss future potential for sentiment-based methods on this task.
BibTeX
@misc{lee2022searching,
  author = {Patrick P. C. Lee and Martha Gavidia and Anna Feldman and Jing Peng},
  title = {{Searching for PETs: Using Distributional and Sentiment-Based Methods to Find Potentially Euphemistic Terms}},
  year = {2022},
  eprint = {2205.10451},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2205.10451},
}

Loading related content...

Figure