Size Sound Symbolism Survives Genealogy Across 27 Languages¶
Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/Anika1703/testing_size_symbolism
Area: LLM Reasoning
Keywords: Sound Symbolism, Cross-Linguistic Generalization, Adversarial Representation Learning, Phoneme Analysis, Genealogical Control
TL;DR¶
Addressing whether sound symbolism is merely an artifact of shared ancestry or language contact, this paper establishes an audio-validated dataset of 810 size adjectives across 27 languages and 13 families, demonstrating through normalized Levenshtein distance transfer tiers and adversarial feature scrubbing that phonological size cues survive genealogical controls and remain consistently above chance.
Background & Motivation¶
Classical linguistic structuralism rooted in Saussurean theory posits the arbitrariness of the sign as a foundational axiom, maintaining that the relationship between phonological form and semantic content is largely arbitrary. In contrast, extensive perceptual and psycholinguistic investigations—ranging from the Bouba–Kiki paradigm to ideophonic systems across diverse languages—have long documented non-arbitrary sound-symbolic associations. In the domain of magnitude and size symbolism, acoustic frequency-code hypotheses suggest that smaller physical cavities produce higher resonant frequencies, linking high front vowels such as /i/ to smallness, whereas larger cavities correspond to lower resonant frequencies, linking low/back vowels such as /o/ and /a/ to largeness.
However, demonstrating genuine cross-family sound symbolism remains notoriously challenging because apparent cross-linguistic regularities can easily be confounded by shared genealogical ancestry and areal contact borrowing. When closely related languages share common lexical roots and phoneme inventories, computational models can readily exploit genealogical shortcuts to achieve high classification accuracy without learning universal perceptual groundings. While prior computational studies have deployed family-level stratification or mixed-effects regression, they rarely subject sound-symbolic representations to rigorous out-of-family stress tests where models are actively prohibited from leveraging source-language and genealogical markers.
To resolve this impasse, this paper reframes cross-family sound symbolism as an adversarial generalization and representation scrubbing stress test, investigating whether magnitude cues persist when genealogical shortcuts are deliberately pressurized away. Core idea: quantify lexical distance via 40-word Swadesh normalized Levenshtein distances (LDN) to construct target-relative cross-family transfer tiers, and combine gradient-reversal adversarial scrubbing against similarity bins and source languages with interpretable phoneme probing to demonstrate that resilient, vowel-dominated size cues survive across 13 distinct language families.
Method¶
Overall Architecture¶
The framework implements a comprehensive evidence ladder to test whether size-predictive phonological cues survive when genealogical shortcuts are suppressed. The pipeline begins with a curated dataset of 810 size adjectives across 27 languages spanning 13 genealogical families, phonemically transcribed in IPA and validated via native-speaker audio recordings. Feature extraction computes bag-of-phoneme frequency vectors using specialized IPA tokenization. Evaluation proceeds along two complementary axes: interpretable linear and tree-based baseline classifiers (Logistic Regression and Decision Trees) evaluated across target-relative LDN similarity tertiles (Most, Somewhat, and Least Similar) under a leave-one-language-out (LOLO) protocol; and a neural adversarial representation scrubber built on a frozen Transformer encoder pretrained on 1.6M WikiPron pronunciations, employing a Gradient Reversal Layer (GRL) to invert gradients from a similarity-bin or 26-way source-language adversary while preserving binary size prediction capability.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
A["Cross-Linguistic Size Adjectives<br/>810 words across 27 languages & 13 families with native audio validation"] --> B["Genealogical Distance Stratification<br/>40-word Swadesh lists & target-relative LDN similarity tertiles"]
B --> C["Interpretable Phoneme Classification<br/>LOLO transfer via Logistic Regression & Decision Tree root splits"]
B --> D["Adversarial Representation Scrubber<br/>Pretrained WikiPron Transformer + Gradient Reversal Layer"]
D --> E["Adversarial Dual-Head Optimization<br/>Suppress similarity-bin & source-language cues under post-hoc probing"]
Key Designs¶
1. Genealogical Distance Stratification: Eliminating Cognate Confounders via Target-Relative Distance Tiers
