There is a quiet war raging inside every dictionary, a struggle for survival that plays out over decades and centuries, hidden in the vast printed record of human expression. When two words carry essentially the same meaning, they rarely settle into peaceful coexistence. Instead, they begin a slow-motion duel, one that can end with a clear victor dominating the language while the loser retreats into obsolescence, poetry, or regional dialect. This phenomenon, known as synonym competition, has long fascinated linguists, but until recently it resisted systematic prediction. Now, a team of researchers has turned to the immense digital archive of Chinese texts spanning more than a century, coupling it with advanced machine learning, to decode the hidden rules that determine which word wins. In a study published in Humanities and Social Sciences Communications, Shuiyuan Wang, Yuesheng Wang, and Hongchao Zhang present a rigorous computational framework that quantifies exactly how frequency patterns and subtle linguistic properties interact to crown a lexical champion. Their findings reveal a frequency-dominant mechanism, where the slow accumulation of usage advantage, captured by statistical moments of a word’s historical trajectory, acts as the primary engine of victory, while intrinsic features like stroke count or character radical play a surprisingly minor, though occasionally crucial, supporting role.
The researchers built their analysis on two monumental resources: the Google Books Ngram Corpus for Chinese, spanning the years 1891 to 2009, and the Chinese Open Wordnet, which provides carefully curated sets of synonyms known as synsets. From the Wordnet, they extracted hundreds of synonym pairs that have competed over the modern era, words like 立刻 and 马上 (both meaning “immediately”) or 高兴 and 快乐 (both meaning “happy”). For each word, they constructed a rich profile consisting of three categories of statistical features and four categories of linguistic features. The statistical features included relative growth, which captures the short-term rate of change in usage frequency; linear extrapolation, which projects a word’s trajectory forward based on its recent trend; and a suite of central moments—mean, variance, skewness, and kurtosis—computed over the entire observed frequency distribution. Central moments are particularly revealing because they distill not just the average popularity of a word, but the shape of its historical ups and downs: variance measures how wildly the word’s usage has fluctuated, skewness indicates whether its peak happened early or late, and kurtosis reveals whether it experienced sudden bursts of fame or plodded along steadily. Together, these statistical features encode different aspects of the cumulative advantage hypothesis, the idea that a word that gets a head start or maintains a more consistent presence will, over time, entrench itself in the minds of speakers and writers, becoming the default choice.
On the linguistic side, the team computed features that have long been hypothesized to affect word processing and memory in Chinese. Stroke count captures the visual and motor complexity of a character; a word written with fewer strokes should, in theory, be easier to write and recognize, giving it a cognitive edge. Radical refers to the semantic classifier component of a character, like the “water” radical in words related to liquids, and the researchers measured whether having a more semantically transparent radical conferred an advantage by making the word easier to learn. Word age tracks the historical pedigree of a lexical item, since older, more established words might have had more time to sink deep roots into the language. Finally, categorial variation measured how flexible a word is in terms of part-of-speech usage; a synonym that can also serve as a noun, verb, or adjective might have a broader range of applications, increasing its exposure and thus its competitive strength. Each of these features was painstakingly quantified, and the entire dataset was then fed into an XGBoost classifier, a gradient-boosted decision tree algorithm renowned for its ability to capture non-linear interactions and resist overfitting. The machine’s task was simple to state but immensely complex to solve: given the historical frequency profiles and linguistic properties of two competing synonyms, predict which one will eventually become the dominant form.
The researchers conducted a comprehensive suite of experiments to dissect the contribution of each feature type. First, they trained models using all seven features together, then models using only statistical features, only linguistic features, and single-feature models, as well as ablation experiments where one feature was systematically removed. The results were striking in their clarity. Statistical features overwhelmingly outperformed linguistic features across all prediction scenarios. When used alone, the best statistical feature, central moments, achieved a prediction accuracy that dwarfed any single linguistic feature. In fact, even the weakest statistical feature, relative growth, still beat the strongest linguistic feature, radical, by a wide margin. This hierarchy held true at different time windows, from the early 20th century to the early 21st century, though the relative importance of individual statistical features shifted as the competition matured. In the early stages of a synonym duel, relative growth proved moderately useful, capturing the initial burst of momentum that a challenger might gain. But as time went on and the usage trajectories stabilized, its predictive power waned. Linear extrapolation, on the other hand, became more influential in the mid-to-late stages, when the long-term trend had solidified and projecting it forward could reliably indicate the likely winner. Yet it was central moments, robust and unshakable, that consistently outperformed all other features, no matter the stage. This suggests that the outcome is largely determined not by a single moment of explosive growth, but by the entire historical shape of the frequency curve—whether it has a high mean, low variance, and a positive skew indicating a late-stage peak that refuses to fade.
