Slovotvir: Grassroots reform of the Ukrainian language reveals the role of drift and selection in language change Authors: Alexey Koshevoy, Oleg Sobchuk, Olivier Morin, Mohammad Atari Published: 2026 ABSTRACT AbstractLanguage change involves a complex interplay of selection and drift (or unbiased copying), shaping the popularity of words over time. This pre-registered study examines lexical evolution in the Ukrainian language using data from Slovotvir, a crowdsourcing platform where over 4,000 users proposed and ranked translations of foreign-origin words over 9 years. We rely on an agent-based model combined with novel generative inference methods to investigate the roles of frequency-dependent selection, selection for brevity, and random drift in shaping word popularity. Our results indicate that Slovotvir users exhibit a preference for shorter words, supporting Zipf’s principle of least effort. However, the popularity of translations, which we approximate using likes on the platform, appears largely frequency-independent, suggesting that users do not disproportionately favor already popular words. However, this conclusion might be limited by the fact that our data consist of low-frequency neologisms that are still evolving in this population of speakers. These findings contribute to a broader understanding of language change by providing empirical evidence that selection for brevity, rather than frequency-dependent selection, drives lexical evolution in this grassroots language reform initiative, which reflects and shaped real-world language change.Keywords: drift, language change, Approximate Bayesian Computation, word length, Ukrainian Language change, like other cultural processes, is shaped by both random chance (drift) and various selective pressures. We investigated these mechanisms using a decade of data from Slovotvir, a large-scale, grassroots Ukrainian online initiative where users proposed and ranked thousands of alternatives for foreign-origin words. This unique dataset allowed us to test linguistic theories on an unprecedented scale. Using a combination of agent-based modeling and generative inference, we show that the success of a new word is primarily driven by a subtle bias for brevity, supporting the idea that communicative efficiency is a key selective pressure. This work highlights how decentralized online platforms can serve as controlled, natural experiments for studying cultural evolution. Introduction Why do language users prefer using some words over their synonyms? For instance, why is the “water-serving vessel” more often called jug rather than carafe, pitcher, or ewer in British English? Is the word jug better than its counterparts in some ways, or is this choice random? Linguists were interested in this question since the late 19th century. Max Müller, commenting on Schleicher’s review of Darwin’s work (1), argued that what appears to be accidental shifts in word usage can often be traced to “intelligible and generally valid reasons” (2). For example, some words become more common due to selective advantages related to their phonology, frequency, or other features. This perspective aligns with modern accounts of word popularity dynamics (3, 4) and broader explanations of linguistic behavior (5, 6). Yet, there is also evidence that random processes, such as random drift,a also play a role in the dynamics of word popularity (7–9). To empirically test the role of selection and drift in language change, we analyze data from Slovotvir (“word creation” in Ukrainian, see https://slovotvir.org.ua/). Slovotvir is a Ukrainian crowdsourcing platform where users rank and propose synonyms (translations) for Ukrainian words borrowed from other languages.b This platform was launched in 2014 in the wake of the Russian invasion of Crimea as a grassroots effort to reclaim the Ukrainian language, which has faced multiple cultural threats, including extensive Russification campaigns since the establishment of the Soviet Union. Since its founding, the website was met with unprecedented success for such an initiative. Language reforms are usually promoted in a top-down manner by state institutions, such as the Académie Française, often with mixed success. The Slovotvir initiative is exceptional in its success and popular support, independent of direct state support. A good example from Slovotvir is the word безхатько (bezkhat’ko, literal translation “without-a-house”), which replaced the previously used Russian lexicalized acronym бомж (bomž), both meaning “homeless person.” It is worth noting that Slovotvir is not designed specifically to propose alternatives to Russian words. In fact, some of the most translated words are borrowings from English and other languages: ксерокс (kseroks—digital scanner, a semantic extension linked to the brand name “Xerox”), куки (kuk’i—web cookies), селфi (selfi—selfie), and баг (bag—omputed bug). As we will show, Slovotvir reflected substantial changes in the Ukrainian language, and many of the replacements promoted by the website gained in prevalence since the website was established. Thousands of Ukrainians participated in this free platform, proposing over 7,000 words for translation, suggesting more than 20,000 translations, and ranking others’ suggestions for over 9 years. Slovotvir offers a rare opportunity to study word popularity dynamics at the populational level, with high-resolution temporal data. In this study, we tested whether two types of selective pressures shape the popularity dynamics of translations in Slovotvir. The first is frequency-dependent selection, which has been suggested to play an important role in determining the popularity of words (4). The second is selection for brevity (6), whereby short words enjoy selective advantage since they accommodate the needs of the speakers to reduce articulatory effort. These two forms of selection will be contrasted with a null hypothesis that assumes changes in popularity are driven by drift rather than selective pressures. Drift-based explanations One of the explanations for why some cultural items (like words) become more popular than others is unbiased copying combined with random drift (10–12). Unbiased copying is a model of cultural transmission where agents are copying items based on their probability of being exposed to them, regardless of any intrinsic advantages that they might have. Random drift is the process by which the frequency of items changes due to sampling noise that accumulates when unbiased copying occurs in finite populations. In other words, unbiased copying can be viewed as a frequency-preserving copying from generation to generation, and drift as the sampling noise that results in the loss of some items. In finite populations without change (addition of new items in the population), this sampling process will result in one item dominating the rest of the population over long-enough time periods. In infinite populations item frequencies will be constant due to the absence of sampling noise, indicating that, in principle, drift can be a property of any finite population. This seemingly trivial process of sampling noise accumulation can predict nontrivial properties like power-law