When language users encounter linguistic material, this facilitates the subsequent production of identical or related material. In experimental psycholinguistics, this facilitation effect is called priming: exposure to a prime stimulus reduces processing costs for a related target. In linguistic interaction, the same process maps onto how Speaker 1 and Speaker 2 use a given construction.

In conversation, multiple mechanisms affect the facilitation effect. Self-priming is the facilitation of a speaker’s own subsequent production by their own prior output. Reciprocal priming occurs when one speaker’s output facilitates the other speaker’s production of the same or a related form. Alignment and convergence refer to the broader process by which interlocutors come to use increasingly similar forms over the course of an interaction, which may arise from reciprocal priming, shared situational pressures, or both (Garrod & Pickering, 2009; Pickering & Garrod, 2004). The absence of a speaker-identity effect in a conversational dataset is consistent with reciprocal priming, but does not rule out parallel self-priming under shared lexical biases.

A similar set of distinctions applies within morphology. The lexical boost is the facilitation of production by recent activation of the same lemma or word form (Pickering & Branigan, 1998). This is distinct from structural priming, which operates on abstract syntactic configurations independently of lexical content (Bock, 1986; Gries, 2005). Within morphological priming specifically, one can distinguish stem-level priming (facilitation of a lemma), affix-level priming (facilitation of a suffix or inflectional pattern across different stems), and whole-inflected-form priming (facilitation of a specific inflected token). Evidence for affix-level priming comes primarily from experimental work on derivational and inflectional decomposition (Amenta & Crepaldi, 2012; Marslen-Wilson, 2007). In naturally occurring speech, it is difficult to distinguish affix-level from whole-form priming, because the same speaker typically re-uses the same verb in the same inflected form. Whole-form priming is often referred to with a theoretically more neutral label, persistence, in conversational data – see Szmrecsanyi (2006) below.

In psycholinguistics, lexical boost is well established, and we have clear evidence for the priming effect of specific suffixes as well (see e.g. Amenta & Crepaldi, 2012). People access and produce stems and affixes more readily after recent exposure.

Rácz and Lukács (2024) and Rácz et al. (2020) demonstrated alignment between language users not only in the use of stems or affixes, but in entire distributions of variation in morphological inflection. Through a series of online word-picking games, they showed that players could converge to the choices of their co-player and that the effects of this interaction persisted after the game in post-testing. The effects themselves were not limited to across-word or across-suffix priming, but rather entailed a shift in lexical distributions at the subword level: the co-player’s word choice influenced the behaviour of similar words and the strength of this effect was a function of their similarity to it.

In models of dialogue as joint action, convergence provides a core coordination mechanism and is a lynchpin of language variation and change (Beckner et al., 2009; Garrod & Pickering, 2004, 2009; Giles & Coupland, 1991; Pickering & Garrod, 2004, 2006). Convergence has been attested in naming preferences (Roberts, 2010), syntax (Gries, 2005), and phonetics (Babel, 2012). In comparison, morphological convergence in natural language has received limited attention. Weiner and Labov (1983) found that a strong predictor of passive use in English-language interviews was the presence of another passive in the previous five utterances. Szmrecsanyi (2006) used a corpus-based approach to look at the English comparative, where a word-formation pattern (friendlier) alternates with an analytical construction (more friendly). He found that the choice of either variant persisted in discourse, meaning that if speakers started using one, they would be more likely to keep using it.

The dearth of similar research work in natural language reflects the challenges of detecting morphological convergence in interactional data. Many analytic languages, like English, provide a shallow pool of morphological variation, which is, in turn, very hard to dredge in corpora. Lab studies, like Rácz et al. (2020), used an experimental battery that exposed participants to a rapid fire of nonwords and a framework that actively encouraged predicting the co-player’s behaviour. This allowed these studies to explore more nuanced convergence effects across lexical distributions. On the flip side, the design was divorced from natural language use, in which convergence usually operates on familiar words and its incentives and mechanisms are more implicit. The natural-language studies, like Weiner and Labov (1983), relied on natural language and identified patterns that were frequent enough in English to constitute a robust, compact sample.

This study uses a set of dyadic interactions recorded in Hungarian. Interactive tasks entail the frequent use of instructions, requests, and commands. Hungarian uses the subjunctive to express these imperative functions (Tóth, 2007). The Hungarian 2sg subjunctive suffix varies between a long and a short form. This structured variation can be tracked across interactions and offers a unique window into how lexical distributions and convergence effects shape word use together in a language with rich inflection and complex word-formation processes.

