Language is in a constant state of flux, yet its evolution is remarkably predictable. If you track the movement of words through semantic space over centuries, you find that vocabulary does not drift at random. Instead, a powerful statistical law governs language change: the more frequently a word is used, the more resistant it is to historical drift. This phenomenon reveals that usage frequency acts as a stabilizing anchor for meaning.
The Cognitive Inertia of High-Frequency Words
To understand why frequent words resist change, we must look at how language is stored and processed in the human brain. Janet B. Pierrehumbert provides a foundational intuition through the lens of episodic memory models in Phonological Representation: Beyond Abstract Versus Episodic. Every time we encounter or use a word, our brain stores a rich contextual and phonetic trace. High-frequency words possess a massive, dense "center of mass" in memory because they are backed by millions of these stored episodes. When a speaker encounters an innovative or non-standard usage, it represents only a tiny fraction of the word's total memory footprint, leaving the core representation virtually unchanged.
Conversely, low-frequency words have sparse representations. A handful of novel exposures can easily shift their entire distributional center. This cognitive inertia explains macro-level historical patterns. Jure Leskovec et al. used diachronic word embeddings in Diachronic Word Embeddings Reveal Statistical Laws of Semantic Change to show that semantic change follows an inverse power-law relationship with usage frequency—a principle known as the Law of Conformity. As Ryo Nagata et al. note in Report on the Theme Session “Linguistics and Natural Language Processing in the Era of Deep Learning”, these statistical laws remain observable across millennium-scale historical data when mapped into vector spaces.
The stabilizing force of frequency is so pronounced that it even applies across synchronous social divides. Hosein Azarbonyad et al. demonstrated in Words are Malleable: Computing Semantic Shifts in Political and Media Discourse that high-frequency words remain semantically stable across conflicting political and ideological viewpoints, whereas lower-frequency words easily warp to fit specific group biases.
The Paradox of Polysemy
If high-frequency words are so stable, why do they also tend to have the most definitions in the dictionary? This apparent paradox is resolved by distinguishing between semantic drift (the shifting of a word's core meaning over time) and semantic diversification (the simultaneous coexistence of multiple senses).
As a word's frequency scales, it naturally appears in a wider variety of contexts. In RuDSI: Graph-based Word Sense Induction Dataset for Russian, Anna Aksenova et al. confirm that the number of corpus-derived senses strongly correlates with frequency. However, this polysemy does not mean the word's core is unstable. Rather, its cognitive entrenchment prevents these diverse senses from causing the word to drift away entirely.
Andreas Baumann et al. explain in Seeing through the mess: evolutionary dynamics of lexical polysemy that because high-frequency words are deeply rooted in cognitive processing, they resist structural replacement. True semantic innovation—where a word completely abandons its old meaning for a new one—is disproportionately concentrated in lower-frequency vocabulary. Lower-frequency words possess the flexibility to acquire and maintain entirely new meanings because they are less constrained by rigid, pre-existing associations.
Structural and Categorical Attractors
Frequency does not operate in isolation; it interacts with a word's structural position within the language ecosystem. One key mechanism is semantic prototypicality. Haim Dubossarsky et al. discovered in A bottom up approach to category mapping and meaning change that words sitting at the mathematical center of their semantic categories act as "attractors." The closer a word's usage is to the core definition of its category, the more resistant it is to being pulled toward new interpretations over time.
Frequency also anchors grammatical and morphological structures. For instance, Shijie Wu validated the classic linguistic hypothesis that morphological irregularity correlates with high frequency in Morphological Irregularity Correlates with Frequency. High-frequency irregular forms (like is/was or go/went) enjoy a cognitive storage advantage. As Timothy J. O’Donnell argues in The evolutionary causes of irregularity: Evidence for an irregularization bias in morphological learning, these common forms are memorized so reliably that they resist the natural pressure to regularize, effectively acting as anchors that preserve complex linguistic paradigms over centuries.
Social Niches and the Illusion of Volatility
Beyond cognitive and structural factors, a word's stability is heavily dictated by its social distribution across a population. Eduardo G. Altmann et al. showed in Niche as a Determinant of Word Fate in Online Groups that on short timescales, a word's "niche"—the number of unique users and topics it spans—is a more robust predictor of long-term survival than raw frequency alone. Words that fail to spread beyond a narrow social clump are highly vulnerable to sudden usage crashes.
This is mirrored in modern digital language. In Language in Our Time: An Empirical Analysis of Hashtags, Yang Zhang et al. found that community-driven labels shared broadly across a diverse user base maintain highly stable meanings, whereas niche labels drift rapidly.
Understanding this frequency-stability relationship is vital for computational linguistics because failing to control for frequency can create illusions of volatility. For example, subjective language like evaluative adjectives (terrific, bad) often feels like it changes at a breakneck pace. However, Julia Rodina et al. demonstrated in Measuring Diachronic Evolution of Evaluative Adjectives with Word Embeddings: the Case for English, Norwegian, and Russian that once you statistically control for usage frequency, evaluative adjectives are just as stable as objective ones. The perceived volatility of subjective language is often an artifact of high usage volume masking its underlying structural persistence.
Key Foundational Work
Characterising Measures of Lexical Distributional Similarity (2004) — contributed 1 concept elements to this direction.
Models of Lexical Decay (1965) — contributed 1 concept elements to this direction.