The linguistic features, while not the main drivers of prediction, revealed fascinating subtleties in their auxiliary role. Radical information turned in the best performance among the linguistic features when each was used in isolation. This finding aligns with cognitive research showing that the semantic radical is a powerful cue during Chinese character recognition, and a word built around a common, meaningful radical might enjoy a processing fluency advantage that incrementally nudges speakers toward it. Stroke count, surprisingly, did not emerge as a strong predictor. One might expect that a simpler character would be preferred, but in the context of modern written Chinese, where typing and digital input have diminished the cost of complex strokes, visual simplicity may have lost its evolutionary edge. Word age also had a limited effect, suggesting that being ancient is no guarantee of victory; a younger coinage can still triumph if it catches the frequency wave. Categorial variation showed minimal independent predictive strength, perhaps because most synonyms in the dataset share similar grammatical profiles, or because flexibility matters less than sheer repetition. Crucially, when all four linguistic features were combined together and added to the statistical feature set, the model’s performance nudged upward, exceeding the accuracy of the best single statistical feature. This indicates a synergistic effect: linguistic factors, while insufficient on their own, provide a slight but significant boost when they work in concert, helping the classifier disambiguate cases where the statistical trajectories are nearly identical. The overall picture is one of a frequency-dominant mechanism with linguistic characteristics acting as subtle tiebreakers, a view that reframes the nature of lexical evolution as a primarily self-reinforcing process governed by cumulative advantage, secondarily shaped by the intrinsic cognitive cost of the words themselves.
To truly appreciate what the XGBoost model was learning, it is helpful to visualize the data landscape it navigated. Imagine a cloud of points in a high-dimensional space, each point a synonym pair at a particular moment in history. The statistical features capture the temporal texture: a word like “电脑” (computer) might show an exponential rise beginning in the late 20th century, its variance and kurtosis spiking dramatically, while an older term like “计算机” (calculating machine) might exhibit a gentler, earlier hump. The central moments encode these differences succinctly. The model learns that a competitor with a high mean frequency, low variance (meaning consistent usage), and a skewness indicating recent ascendancy is almost unstoppable. Meanwhile, linguistic features add small but consistent offsets: among words with nearly identical frequency curves, the one with a more transparent radical might get a fractional boost in the predicted probability of winning, reflecting the cumulative effect of millions of small cognitive favors over decades. The ablation experiments confirmed this interpretation by showing that removing central moments led to catastrophic drops in accuracy, whereas removing any single linguistic feature caused only minor, barely perceptible dips. The researchers also examined specific case studies where the model succeeded or failed, finding that errors often involved synonyms that were near-perfect mirrors in frequency, or cases where a sudden cultural shift—like a government language reform or a technological invention—abruptly altered the competitive landscape in a way not captured by the historical trajectory alone.
The study’s reliance on the Google Books Ngram Corpus deserves special attention, as it is both a strength and a limitation. This corpus, derived from millions of books scanned by Google, offers an unparalleled window into the history of written language. For Chinese, it provides yearly frequency counts for words and n-grams stretching back to the late Qing dynasty, allowing researchers to trace the rise of modern Standard Mandarin, the influence of the May Fourth Movement’s vernacular revolution, the vocabulary shifts of the Mao era, and the explosion of new terms during the Reform and Opening Up period. However, the corpus is not a perfect reflection of spoken language or of the full diversity of registers. It over-represents formal, published prose and under-represents colloquial speech, personal letters, and online communication. Synonyms that compete fiercely in daily conversation might appear differently in books, where editors and stylistic norms act as gatekeepers. The researchers acknowledge this, noting that their model predicts the outcome of competition specifically within the written, book-based ecosystem. Extending the analysis to other corpora—social media, newspapers, television transcripts—would test the generality of the frequency-dominant mechanism and might reveal that in more informal domains, linguistic features like stroke count, freed from the conservatism of print, play a larger role.
Nevertheless, the technical achievement of this work is undeniable and points toward a new era in historical linguistics. XGBoost, with its ensemble of decision trees trained on gradient-boosted residuals, excels at modeling the kind of non-linear, threshold-based dynamics that characterize real-world language change. A word might limp along at low frequency for decades, then cross an invisible threshold of familiarity after which it accelerates via a network effect, as speakers hear it more often and start using it themselves. Decision trees naturally capture such step functions, and XGBoost’s boosting process sequentially corrects for the errors of previous trees, homing in on the precise combinations of central moments and growth metrics that signal a coming takeover. The model’s interpretability tools, such as SHAP (SHapley Additive exPlanations) values, allowed the researchers to peek inside the black box and confirm that the most important features were indeed the statistical descriptors of long-term frequency patterns. This combination of predictive power and explanatory transparency is rare in machine learning applications to the humanities, making the paper a methodological landmark.