frequency distributions of cultural items (13). Drift with realistic rates of change will produce a frequency distribution of items where a handful of them will dominate the population. The ability of drift to yield such distributions has been featured in the study of the distribution of baby names (14), as well as the distribution of pottery designs in archaeological excavations (15, 16) or the popularity of dog breeds (17). Moreover, natural languages are also known to have power law dynamics in word frequency distributions, such as Zipf’s law—the inverse relationship between a word’s frequency rank and frequency (6, 18). This fact has led some researchers to view drift as the primary mechanism that explains the dynamics of word popularity (8, 9). In linguistics, this simple mechanism has been argued to provide surprisingly accurate explanations of word popularity dynamics. For instance, Reali and Griffiths (8) and Ruck et al. (9) demonstrated that models of drift explain many aspects of natural languages, including the frequency distributions of words, and the dynamics of word popularity over time. Furthermore, Newberry et al. (19) used statistical tests adapted from population genetics, such as the Frequency Increment Test (FIT, see Ref. (20)), to detect signatures of drift or selection in the frequency changes of 36 English verbs. The results in Newberry et al. (19) showed that only six of these verbs provided strong evidence of selection, while the remaining 30 were more consistent with random drift. Because FIT cannot distinguish between specific selection pressures, the authors speculated that they detected signs of selection for irregular forms. Verbs are easier to remember when their present and past tense rhyme (e.g. dive–dove), which is a common feature of irregular verb forms. Subsequently, these data were reanalyzed in Karjus et al. (21) and Karsdorp et al. (22), confirming the overall validity of findings in Newberry et al. (19). However, they noted that the results were highly sensitive to frequency binning—that is, the time window used to measure aggregated word frequency in the historical corpus. Similarly, Blythe and Croft (23) analyzed S-shaped adoption curves and argued that such dynamics are typically associated with selection acting on a particular linguistic variant. Crucially, these approaches operate at the level of individual linguistic units or small sets of directly competing alternatives, modeling their frequency trajectories over time, whereas our study aims at estimating transmission biases at the population level across many competing items. Selection-based explanations Selection-based approaches to word popularity can be divided into two distinct groups. The first approach aims at explaining the popularity of certain words over others due to their intrinsic properties like word length or grammatical regularity. For instance, G.K. Zipf, although indirectly, provided a selection-based account of word popularity. In Zipf (6), he introduced his general principle of least effort—the idea that living organisms try to conserve energy as much as possible. He hypothesized that certain linguistic regularities can be explained using this principle. One such regularity is the negative correlation between word length and frequency, otherwise known as the law of abbreviation. His principle of least effort predicts that speakers aim to reduce their average articulatory effort while maintaining intelligibility. As a result, words that are pronounced more often should be short. For instance, the word jug might be more popular across language users because it is easier to pronounce as it only has three sounds, while the words flagon and carafe have six and five sounds each, respectively. In general, Zipf’s idea that the benefits of communication leverage its costs has become very influential in the study of language (24–26). For instance, Mahowald et al. (5) demonstrated that speakers tend to prefer shorter words, such as chimp over chimpanzee, in predictive contexts. This suggests that both the costs and benefits of communication influence word choice, with speakers favoring shorter words when brevity does not compromise communicative success. The second type of selective pressure that is used to explain the popularity of certain cultural items is frequency-dependent selection. In the case of frequency-dependent selection, the frequency of a variant leads to its disproportionate copying either because it is common (positive frequency-dependent selection) or rare (negative frequency-dependent selection, see (11, 27, 28)). Because unbiased copying tends to preserve an item’s frequency in the population (and does so perfectly when the population is infinitely large), it bears a superficial resemblance to frequency-dependent selection. The two, however, are fundamentally different. Under pure drift, agents simply copy items from whomever they encounter. As a result, the probability of encountering an item tends to match its frequency in the population. Under frequency-dependent selection, in contrast, agents exaggerate the effect of frequency on cultural transmission. Positive frequency-dependent selection was proposed as the main driving force behind word popularity in Pagel et al. (4). In this article, the authors examined thousands of word choices across 400 meanings among American English speakers extracted from Linguistic Atlas Project (29). The Linguistic Atlas Project aimed to study sociolinguistic variation among English speakers across different regions of the United States. They found that the observed ranking of alternatives for each meaning is best explained by speakers being disproportionally more likely to use frequent words. Moreover, they also found weak effects of word-intrinsic properties, such as word length or phonotactic complexity, indicating that shorter and more phonotactically sound words also enjoyed a slight advantage over their counterparts. In another study, Carr and Rastle (30) showed that both positive and negative frequency-dependent selection influenced the prevalence of English spelling variants at different periods over the last millennium. For instance, they found strong evidence of frequency-dependence in the 18th century, which aligns with the onset of mass schooling and the standardization of English spelling, where the most frequent spelling variants often became the norm. The puzzle: what drives word popularity? Languages evolve under various pressures; detecting and identifying these pressures remains a significant challenge. Previous research has produced mixed findings, with some studies emphasizing the role of drift in shaping word popularity, while others suggest that selective pressures are also at play. This apparent divide may arise from multiple factors. Much prior work has focused on either (i) demonstrating that certain patterns in natural languages, such as power laws and word turnover rates, can be explained by drift (8, 9) or (ii) identifying signatures of drift or selection from time series data on individual linguistic units (19, 21, 22, 31). However, in the case of (i), many studies have relied primarily on computational models without