No three productions alike: Lexical variability, situated dynamics, and path dependence in task-based corpora

2.1 Factors of variability and path dependence

Corpus linguistics has viewed variability primarily through one of two lenses: (1) as a function of language dynamics on a personal or population-based trajectory of language change, including variability in L1 or L2 acquisition (Ädel 2015, Kachkovskaia et al. 2017, Dirdal 2022, Kerz and Wiechmann 2020, Lowie and Verspoor 2019, Vyatkina et al. 2015, Lowie and Verspoor 2015, Lowie 2017, Verspoor et al. 2021, Larsen-Freeman 2019) or language attrition (Martin Villena 2023, Matos and Flores 2024), and language contact situations and diachronic change (Hilpert 2017, Hernandez-Campoy 2021, Karjus et al. 2020). Since language development and change are not the subject of this article, these perspectives will not be further discussed. Or (2), as a function of situational, regional, and social language stratification or differentiation. This branch includes the variationist approaches to corpus linguistics (Cacoullos and Travis 2019, Ludeling 2017, Szmrecsanyi 2017, Biber et al. 2016), in which variability is attributed to diatopic (e.g., regional/dialectal) and diastratic differences (age, level of education, etc.), or intra-individually functional, inter-situational aspects, such as differentiation by register (Pescuma et al. 2023, Szmrecsanyi 2019, Biber 2012). Recent psycholinguistic work further suggests that individual grammars may not cover the full syntactic range of a language community, but vary in accordance with factors such as exposure to written language or level of reading engagement (Da̧browska 2014, 2018, 2019, Petre and Anthonissen 2020).

All of these factors may fluctuate to some degree over a lifetime, but are expected to be largely stable and controllable in elicitations. The variationist model can therefore be described as a factorized function of language production, in which each factor that contributes to the choice of a particular linguistic expression is generalizable beyond the current situation (either to the speaker or to the type of situation). This includes an amount of free or random variation that emerges from ambiguities within the system itself, such as parallel structures and their reanalysis in language contact and beginning language change. The variationist model is therefore a typology of systemic variation relative to the system of a speaker or the language itself – differences in linguistic expression as we find them in corpora are viewed as classificatory of underlying, causal differences in features or functional choices.

More recently, the field has seen an increase in attention towards intra-situational inter-individual differences in language production within identifiable, factorially homogeneous strata of speakers, i.e., speaker groups that are controlled by age, regional, and societal aspects such as urbanicity and socio-economic factors. (Özsoy and Blum 2023, Tsehaye et al. 2021, De Smet 2020, Zerbian et al. 2022, Alexiadou et al. 2022, 2023, Shadrova et al. 2021, Wiese et al. 2022). Naturally, it is impossible to control for all potential factors of variation that come to mind. Some of the variability observed in these studies may thus still be attributable to stable and predictive, but currently unaccounted factors.

Beyond this, there is evidence to suggest that a considerable amount of variability may be linked to effects of local path dependence, some of which are conceptually well known in linguistics. In a path-dependent system, what has previously happened and influences what is going to happen later. Path-dependent phenomena are ubiquitous in human cognition, for instance, in priming (the persistent activation of concepts once they are introduced and the temporary co-activation of related concepts, Baayen and Smolka 2020, Kidd 2012, Melinger and Dobel 2005, Arias-Trejo and Plunkett 2013, Kaschak et al. 2014, Lucas 2000, Hutchison et al. 2008), categorization (e.g., family resemblance, in which members are assigned to a category based on the features of other, previously assigned members of the same category, Wittgenstein 1953, Rosch and Mervis 1975), or the ongoing construction of concepts and mental models (in which categories or paradigms, once established, define their own idiosyncratic constraints and plausible heuristics therein, see Kuhn 1970, Lakoff 1987, 1999, Tversky and Kahneman 1974, Neth and Gigerenzer 2015). In dialogue, path dependence is created by alignment or convergence between speakers on (morpho-)syntactic, phonological, semantic, lexical, and conceptual levels (Pickering and Branigan 1998, Garrod and Pickering 2007, Watson et al. 2004, Steels and Loetzsch 2006, Dobnik et al. 2015, Branigan et al. 2000, 2007, Reitter and Moore 2014, and others). In a similar sense, an ongoing process of synchronization between speakers in dialogue and speaker communities allows for temporary and locally stable meaning mapping (such as a locally known way of referring to a certain place), thereby creating idiosyncratic language bundles that may diverge from the larger speaker population. Synchronizations like these have been suggested as initialization points for grammaticalization and language change (Schmidt and Herrgen 2011), although that would first require their stabilization and re-convergence between communities.

