Size Sound Symbolism Survives Genealogy Across 27 Language Families¶
Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/Anika1703/testing_size_symbolism
Area: LLM Agent
Keywords: Sound Symbolism, Phonological Iconicity, Genealogical Control, Adversarial Debiasing, Cross-Linguistic Generalization
TL;DR¶
By assembling a curated cross-linguistic dataset of 810 audio-validated size adjectives across 27 languages spanning 13 families, this work combines distance-stratified leave-one-language-out transfer with adversarial debiasing to demonstrate that phonetic size cues (/i/ for smallness and /o, a/ for largeness) robustly survive the suppression of genealogical and source-language shortcuts.
Background & Motivation¶
Classical structural linguistics has maintained since Ferdinand de Saussure that the relationship between linguistic signifier and signified is fundamentally arbitrary, treating exceptions as marginal curiosities. However, robust cognitive and psychological phenomena like the Bouba-Kiki effect demonstrate that cross-sensory iconic intuitions consistently bridge acoustic properties and semantic concepts across cultures. In the domain of magnitude and spatial deixis, phonetic theories such as Ohala's acoustic frequency code posit that high-frequency acoustic properties like high-front vowels reflect smaller resonating physical cavities, whereas low-frequency back or open vowels naturally signal larger physical volume.
Despite widespread anecdotal and experimental observations, proving the cross-linguistic universality of size sound symbolism faces profound genealogical confounds and areal contact barriers. When unrelated or distantly related languages exhibit similar phoneme-meaning pairings, it is notoriously difficult to disentangle whether such patterns reflect universal cognitive regularities or merely shared deep ancestry, historical lexical borrowings, or idiosyncratic phoneme inventory overlaps. Previous computational linguistic inquiries have attempted to mitigate shared heritage via mixed-effects modeling or genealogical stratification, yet they rarely subject the learned representations to active adversarial pressure to test whether size-predictive signal genuinely survives when language identification shortcuts are made inaccessible.
To overcome these historical limitations, this paper frames cross-family sound symbolism as an adversarial out-of-family generalization stress test: pairwise genealogical distances across 27 languages from 13 families are quantified via normalized Levenshtein distances (LDN) over 40-word Swadesh lists, partitioning candidate training sets into strictly separated similarity tertiles relative to held-out target languages, while gradient reversal adversarial heads penalize the encoding of language-identity shortcuts. Core idea: formalize the cross-linguistic validation of size sound symbolism as a genealogy-aware adversarial stress test, combining distance-stratified leave-one-language-out transfer with gradient reversal debiasing to verify whether phonological size mappings survive once genealogical and source-language cues are actively suppressed.
Method¶
Overall Architecture¶
The experimental methodology establishes an evidence ladder designed to rigorously isolate phonetic size cues from genealogical inheritance. The pipeline begins with 810 size adjectives gathered across 27 languages spanning 13 distinct language families, validated via native-speaker audio recordings and phonemically segmented into IPA tokens using ipatok. Pairs of languages are stratified into target-relative similarity tertiles using Swadesh-based normalized Levenshtein distances. Two complementary modeling paradigms are evaluated: interpretable Bag-of-Phonemes linear and tree-based classifiers for phonological feature attribution, and a frozen 2-layer WikiPron-pretrained Transformer encoder equipped with a gradient reversal layer (GRL) that jointly optimizes size prediction while driving adversarial language/bin discriminators toward chance.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
A["Cross-Linguistic Size Adjectives<br/>27 Languages / 13 Families / 810 Words"] --> B["Audio Validation & IPA Segmentation<br/>XPF Pipeline + Native Audio + ipatok"]
B --> C["Genealogical Stratification Protocol<br/>Swadesh-40 Pairwise LDN Tertile Split"]
C --> D["Interpretable Bag-of-Phonemes Baselines<br/>Logistic Regression Coefficients & Decision Tree Roots"]
C --> E["Frozen Phonetic Transformer Encoder<br/>WikiPron 1.6M Pretrained 2-Layer Backbone"]
E --> F["Dual Adversarial Debiasing Architecture<br/>GRL Gradient Reversal / Bin & 26-Way Source Adversary"]
D --> G["Held-Out Cross-Family Generalization<br/>LOLO Evaluation / Suppression of Language Shortcuts"]
F --> G
Key Designs¶
1. Audio Validation & IPA Segmentation: Mitigating Orthographic Artifacts and Dialectal Noise
Cross-linguistic lexical datasets often suffer from orthographic irregularities and non-standardized phonetic transcriptions in bilingual dictionaries. To ensure consistent phonological representations, the authors translate 50 English seed adjectives into 27 languages using commercial translation engines and authoritative bilingual lexicons, manually filtering loanwords and transliterations to preserve native roots (yielding exactly 15 small and 15 large adjectives per language). The initial IPA transcriptions generated via rule-based XPF mappings are subsequently vetted against raw audio recordings collected from native speakers via language-exchange platforms (Tandem). A trained phonetician reviews each recording with acoustic analysis tools, manually correcting allophonic discrepancies, stripping language-idiosyncratic suprasegmentals (tone, pitch, and lexical stress), and tokenizing sequences into clean IPA segmental phone counts via ipatok.