direct comparisons to empirical data. Concerning (ii), most analyses have been based on relatively small samples of words, and specific hypotheses about selective pressures have seldom been tested (although see Ref. (4)). Methods like the FIT, for example, can indicate whether selection is occurring but cannot determine which specific selection pressures are driving changes in frequencies. As a result, hypotheses about particular selective pressures have often been proposed on an ad hoc basis rather than tested empirically. In this study, we contrast specific selective pressures with random drift using an agent-based model rooted in cultural evolutionary theory. By simulating various scenarios—drift, selection for brevity, and frequency-dependent selection—and fitting these simulated data to the observed data from Slovotvir, we aim to identify which exact forces drive word popularity. For this, we use generative inference (or simulation-based inference). It is widely used in cultural evolutionary research (32–36), including studies on individual-level biases in fake news sharing (37) or the transmission of bird vocalizations (38). However, its application to linguistics has been constrained by the lack of high-resolution temporal data with individual-level information, which is critical for employing the generative inference framework. Notable exceptions include Karsdorp et al. (22) and Pagel et al. (4), although these studies focus on very limited samples and/or only partially employ simulation-based inference. Our goal is to provide a more comprehensive understanding of the dynamics influencing language change and offer new insights into how selective pressures and random drift shape linguistic patterns. Unlike previous studies that relied on small-scale datasets, our study utilizes a large, high-resolution dataset, allowing us to apply simulation-based inference more effectively to disentangle the roles of selection and drift in language change.c Data When referring to Slovotvir data, we will be using the terms below in a specific sense, similar to the terminology employed in the resource itself. Word—a target word (usually a borrowing) for which alternatives are being proposed by users of the Slovotvir platform. Words are added to the platform by the users. Word—a target word (usually a borrowing) for which alternatives are being proposed by users of the Slovotvir platform. Words are added to the platform by the users. Translation—alternative words that are being proposed by the users as a replacement for the target word. Translations of each word can receive approval from the users in the form of “likes,” and the number of likes determines the position of the translation in the list, with more popular translations displayed on top of the list. Translation—alternative words that are being proposed by the users as a replacement for the target word. Translations of each word can receive approval from the users in the form of “likes,” and the number of likes determines the position of the translation in the list, with more popular translations displayed on top of the list. Preprocessing The data from Slovotvir come in the form of an SQL database that was generously provided to us by the administrators of the website. It contains information on every word, translation, and like, that has been logged on this website, with precise timestamps and anonymized user information. The data collection began in April 2014, and the database was retrieved in June 2023, resulting in more than 9 years’ worth of observations. We have extracted the data on words, translations, and their lengths, as well as the number of likes that they received from the database and removed both words and translations that consisted of more than one word. Moreover, we removed all the words and translations that were one character long. We stored the length of translations in characters, since this measure is often used as a proxy for the length of a phonetic word (39). Moreover, Ukrainian orthography is highly transparent, meaning that the length of the written form closely corresponds to the number of sounds in a word (see Supplementary material 3.1). As a result, this measure closely approximates spoken word duration. This resulted in 42,044 translations, corresponding to 7,606 words that were being translated. Each word has a median of 4 translations, with a minimum of 1 and a maximum of 155 translations per word. These translations have received 129,974 likes from 4,700 unique users in total. To reduce the computational costs related to the generative model, we divided the data into time-steps: each year was divided into three parts, corresponding to ∼4 months. Ultimately, we arrived at 30 equal bins. Figure 1 shows the cumulative distribution of the number of unique translations, words, and likes over the 30 4-month bins, as well as the distribution of translation lengths and the growth of the number of users. Binning serves two technical purposes: (a) to align simulated activity of agents with the empirical timing of user participation and (b) to make the simulation computationally tractable over a 9-year period. Bin size and number do not affect the output in any significant way (see Section S1.2.). Summary statistics of the Slovotvir data. a) Distribution of translation lengths in number of characters. b) Distribution of the number of likes. c) Cumulative growth of the number of words, translations, likes, and unique users for 30 4-month bins. The x-axis begins on 2014 April 29 and ends on 2023 June 15. The left-side y-axis corresponds to the number of unique words, translation and likes, and the right-side y-axis corresponds to the cumulative number of unique users. Methods To examine the respective roles of frequency-dependent selection and selection for short words influencing the liking decisions among the users of Slovotvir, we rely on amortized Bayesian inference with invertible neural networks, also known as BayesFlow framework (see Refs. (35, 40, 41) for detailed description of this method). In contrast with the standard Approximate Bayesian Computation approaches that relied on Markov chain Monte Carlo (MCMC), BayesFlow uses neural networks, which allows increasing the speed and the precision of inference. ABC methods are often used with agent-based models, and can help researchers uncover the causal structure behind the observed data. These methods allow to estimate posterior distribution for models where the likelihood function cannot be calculated analytically, which is often the case with complex agent-based models. BayesFlow workflow consists of a generative model, an inference network and a summary network. The model generates synthetic datasets by sampling different values from the prior distribution for the parameters of interest. The inference network estimates the data-generating parameters from data, while the summary network generates aggregated summary statistics from the output of the generative model. These two networks are trained simultaneously: the inference network learns to infer the parameters, and the summary network learns to create summary embeddings which increase the accuracy of parameter estimation. Once