Path dependence can also be created via morphological self-priming, morphophonotactic preferences, or collocational associations. For instance, in other analyses of the essay corpora also used in this study, Shadrova et al. (2021) found clustered occurrences of split particle and prefix verb types within 50 tokens after the first verb of the same morphological type was used. This persistence does not appear to be driven by graphematic similarity or semantics,^[6] thus introducing many new and unrelated lexemes. For coselectional path dependence, Shadrova (2024) found a tendency in German L1 speakers to prefer polysyllabic verb + noun combinations (4+ syllables for both elements). The selection of a polysyllabic verb or noun then raises the local probability of also selecting the other, creating a path dependence. Similarly, the choice of one collocate has been argued to increase the local probability of the other in coselectional association (Gries 2013, 2023, Brezina 2018, Laufer and Waldman 2011, Hoey 2005 and others).

Path dependence also follows from the choice of discourse topic on various levels of resolution. Even where macro-level variability is limited by prompt or purpose, speakers are generally free to choose their content for meso- and micro-levels. This may sound trivial, but the choice of the many subdiscourses related to a general thematic field (such as, in the case of a better life for younger generations?, the subtopics education or family life) and the various levels of granularity (e.g., amounts of homework in primary school vs how school prepares students for life) have far-reaching consequences for lexical distributions and between-speaker overlap. This is a common issue with prompt-based elicitations, e.g., in learner corpora. Observable repercussions of path-dependent processes on the frequencies of lexical and morpholexical categories in corpus elicitations, including strong task effects, have been abundantly found in this line of research (Lüdeling et al. 2017, Alexopoulou et al. 2017, Vyacheslavovna and Il’Inichna 2022, Pérez-Paredes and Bueno-Alastuey 2019, Gilquin 2021, Shadrova et al. 2021, Shadrova 2020). We will return to this in Section 4.2.

Path dependence and local non-stationarity are generally intrinsic features of complex adaptive or complex dynamic systems (CAS/CDS) – multi-agent systems that enter states of dynamic equilibrium through the ongoing ecological or functional adaptation of their agents without centralized control loops.^[7] Usage-based linguistics has a tradition of viewing language as a CAS (van Trijp 2024, Al-Hoorie et al. 2023, Verspoor et al. 2021, Lowie and Verspoor 2019, Lowie 2017, Beckner et al. 2009, Steels 2000). CAS expects a high degree of variability at all times and local convergence of speaker performance is not presumed, although the total system could in principle still converge in its full distributional realization over time/on average, e.g., in cyclical developments. Empirical approaches to CAS hypotheses are relatively young, and no explicit lexicological theories or analyses of lexical production in context appear to have been published to date. In an individual speaker’s CAS, an interaction between lexical retrieval, the organization of the mental lexicon, and other situational factors as named above would be expected.

Lexical retrieval has been psycholinguistically shown to be influenced by many factors, including semantic field, semantic similarity, lexical ambiguity, and the set of alternatives available in context (Nickels et al. 2022, Anders et al. 2015, Jolsavi et al. 2013), morphological relationship and phonological and graphematic neighborhood (Li et al. 2024, Vitevitch et al. 2014, Smolka et al. 2014, 2019, Baayen et al. 2006, Luce et al. 2000), multilingualism (Sullivan et al. 2018, Kambanaros et al. 2013), frequency of a word and frequency of co-occurrence with other words (Agustin-Llach and Rubio 2024, Jacobs et al. 2017, Mark Knobel and Caramazza 2008, Ellis 2002), attentional control (Roelofs 2008), and phonological aspects (Vitevitch 2022), among others. This line of research derives properties of the organization of the mental lexicon from measures of word production, association, and recognition from experimental settings. Contextualized production is not usually considered due to the complex challenges of factor isolation in situated context. Although certainly meaningful for situated results, the operationalization of those properties for corpus-based studies is not straightforward. While the process of language production must be mediated by the structure of the mental lexicon, it may be determined by many additional factors. Furthermore, the mental lexicon itself appears to be highly fluid and malleable (even between experiment trials according to Heitmeier et al. 2023), raising the likelihood for situational dynamics to dominate the scene. To derive concrete hypotheses for production in context, a synthesis with the interactional, communicative, and cognitive factors named above as well as a more general theory of mindwandering or cognitive association in context (which is generally understudied, see Fossa et al. 2019), would be required. This would be an immense undertaking and does not appear to have been tackled yet.