To prevent models from relying on shared linguistic ancestry, the methodology departs from arbitrary geographic groupings and adopts normalized Levenshtein distance (LDN) computed over standardized 40-word Swadesh lists. For two words \(\alpha_i\) and \(\beta_j\), word-level distance normalizes edit distance by the longer word length \(D(\alpha_i, \beta_j) = D_l(\alpha_i, \beta_j) / L(\alpha_i, \beta_j)\), with language-level LDN defined as the mean over aligned concepts: $\(D(\alpha, \beta) = \frac{1}{M}\sum_{i=1}^M D(\alpha_i, \beta_i)\)$ Addressing prior debates regarding further-normalized LDND, the authors replicate correlation analyses across all 351 language pairs in their sample, showing that background non-cognate similarity \(\Gamma\) correlates strongly with lexical distance (\(C(\Gamma, D_s) = 0.49\)), confirming that dividing by \(\Gamma\) inadvertently removes genuine genealogical signals. Consequently, LDN is retained. For each held-out target language \(\ell\), the remaining 26 languages are ranked by LDN and split into three equal-sized similarity tertiles (Most, Somewhat, and Least Similar), establishing a rigorous gradient of phylogenetic distance. Crucially, inventory distance analysis using articulatory PanPhon features confirms that phoneme inventory overlap correlates only weakly with LDN (\(r = 0.255\)), validating that the tertiles cleanly control for historical relatedness rather than accidental inventory matching.
2. Interpretable Phoneme Classification: Disentangling Cross-Family Vowel Dominance from Local Consonantal Cues
To inspect individual phonological cues without black-box distortion, words are segmented using ipatok into bag-of-phoneme frequency vectors over attested IPA segments, excluding language-specific suprasegmentals (tone, stress). Under a leave-one-language-out (LOLO) cross-validation protocol across 81 language-bin models, Logistic Regression (L2-regularized, \(C=1.0\)) and Decision Trees (max depth 6) are trained within specific similarity tertiles. The analysis filters for phonemes demonstrating consistent coefficient signs across \(>60\%\) of models and conducts 1,000 bootstrap resamples within each language. Decision tree root splits provide an orthogonal measure of feature gain, identifying the single most informative phonological cut separating small from large words across language families.
3. Adversarial Representation Scrubber: Inverting Gradients to Prevent Exploitation of Genealogical Shortcuts
To prevent neural representations from implicitly encoding and exploiting source-language provenance, the authors introduce a dual-task adversarial architecture. The base network comprises a 2-layer, 128-hidden-unit Transformer encoder pretrained on 1.6M IPA words from WikiPron via masked language modeling and subsequently frozen as a fixed feature extractor. A trainable 64-dimensional linear projection layer \(E\) receives outputs and branches into two task heads: a binary size classifier \(C\) (2-layer MLP with ReLU and dropout) and an adversary \(A\) (2-layer MLP). Connected to \(E\) via a Gradient Reversal Layer (GRL) scaling inverted gradients by \(-\lambda\), the system optimizes the minimax objective: $\(\min_{E, C} \max_A \mathcal{L}_{\text{size}}(C(E(x)), y_{\text{size}}) - \lambda \cdot \mathcal{L}_{\text{bin}}(A(E(x)), y_{\text{bin}})\)$ Two adversarial conditions are imposed: a 3-way similarity-bin adversary (\(\mathcal{L}_{\text{bin}}\)) and an even stricter 26-way source-language adversary (\(\mathcal{L}_{\text{lang}}\)). By ramping \(\lambda\) smoothly from 0 to 1 over 20 epochs, the projection \(E\) is actively forced to unlearn language-specific phonetic fingerprints, preserving only those representations that support size categorization under severe adversarial pressure.
4. Targeted Phonetic Ablation and Residual Probing: Verifying Acoustic Grounding and Informational Usability
To guarantee that classification successes reflect genuine semantic-phonetic mappings rather than distributional artifacts, the authors implement targeted feature ablations and post-hoc probing. Phonetic ablations systematically remove vowels, restrict features to 23 typologically frequent cross-linguistic segments, contrast plosives against nasals, scramble size labels to establish chance baselines, and control for Romance inflectional morphology (excluding Spanish and Portuguese to eliminate masculine suffix /-o/ confounds). At the representation level, a post-hoc linear probe is trained on the frozen embeddings produced by the adversarial scrubber. While the joint adversary is driven near chance (34.0% for bins; 5.1% for source languages), the post-hoc probe recovers source-language identity with 43.4% accuracy, formally characterizing the adversarial framework as a test of cue usability rather than naive information removal.
