💬 NLP

When Semantic Overlap Is Not Enough: Cross-Lingual Euphemism Transfer Between Turkish and English

This study investigates cross-lingual euphemism transfer between Turkish and English, revealing that semantic overlap alone is insufficient to guarantee positive transfer and that performance can sometimes improve with non-overlapping terms due to differences in label distribution and pragmatic context.

Original authors: Hasan Can Biyik, Libby Barak, Jing Peng, Anna Feldman

Published 2026-02-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hasan Can Biyik, Libby Barak, Jing Peng, Anna Feldman

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot how to understand human "soft talk."

In our daily lives, we often use euphemisms—polite words to replace harsh or embarrassing truths. Instead of saying "He was fired," we say "He was let go." Instead of "He died," we say "He passed away." These words are tricky because they rely heavily on culture and context, not just dictionary definitions.

This paper asks a big question: If we teach a computer to spot these polite lies in English, will it automatically know how to spot them in Turkish?

To find out, the researchers built a special experiment using two languages: English (a language the computer knows very well) and Turkish (a language the computer knows less well).

The Two Types of "Soft Talk"

The researchers realized that not all polite words are created equal. They sorted them into two buckets:

The "Twins" (Overlapping PETs): These are concepts where both languages use a similar "soft" strategy.
- Example: In English, we say someone "passed away." In Turkish, they say "vefat etmek" (which literally means "to fulfill a duty" or "to depart"). Both languages treat death like a journey or a departure. The computer can easily see the pattern here.
The "Strangers" (Non-Overlapping PETs): These are concepts where one language has a polite word, but the other just uses the blunt, direct word.
- Example: In English, if someone is unemployed, we might say they are "between jobs." In Turkish, there isn't a cute, idiomatic phrase for this; they just say "unemployed." If the computer learns the English "between jobs" trick, it won't know what to do when it sees the direct Turkish version because the "soft" version doesn't exist there.

The Big Surprise: Knowing More Doesn't Always Help

The researchers trained a smart AI model (XLM-R) to spot these words. Here is what they found, explained through a few analogies:

1. The "Overfitting" Trap (When learning too much hurts)
Imagine you are a student who memorizes the answers to a specific math test in English. If you take the same test in Turkish, you might do okay because the math is the same. But if you study too hard specifically for the English test, you might actually forget the general rules of math!

The Result: When the AI was trained specifically on English "soft talk," it sometimes got worse at spotting Turkish "soft talk" than if it had just been given a general, untrained brain. It learned English-specific tricks that didn't apply to Turkish.

2. The One-Way Street
The transfer of knowledge was very uneven.

English → Turkish: The AI was great at this. Because English has so much data, the AI learned the concept of being polite so well that it could spot it in Turkish, even when the Turkish words were totally different.
Turkish → English: This was a disaster. When the AI tried to learn from the smaller Turkish dataset and apply it to English, it struggled badly. It's like trying to teach a child to drive a Ferrari using only a bicycle. The Turkish data wasn't "rich" enough to teach the AI the complex rules of English politeness.

3. The "Category" Problem
The AI did okay with general topics like "Death" (because both cultures have similar polite ways to talk about it). But it completely failed on specific topics like Politics or Employment.

Why? In politics, English has very specific, culture-bound euphemisms (like "regime change" or "between jobs") that simply don't exist in Turkish. The AI tried to force a connection where there was none, leading to confusion.

The "Magic" vs. The "Robot"

The researchers also tested a giant, famous AI (GPT-4o) without training it on their specific data.

The Giant AI: It was good at guessing, but it had a weird bias. It just assumed everything was a euphemism because it was too eager to be polite. It got lucky on some tests because its "guessing bias" matched the data, not because it truly understood the language.
The Trained Robot: The smaller, trained model was actually more reliable. It proved that you don't need a massive, expensive brain to do this job; you just need to train a smaller brain carefully on the right kind of data.

The Takeaway

The main lesson of this paper is: Just because two languages share a meaning, it doesn't mean a computer can easily translate the "vibe" between them.

If you teach a computer to be polite in a language with lots of data (English), it can often figure out how to be polite in a language with less data (Turkish).
But if you try to teach it from the small language to the big one, it often fails.
And sometimes, teaching it too specifically about one language makes it forget how to generalize to others.

In short, teaching a computer to understand human "soft talk" is less like translating a dictionary and more like teaching it to understand the unspoken rules of a party. You can't just translate the words; you have to understand the culture, and that's much harder to do across different languages.

1. Problem Statement

Euphemisms are culturally and pragmatically dependent expressions used to soften taboo or sensitive topics (e.g., death, employment, politics). While they serve a universal function, their specific linguistic realizations vary significantly across languages. This creates a challenge for Natural Language Processing (NLP): Can multilingual models learn the abstract concept of a euphemism in one language and apply it to another, or is detection strictly bound by cultural familiarity?

Existing research has explored multilingual euphemism detection but has not systematically investigated how semantic overlap (the existence of functional equivalents across languages) influences zero-shot transfer. Specifically, it is unclear whether training on overlapping terms (OPETs) guarantees better cross-lingual generalization than training on non-overlapping terms (NOPETs), particularly between typologically distinct languages like English (analytic, high-resource) and Turkish (agglutinative, lower-resource).