Another underconsidered source of variability is the general diversity in situational perception and interpretation. As a classic linguistic example, Chafe’s (1980) pear stories show that speakers focus on different aspects of a story in their narrations of events witnessed in a video (in which a character picks pears, which are then stolen). This results in major differences in narrations and, consequently, divergent lexical expression. As perception and sensory interpretation is fundamentally selective, filtered through an individual’s physiological/neurological, cognitive, cultural, social, psychological, and biographical setup, current interest, communicative goals, and other factors, variability in interpretation can be expected with respect to all aspects of a communicative situation (Lamy et al. 2012, Kemmerer 2023, Blomberg and Zlatev 2021).

Unlike globally stratified variation (by register/genre, age, social and regional factors), path-dependent variability is expected to occur inter- as well as intra-individually in near-identical contexts. A person may generally have a particular perceptual, lexical, collocational, etc. preference, and they may also situationally modulate their attention to include or exclude certain aspects or temporarily prime themselves to favor a particular expression.

In corpus linguistics, path-dependent phenomena and other production phenomena have found little attention until now, with the exception of some discussion of persistence/structural self-priming (Gries and Kootstra 2017, Gries 2005, Reitter 2008, Szmrecsanyi 2006). Usage-based lexicology has instead primarily been shaped through debates from phraseological and constructionist models of language. These will be reviewed next.

2.2 Phraseological and constructionist models of lexical production

Phraseological and constructionist views take a largely lexicographic perspective and tend to provide structural, rather than procedural, descriptions. However, as both approaches make claims towards acquisition and processing, they programmatically position themselves as partially cognitive theories of language. Thus, predictions for active language production should be derivable from the interplay of their proposed structures. This is complicated by the fact that some of their core aspects remain vague and underspecified 30 years into their development (Diessel 2023, Ungerer 2024). Some discussion of their underlying principles is hence required for an adequate embedding of the present study, especially in the absence of other, more explicit theories of the lexicon in situated use.

Phraseological and constructionist models of the lexicon agree on some core assumptions, such as the rejection of separability of lexicon and syntax and the importance of lexical material in the acquisition of linguistic structure from input (Bates and Goodman 1999, Boas 2010, 2013, Bybee and Torres Cacoullos 2009, Bybee 2013, Goldberg 2013). In their methodological tradition, both approaches overwhelmingly operate from a bag-of-words approach to lexical production, which discards the order and context of words and constructions from the analysis of corpus data. This appears to be a pragmatic choice that serves the simplification of analytical procedures, e.g., for more accessible normalization against the total number of lexemes or to avoid the heavy load of manual annotation of semantic context. However, this practice also carries epistemological weight, as it marks a dimensionality reduction of a linguistic macrostructure such as a paragraph or a text, which is defined by production time, speaker, situation, etc., into an unstructured set of mutually independent words. Although not explicitly stated, the general acceptance of this detachment and isolation of the occurrences of phrases and constructions from the context of their realization implies a model of lexical production as a series of independent actualizations of global generative patterns represented elsewhere (e.g., the mental lexicon).^[8]

Through these similarities in modeling and empirical study, the two models have found some convergence in the postulation of the phraseological continuum (Granger et al. 2008; Gyllstad and Wolter 2016), and it is not uncommon for research within the field to be rooted in both accounts or view them as complementary or as simultaneously operating on different levels of linguistic abstraction, e.g., as form- vs schema-based perspectives (Ordines 2023, Mellado Blanco et al. 2022, Pavlova 2020, Ziem 2018, Stumpf 2016, Granger et al. 2008, Granger and Meunier 2008, Benigni et al. 2015). However, for the purposes of this study, their philosophical differences with respect to the role of conventionality and their predictions for lexical variability in context are of particular interest. These differences will be highlighted here in spite of the shared community of the two approaches.