Loss & Training¶
The neural scrubber minimizes standard cross-entropy loss for size classification while maximizing cross-entropy loss on adversarial heads through gradient reversal. The base Transformer is pretrained for 2 epochs on WikiPron before freezing. During downstream adversarial fine-tuning, dropout is set to 0.1, and \(\lambda\) scales linearly from 0 to 1 across 20 epochs. Optimal model checkpoints are selected at epochs \(\ge 8\) where the discrepancy between training size accuracy and adversary accuracy is maximized, ensuring robust adversarial suppression. Baseline linear models are fitted with the L-BFGS optimizer with max 1,000 iterations.
Key Experimental Results¶
Main Results¶
Table 1 summarizes classification performance across models and LDN similarity bins averaged over all 27 languages. Decision Trees sustain statistically significant above-chance accuracy even in the most distant cross-family transfer bin (Least Similar).
| Similarity Bin | Model | Accuracy (%) | Precision (%) | Recall (%) | Macro F1 (%) | F1 Small (%) | F1 Large (%) |
|---|---|---|---|---|---|---|---|
| Most Similar | Logistic Reg. | 59.3*** | 60.8 | 59.3 | 58.1 | 58.3 | 57.8 |
| Decision Tree | 60.2*** | 63.2 | 60.2 | 58.2 | 57.1 | 59.3 | |
| Somewhat Similar | Logistic Reg. | 57.0** | 58.0 | 57.0 | 55.9 | 58.2 | 53.7 |
| Decision Tree | 56.4*** | 55.7 | 56.4 | 50.2 | 53.1 | 47.3 | |
| Least Similar | Logistic Reg. | 52.2 | 52.4 | 52.2 | 50.0 | 49.5 | 50.5 |
| Decision Tree | 55.8** | 56.5 | 55.8 | 50.7 | 47.5 | 54.0 |
(Note: *** \(p < 0.001\), ** \(p < 0.01\); one-sample t-test against 50% chance baseline across 27 held-out language accuracies.)
Under adversarial scrubbing: - Similarity-Bin Adversary: The jointly trained adversary falls to 34.0% (chance is 33.3%; unconstrained baseline reaches 54.4%), while held-out size accuracy remains robust at 54.4% (relative to 55.4% unconstrained baseline; \(p = 0.09\) for difference, indicating no significant loss of size information). Across the 25 languages with non-zero variance, 15 languages (60%) achieve above-chance size accuracy under Benjamini-Hochberg FDR correction (\(q < 0.05\)) while bin accuracy is suppressed to chance, spanning 9 of the 13 language families. - 26-Way Source-Language Adversary: Training language prediction plummets from 23.9% to 5.1% (near 3.8% chance), yet held-out size accuracy is entirely preserved at 55.8% (versus 54.2% baseline).
Ablation Study¶
Table 2 details bin-wise mean accuracies across phonological ablation conditions using interpretable classifiers.
| ID | Ablation Condition | LR Mean (%) | DT Mean (%) | LR by Bin (Most / Some / Least %) | DT by Bin (Most / Some / Least %) |
|---|---|---|---|---|---|
| 0 | Baseline (All phonemes) | 56.1 | 57.5 | 59.3 / 57.0 / 52.2 | 60.2 / 56.4 / 55.8 |
| 1 | Scrambled Labels | 48.8 | 50.4 | 48.6 / 47.9 / 50.0 | 50.9 / 50.0 / 50.4 |
| 2 | No Vowels | 54.2 | 54.2 | 58.6 / 52.8 / 51.1 | 55.8 / 54.6 / 52.2 |
| 3 | High-Frequency (23 phonemes) | 56.5 | 58.5 | 58.5 / 55.6 / 55.4 | 60.2 / 57.8 / 57.4 |
| 4 | Only Plosives | 54.5 | 56.4 | 56.7 / 53.5 / 53.2 | 58.6 / 55.9 / 54.6 |
| 5 | Only Nasals | 51.6 | 52.5 | 52.0 / 51.0 / 51.8 | 53.0 / 51.6 / 53.0 |
Key Findings¶
- Vowels carry the primary cross-family size signal: Eliminating vowels causes Logistic Regression performance in the Least Similar bin to collapse to 51.1%, while Decision Trees drop to 52.2%. Consonants provide localized cues within close genealogical bins but fail to transfer reliably across language families.