trained, the inference network can be used to sample the posterior distributions of the parameters based on the observed data. The main advantage of amortized inference on neural networks is that it does not require researchers to make a subjective choice among different summary statistics. These are replaced by learned summary embeddings from the raw output of the generative model, reducing researcher’s influence on the inference process. Another advantage lies in the fast estimation time and a lower probability of over-fitting. This method has previously been used to model COVID-19 outbreak dynamics (42) and age-related effects on response time in decision-making tasks (43), and overall has demonstrated results that surpassed traditional ABC methods relying on the MCMC sampling algorithms. Moreover, this method has recently been used to study transmission biases in birds (44). The generative model and the predictions for the parameter values were preregistered. We had hypothesized that the behavior of Slovotvir users is characterized by selection for short words, whereby shorter forms enjoy selective advantage. For the frequency bias, based on previous literature (4), we predicted that translations in Slovotvir undergo positive frequency-dependent selection. The preregistration document can be found hered: https://osf.io/jasvb/?view_only=df85bfad375543eab8dc86d0a44d03f3. Generative model Our generative model relies on limited input from the Slovotvir data. This input ensures that the structure of the simulated data, including the number of words, translations, and possible word lengths, is consistent between the empirical and generated data. Slovotvir data were divided into 30 bins, each of which correspond to roughly 4 months of observations (see Data section). We extracted the total number of words that appeared on the platform, the total number of translations and the number of likes given by each active user during each of the 30 bins. Given this data, the model works as follows. During each of the 30 periods, the number of words appearing in the model and the corresponding number of translations are directly taken from empirical data for the same time period, with the translations randomly distributed among the words.e The length of each new translation is randomly sampled from the distribution of lengths of the translations in Slovotvir data, and the counter of likes for each newly added translation is set to zero. Then, artificial agents (their number is equal to that of the real users of Slovotvir at the same time-period), become active and perform the following actions, which are controlled by three parameters in the model. The length bias (α) influences the probability of an agent liking a translation based on its length. The other two parameters, frequency bias (β) and exploration bias (τ), are associated with the number of likes each translation receives. Frequency bias determines which translations will be liked by the agent based on their number of likes. Finally, exploration bias determines which word the agent chooses to explore for its corresponding translations, as words are ranked based on the number of likes their most popular translation has received on Slovotvir. Exploration bias is implemented in the same way as frequency bias, but it influences which words users explore, rather than which translations of a word they prefer. Specifically, these actions are performed as follows. Each agent first needs to choose a word from the list of words available at time t. In other words, if a word has three translations, which have 13, 20, and 100 likes, then the associated number of likes will be set to 100. The probability of selecting a word from the list of available words is then computed as follows: Lw (likes per word) corresponds to the maximum number of likes a translation for each word has received. Under the formulation in 1 the parameter τ (exploration bias) corresponds to drift at τ=1 (meaning that agents are exploring words proportionally to the top-most number of likes their most popular translations has received). If τ<1, then the agents will choose words whose translations are less popular, indicating negative frequency-dependence. Finally, if τ>1 they will only explore words with highly liked translations, which will correspond to positive frequency-dependent selection. The denominator, representing the sum of all values, is necessary to normalize the outcome values, so that their sum is equal to one, and they can be interpreted as probabilities for the purposes of sampling. Overall, at this stage, words are selected according to a probability distribution based on their current popularity (through τ). This distribution is computed once, at the beginning of the voting phase of the epoch, and all word targets for that epoch’s votes are drawn from it. In that sense, the model compresses what would be many individual word-selection events into a single draw from a fixed distribution. After a word is chosen, its translations are retrieved, and the probability of an agent liking each translation as a function its length and number of likes is computed as follows: Equation 2 determines the probability of liking a translation by leveraging either the length of the translation (length bias, α) or its number of likes (frequency bias, β). We used raw values for these two inputs without applying any transformations. In 2, Bt corresponds to the lengths of translations, Lt corresponds to the number of likes that each translation has received at the time that the agent has chosen the word. Here, α<0 corresponds to selection for short words, and α>0 corresponds to the selection for long words. α=0 corresponds to the absence of selection for length, as it means that each word will have an equal contribution, independently of its length. The β parameter functions exactly as the τ parameter described above. In general, both β and τ are used to test our hypothesis about frequency-dependent selection influencing word popularity. The α parameter is used to test the role of selection for length. The index of a word which will be liked by the agent is sampled according to the resulting distribution Plike. This process is repeated the number of times that the corresponding user in Slovotvir has liked words within each time bin. Contrary to the first stage, where words are selected in bulk, the choice of translations is implemented sequentially. For each vote assigned to a translation, the number of likes is immediately updated. Therefore, if the same translation is selected multiple times within an epoch, later translation choices for that translation do reflect the updated state. Importantly, τ is only applied at the first stage of the model, when words are selected, and α and β are only applied at the second stage, when agents are liking translations. A graphical overview of this two-step process of agent behavior in our model is demonstrated on Fig. 2 (a and b). An example of an agent’s two-step action in the model. a) An agent is choosing the word based on the number of likes that the word’s most liked translation has received. b) After selecting the word, the agent can evaluate all of its