Today’s phraseological accounts originated from earlier lexicographic catalogues of idiosyncratically, often non-compositionally combined lexical material, for instance, in the Russian school of phraseology (Zykova 2016, Guliyeva 2016) and in the early computational lexicography of English collocations (Berry-Rogghe 1970, Sedelow 1985). Since then, they have synthesized into a common research paradigm with psycholinguistic and cognitive subdisciplines, which views language production as primarily defined by the reuse of prefabricated lexical material (Siepmann 2012, Siyanova-Chanturia 2015, Wray 2002, 2012, 2013, Pawley and Syder 1983, Ägel 2004, Poulsen 2022, Gyllstad and Wolter 2016, Granger 2005, Schmitt 2004, Conklin and Schmitt 2012, Wood 2015, Frath and Gledhill 2005, Jolsavi et al. 2013, Tavakoli and Uchihara 2019, among others. See Singleton and Leśniewska 2021 and Shadrova 2020 for recent synopses). This programmatic view has perhaps been most prominently formulated in Sinclair’s widely cited idiom principle, which states that “a language user has available to him or her a large number of semi-preconstructed phrases that constitute single choices, even though they might appear to be analyzable into segments” (Sinclair 1991, 110). To Sinclair, communication outside of highly specialized language such as legal registers or poetry can explicitly be attributed almost exclusively to phraseological material (p. 114). Similarly, Wray (2002, 119) formulates ‘formulaic processing as the default’ and ‘construction out of, and reduction into, smaller units by rule [as occurring] only as necessary’. Similar positions were previously formulated by Bolinger (1979), as well as Pawley and Syder (1983), to whom native-like fluency and native-like selection are ‘two puzzles of language use’: too challenging to solve if language was reconstructed from the lexicon for each utterance.

At first glance, it may appear as though phraseology is primarily a study of ‘words that go together’, rather than one of which words occur at all. However, functional arguments presented within the field and the scope of its summarizing contributions suggest that phraseology views itself as a philosophical approach seeking “to answer the age-old question of whether human beings are ‘primarily like buses, which travel along regular routes’ or ‘like taxis, which move about freely”’ (Ding 2018, 144). Moreover, as recombination is a function of selection, there is a strict entailment of the predictable occurrence of particular lexemes in the realization of the chunks that include them, as well as in the statistical associations of words and their coselected elements (Section 2.3). To simplify, if I predictably choose two words together, each one of them will also occur. The fact that the phraseological paradigm does not view them as individually selected does not change that.

Phraseological accounts do not typically make statements regarding lexical invariability or precise predictions as to which chunks will be reused in similar contexts. Rather, they explicitly postulate only the availability of such formulaic material. However, two much cited arguments suggest functionality of limitation, i.e., the reuse of a small and predictable set of chunks or otherwise coselectionally constrained elements.

First, phraseological material is assigned semiotic status. This means that not words, but identically recurrent combinations are viewed as the basic communicative signs (Wray 2002, Sinclair 1991). This more complex composition of communicative signs is suggested to be of functional advantage due to a limitation of choice in context: rather than picking words from sets of many near-synonyms, which may all carry slightly divergent connotations, a community of speakers can settle on one conventional way of expression. Although this necessarily limits referential precision, as situations never recur in truly identical ways, it is argued that speakers and hearers profit from a limitation of semantic decomposition effort and lower contextual ambiguity (Siepmann 2012, Wray and Perkins 2000). A semiotic view of chunks as primary communicative units has also been prominent in first language acquisition research (Tomasello 2000, 2009, MacWhinney 2014, Bates et al. 1988, Borensztajn et al. 2009, Dąbrowska and Lieven 2005, Ellis et al. 2008, Ellis 2006, and others). This functional argument would fail if there was in fact no limitation of choice in context, i.e., if many chunks competed for actualization. In fact, the availability of many chunks for the same situation and the incorporability of multiple near-synonyms into their slots would exacerbate the initially proposed problem: a competition between the full range of semantically plausible chunks is simply a more complex and combinatorially larger version of a competition between the full range of semantically plausible words.