The Hungarian verb is marked for person, number, tense, mood, and definiteness. The subjunctive, used both in the imperative and in certain subordinate constructions (Tóth, 2007), is marked in the present tense and has a different paradigm for definite and indefinite verbs. The 2sg form varies in both paradigms, with a long and a short form. In this article, I focus on the more frequent indefinite. Example (1) shows the long and the short form.

The short form is an underlying -j which can assimilate to the preceding stem, while the long form adds -jál/-jél, depending on the backness of the final stem vowel, following the general patterns of Hungarian vowel harmony (Siptár & Törkenczy, 2000) (hence, keltsél/*keltjél wake-sbjv-2sg-indef). The long/short variation is attested, in some form, with all types of verbs, including irregular forms (see e.g. légy/legyél be-sbjv-2sg-indef, gyere/jöjjél come-sbjv-2sg-indef). This variation has been present in Hungarian since at least the 15th century (László, 2001), but no systematic, large-scale, quantitative study exists of the stylistic and social factors that condition it.

I drew corpus data from a frequency list based on the second Hungarian Webcorpus (Nemeskey, 2020; Rácz, 2025), which was built from Common Crawl and has a size of 9 billion words. I used transcribed conversation data from two sources. The Akaka Maptask Corpus consists of five hours of recordings of 24 task-oriented dialogues recorded using head-mounted microphones, with pairs of participants completing a map task (Molnár et al., 2023). The dataset has recordings from 46 participants (24 women, 22 men, mean age 20). The Budapest Games Corpus consists of 9 hours of 36 dialogues of pairs of participants. The dataset has recordings from 12 participants (five women, seven men), working together in an object identification task. Participants were recruited using convenience sampling, were matched in age group between the ages of 20 and 60 and knew each other (Mády et al., 2023). For more details, see Mihajlik et al. (2024).

The distributions of the raw data can be seen in Figure 1. We see the number of long/short forms used by speakers across conversations on the left and the number of long/short forms per verb lemma in total on the right. Both show a Pareto-distribution. While many conversations see only sporadic use of the subjunctive, in some, participants use the subjunctive extensively. Some verbs, like ‘wait’, ‘go’, or ‘watch’, are overrepresented in subjunctive use.

For details on the transcription and annotation of the original data, see Mády et al. (2023) and Molnár et al. (2023). I used force-aligned Praat textgrids and extracted the word tier and speaker identifier from each textgrid file. I filtered the resulting dataset to only include indefinite subjunctives and checked the data to avoid overlaps between subjunctive forms.

I drew a dataset of subjunctive forms from the webcorpus and calculated log(freq(long variant)/freq(short variant)) for each lemma.

I joined these data with the subjunctive dataset from the conversations to add word-level information. Three observations, for three different low-frequency verbs (tenyerel ‘palm’ kalauzol ‘guide’, tologat ‘push habitually’) in the conversation data had no long form in the corpus. I excluded these from the final analysis.

The final dataset consisted of 334 subjunctives from 22 recordings in the Akaka Maptask Corpus and 27 recordings in the Budapest Games Corpus. This is 0.04% of all transcribed words in the recordings. The sample was hand-checked for accuracy.

I analysed the data in R (R Core Team, 2025), fit models in brms (Bürkner, 2021), and used the ggplot2 and sjPlot packages for visualisations (Lüdecke, 2025; Wickham, 2016).

The analysis compares each subjunctive verb form (the target) with the last subjunctive verb form (the prime) in the conversation. The distance between the prime and the target can vary. Out of the 334 observations, 3 are infrequent verbs that have no long form in the webcorpus and 52 are single mentions or first mentions in a conversation, meaning that there is no preceding variant to compare them with. This leaves 279 observations which belong to 26 verbs. The top five verbs (megy ‘go’, vár ‘wait’, figyel ‘watch’, jön ‘come’, and indul ‘start’) are responsible for 80% of all observations, with a Gini coefficient of 0.76 (see Figure 1). The Gini coefficient quantifies distributional inequality: here, it indicates that most tokens belong to few types.