translations that are available at the time. The agent chooses which translation to like based on its length and the number of likes. Ukrainian alternatives for the word дедлайн (deadline, a borrowing from English, possible trough Russian) are given as illustration. The functions determining the probability of choosing a word and liking a translation are given below the corresponding steps. There, Lw (likes per word) corresponds to the maximum number of likes among translations of all words, Lt corresponds to the number of likes a translation has received at the time of choice, and Bt corresponds to the length of the translation (in characters). The parameter τ determines exploration bias at the word level, β determines frequency bias, and α determines length bias, both applied when agents select which translation to like. Under this formulation, τ=1 and β=1 correspond to drift, τ>1 or β>1 correspond to positive frequency-dependent selection, τ<1 or β<1 correspond to negative frequency-dependent selection, and α<0 corresponds to selection for short words. To illustrate the interaction of α and β parameters in the model, we ran it for every combination of these parameters within the [−5,5] range, while fixing τ at 1. This setting means the exploration bias is set to drift, so the agent selects words strictly proportionally to the number of likes their most popular translation has received (Fig. 2a). The [−5,5] range was chosen since it includes both negative and positive frequency-dependence for β, and selection for short and long words for α. Moreover, this range includes extreme values for both parameters, allowing us to explore the behavior of our model under very strong selection for either length or frequency. For each run of the model, we computed the average length of translations weighted by the number of likes, and the Gini index of the vector of lengths gini(Lt). These statistics are closely related to the model outputs which will be used to estimate model parameters, see Parameter estimation section. The average length of translations weighted by the number of likes is computed as ∑i=1nBti×Lti∑i=1nLti. This measure indicates which word lengths are more commonly associated with the most-liked translations, reflecting the tendency for longer or shorter translations to be preferred, based on their popularity. The result of these simulations are shown in Fig. 3. Both heat-maps indicate that the model is sensitive to the values α and β, but that different measures applied to the output provide different information. The Gini index is clearly more sensitive to changes in β, while the expected length is sensitive to both α and β. Weighted average length clearly distinguishes between selection for short words (α<0) and for long words (α>0). When the value of α is below zero, shorter words become more popular, while the inverse happens when the value of α is above zero, see Fig. 3b. Similarly, Gini index clearly distinguishes positive (β>1) and negative frequency-dependent selection (β<0). Under positive frequency-dependent selection, the Gini index is approaching 1, indicating extremely high inequality in the number of likes, as only a handful of translations dominate the rankings due to the self-reinforcing dynamics of this process, see Fig. 3a. Variation in the model output as a function of α and β values. Red axis values correspond to drift for the frequency bias (β=1), and the absence of selection for the length bias (α=0). Both measures are computed based on the final output of the generative model; the number of likes each translation received (Lt) and their corresponding lengths (Bt). a) Values of the Gini index of the number of likes. Higher values indicate higher inequality. b) Values of average length (in characters) weighted by the number of likes. High values of average length indicate that longer words are preferred over shorter ones, while low values indicate the inverse. Parameter estimation To estimate the parameters of the generative model that correspond best to the empirical data, we used two types of output from the generative model. First, we relied on the rank—number of likes distribution for translations. In addition to that we included the rank—weighted length distribution as well. The weighted length is computed as Lt×Bt, meaning that the number of likes will be multiplied by the corresponding length for each translation. Note that this value is proportional to the weighted average length used in Fig. 3, just without the normalization parameter. Both of these distributions were binned using logarithmic (log2) bins, with each bin containing the sum of the respective values in the corresponding interval. This means that all values from the first bin to the second will be summed, all the values from the second to the fourth, and so on (see Section S1.2). The binning procedure was necessary to reduce the size of the output of the generative model, so that it can be used as the input for the summary network. The combination of both of these distributions served as the input for the BayesFlow (35) estimator, which allowed us to infer the values of α, β, and τ with greater precision. Note that while we subjectively chose which data to collect from the generative model, the summary network transforms it in a way that maximizes estimation accuracy. For the frequency bias (β) we used an uninformative Uniform(−2,2) prior, as it contains both positive (β>1) and negative (β<0) values, which corresponds to positive and negative frequency-dependent selection, respectively. This range is more restricted than the one used in the demonstration above, since values higher than 2 were rarely found in previous research with similar methodology (37, 38). And we were relying on an uninformative Uniform(−5,1) prior for the length bias (α), since it allows for the exploration of both the selection of short words (α<0) and long words (α>0). This prior is restricted at one, since we do not expect our data to show cases of strong selection for long words. Finally, we relied on an informative LogNormal(0,0.5) prior for the τ parameter, with the highest probability density centered around 1, i.e. the value that corresponds to drift. This prior reflects previous results suggesting that the behavior of users while selecting from a list is consistent with drift (45). Overall, the parameter estimation procedure consists of three steps Generation: the agent-based model is used to generate 100,000 simulated datasets based on the prior distributions of α, β and τ described in Table 1. Each simulated dataset contains (i) the binned rank—number of likes distribution and (ii) binned rank—weighted lengths distribution, and has 34 entries (17 bins from each of the binned distributions). Additionally, 10,000 test datasets, and 10,000 validation datasets were generated for model evaluation. Generation: the agent-based model is used to generate 100,000 simulated datasets based on the prior distributions of α, β and τ described in Table 1. Each simulated dataset contains (i) the binned rank—number of likes distribution and (ii) binned rank—weighted lengths distribution, and has 34 entries (17 bins from each of the binned