Furthermore, the concept of conventionalization by definition implies the limitation of expression to only a subset of the full generative potential. Even if there may be a few equally conventional ways to refer to a particular event, expanding this to the entire semantically plausible lexical set undoes the phraseological convention. To say that in a particular context, any of the semantically available words are equally conventional is to say that the explanatory power of conventionalization does not extend beyond semantic plausibility.

Second, phraseological accounts of the lexicon have underscored their functional interpretation with a range of cognitive arguments derived from psycholinguistic studies, arguing that formulaic language offers decisive processing advantages in production, perception, and sociopragmatic interpretation by retrieval of lexicalized chunks vs analytical or composite processing (e.g., Conklin and Schmitt 2012, Jiang and Nekrasova 2007, Carrol and Conklin 2020, Schmitt 2004, Weinert 1995, Underwood et al. 2004, Siyanova-Chanturia 2015). The functionality of this minimization of morphosyntactic processing and recombination effort is only given if it is not outweighed by the effort of selection. It is well evidenced that lexical retrieval from large sets of alternatives is slow due to higher inhibition demands (Sullivan et al. 2018, Kroll et al. 2013, Dijkstra 2003, Fargier and Laganaro 2023, Wulff et al. 2016), especially for competing, partially overlapping forms (as would be the case for similar chunks of different lexemes). Thus, if a processing advantage exists for formulaic sequences, it could only persist within a lexically limited space, although perhaps less strictly so than in the semiotic argument.

A global scope of the phraseological view, i.e., the idea that formulaic material is used whenever possible, is further suggested in estimations of the extent of formulaic material, which has been proposed to reach 50% of regular discourse (Conklin and Carrol 2021; Nelson 2018; Erman and Warren 2000) and up to 80% or even 90% for local phenomena (e.g., demonstrative cleft phrases, Calude 2008).^[9] If processing advantages explain phraseological tendencies, they should not be limited to phraseological niches, but expand to cover most everyday situations, as this would provide the largest cumulative advantage (see also Ellis 2008).^[10]

Constructionist approaches and Construction Grammar (CxG, Goldberg 1995, 2006, 2013, Sag 2012, Boas and Sag 2012, Steels 2006, Croft 2001) are of a more generative nature compared to phraseological accounts. They view the lexicon as a set of schematic templates for higher-level abstractions that allow for the blending of exemplars and constructions with one another.

At the same time, the semantic model of Construction Grammar presumes a distribution of words into exact semantic niches. This is explicitly stated for Construction Grammar in the principle of no synonymy (Goldberg 1995, 67): “if two constructions are syntactically distinct, they must be semantically or pragmatically distinct,” and similarly in Croft’s (2001, 111) principle of contrast (see also Leclercq and Morin 2023, Uhrig 2015 for more discussion). The principle of no synonymy is equivalent to the structural model of a very detailed semantico-pragmatic map that connects each linguistic form to clearly delineable meaning including socio-pragmatic information on where and when it can be used. This map, termed Constructicon in CxG, is modeled to contain detailed information on the distributional niches for alternations, verb argument structure preferences, and near-synonyms, and is proposed to emerge from exposure to statistical patterns in use (Gries 2010, 2013, 2015, 2019, Kerz and Haas 2009, Wulff and Gries 2021, Faulhaber 2011, Zeldes 2013, 2012, Diessel 2023, Shadrova 2020).

Although the principle of no synonymy was originally formulated for abstract constructions, it must extend to words, as a continuity between lexicon and syntax (or abstract and concrete constructions) as form-meaning pairs of more or less identical buildup is one of the core principles of the constructionist theory (Bybee 2002, Bybee and Torres Cacoullos 2009, Ellis 2008, Ellis et al. 2008).

As Uhrig (2015) points out, the creation and sustenance of this map entails global invariability: if speakers are highly sensitive to very fine-grained semantic differentiation, they can only convey their intended meaning by realizing the correctly mapped lexemes. If this map is further created from usage and exposure, it would lose precision from variable use by different speakers in the same semantic situation.