2. Genealogical Stratification Protocol: Quantifying Lexical Distance without Information Loss To establish an empirical transfer ladder across language families, the framework computes pairwise normalized Levenshtein distances (LDN) using standard 40-word Swadesh lists compiled from the ASJP database. For two words \(\alpha_i\) and \(\beta_j\) sharing meaning \(i\) across languages \(\alpha\) and \(\beta\), the normalized edit distance is defined as: $\(D(\alpha_i, \beta_i) = \frac{D_l(\alpha_i, \beta_i)}{L(\alpha_i, \beta_i)}\)$ The macro distance \(D(\alpha, \beta)\) averages these values across all shared Swadesh concepts. For each held-out target language \(\ell\), the remaining 26 languages are ranked by LDN and split into three equal-sized tertiles: Most Similar, Somewhat Similar, and Least Similar. The authors deliberately reject the further-normalized LDND metric: replicating Petroni & Serva (2010), they demonstrate that the unrelated-concept normalizer \(\Gamma\) retains significant positive correlation with lexical distance (\(r = 0.49\) across all 351 language pairs), meaning LDND inadvertently strips genuine genealogical signal. They further verify that LDN distances exhibit only weak correlation with articulatory PanPhon inventory distances (\(r = 0.255\)), confirming that lexical heritage and phonetic inventory availability represent distinct axes of variation.
3. Interpretable Bag-of-Phonemes Baselines: Decoupling Cross-Family Vowels from Local Consonants To directly examine which individual phonetic segments drive predictive performance, the framework evaluates L2-regularized Logistic Regression and shallow Decision Trees (depth 6) trained on raw phoneme frequency vectors. Feature attribution across all 81 language-tertile models reveals that vowel associations are exceptionally consistent across families: /o/ exhibits positive coefficients for largeness in 100% of models (mean coefficient \(+0.67\)), /a/ trends positively (\(+0.26\)), and /i/ displays negative coefficients indicating smallness in 99% of models (mean \(-0.44\)), robust across 1,000 bootstrap resamples. In Decision Trees, /a/ and /i/ account for 51% of all root splits across 13 families. Conversely, high-magnitude consonantal coefficients (e.g., /H/ at \(+1.32\), /หe/ at \(+1.12\), /nj/ at \(-0.85\)) are confined to restricted genealogical subsets, confirming that vowels anchor universal cross-family mappings while consonants function as lineage-conditioned local cues.
4. Dual Adversarial Debiasing Architecture: Neutralizing Source-Language Shortcuts To ensure that neural models cannot exploit source-language identity or family clustering as classification crutches, the authors pretrain a 2-layer, 128-hidden-unit Transformer on 1.6M IPA words from WikiPron using masked language modeling. With the encoder weights frozen as a general phonotactic feature extractor, a trainable 64-dimensional linear projection \(E\) feeds both a binary size classifier \(C\) and an adversarial discriminator \(A\) via a Gradient Reversal Layer (GRL). The system optimizes the minimax objective: $\(\min_{E, C} \max_{A} \mathcal{L}_{\text{size}}(C(E(x)), y_{\text{size}}) - \lambda \cdot \mathcal{L}_{\text{adv}}(A(E(x)), y_{\text{adv}})\)$ where \(\lambda\) smoothly scales from 0 to 1 over 20 epochs. The framework implements two increasingly rigorous adversarial constraints: a 3-way similarity-bin adversary (chance 33.3%) and a stringent 26-way exact source-language adversary (chance 3.8%). Under this competitive dynamic, the projection layer is forced to discard language-specific phonetic fingerprints while preserving residual size-predictive acoustic cues.
