Abstract
This study investigates the computational processing of euphemisms, a universal linguistic phenomenon, across multiple languages. We train a multilingual transformer model (XLM-RoBERTa) to disambiguate potentially euphemistic terms (PETs) in multilingual and cross-lingual settings. In line with current trends, we demonstrate that zero-shot learning across languages takes place. We also show cases where multilingual models perform better on the task compared to monolingual models by a statistically significant margin, indicating that multilingual data presents additional opportunities for models to learn about cross-lingual, computational properties of euphemisms. In a follow-up analysis, we focus on universal euphemistic "categories" such as death and bodily functions among others. We test to see whether cross-lingual data of the same domain is more important than within-language data of other domains to further understand the nature of the cross-lingual transfer.
Introduction
Euphemisms are a linguistic necessity used to navigate sensitive, harsh, or taboo topics—such as using "collateral damage" instead of "civilian deaths" or "between jobs" instead of "unemployed." While the impulse to soften language is a human universal, the specific lexical choices are deeply rooted in culture and context. In Natural Language Processing (NLP), this creates a significant disambiguation challenge. For instance, the phrase "let go" is a euphemism for termination in a corporate context but literal when describing a physical action. Most previous research has focused on English, leaving a gap in understanding how these linguistic "softeners" behave across diverse languages and whether models can learn the underlying concept of a euphemism regardless of the language.
Methodology
The researchers employed XLM-RoBERTa-base (XLM-R), a transformer model pre-trained on a massive multilingual corpus. The problem was framed as a binary classification task: given a context, is a specific "Potentially Euphemistic Term" (PET) being used euphemistically (1) or literally (0)?
Experiments were conducted across three training configurations:
- Monolingual: Training and testing on the same language.
- Zero-shot: Training on one language and testing on a completely different one.
- Multilingual: Training on a mix of languages to identify potential cross-lingual synergy.
The study utilized datasets for four languages: American English (EN), Mandarin Chinese (ZH), Spanish (ES), and Yorùbá (YO). The authors improved existing Chinese and Spanish datasets and added boundary tokens to the Yorùbá data to accommodate its unique word-order properties.
Key Results and Multilingual Synergy
A primary finding of the study is the "Synergy Effect": multilingual models frequently outperformed monolingual ones. This suggests that the model identifies computational properties of euphemisms shared across cultures. As shown in the results table, the "All" model (trained on all four languages) consistently provided the strongest performance. For example, the Chinese test set achieved a Macro-F1 score of 0.895 with the multilingual model, compared to 0.879 for the monolingual version.
Furthermore, the models demonstrated significant zero-shot capabilities. An English-trained model achieved a 0.653 Macro-F1 on Chinese data, far exceeding the majority-class baseline of 0.426. The performance boost from including data from all four languages was statistically significant (p < 0.05) for Chinese, English, and Spanish.
Semantic Category Analysis
A novel component of the methodology involved isolating four semantic domains: Physical/Mental Attributes (ATTR), Bodily Functions (BODY), Death (DEATH), and Sexual Activity (SEX). The goal was to determine if training on the same topic in a different language was more effective than training on a different topic in the same language.
The results indicated that for certain categories, such as "Attributes" in Chinese, training the model on the same category in English, Spanish, and Yorùbá was more effective than training it on different euphemism categories within Chinese. This suggests that certain taboo topics have a cross-lingual semantic signature that the model can recognize and transfer across languages.
Discussion and Limitations
While the results are promising, Yorùbá proved to be a significant outlier, showing lower performance. This is likely due to the fact that XLM-R was not pre-trained on Yorùbá, highlighting the ongoing challenges for low-resource languages in modern NLP. Additionally, the semantic category experiments were conducted on relatively small data subsets; larger, more balanced datasets are required to confirm the universality of "topic-transfer."
Future work will explore larger models and deeper linguistic analyses to further decode the mechanics of cross-lingual euphemism transfer.
Summary updated