The expected invariability of semantic niches raises questions with respect to the place and role of schema-based innovation. Since schemas only ever occur with lexical realizations in language production (one cannot realize an accusative object without using words), schema-based entrenchment entails token entrenchment within the schema (Schmid 2018), which should stabilize as a phraseological object in the semantic map. On the other hand, innovation through the blending and reanalysis of constructions has traditionally been at the center of constructionist modeling. For instance, a speaker may choose to say “she sneezed the napkin off the table,” although they would certainly have the option to express very similar meaning from existing chunks, e.g., [she sneezed][and][caused][the napkin][to fall off the table]. It is not very clear when, if, or why a speaker would be pragmatically required to apply a particular constructional blend to express meaning in context, i.e., whether particular meanings demand innovative realizations. There are no explicit predictions on the type of situation that would necessitate, justify, warrant, or license the use of the causative construction in she sneezed the napkin off the table as opposed to other, e.g., phraseological, variants. In the context of the phraseological continuum, a division between everyday phraseological vs out-of-the-ordinary constructional activity in lexical production could be presumed; or constructional blending may be seen as the generative mechanism contributing to a phraseological potential for conventionalization. Similar questions regarding terminological rigidity and situational prediction vs linguistic abstraction have recently been raised from within the framework itself (Michaelis 2024, van Trijp 2024). Another underdiscussed aspect with repercussions on the lexicon in use concerns the methodological paradigm of both approaches. This will be discussed next.

2.3 The statistical paradigm

Lexicology within usage-based linguistics is marked by the dominance of a statistical paradigm, especially in corpus linguistics (Larsson et al. 2022, Larsson and Biber 2024). This includes in particular a practice of accepting lexicological evidence from probability-based metrics computed from relative corpus frequency. These metrics count the occurrence of words within some particular context and derive conclusions to either the properties of the context in question or the properties of the words themselves in a lexicographic sense. The most influential among them are the following:

1. Lexicographic metrics, including measures of collostructional or word association, mutual information, odds ratio; or productivity measures of types vs tokens. These are usually transformations of the relative corpus frequency of a lexeme or a combination of lexemes against the total number of words (Stefanowitsch and Gries 2003, 2013, 2023, Evert et al. 2017, Pecina 2010, Shadrova 2020). Context is defined by the co-occurring lexemes or constructions within a small window of the target words (e.g., ‘strong’ + ‘coffee’ within five tokens) or by other words of the same class (e.g., the other nouns for noun productivity). Some of these measures are purely descriptive and do not make explicit probabilistic or inferential claims (e.g., type/token ratio), while others argue from an explicitly stochastic perspective with inferential claims towards generalizable features (e.g., significance testing with the Fisher–Yates exact test for collostructional analysis; Bayes’ theorem of conditional probability for Δ P , see below).

2. Regression and mixed-effect modeling, a group of stochastic techniques that probabilistically predict outcomes from combinations of influential factors (Gries and Otani 2010, Gries 2015, Siyanova-Chanturia and Spina 2020, Speelman et al. 2018, Barth and Kapatsinski 2018, Bates et al. 2015, and others). A function is inferred from scales or levels of factors and mapped to a distribution of data points to match it as closely as possible. If the inclusion of a factor improves the congruence between distribution and function, it is interpreted as adding probabilistic prediction value to the model. Context is defined as the environment of factors that predict a variable. In corpus linguistics, factors have included speaker attributes as well as syntactic and semantic environment and concrete lexical variables (e.g., to infer a higher rate of realization of ‘that’ in subclauses that complement particular verb lexemes, see Gries 2021).

3. Clustering techniques, such as topic modeling, multidimensional analysis, and certain applications of distributional semantics (Biber 2019, Biber and Jones 2009, Bruni et al. 2014, Vogt et al. 2023, Shadrova 2021), which infer from a set of co-occurring words to a higher-level linguistic concept such as meaning, genre, or semantic field/topic. Context is not categorially defined in these techniques but emerges as a latent construct from the clustering procedure itself. For this, words and other features are randomly assigned to groups and iteratively reordered until a mathematical function is maximized or converges. The resulting division of data is viewed as linguistically meaningful. The probabilistic aspect here usually refers to the likelihood that a particular combination of words represents one class (topic, genre, register etc.) over another.

