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

Technical Deep Dive: When Semantic Overlap Is Not Enough

Overview

This research investigates the limitations of cross-lingual transfer in detecting euphemisms, examining how semantic overlap between languages influences the performance of multilingual language models like XLM-RoBERTa (XLM-R). The key findings are:

• Semantic overlap has limited impact when transferring from high-resource to low-resource languages. XLM-R models trained on English OPETs (Overlapping Potentially Euphemistic Terms) show minimal degradation when applied to Turkish OPETs and NOPETs (Non-Overlapping PETs).
• However, a substantial asymmetry exists in the opposite direction, with Turkish-to-English transfer degrading significantly, particularly for certain semantic categories like Employment and Politics.
• Error analysis reveals that successful transfer occurs when both languages use similar expressions for the same taboo concept, while failures stem from culture-specific mappings absent in the source language.

Problem & Context

• Euphemisms are indirect expressions that soften harsh or taboo concepts, but their usage is highly dependent on cultural and pragmatic context.
• This creates unique challenges for multilingual language models: can a model learn the concept of a euphemism in one language and apply it to another, or is detection strictly bound by cultural familiarity?
• Beyond the theoretical interest, cross-lingual euphemism detection has practical implications for applications like content moderation and hate speech monitoring.

Methodology

• The researchers categorized PETs in English and Turkish into Overlapping (OPETs) and Non-Overlapping (NOPETs) subsets based on functional, pragmatic, and semantic alignment across the two languages.
• They evaluated XLM-R's zero-shot cross-lingual transfer performance, training on one language subset and testing on the other.
• The study also included a comparison to the zero-shot performance of the GPT-4o model.

Data & Experimental Setup

• The researchers used and expanded the existing English and Turkish PETs datasets, which contain words/phrases that can be used euphemistically.
• They achieved high inter-annotator agreement (κ = 0.96) in categorizing the PETs into OPETs and NOPETs.
• Experiments used 10-fold cross-validation to ensure robust evaluation across diverse train-test-val splits.

Results

1. Overlap Has Limited Impact when Transferring from High-Resource to Low-Resource:
   • XLM-R models trained on English OPETs show minimal degradation when applied to Turkish OPETs (0.68 vs. 0.72 baseline) and Turkish NOPETs (0.70 vs. 0.68 baseline).
   • This suggests that extensive pretraining on English provides more robust cross-lingual transfer, regardless of overlap status in the target language.
2. Substantial Transfer Asymmetry, Most Pronounced at the Category Level:
   • English-to-Turkish transfer remains robust across domains (Employment F1=0.90, Death F1=0.86).
   • Turkish-to-English transfer degrades substantially (Employment F1=0.36, Death F1=0.38), with performance gaps exceeding 0.50 F1 points in some categories.
   • This asymmetry is consistent with resource imbalance in multilingual pretraining, though language-specific factors may also contribute.
3. Error Analysis Reveals Cultural Gaps and Potential Lexical Memorization:
   • Failures stem from culture-specific mappings absent in the source language (e.g., "between jobs" in English has no Turkish equivalent).
   • Some errors suggest the model relies on specific tokens rather than context (e.g., over-predicting the Turkish term "dört kollu" as euphemistic in literal contexts).
   • Successful transfer occurs when both languages use semantically similar expressions for the same taboo concept (e.g., "pass away" and "vefat etmek" for death).

Interpretation

• The results suggest that while semantic overlap can provide some benefit, the robustness of cross-lingual transfer is primarily determined by the relative resource availability and pretraining coverage of the source and target languages.
• The asymmetry in transfer performance highlights the challenges of deploying multilingual models in real-world settings, where the availability and distribution of training data can significantly impact generalization across languages.
• The OPET/NOPET framework offers a nuanced lens for analyzing cross-lingual transfer dynamics, revealing that even with shared multilingual representations, task-specific adaptation can reduce generalization abilities.

Limitations & Uncertainties

• The study was limited to English and Turkish, which differ typologically and in pretraining resource availability. Extending the analysis to additional language pairs could help disentangle the factors driving the observed asymmetries.
• The categorization of PETs into OPETs and NOPETs relies on functional and pragmatic equivalence, which may be affected by sociolinguistic variation across regions and dialects.
• Some semantic categories were underrepresented in the datasets, limiting the statistical power of the category-level analysis.
• The study focused on binary classification and did not explore more nuanced measures of euphemistic strength or speaker intent.

What Comes Next

• Expand the OPET/NOPET framework to additional language pairs with varying typological and resource characteristics (e.g., English-Spanish, Turkish-Azerbaijani) to better understand the generalizability of the observed patterns.
• Investigate whether the transfer asymmetry persists across different model architectures (e.g., mBERT, mT5) to assess the robustness of the findings.
• Develop training strategies that reduce lexical memorization and improve context-dependent classification of euphemisms.
• Explore more fine-grained approaches to euphemism detection, such as predicting euphemistic strength or modeling speaker intent.

Sources:

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