Abstract
This paper presents a linguistically driven proof of concept for finding potentially euphemistic terms (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 "passed away" for "died" or "enhanced interrogation" for "torture"—present a significant hurdle for Natural Language Processing (NLP). These terms are inherently ambiguous, often possessing literal meanings in other contexts, and represent a "conscious lexical choice" to convey politeness or obscure harsh realities. While previous research has often focused on narrow domains like drug-related slang, this study proposes a generalized, sentiment-driven pipeline to identify PETs across diverse sensitive topics.
Methodology
The authors developed a four-stage pipeline to extract and rank PETs from input sentences:
- Phrase Extraction: Using Gensim’s
Phrasesmodel, the system identifies collocations (bigrams and trigrams) within the sentence (e.g., "mentally_disabled"). - Phrase Filtering: Phrases are filtered based on topical relevance by calculating the cosine similarity between the phrase's word2vec embedding and a list of sensitive topics (e.g., death, politics).
- Phrase Paraphrasing: For each quality phrase, the system identifies the top 25 most similar words via word2vec to act as "near-synonym" candidates representing literal interpretations.
- Phrase Ranking: A RoBERTa model calculates sentiment and offensiveness scores for the original sentence and its replacements. Phrases causing the greatest shift toward negative or offensive scores are ranked highest.
The system was tested on the Euphemism Corpus (1,382 annotated sentences) and utilized the GloWbE (Corpus of Global Web-Based English) to train the underlying distributional models.
Results and Insights
The core innovation is the Sentiment Shift Hypothesis: replacing a PET with its literal counterpart should result in a measurable spike in offensiveness or negative sentiment. The algorithm correctly identified the target PET as one of the top two candidates in 52.5% of cases (725 out of 1,382 sentences), significantly outperforming a random baseline of 26.3%.
Interestingly, the model often discovered valid euphemisms not originally marked by human annotators. For instance, in a sentence where "mentally disabled" was the target, the model also correctly identified "intoxicated person" as a high-ranking PET. Out of the 725 successful detections, 64.5% were ranked as the top candidate.
Discussion and Limitations
Despite its success, the pipeline faces several bottlenecks. The reliance on pre-identified collocations means multi-word expressions not in the training data are missed. Additionally, common words used euphemistically (e.g., "expecting") are sometimes filtered out if they don't meet the similarity threshold for sensitive topics. Future work will explore contextualized embeddings (like BERT or RoBERTa) to improve phrase extraction and better handle the nuances of situational meaning.
Summary updated