Probabilistic reasoning is often introduced to a model in order to soften deterministic predictions and resolve dichotomies, i.e., to allow for proportional and gradient categories with fuzzy boundaries and to include tendencies and potentialities beyond strictly rule-based processes. However, for the lexicon, the mathematical repercussions of interpreting relative corpus frequencies as probabilities have quite the opposite effect, as they integrate the entire lexicon into a single coherent, mutually interdependent, and globally predictive system.^[11] This results in more specific predictions to lexical production in context than may be philosophically intended:

• Generalizability: in frequentist statistics, probabilities are an objective property of the elements of a stochastic system to approximate expected values over time, i.e., in infinite repetitions of an experiment.^[12] The objectivity of that property is often tied to tangible and measurable physical properties. For instance, fair dice are expected to roll an equal number of times onto each of their sides, if – or rather: because – those sides are the same size and shape. This means that words, if their frequency is interpreted as probability, have an objective reason to occur at particular rates independently of the situational parameters (as long as the stochastic system remains intact).

• Stationarity: since probabilities are inherent features of the variable they describe, they must remain the same across time and space, for as long as they belong to the same stochastic system. The interpretation of relative corpus frequency of lexemes as probability strictly entails that elements will recur, within a limited degree of variance/noise, at approximately the same rates and in the same general distributions in comparable samples. Statistical inference is generally a tool to test the predictable recurrence of stable distributional assumptions. If a system is not stationary, i.e., if today’s corpus frequency says nothing about tomorrow’s, samples are not comparable and any metrics derived from the sample are meaningless for the inference from one corpus to the next or from corpus to language (this does not mean that certain metrics could not still be meaningful for description or analysis). Obviously, language is overall not stationary, as it continues to change. However, the statistical paradigm strictly presupposes stationarity of the frame of reference it chooses, e.g., a collection of texts from a particular time, genre, type of elicitation or other class, to remain interpretable.

• Systemhood: since probabilities are defined as proportions of occurrence of variables within a system, an interpretation of corpus frequency as probability entails the systemhood of all words in a corpus. Similarly, assigning probabilities to lexemes in a speaker’s lexicon entails systemhood of their entire lexicon. Words cannot have probabilities independently of all other words, they must be part of a distribution of probabilities that sums up to 1 for the totality of all realizations. If the relative frequencies of words in a corpus are interpreted as probabilities, this means all of the words in the corpus share the same space of potential realization, or simply put, they all compete with one another.^[13]

• Mutual (in-)dependence: the interpretation of the rates of co-occurrence of words with other words, constructions, genres, etc. as conditional probability entails a bidirectionality by which the other words, constructions, genres, etc. also predict the word itself. This is due to Bayes’ theorem of conditional probability, which states that two variables are either mutually independent or mutually dependent. Thus, words and constructions, genres, etc. in this model are predictive of each other.^[14] It is mathematically impossible for a lexeme to predict a construction, register, genre, or other linguistic situation under conditional probability without the construction, genre, register, or other linguistic situation also probabilistically predicting the occurrence of the lexeme. This would require other mathematical techniques, such as Markov Chain Modeling, which carry other entailments and are not currently common in linguistic research.

To accept evidence from the metrics mentioned earlier in this section is to accept these entailments for lexical data. This cannot be avoided with a limitation of scope, as metrics would become ill-defined and any analyses based on them invalid if assumptions were not approximated to an acceptable degree.^[15] Or, to simplify, within the bounds of some expectable variance/noise, accepting this methodological tradition means that, upon knowing the composition of a corpus, I should be able to predict how many of which lexemes will on average occur.

3.1 Data

The present analysis is based on a total of five corpora that were chosen from the currently available corpora of German composed from productions of homogeneous speaker groups in controlled elicitations:

1. Two corpora of argumentative essays, Kobalt (Zinsmeister et al. 2012) and Falko (Reznicek et al. 2012);

2. Two dialogue corpora, BeMaTaC (Berlin Map Task Corpus, Sauer and Lüdeling 2016, map descriptions) and BeDiaCo (Berlin Dialogue Corpus, Belz and Mooshammer 2023, discussions of differences between pictures);

5. The RUEG corpus (corpus of the Research Unit Emerging Grammars in Language Contact Situations, Lüdeling et al. 2024, Wiese 2020, witness reports in four settings – formal written, formal spoken, informal written, informal spoken). Only the RUEG corpus contains parallel elicitations in English and Russian, which are also analyzed in this study.