Benchmark & Evaluation Protocol¶
The benchmark protocol evaluates out-of-family generalization across 27 typologically diverse languages using leave-one-language-out (LOLO) cross-validation against a 50.0% chance baseline. Statistical significance across target languages is established via one-tailed one-sample t-tests with Benjamini-Hochberg False Discovery Rate (FDR) corrections over the 25 languages with non-zero cross-run variance. Furthermore, an extreme bidirectional cross-family partition splits the data into 11 non-Indo-European training languages versus 16 Indo-European test targets, and vice versa.
Key Experimental Results¶
Main Results¶
The primary experimental results demonstrate that statistical classifiers consistently detect size symbolism above chance, with neural models retaining statistically significant size discrimination even under aggressive adversarial scrubbing of genealogical and language cues.
| Model / Evaluation Setting | Genealogical Distance Tertile | Accuracy (%) | Precision (%) | Recall (%) | Macro F1 (%) | F1-Small (%) | F1-Large (%) |
|---|---|---|---|---|---|---|---|
| Logistic Regression (Baseline) | Most Similar | 59.3*** | 60.8 | 59.3 | 58.1 | 58.3 | 57.8 |
| Logistic Regression | Somewhat Similar | 57.0** | 58.0 | 57.0 | 55.9 | 58.2 | 53.7 |
| Logistic Regression | Least Similar | 52.2 | 52.4 | 52.2 | 50.0 | 49.5 | 50.5 |
| Decision Tree (Baseline) | Most Similar | 60.2*** | 63.2 | 60.2 | 58.2 | 57.1 | 59.3 |
| Decision Tree | Somewhat Similar | 56.4*** | 55.7 | 56.4 | 50.2 | 53.1 | 47.3 |
| Decision Tree | Least Similar | 55.8** | 56.5 | 55.8 | 50.7 | 47.5 | 54.0 |
| Neural Baseline (\(\lambda = 0\)) | Macro Average (All 27) | 55.4 | - | - | - | - | - |
| Similarity-Bin Adversary | Macro Average (All 27) | 54.4* | Adv Head: 34.0 (Base: 54.4, Chance: 33.3) | - | - | - | - |
| 26-Way Source-Lang Adversary | Macro Average (All 27) | 55.8* | Lang Head: 5.1 (Base: 23.9, Chance: 3.8) | - | - | - | - |
| Non-IE \(\to\) Indo-European Split | Decision Tree (11 Non-IE \(\to\) 16 IE) | 57.5*** | Significantly above chance (\(p = 0.001\)) | - | - | - | - |
| Non-IE \(\to\) Indo-European Split | Logistic Regression (11 Non-IE \(\to\) 16 IE) | 57.9** | Significantly above chance (\(p = 0.010\)) | - | - | - | - |
(Note: Asterisks denote significance over 50.0% chance: *** \(p < 0.001\), ** \(p < 0.01\), * \(p < 0.05\).)
Ablation Study¶
Ablations on interpretable feature representations systematically dissect the specific contribution of phonetic subclasses, feature sparsity, and morphological confounds.
| ID | Ablation Condition | LR Mean (%) | DT Mean (%) | LR by Bin (Most / Some / Least) | DT by Bin (Most / Some / Least) | Phonetic Rationale & Mechanism |
|---|---|---|---|---|---|---|
| 0 | Baseline (All Phonemes) | 56.1 | 57.5 | 59.3 / 57.0 / 52.2 | 60.2 / 56.4 / 55.8 | Complete phonological feature set across all distances |
| 1 | Scrambled Labels | 48.8 | 50.4 | 48.6 / 47.9 / 50.0 | 50.9 / 50.0 / 50.4 | Collapses to chance; verifies absence of dataset artifact leaks |
| 2 | No Vowels | 54.2 | 54.2 | 58.6 / 52.8 / 51.1 | 55.8 / 54.6 / 52.2 | Distant transfer collapses; proves vowels carry cross-family iconicity |
| 3 | High-Frequency Phonemes | 56.5 | 58.5 | 58.5 / 55.6 / 55.4 | 60.2 / 57.8 / 57.4 | Retains 23 common phones; regularizes noise and matches full model |
| 4 | Only Plosives | 54.5 | 56.4 | 56.7 / 53.5 / 53.2 | 58.6 / 55.9 / 54.6 | Acoustic burst/closure provides modest cross-lingual discrimination |
| 5 | Only Nasals | 51.6 | 52.5 | 52.0 / 51.0 / 51.8 | 53.0 / 51.6 / 53.0 | Continuous nasal resonance exhibits negligible magnitude association |
| - | Romance Morphology Control | - | - | /o/ mean \(+0.48\) (75/75 pos) | /i/ mean \(-0.43\) (98.7% neg) | Excludes Spanish/Portuguese; rules out masculine suffix /-o/ bias |
Key Findings¶
- Vowels Drive Universal Generalization while Consonants Mediate Local Family Structure: Removing vowels causes cross-family generalization in the Least Similar bin to collapse toward chance (51.1% for Logistic Regression), whereas restricting features to 23 typologically frequent phonemes maintains 57.4% accuracy in Decision Trees. High-front vowels (/i/) and low/back vowels (/o, a/) anchor the universal acoustic-magnitude mapping, while consonants provide family-specific predictive utility.