2. Methodology

Dataset Construction and Annotation

The authors expanded an existing Turkish Potentially Euphemistic Terms (PETs) dataset to create a balanced, cross-lingual resource.

Expansion: The Turkish dataset was expanded from 30 to 70 PETs using the Turkish Web 2020 Corpus (trTenTen20), ensuring a balanced distribution of euphemistic and non-euphemistic instances.
Categorization: PETs in both English and Turkish were rigorously annotated into two subsets based on functional and pragmatic equivalence (not just literal translation):
- Overlapping PETs (OPETs): Terms where both languages share a functional euphemistic strategy for the same taboo concept (e.g., pass away $\approx$ vefat etmek for death).
- Non-Overlapping PETs (NOPETs): Terms where the euphemistic mapping exists in the source language but lacks a functional equivalent in the target language (e.g., between jobs for unemployment in English has no direct functional Turkish equivalent).
Reliability: Inter-Annotator Agreement (IAA) was measured using Cohen's Kappa ( $\kappa = 0.96$ ), indicating near-perfect reliability.

Experimental Setup

Model: The study utilizes XLM-RoBERTa (XLM-R), a multilingual masked language model pretrained on 100 languages.
Protocol:
1. Baseline: Evaluation of frozen XLM-R embeddings (no fine-tuning) to assess inherent cross-lingual representation.
2. Fine-tuning: XLM-R was fine-tuned separately on four subsets: English OPETs, English NOPETs, Turkish OPETs, and Turkish NOPETs.
3. Zero-Shot Transfer: Models trained on a source language subset were evaluated directly on target language subsets without further training.
4. Comparison: Results were compared against a zero-shot generative model (GPT-4o) using prompt engineering.
Validation: 10-fold cross-validation was used, with splits performed at the instance level.

3. Key Contributions

First Cross-Lingual Euphemism Dataset: Introduction of the first dataset explicitly categorized into OPET and NOPET subsets for English and Turkish, accompanied by detailed annotation guidelines.
Discovery of Transfer Asymmetry: The paper reveals a counterintuitive transfer asymmetry:
- English $\to$ Turkish: Transfer remains robust across domains (e.g., Employment F1=0.90).
- Turkish $\to$ English: Transfer degrades significantly in specific categories (e.g., Employment F1=0.36, Politics F1=0.18).
Semantic Overlap Limitations: The study demonstrates that semantic overlap is insufficient to guarantee positive transfer. In the low-resource Turkish-to-English direction, performance on overlapping terms can actually degrade compared to a frozen baseline, and in some cases, NOPET-based training yields better results than OPET-based training.
Analysis of Label Distribution: The authors identify that differences in label distribution (the proportion of euphemistic vs. literal instances) within specific categories explain many of the counterintuitive results. For instance, Turkish NOPETs in certain categories have a higher proportion of euphemistic labels, aligning better with model biases.

4. Key Results

Performance Trends

Fine-tuning vs. Baseline: While fine-tuning improves in-domain performance significantly (e.g., English OPETs F1 rose from 0.71 to 0.85), it can harm cross-lingual transfer. Training on English OPETs resulted in a lower F1 score (0.68) on Turkish OPETs compared to the frozen baseline (0.72), suggesting task-specific fine-tuning may overfit to language-specific cues.
Directional Asymmetry:
- EN $\to$ TR: Models trained on English (high-resource) transfer well to Turkish, even for NOPETs.
- TR $\to$ EN: Models trained on Turkish (low-resource) struggle to transfer to English, showing substantial degradation regardless of overlap status.
Category-Level Variance: Transfer failure is most pronounced in Employment and Politics. English-specific concepts like between jobs or regime change lack Turkish functional equivalents, causing F1 scores to drop to 0.36 and 0.18 respectively in the TR $\to$ EN direction.

Error Analysis

Cultural Gaps (False Negatives): Transfer fails when a euphemism relies on conceptual mappings absent in the source language (e.g., the baseball metaphor second base for sexual activity).
Lexical Memorization (False Positives): Models sometimes rely on specific tokens rather than context. For example, the Turkish term dört kollu (four-armed), used as a euphemism for a coffin, was incorrectly flagged as a euphemism in literal contexts (e.g., describing a goddess).
Generative Models: GPT-4o showed a systematic bias toward predicting "Euphemistic" (over-prediction), particularly in Turkish. While its F1 scores were competitive, they were driven by label distribution alignment rather than genuine linguistic generalization.

5. Significance and Implications

Theoretical Insight: The study challenges the assumption that semantic overlap is the primary driver of successful cross-lingual transfer in figurative language tasks. It highlights that resource imbalance and label distribution play critical roles, often outweighing semantic alignment.
Practical Application: For content moderation and hate speech detection on global platforms, relying solely on high-resource language models (like English) to detect euphemisms in low-resource languages may be insufficient, especially for culture-specific or non-overlapping terms.
Future Directions: The authors suggest extending the OPET/NOPET framework to other language pairs (e.g., Turkish-Azerbaijani, English-Spanish) to disentangle typological effects from resource effects, and developing training strategies that reduce lexical memorization in favor of contextual understanding.

In conclusion, the paper establishes that while multilingual models possess shared semantic spaces, the pragmatic and cultural nuances of euphemisms create significant barriers to zero-shot transfer, particularly when moving from low-resource to high-resource languages, and that semantic overlap alone does not solve this challenge.