The implications for our understanding of language as a complex adaptive system are profound. Language is a classic case of self-organization, where the global regularities we observe—grammar, vocabulary, pronunciation norms—emerge from the local interactions of millions of speakers without any central planner. Synonym competition is a microcosm of this process. Each time a speaker chooses between “立即” and “马上,” they cast a tiny vote, and the accumulation of these votes over time shapes the probability that future speakers will make the same choice, because we are all influenced by the frequencies we perceive. This positive feedback loop, known as the frequency effect or the Matthew effect, can amplify small initial advantages into overwhelming dominance. The study’s finding that central moments are the best predictors aligns beautifully with this theory, because the moments capture the entire history of the amplification process. In contrast, the limited role of linguistic features suggests that purely functional explanations of language change—the idea that words evolve to be easier to say, hear, or process—while not false, operate as a weak background force, easily overpowered by the brute momentum of popularity.
Yet the study also demonstrates that these weak background forces are not negligible. When the team combined linguistic features with statistical ones, the model outperformed any pure statistical model, albeit by a small margin. This is a crucial insight: in the tangled web of causation that drives language change, the primary driver is random drift and cumulative advantage, but natural selection in the form of cognitive biases still sifts the variants, nudging the system toward forms that are slightly more learnable, more memorable, or more easily integrated into the existing grammatical network. The radical feature’s relative success among linguistic predictors hints at the deep role of the Chinese writing system’s structure in shaping lexical evolution. Because radicals provide a semantic hook, words that leverage this hook might have a tiny but consistent edge in the competition for mental real estate, an edge that only becomes visible when the noise of frequency is partially controlled for. The stroke count’s failure, on the other hand, may reflect a genuine shift in the cost landscape of Chinese writing. In an era of keyboards and touchscreens, the motor act of writing has been replaced by selection from a list of homophones, neutralizing what was once a potent selective pressure.
The researchers are careful not to claim that their model captures all the forces at play. Social factors, prestige dynamics, language policy, and sheer chance all influence which words rise and fall. The standardization of Mandarin, the influence of media, the spread of internet slang—these macro-level forces can suddenly disrupt the smooth working of cumulative advantage, as when a regional term gets catapulted into national consciousness by a viral video. Future work, they suggest, could incorporate such event data, perhaps as external shock variables added to the temporal features. Additionally, the current study focused on a set of synonyms drawn from the Chinese Open Wordnet, which, while expertly curated, is limited in size. Expanding to the full set of tens of thousands of synonym pairs in the language would provide a more comprehensive picture, and might reveal that the relative importance of linguistic features is higher in certain semantic domains, such as concrete nouns versus abstract verbs. The team also expressed interest in pushing the analysis back in time, using historical databases of classical Chinese to examine whether the frequency-dominant mechanism is a universal constant or a product of the mass-print era. In pre-modern contexts, where literacy was restricted and texts were hand-copied, the dynamics of cumulative advantage might have operated differently, with linguistic and social prestige factors playing a larger role in the absence of the homogenizing force of print.
The methodology itself is a gift to the broader field of digital humanities. By demonstrating that gradient-boosted trees can be effectively trained on diachronic corpus data to predict lexical outcomes, Wang and colleagues have provided a template that can be adapted to any language with a sufficiently large historical text corpus. English, with its enormous Google Ngram dataset and resources like WordNet, would be a natural next target. Does the same frequency-driven logic govern the competition between “gotten” and “got,” or “whom” and “who”? Are linguistic features like word length and regularity of inflection stronger or weaker in languages with different structural properties? The XGBoost framework, with its ability to handle mixed data types and missing values, is ideally suited for cross-linguistic comparisons. Moreover, the careful distinction between statistical and linguistic features could be productively applied to other domains of cultural evolution, such as the competition between baby names, fashion trends, or scientific terminology. In each case, the question is parallel: to what extent does success breed success through mere exposure, and to what extent does intrinsic quality matter? The answer, as this study hints, might often be the same: exposure is king, but quality can be a queenmaker.