distributions). Additionally, 10,000 test datasets, and 10,000 validation datasets were generated for model evaluation. Training: the summary and inference networks from BayesFlow are trained for 200 epochs, with a batch size of 32 and the learning rate of 0.001. A deep permutation-invariant network (46, 47) with a summary dimension of size 4 is used as the summary network, and an invertible neural network is used as the inference network. Inference network is used for inferring the values of α, β, and τ given the data (which is transformed using the summary network). Both the summary and the inference networks are trained following the standard practices suggested in Radev et al. (35). Training: the summary and inference networks from BayesFlow are trained for 200 epochs, with a batch size of 32 and the learning rate of 0.001. A deep permutation-invariant network (46, 47) with a summary dimension of size 4 is used as the summary network, and an invertible neural network is used as the inference network. Inference network is used for inferring the values of α, β, and τ given the data (which is transformed using the summary network). Both the summary and the inference networks are trained following the standard practices suggested in Radev et al. (35). Inference: the trained inference and summary networks are used to sample 10,000 posterior distributions of α, β, and τ given the Slovotvir data as input. Inference: the trained inference and summary networks are used to sample 10,000 posterior distributions of α, β, and τ given the Slovotvir data as input. Table 1Model parameters and their prior distributions that were used for parameter estimation.ParameterDescriptionPrior α Length biasUniform(−5, 1) β Frequency biasUniform(−2, 2) τ Exploration biasLogNormal(0, 0.5) Model parameters and their prior distributions that were used for parameter estimation. After training, the inference network was able to recover the values of the parameters with very high accuracy based on the data from the validation set, i.e. the dataset that the model was not exposed to during training. The respective values of the coefficient of determination (R2) for each of the model parameters are as follows: Rα2 = 0.855, Rβ2 = 0.883, and Rτ2 = 0.989. The observed values of R2 for the parameters of the generative model indicate high fidelity in recovering the true values of the parameters given the simulated dataset (see Section S2.2). We also examined the loss curve of the model, and found that there is no over-fitting, since both the training and the testing losses are decaying without a visible divergence anywhere throughout learning (see Section S2.1). Results Length bias The mean value of the estimated parameter for length bias (α) is −0.662 (95% HDI: [−0.870, −0.464]). This indicates that the users of the Slovotvir platform have a preference toward liking shorter translations from a set of alternatives. The 95% CI for the α-parameter does not include 0, indicating that selection for length is present in our data. The posterior distribution, as well as the corresponding prior distribution are shown in Fig. 4. Posterior estimation of the length bias parameter (α). The x-axis corresponds to parameter values (Θ). The black solid curve corresponds to the posterior distribution, and the black dotted curve correspond to the prior distribution. The blue vertical line indicates the absence of selection for length. The red arrow indicates the direction of selection for short words. This figure shows the posterior estimate of the length bias parameter α. The black solid curve is concentrated below 0, indicating selection for shorter words. The dotted curve shows the prior distribution, the blue vertical line marks no length selection, and the red arrow indicates the direction of selection for short words. Frequency bias and exploration bias The posterior estimations of the frequency bias (β) and exploration bias (τ) parameters are shown in Figure 5. These results partially conform with our initial predictions. For the frequency bias (β), the behavior of the users is consistent with frequency-independence, as the posterior mean is equal to −0.067 (95% HDI: [−0.423, 0.270]), which includes zero. This means that the decision to like a translation is probably not influenced by how many likes it has already received. Instead, this suggests all translations have an equal chance of receiving a like, regardless of their popularity. The posterior for the exploration bias (τ) has the mean of 1.025 (95% HDI: [0.957, 1.086]), which corresponds to drift. This means that Slovotvir users choose words to explore proportionally to the number of likes their most popular translation has received, without signs of either negative of positive frequency-dependent selection. a) Posterior estimations of the parameters that model the number of likes a word’s translation is likely to get: the frequency bias (β) and exploration bias (τ). The x-axis corresponds to the respective parameter values (Θ). Black solid curves correspond to the posterior distribution, and black dotted curves correspond to priors. The red arrow indicates the direction of negative frequency-dependent selection. The orange arrows indicate the direction of positive frequency-dependent selection. The blue vertical line corresponds to frequency independence. The green vertical line corresponds to drift. b) Functions illustrating the different strategies considered in our model and highlighted on previous panel. X-axis corresponds to the number of likes, and the Y-axis corresponds to the probability of action (either liking a translation or selecting a word). Θ=1 corresponds to drift, Θ=2—positive frequency-dependent selection, Θ=−1—negative frequency-dependent selection and, finally, Θ=0 to frequency independence. Posterior checks To explore the posterior predictions in more detail, we performed two exploratory analyses. First, we simulated the rank–number of likes distribution running the generative model using the posterior estimates of α, β, and τ, and we compared it with the same distribution extracted from Slovotvir data. We also ran our generative model using the prior distributions of the parameters. Figure 6a plots the results of this analysis. The rank-number of likes distributions from models where the parameters are sampled from the posterior distribution are very closely matching the empirical data. Moreover, there is a highly visible reduction of variability in comparison with the model outputs where the parameters are drawn from the prior distributions (see Table 1), highlighting that the posterior estimations capture the structure of the data well. a) Slovotvir rank–number of likes distribution (black), compared to the posterior predictive distribution (yellow), and the prior distribution (gray). X-axis corresponds to rank, y-axis corresponds to the number of likes. b) Average length weighted by the number of likes extracted from the model initiated with the parameter values from the posterior distribution (yellow), and the model that corresponds to the absence of selection for length (blue). Diamonds correspond to the mean values, gray scatter corresponds to the exact values extracted from 500 