I fit a series of eight hierarchical generalised linear models with a binomial error distribution and a logit link function, predicting p(target is long). All models included conversation, speaker, and verb lemma as grouping factors (random intercepts). Fixed effects varied across models: the baseline model included only the verb’s corpus log-odds of the long form; subsequent models added the prime (long/short), pairwise interactions of the prime with same/different verb, same/different speaker, distance, and corpus frequency, as well as a three-way interaction and a model with all two-way interactions. The full set of models and their formulae are reported in the Appendix (Table A1).

I placed Normal(0, 3) priors on intercepts, Normal(0, 2) priors on fixed-effect coefficients, and Exponential(1) priors on random-effect standard deviations. These are weakly informative priors that allow large effects on the log-odds scale, while regularising against implausible extremes, following standard recommendations for logistic regression (Gelman et al., 2008). I compared models using approximate leave-one-out cross-validation (LOO-CV; Vehtari et al., 2017) and Bayes Factors computed via bridge sampling (Gronau et al., 2020). Residual autocorrelation was inspected manually and was not detected in any model.

I report the best-fitting model below. Where relevant, I convert log-odds estimates to probabilities using the inverse logit function (p = exp(x)/(1 + exp(x)), implemented as plogis() in R) to aid interpretation. I use three significant figures throughout (unless further precision is informative), matching Table 1.

The estimates of the best model can be seen in Table 1. Both priming effects and lexical effects are visible.

The best-fitting model (fit3 in the Appendix) includes the prime, whether the prime is the same or a different verb, their interaction, and the verb’s corpus log-odds. Model comparison supports the following answers to the research questions.

Going back to the best model, the credible intervals capture how the groups are related to the intercept (short primes). How they are related to each other is best understood looking at the raw data in Figure 2 and the model predictions in Figure 3.

If a verb prefers the long form in the webcorpus (x axis), it will do so in the conversations (y axis). We see a strong priming effect within the same verb: a long prime is followed by a long target. In the absence of a long prime, the short form predominates. There is no clear priming pattern across verbs.

Both the lexical effect and the priming effect persist in the best model. If we transform log odds to probabilities using the inverse logit function, we can say that verbs at the highest end of the (scaled) corpus distribution of long forms are about 30x more likely to pick a long form over verbs at the lowest end (plogis(–4.85 + 4.15) = .332 over plogis(–4.85) = .008). Within verbs, a long prime variant means that a long target variant is about 3x more likely (plogis(–4.85 + 1.68) = .040 over plogis(–4.85) = .008). The two effects are additive, not interactive, in the model. Model predictions can be seen in Figure 3.

Based on model selection, the best-fitting model has two plausible alternatives. First, target verb length may reflect lexical preferences alone, with no priming effect present. There is robust evidence against this model (BF = 62.61).

A second alternative model includes an interaction between prime and lexical preference. This is, again, best understood by looking at the predictions of this alternative model (Figure 4). In this model, it remains true that the prime only has a clear effect on the target if it is the same verb. However, this effect interacts with the verb’s lexical preference: if a verb is unlikely to be long based on the webcorpus, it will remain short in the conversation, irrespective of the prime. If the verb is more likely to be long in the corpus, this is more prominent in the conversation. Evidence for this interaction is indeterminate (BF = 0.82 relative to the simpler model), meaning the data neither support nor rule out a role for lexical expectations in modulating priming. If this interaction proved robust in larger samples, it would be consistent with expectation-based models of production. I return to this in Section 3.

Leave-one-out cross-validation fails to discriminate between models (identical LOO scores across all three candidates), likely reflecting data scarcity: with relatively few observations and few verb types, out-of-sample prediction is uninformative. As noted above, Bayes Factor comparison favours the priming model over the null, but remains indeterminate between the priming model and the more complex alternative which assumes a lexical preference : priming interaction. This indeterminacy may also reflect insufficient data to justify additional parameters and could be tested on a larger interactional dataset.

Since it was flagged by cross-validation, I used a permutation test to further confirm the priming effect (the reported model). I subsetted the data to observations where the target was the same verb as the prime (138 observations). I randomly shuffled the prime (long/short) 10,000 times and calculated a simulated ratio for the number of long targets over short targets. The mean of these ratios was 0.49. The true mean of long targets following long primes was much higher: 0.78. No simulated mean was higher than this. This means that the priming effect across identical primes and targets is unlikely to be due to chance alone (p < 0.001). This simple test does not replicate the entire interaction in the best model, but it does lend further support to the robustness of the priming effect.