- Stable polarity of /i/ versus /o/ and /a/: Across 81 language-bin models, /o/ is positively associated with largeness in 100% of models (mean coefficient +0.67), while /i/ associates with smallness in 99% of models (mean coefficient -0.44). In Decision Trees, /a/ accounts for 27% and /i/ for 24% of root splits, spanning diverse families including Sino-Tibetan, Afroasiatic, Indo-European, and Turkic.
- Robust transfer from non-Indo-European to Indo-European: Training on 11 non-Indo-European languages and testing on 16 Indo-European languages yields statistically significant above-chance accuracy for both Logistic Regression (57.9%, \(p = 0.010\)) and Decision Trees (57.5%, \(p = 0.001\)), ruling out single-family phylogenetic skew as an explanation.
- Morphology does not account for the vowel pattern: Excluding Spanish and Portuguese to eliminate masculine suffix /-o/ confounds leaves the positive association of /o/ intact across all 75 remaining models (mean coefficient +0.48).
Highlights & Insights¶
- Reframing cognitive linguistic debates into distribution-shift and adversarial robustness benchmarks: Rather than relying solely on observational correlational statistics, the authors demonstrate that size-sound mappings survive active adversarial scrubbing of language identity.
- Triangulation between interpretable symbolic models and deep neural scrubbers: By pairing transparent Decision Tree root splits and bootstrap coefficient analyses with neural minimax optimization, the paper links abstract learned representations directly to concrete phonetic mechanisms.
- Nuanced interpretation of adversarial representation scrubbing: Through post-hoc linear probing, the paper provides a cautionary methodological lesson: adversarial networks disrupt the model's immediate usability of protected attributes without achieving total information erasure.
Limitations & Future Work¶
- Constrained vocabulary size due to audio validation overhead: Recruiting and verifying native-speaker pronunciations across 27 typologically diverse languages limited the dataset to 30 adjectives per language (810 total), precluding large-scale end-to-end model fine-tuning.
- Loss of acoustic and suprasegmental granularity: The symbolic bag-of-phonemes representation abstracts away token position (word-initial vs. word-final) and excludes prosodic features such as tone and lexical pitch-accent, which carry significant iconic meaning in many languages.
- Incomplete removal of language provenance: Post-hoc probes confirm that source-language identity remains recoverable at 43.4%, indicating that the representation is not fully language-invariant.
Related Work & Insights¶
- vs Sapir (1929) & Ohala (1994) Classical Sound Symbolism: While foundational psycholinguistic experiments examined synthetic nonce words or single languages, this work verifies the acoustic frequency code across natural lexicons in 27 typologically distinct languages.
- vs Blasi et al. (2016) & Pimentel et al. (2021) Statistical Cross-Linguistic Studies: Prior computational investigations primarily employed mixed-effects regression over broad lexical databases, leaving residual contact confounds unchecked; this paper establishes a harder generalization test via LDN transfer tiers and adversarial scrubbing.
- vs Elazar & Goldberg (2018) Adversarial Attribute Removal: Consistent with findings in NLP fairness literature, this study confirms that adversarial training effectively impedes the primary model from exploiting protected traits (language identity) while linear probes still retain residual signal.
Rating¶
- Novelty: ⭐⭐⭐⭐⭐ Formulates an innovative genealogical stress test combining distance-based LOLO splits with adversarial representation scrubbing for cognitive linguistics
- Experimental Thoroughness: ⭐⭐⭐⭐⭐ Rigorous methodology featuring native-speaker audio validation, multi-family distance stratification, extensive phoneme ablations, and morphological controls
- Writing Quality: ⭐⭐⭐⭐⭐ Exceptionally clear narrative progression with balanced empirical interpretations and transparent reporting of statistical nuances
- Value: ⭐⭐⭐⭐⭐ Provides rigorous computational evidence for universal sound-meaning iconicity, informing cross-lingual grounding and representation learning