When we look up a word in the dictionary, we see a static snapshot, a definition frozen in time. But the reality, as this work makes viscerally clear, is that every word is a historical entity, carrying within its usage patterns the echoes of a long struggle for existence. The winning synonym, the one we use without thinking today, is not necessarily the most elegant, the easiest to write, or the most ancient. It is simply the one that, for a complex tangle of reasons both measurable and stochastic, managed to accumulate enough early usage to tip the frequency feedback loop in its favor. The central moments of its frequency curve—its mean, variance, skewness, and kurtosis—are the mathematical fossils of that struggle, and a machine learning model like XGBoost can read those fossils with startling accuracy. This does not mean language change is deterministic; there is plenty of room for contingency, for the path not taken. But it does mean that once a trajectory is established, it is exceedingly difficult to reverse without a major external shock. The words we speak are both monuments to the past and weapons in an ongoing, silent war, and we are all, through our every utterance and keystroke, enlisted as footsoldiers in that war, deciding the future of the language one word at a time.
The study’s broader narrative challenges the romantic notion of language as a purely creative, free human endeavor and replaces it with a view that is both humbling and exhilarating: language is a vast, self-organizing statistical machine, and we are the computing elements that run its algorithms. The XGBoost model is, in a sense, a meta-machine that learns the rules of this organic computer by watching its output over a century. The fact that simple statistical aggregations of frequency data can so robustly predict the fate of words speaks to the deeply collective, unconscious nature of language evolution. No single author, no matter how influential, can dictate which synonym wins. Even a widely admired writer can only nudge the frequencies a little; the ultimate decision is made by the aggregate behavior of the reading and writing public over generations. In that sense, every time we choose a word, we are participating in a form of distributed problem-solving, collectively deciding which verbal tools are fittest for our communicative needs. The XGBoost model, having crunched the numbers, simply reveals the outcome of that distributed process earlier than our conscious awareness can.
Looking ahead, the integration of neural network models, such as transformers that capture semantic context, might further refine the prediction. The current study treats each synonym pair in isolation, but words are embedded in a rich network of associations and collocations. A word might win not just because of its own frequency curve, but because it becomes entrenched in common phrases, idioms, and constructions. Future models could incorporate contextual embeddings derived from BERT or similar architectures, adding a layer of semantic and syntactic context to the predictions. There is also the tantalizing possibility of applying reinforcement learning simulations: if we can model the competition as a game where each use of a word reinforces that word’s probability of future use, we could run counterfactual experiments, rewinding the tape of history and seeing whether a small perturbation could have led to a different victor. The discovery that central moments are the key predictors already hints that the system has a kind of momentum, an inertia, making it relatively stable against small perturbations unless they occur at critical junctures where the frequency trajectories are still malleable.
For language learners, lexicographers, and AI developers, the implications are practical. Knowing the statistical profile of a winning word can inform which vocabulary items to prioritize in language teaching materials, or how to design natural language processing systems that can gracefully handle lexical variation over time. The Chinese language, with its unique writing system and its dramatic social transformations in the 20th century, offers a particularly rich laboratory for these studies. The fact that stroke complexity, despite being a perennial topic in debates over character simplification, did not strongly predict synonym survival suggests that the massive character simplification reforms undertaken in the mid-20th century, while successful in reducing writing burden, may not have significantly altered the competitive dynamics among synonyms within the simplified system. The winners and losers in that reform were determined by political fiat, not by organic competition, but among the survivors, the organic competition continues, and it runs on frequency, not just form.
In summary, Wang, Wang, and Zhang have given us a powerful demonstration of how diachronic language data, when subjected to the discerning eye of machine learning, can yield quantitative laws of linguistic evolution. The dynamic competition of Chinese synonyms over a century is revealed to be a frequency-driven process, with long-term cumulative advantage encoded in statistical moments serving as the primary determinant of victory, while linguistic features like radical transparency provide a subtle modulatory influence. This work not only illuminates a specific corner of Chinese linguistics but also offers a generalizable framework for studying lexical change in any language, marrying the rich theoretical traditions of historical linguistics with the predictive muscle of modern computational tools. The words we use are the living fossils of this relentless competition, and we now have the tools to read their evolutionary history with unprecedented clarity.
Subject of Research: The dynamic competition and prediction of winning synonyms in Chinese language using diachronic corpus data and machine learning.
Article Title: The dynamic competition of Chinese synonyms over a century
Article References: Wang, S., Wang, Y. & Zhang, H. The dynamic competition of Chinese synonyms over a century. Humanit Soc Sci Commun 13, 956 (2026). https://doi.org/10.1057/s41599-026-07135-w
Image Credits: AI Generated
DOI: https://doi.org/10.1057/s41599-026-07135-w
Keywords: Chinese synonym competition, language change, XGBoost, central moments, frequency patterns, diachronic corpus analysis, Google Books Ngram Corpus, Chinese Open Wordnet, linguistic features, cumulative advantage
