Euphemisms substitute socially sensitive expressions, often softening or reframing meaning, and their reliance on cultural and pragmatic context complicates modeling across languages. In this study, we investigate how cross-lingual equivalence influences transfer in multilingual euphemism detection. We categorize Potentially Euphemistic Terms (PETs) in Turkish and English into Overlapping (OPETs) and Non-Overlapping (NOPETs) subsets based on their functional, pragmatic, and semantic alignment. Our findings reveal a transfer asymmetry: semantic overlap is insufficient to guarantee positive transfer, particularly in low-resource Turkish-to-English direction, where performance can degrade even for overlapping euphemisms, and in some cases, improve under NOPET-based training. Differences in label distribution help explain these counterintuitive results. Category-level analysis suggests that transfer may be influenced by domain-specific alignment, though evidence is limited by sparsity.
Figures & tables
EN
TR
ZH
ES
YO
PETs
144
70
149
233
157
Total
3098
2436
3211
2952
2598
Euphemistic
1841
1457
2213
1955
1689
Non-Euph.
1257
979
998
997
909
Table 1: Statistics of PETs datasets across languages. The Turkish (TR) column reflects our newly expanded dataset (70 PETs). For English, we use a subset of 71 PETs from the original 144 PETs to create a balanced sample for both languages.
Model
Previous F1
New F1
BERTurk
0.84
0.84
Electra
0.86
0.83
mBERT
0.80
0.76
XLM-R
0.82
0.81
Table 2: Comparison of old vs. new F1 scores after expanding the Turkish PETs dataset.
Table 3: Distribution of PETs across OPET/NOPET subsets (a) and semantic categories (b). Category values show number of PETs (number of instances). Terms marked "Always Euph." are used exclusively in euphemistic contexts; "Never Euph." appear only in literal uses.
English
Turkish
Metric
OPET
NOPET
OPET
NOPET
F1 Score
0.71
0.61
0.72
0.68
Table 4: Baseline performance (F1) using frozen XLM-R embeddings.
Test on English
Test on Turkish
Train
Subset
OPET
NOPET
OPET
NOPET
Baseline
(XLM-R)
0.71
0.61
0.72
0.68
EN
OPET
0.85 *
0.65
0.68
0.70
NOPET
0.59
0.62 *
0.52
0.56
TR
OPET
0.64
0.58
0.80 *
0.68
NOPET
0.66
0.60
0.70
0.79 *
Table 5: Cross-lingual transfer performance showing F1 scores and standard deviations across 10-fold cross-validation. Models are trained on source language subsets (rows) and tested on target language subsets (columns). (*) indicates in-domain (same language, same subset) performance.
TR→EN
EN→TR
Category
OPET
NOPET
OPET
NOPET
Death
0.67
0.38
0.74
0.86
Sexual Activity
0.86
0.44
0.79
0.84
Employment
0.35
0.36
0.48
0.90
Phys./Ment. Attr.
0.79
0.59
0.68
0.70
Politics
0.23
0.18
0.50
N/A
Table 6: Category-specific F1 scores for cross-lingual transfer. Models are trained on all source language OPETs (or all NOPETs), then tested on target language subsets, with F1 computed only for instances from each category.
Model
EN OPET
EN NOPET
TR OPET
TR NOPET
Avg
Frozen XLM-R (baseline)
0.71
0.61
0.72
0.68
0.68
Fine-tuned XLM-R
0.85
0.62
0.80
0.79
0.77
Zero-shot GPT-4o
0.75
0.71
0.73
0.80
0.75
Table 7: In-domain performance comparison across approaches. Fine-tuned XLM-R results show models trained and tested on the same subset. Bold indicates best per column.
Split
Pred. %
Label %
Δ
TR OPET
67.3
51.2
+16.1
TR NOPET
77.0
59.0
+18.0
EN OPET
64.0
52.0
+12.0
EN NOPET
59.2
49.4
+9.8
Table 8: GPT-4o prediction bias. Δ shows over-prediction of “Euphemistic” compared to actual label distribution.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Category
TR OPETs
TR NOPETs
EN OPETs
EN NOPETs
PETs
Instances
PETs
Instances
PETs
Instances
PETs
Instances
Death
16
484
5
177
14
528
1
52
Body Functions/Parts
5
290
9
179
3
27
1
5
Sexual Activity
4
134
4
92
5
118
5
147
Employment/Finances
5
122
2
45
8
292
6
138
Physical/Mental Attributes
5
89
6
168
5
132
14
421
Appendix
Table 9: Category statistics for Turkish (TR) and English (EN) PETs OPETs and NOPETs subsets.
English PET
Turkish PET
Taboo Concept
Label
Notes
comfort woman
hayat kadını
Sex work
OPET
Different expressions, same euphemistic function.
pass away
vefat etmek
Death
OPET
Functional equivalence across languages.
the birds and the bees
—
Sex education
NOPET
Non-Overlap.
—
fındık kırmak
Sexual activity
NOPET
Turkish slang; no English equivalent.
Appendix
Table 10: Examples of OPET and NOPET classifications. While lexical forms may differ, OPETs share a functional role in euphemizing the same taboo concept.
Cross-lingual transfer is central to low-resource machine translation, but its behavior within closely related language families remains insufficiently characterized. We study transfer among five Turkic languages; Turkish, Azerbaijani, Uzbek, Kazakh, and Kyrgyz; using pairwise transfer matrices. In this setting, each model is fine-tuned with one transfer source and evaluated on a different transfer target while the translation target remains the same. Across mT5 experiments, we find that transfer is strongest between closely related Turkic pairs, especially Turkish-Azerbaijani and Kazakh-Kyrgyz. We also show that transfer direction matters, and that the same transfer source-transfer target pair can behave differently when the translation target changes. Latinization improves BLEU and chrF in several script-mismatched settings, but its effect is not uniform across metrics. Additional analyses show that transfer sources are mostly stable across different datasets and model settings.
Omer Burak Cinar, Mehmet Mert Dalkilic, Cagri Toraman
Middle East Technical University · Computer Engineering Department
Paralinguistic speech tasks are often considered relatively language-agnostic, as they rely on extralinguistic acoustic cues rather than lexical content. However, prior studies report performance degradation under cross-lingual conditions, indicating non-negligible language dependence. Still, these studies typically focus on isolated language pairs or task-specific settings, limiting comparability and preventing a systematic assessment of task-level language dependence. We introduce the Cross-Lingual Transfer Matrix (CLTM), a systematic method to quantify cross-lingual interactions between pairs of languages within a given task. We apply the CLTM to two paralinguistic tasks, gender identification and speaker verification, using a multilingual HuBERT-based encoder, to analyze how donor-language data affects target-language performance during fine-tuning. Our results reveal distinct transfer patterns across tasks and languages, reflecting systematic, language-dependent effects.
Pol Buitrago, Oriol Pareras, Federico Costa +1
Barcelona Supercomputing Center (BSC), Spain · Universitat Polit`ecnica de Catalunya (UPC), Spain
We study cross-lingual transfer by fine-tuning seven large language models (4B--671B parameters) on Arabic and evaluating zero-shot reading comprehension on Semitic languages and non-Semitic controls. Across dense and Mixture-of-Experts architectures, we find no evidence of Semitic-specific transfer: models with weak baselines improve dramatically across all languages, while strong-baseline models show only marginal gains regardless of language family. A chain-of-thought ablation reinforces this finding -- the same models that benefit most from fine-tuning benefit equally from inference-time reasoning, suggesting both mechanisms address task-format alignment rather than cross-lingual knowledge transfer.