- Size Cues Survive Active Genealogical and Language Suppression: Under 3-way bin adversarial training, bin discriminator accuracy is suppressed from 54.4% down to 34.0% (virtually at the 33.3% chance baseline), yet held-out size accuracy remains at 54.4% (an 8.8% relative error reduction over 50%). Under the stricter 26-way source-language adversary, source identification falls to 5.1% (chance 3.8%) while size accuracy holds at 55.8%. Across 25 variable languages, 15 languages spanning 9 distinct families exhibit statistically significant above-chance size accuracy after Benjamini-Hochberg FDR correction.
- Adversarial Training Attenuates Cue Usability Rather than Guaranteeing Total Removal: Post-hoc linear probing on the learned representations recovers source-language identity at 43.4% accuracy. This clarifies that adversarial optimization successfully blocks the primary task network from relying on language-specific shortcuts (cue usability) without completely purging the underlying geometric footprint of language provenance.
Highlights & Insights¶
- Reframing Linguistic Debates as Machine Learning Stress Tests: Instead of relying solely on observational correlational statistics or mixed-effects regression, the paper introduces an active minimax gradient reversal formulation to rigorously test whether linguistic universals withstand adversarial confound scrubbing.
- Audio-Grounded Phonemic Transcription Rigor: Rather than accepting noisy dictionary spellings, the authors enforce an audio-grounded verification loop combining native-speaker recordings from language-exchange platforms with acoustic analysis by trained phoneticians.
- Disentangling Genealogical Distance from Phoneme Inventory Overlap: Empirical correlation checks show that LDN lexical edit distances correlate weakly with PanPhon articulatory inventory distances (\(r = 0.255\)), validating the distinction between shared ancestry and phonemic repertoire constraints.
Limitations & Future Work¶
- Constrained Sample Size and Single Part-of-Speech Focus: Logistical hurdles in recruiting native speakers for low-resource languages limited the dataset to 30 adjectives per language, leaving open the question of whether size symbolism extends to nouns, verbs, or ideophones in these families.
- Discrete Phonemic Abstraction Discards Acoustic Nuance: Representing words as discrete IPA segment sequences abstracts away from fundamental frequency (\(F_0\)), formant bandwidths, duration, and fine-grained allophonic variation across native speakers.
- Absence of Suprasegmental Modeling: Suprasegmental features such as tone, stress, and pitch accent were excluded due to annotation complexity, despite known interactions between pitch register and perceived physical magnitude in tonal language families.
Related Work & Insights¶
- vs Blasi et al. (PNAS 2016): While Blasi et al. analyzed large-scale lexical databases using mixed-effects regression to identify form-meaning associations, this paper introduces an out-of-family machine learning paradigm with active adversarial neutralization of genealogical shortcuts.
- vs Ohala (1994) Acoustic Frequency Code: The empirical findings offer modern cross-linguistic computational validation for Ohala's evolutionary frequency code, showing that the acoustic mapping of high resonant frequencies to smallness and low frequencies to largeness persists across 13 diverse language phyla.
- Implications for Multimodal and Low-Resource LLMs: The demonstrated persistence of sub-lexical phonological iconicity indicates that sound-symbolic inductive biases can serve as a natural grounding mechanism for zero-shot cross-lingual transfer in low-resource speech-text foundation models.
Rating¶
- Novelty: โญโญโญโญโ Introduces an innovative adversarial stress-testing paradigm to evaluate linguistic form-meaning universals.
- Experimental Thoroughness: โญโญโญโญโญ Audio validation across 27 languages, 13 families, LDN tertile ranking, feature ablations, and dual adversarial probes.
- Writing Quality: โญโญโญโญโญ Clear theoretical framing, rigorous statistical reporting, and nuanced discussion of adversarial representation dynamics.
- Value: โญโญโญโญโ Provides an empirical and methodological bridge connecting cognitive linguistics, phonetics, and robust representation learning.