For over a century, modern linguistics has operated under a foundational axiom popularized by Ferdinand de Saussure: the arbitrariness of the sign. According to this dogma, the relationship between a word’s form (its sound or spelling) and its actual meaning is purely conventional. There is no intrinsic reason why a furry, four-legged barking animal is called a dog in English, a chien in French, or a perro in Spanish. Arbitrariness is highly efficient; it allows human language to map an infinite space of meanings onto a finite set of sounds without being constrained by physical resemblance.
Yet, if you ask a human whether a jagged, spiky shape should be named Kiki or Bouba, nearly everyone chooses Kiki. This cross-modal mapping—where sharp visual angles intuitively align with abrupt, high-frequency speech sounds—suggests that language is not a completely blank slate of random conventions.
Recent computational and statistical research is turning this long-standing debate into a quantitative science. By analyzing thousands of languages simultaneously through information theory and machine learning, researchers have shown that while language is overwhelmingly arbitrary, it contains a small, persistent, and highly structured signal of form-meaning systematicity.
The Information-Theoretic Footprint
To prove that the sound of a word carries information about its meaning, we must be able to measure it mathematically. How much does knowing what a word means reduce our uncertainty about how it sounds?
Pimentel et al. (2019) addressed this by training neural language models on 106 diverse languages. They found that across almost every language, a word's semantic context provides a small but statistically significant amount of information about its phonetic form. While this reduction in uncertainty is typically less than one percent—confirming that arbitrariness remains the primary structural characteristic of the lexicon—the signal is undeniably real.
[Semantic Meaning] ---> (Statistical Bias) ---> [Phonetic/Orthographic Form]
|
~1% Uncertainty Reduction
This global signal, however, is not distributed evenly. In a follow-up study spanning over 5,000 languages, Pimentel et al. (2021) discovered that roughly 25% of basic vocabulary concepts exhibit strong, non-arbitrary sound-meaning associations that persist globally across unrelated language families. These non-arbitrary associations are heavily concentrated in specific semantic categories, such as body parts (e.g., tongue) and environmental terms. This suggests that universal biological constraints or direct physical interactions with the world drive convergent phonetic evolution.
Pockets of Systematicity
Rather than a uniform law across the entire dictionary, systematicity manifests as dense local "pockets" within a language.
Consider the English cluster of words starting with gl- (gleam, glimmer, glow, glisten, glint), which all relate to light or vision. These sub-morphemic units carry a distinct semantic flavor despite not being formal prefixes. To detect these patterns automatically, Gutiérrez et al. (2016) introduced an algorithm called String-Metric Learning for Kernel Regression (SMLKR). By learning which string edits matter most for predicting semantic similarity, their model demonstrated that these local pockets contribute to a much stronger global structure in English than traditional theory admits.
This structural robustness is also deeply integrated across time. Meng et al. (2025) investigated diachronic rhyme books in Chinese and found that phonologically related characters have exhibited significant semantic congruence across centuries. This diachronic persistence proves that sound symbolism is not a fleeting linguistic accident or a marginal remnant of primitive speech; it is a continuously evolving, stable feature of language.
Sensory-Motor and Cross-Modal Biases
Why do these patterns emerge? The answer lies in how the human brain maps sensory experiences across different modalities.
Our brains naturally link acoustic properties to physical dimensions like size, shape, and gender. For instance, the authors of Sound-Symbolic Semantics of Written Japanese Vowels in a Paper-Based Survey Study demonstrated that back vowels (such as /a/, /u/, /o/) are consistently perceived by humans as physically larger and thicker than front vowels (like /i/, /e/).
This sound symbolism is so pervasive that it can be used to predict human perception in entirely novel contexts:
- Brand Naming: Moorthy et al. (2018) built a machine-learning model trained on human first names to show that phonetic features alone can reliably predict the perceived "gender" of brand names and neologisms.
- Speech Processing: Babel et al. (2012) demonstrated that these subtle sound-meaning connections interact directly with higher-level cognitive tasks, such as identifying a speaker's gender based on vowel acoustics.
- Visual Languages: This grounding is even more transparent in sign languages. Bradley et al. (2022) found that individuals with absolutely no knowledge of American Sign Language (ASL) could accurately intuit whether a verb was transitive (involving an object) or intransitive simply by observing the visual form of the sign.
Remarkably, these human cognitive biases are mirrored inside modern artificial intelligence. The authors of Kiki or Bouba? Sound Symbolism in Vision-and-Language Models discovered that models like CLIP and Stable Diffusion exhibit the kiki-bouba effect. Even though these models have never "heard" a physical sound, they associate specific text strings and phonetic patterns with visual attributes like sharpness or roundness. This occurs because human cross-modal biases are deeply embedded within the text descriptions used to train these models.
Visual Shape (Sharp/Angular) <---> Textual Description <---> Phonetic Pattern ("Kiki")
^
|
Captured by Multimodal AI
Cultural and Structural Evolution
When form-meaning systematicity takes root, it actively shapes how languages and cultures evolve over time.
On a structural level, the physical ease of pronouncing a word can dictate its semantic capacity. Yamshchikov et al. (2020) applied graph theory and curvature metrics to synonym networks, proving that phonetically simple words—those that are easier to produce and robust to transmission errors—tend to accumulate significantly more meanings (polysemy) over time. Furthermore, in simulated language evolution experiments, Carr et al. (2017) showed that when agents are forced to communicate about a continuous world, shape-based sound symbolism consistently emerges as a fundamental tool to bootstrap structure from scratch.
On a cultural level, phonetic properties can alter the social value of words and objects through homophonic associations:
- Grammatical Gender: Williams et al. (2019) used Canonical Correlation Analysis to show that grammatical gender assignments for inanimate objects across 18 languages are not random historical accidents; instead, they correlate significantly with the core semantic properties of the objects themselves.
- Cultural Numerology: Tan (2016) documented how linguistic homophones transform abstract digits into deeply lucky or unlucky symbols (e.g., numbers that sound identical to the words for "death" or "prosperity" in Eastern and Western cultures).
- Artifact Sustainability: Li et al. (2025) demonstrated that when the name of a traditional craft phonetically aligns with auspicious phrases in a local dialect, it embeds the artifact into the social fabric, driving its long-term cultural preservation.
How Computational Models Capture the "Vibe"
If human language relies on this subtle interplay between arbitrary efficiency and non-arbitrary systematicity, can computational models leverage it when encountering completely unfamiliar words?
Waal et al. (2025) tested Large Language Models (LLMs) using novel pseudowords (like puyfred) that have no established dictionary definitions. They discovered that LLMs assign consistent emotional valences to these nonsense strings based entirely on their sub-lexical form, and that these "vibes" interact with sentence contexts in ways that closely mirror human psychological responses.
When we expand our computational view to include multimodal data, the picture becomes even clearer. By combining textual data with visual features extracted from images, Soeteman et al. (2020) demonstrated that iconicity—where a word's form is directly motivated by its perceptual characteristics—becomes highly measurable.
Ultimately, this research direction reveals that language is a hybrid system. While arbitrariness gives language its infinite flexibility, non-arbitrary systematicity provides the essential cognitive scaffolding that grounds abstract symbols in our physical, sensory, and emotional reality.
Key Foundational Work
Questioning Arbitrariness in Language: a Data-Driven Study of Conventional Iconicity (2016) — contributed 3 concept elements to this direction.
Gendered sound symbolism and masking effects in speech processing (2012) — contributed 2 concept elements to this direction.
Finding Non-Arbitrary Form-Meaning Systematicity Using String-Metric Learning for Kernel Regression (2016) — contributed 1 concept elements to this direction.