simulations for each type of model output. Vertical line corresponds to the observed average weighted length from Slovotvir. Finally, to examine whether the data in Slovotvir is indeed consistent with the scenario of selection for short words, we performed the following analysis. We extracted the number of likes and the length of each translation from the final stage of the generative model, which was run for 500 iterations using the posterior parameter estimates. Additionally, we analyzed a model where all parameters were drawn from the posterior, except for α (length bias), which was fixed at 0 to represent the absence of selection for length. This simulation was also performed 500 times. Using this data, we computed the average length weighted by the number of likes for each of the model outputs (exactly the same measure is used in Fig. 3b), which yielded 500 values for each run of the models. Figure 6b plots the result of these simulations. The distribution of weighted average lengths extracted from the model where the parameters are drawn from the posterior distribution has a mean of 7.67 characters, while this value for Slovotvir data corresponds to 7.66 characters. However, the average length for a scenario in which the selection for length would be absent is 8.1 characters. This confirms that Slovotvir users display a weak but persistent tendency for choosing shorter words, which corresponds to a preference for translations roughly half a character shorter that the neutral baseline. Discussion In this article, we analyzed over 9 years of data from thousands of users on the Slovotvir platform, where users rank and propose translations for Ukrainian words of foreign origin. Using generative inference, we identified various pressures shaping word popularity dynamics in Slovotvir. Our analysis revealed that translation popularity is influenced by multiple factors. Notably, user behavior on the platform aligns with selection for shorter words, consistent with previous research on efficient communication (5, 6). However, information about the number of likes that translations receive plays little to no effect, indicating frequency-independence. At the level of word exploration, i.e. the choice of words whose translation the user will explore, the behavior seems to be consistent with drift. Users explore words in proportion to the number of likes received by a word’s top translation. The number of likes a word’s top translation receives influences how words are displayed on the platform’s main page. This aligns with findings in Salganik and Watts (45), where users were more likely to engage with items at the top of a list. In sum, our results suggest that while drift contributes to the patterns observed in Slovotvir data, it is not the only factor shaping word popularity. Selection also plays a significant role; in the case of our data, users are clearly biased toward shorter words. In our generative model, we have parametrized the behavior of users with relation to word popularity in two ways. One way corresponds to the bias in how users select which translations to like (frequency bias, β). We measured this by using the number of likes a translation had received previously as an indicator of its popularity. This number then helped determine whether a user would like the translation after observing it. Here, we found that the users of the platform likely display behavior consistent with frequency-independence. This means that they do not consider the number of likes when deciding which translation to like. Instead, they focus on other features, such as length, preferring shorter words. However, another parameter of the model related to the number of likes, namely exploration bias (τ), is consistent with random drift. Exploration bias determines which words agents choose to explore from the list of all words available for translation. In the case of Slovotvir, the probability of choosing which word to explore is proportional to the number of likes its most popular translation has received. We did not find evidence that users displayed behavior consistent with frequency-dependence in either of these cases, contrary to what was suggested in Pagel et al. (4). One reason for this behavior at the level of word exploration could be that the list of words users need to explore has grown rapidly over the platform’s lifetime, with many new words being added each year. Therefore, it is likely that users explore the most popular words more, since they are listed first on the starting page. This finding is consistent with other studies of user behavior on online ranking platforms (45), as users tend to click on options at the top of the list most of the time, regardless of their underlying quality. The list of translations for a word, however, is much shorter, with the average number of translations per word being around five (number of translationsnumber of words=42,0447,606≈5.53). Therefore, it is likely that the users were exploring the full range of translations for a word, and that the selection for short words was playing the major role at this stage. Another reason why we do not observe signatures of frequency-dependent selection at the level of translation popularity is their frequency in the Ukrainian language. The translations posted on Slovotvir are typically neologisms, either representing relatively infrequent words in Ukrainian or newly coined terms introduced by Slovotvir users. Out of all 42,044 translations (33,959 of which are unique translations that do not repeat across words) available on Slovotvir, 6,409 can be found in the 2024 Ukrainian news corpusf from Biemann et al. (48), which contains more than three million words. For these 6,000 translations, the median word count is 10 occurrences, with a minimum of 1 and a maximum of 367 occurrences per three million. In sum, this means that while a large portion of translations proposed on Slovotvir are already in use in Ukrainian, which is no insignificant achievement, their frequency remains relatively low, as expected given their recent emergence. This factor might explain why we do not detect signatures of frequency-dependent selection, as this force is thought to primarily affect mostly high-frequency words (19, 49), while less frequent items are subject to drift or, as in our case, frequency independence. Another potential limitation lies in the design of our generative model. In our study, we focus only on the process of selecting words and liking their translations, while we ignore the process of proposing translations entirely. For instance, Slovotvir users might add translations to words that already have popular translations, or they might instead look at words that have not been translated yet. That is, while in our model translations are randomly assigned to words, this process is likely nonrandom on Slovotvir. This may affect our inference about frequency-dependent effects, since in our model all words are equally likely to receive new translations, whereas in reality more popular words may attract more new variants. We made this methodological choice for two reasons. First, our goal is to address the ongoing debate about drift and selection as sources of word popularity differences in languages. The process of deciding which word to translate, and how to translate it, may be driven by different and multifaceted mechanisms that are not directly related to the selection processes we aim to study. Second, we favor parsimony in models, especially when using the ABC approach, to avoid overfitting, since adding a translation mechanism would require at least one additional parameter. Our model already provides a good fit to the data, suggesting limited gains from increasing its complexity. Future work could explore alternative innovation mechanisms to assess the robustness of our conclusions. Can our findings about user behavior on Slotovir be extended to offline behaviors of language users? Slovotvir in itself resembles datasets like the Linguistic Atlas Project used in Pagel et al. (4). However, it also has several advantages. First, this dataset includes novel words, and their popularity trajectory on the platform likely mirrors real processes of language change in Ukrainian, suggesting that their dynamics were shaped either within Slovotvir itself or, much more likely, reflected broader real-time linguistic changes. While datasets like the Linguistic Atlas Project are likely influenced by various sociolinguistic pressures, such as prestige or community size, Slovotvir provides a unique case where we have complete information on the dynamics and individual user behavior that shaped the observed data. Moreover, the data on Slovotvir reproduces known linguistic regularities like Zipf’s law of abbreviation; on Slotovir, shorter words tend to receive more likes that their longer counterparts. This is reflected in the robust negative correlation between number of likes and length of translations (Pearson’s ρ=−0.13, p<1.2e−102, see Section S3.2), and also indirectly reflected in our finding that the users of this platform have a preference for shorter words. The relatively weak correlation is comparable to those found in natural languages (50), suggesting that Slovotvir data adequately mirrors these patterns. Additionally, the rank—number of likes distribution used in our analyses exhibits language-like coefficients in Zipf’s law of frequency ((18), also see Section S5), further supporting the idea that this data aligns with behavior characteristic of natural languages, and warranting our use of number of likes as a proxy for frequency. In sum, our approach allowed us to couple individual-level biases, such as the frequency bias or the length bias, with population-level observations of the popularity of translations on the Slovotvir platform. This approach to studying the roles of drift and selection in language change is novel, as most previous studies have focused on the dynamics of individual words while often overlooking individual-level biases due to data limitations or modeling assumptions (19, 21, 31). Our approach is more straightforward, as it focuses on language users themselves, commonly viewed as the main source of change in language (51, 52). Our method enables us to infer the popularity dynamics of linguistic units based on individual user preferences and to map these preferences to a large sample of words in our data. Moreover, having access to detailed data on each individual user’s decisions at every time-step makes the predictions from our model highly reliable. For instance, we know exactly how many words each user has liked at each time-step, and this information was incorporated into our model. In this regard, our study complements the results from Pagel et al. (4), as they relied on similarly structured data on word naming. While we have not found evidence of positive frequency-dependent selection like they did, we have found evidence of selection for short words. In summary, our study shows how individual-level biases influence the population-level structure of language. Additionally, it encourages the use of cultural-evolutionary methods in studying language change, demonstrating their potential to distinguish between the effects of drift and selection on word popularity. Supplementary Material Acknowledgments The authors thank Mason Youngblood and three anonymous reviewers for their comments and suggestions. We also extend our gratitude to Lev Lukomskij for generously providing us with Slovotvir data. Notes We are not using the word “drift” to mean a somewhat deliberate and directional change in language, which is how this word is often used in linguistics following Sapir (53). Rather, we use “drift” in the cultural evolutionary sense, to refer to stochastic changes in cultural item frequencies that arise from unbiased copying during transmission, in the absence of systematic selection pressures. We have provided a detailed description of Slovotvir’s interface in the Supplementary material. An earlier version of this article appeared in the first author’s doctoral thesis (54). Note that the preregistration document contains an addendum highlighting the changes that were made to the model after the original preregistration document was published. The changes were made to address initial issues with the generative model’s design and data handling. The choice of the number of bins is arbitrary and does not affect the behavior of the model in any significant way, see Section S1.2. See https://wortschatz.uni-leipzig.de/en/download/ukr. Contributor Information Alexey Koshevoy, Laboratoire de Sciences Cognitives et Psycholinguistique, ENS, CNRS, UMR 8554, PSL University, 29, rue d’Ulm, Paris 75230, France. Oleg Sobchuk, Department of Human behavior, Ecology and Culture, Max Planck Institute for Evolutionary Anthropology, Leipzig 04103, Germany; Max Planck Institute for Empirical Aesthetics, Frankfurt 60322, Germany. Olivier Morin, Institut Jean Nicod, Département D’études Cognitives, ENS, EHESS, CNRS, UMR 8129, PSL University, 29, rue d’Ulm, Paris 75230, France. Supplementary Material Supplementary material is available at PNAS Nexus online. Competing Interest The authors declare no competing interests. Funding This work has received support under the Major Research Program of PSL Research University “CultureLab” launched by PSL Research University and implemented by ANR with the references ANR-10-IDEX-0001 as well as ANR grant UNPREDICTABLE to Olivier Morin (ANR-25-CE28-7755). Author Contributions Alexey Koshevoy (Conceptualization, Data curation, Software, Formal analysis, Validation, Investigation, Visualization, Methodology, Writing—original draft), Oleg Sobchuk (Conceptualization, Data curation, Supervision, Validation, Methodology, Writing—review and editing), Olivier Morin (Conceptualization, Data curation, Supervision, Validation, Methodology, Writing—review and editing) Data Availability All the data and code needed to reproduce our analyses, as well as the pre-registration documents are available here: https://osf.io/jasvb/?view_only=df85bfad375543eab8dc86d0a44d03f3. References Associated Data Supplementary Materials Data Availability Statement All the data and code needed to reproduce our analyses, as well as the pre-registration documents are available here: https://osf.io/jasvb/?view_only=df85bfad375543eab8dc86d0a44d03f3.