Table 1 provides an overview of the corpora. Type and token numbers are provided in Table 2 in Section 4.1. All corpora are task-based, i.e., elicited and compiled under strictly controlled conditions. Written prompts are provided in Table 1. Examples of prompt images are provided in Figures 1 and 2. All corpora are small to medium-sized and deeply annotated. Normalizations and annotations have been performed and/or corrected manually in the context of the respective corpus projects (references above and in Table 1). Some corrections have been added by me, and all errors remain my own. All corpora are freely available for download and analysis and can be expanded and enriched with further annotations (see under the respective references).

Figure 1

Diapix Street 1 Baker and Hazan (2011) (reprinted from the public domain with author’s permission, available through https://www.phon.ucl.ac.uk/project/kidLUCID/diapix.php). Participants are prompted to identify differences between the two pictures, such as the missing sign on the letting shop or the shop keepers different suit colors. Diapixes used in BeDiaCo were translated to German (see Belz and Mooshammer 2023).

Figure 2

BeMaTaC instructor and instructee maps. Participants are asked to negotiate map directions based on divergent landmarks in their respective maps, such as the missing box at the top right of the instructor map or the pile of nails vs the chimney sweeper to the left of the map (reprinted from the public domain, available through https://www.linguistik.hu-berlin.de/de/institut/professuren/korpuslinguistik/forschung/bematac/bematac).

Some of the corpora were originally designed for L1/L2 or mono- and bilingual contrastive studies, but for this study, only data from monolingually raised L1 speakers were used.^[20] All participants were residents of a country where the respective language represents the majority societal language at the time of elicitation (Germany, Russia, the US). Speaker groups are homogeneous with respect to age, region, and urbanicity in all corpora. For the essay corpora, speaker groups are also homogeneous with respect to socio-economic status (SES) and education background (Table 1). For instance, speakers in Kobalt are all 12th grade students of the same school in an affluent part of Berlin, all taking the same German class with the same teacher. In that particular case, it can be assumed that speakers also act as mutual interlocutors outside the elicitation, i.e., share a community space of lexical usage or a variety.^[21]

3.2 Hypotheses

As virtually no explicit hypotheses towards lexical production in context have been proposed in the literature, they have to be deduced from the structural principles and mechanisms of the approaches discussed in the previous sections. Lexical production is likely shaped to some degree by all of the phenomena reviewed above – conventionalization/distributional features of lexical elements, constructional activity, situational dynamics, and the impact of variationist factors. In order to assess the contribution of each tendency, it is worthwhile to formulate maximal hypotheses for each of the models. Hypotheses are formulated and labeled according to their operationalization based on observable linguistic items in the corpus, not their theoretical status (degree of lexicalization, phraseological weight, etc.).

1. form-based/holophrastic convergence: Phraseological accounts suggest form-based convergence. If tendencies of between-speaker convergence on the level of surface form dominate, speakers would be expected to use the same words for the expression of similar meaning. They would be more convergent in the elicitation settings requiring description of identical material, i.e., the dialogical and witness report corpora. Speakers should overall avoid derivational word formation if similar meaning can be expressed easily with paraphrase, as newly derived words would always be rarer than frequent words used for simple paraphrase.^[22] Overall, a high degree of overlap between speakers would be expected. As speakers are often encouraged to avoid repetitions of identical lexemes in educational settings, and corpus elicitations might be perceived as educational or testing environments by some, it is possible that speakers consciously add lexemes for variety. This should increase convergence rates, as speakers would then realize a larger proportion of the situationally expected set, allowing for more overlap.

2. distributional niches: If lexical production is primarily guided by the optimization of semantic precision, speakers would be expected to be most convergent in semantically fine-grained distributional niches. Where speakers choose similar sub-discourses in the essay corpora, and where they choose to describe identical details, they would choose identical words. Near-synonyms should be rare, unless they transmit contrasting details in context.

3. distribution by external factor: If distributional tendencies by external factors dominate, as is suggested by variationist perspectives, speaker productions would follow hypothesis 1 or 2, but vary intra-individually between elicitation conditions (only applies to the witness report corpus RUEG). For instance, near-synonyms would be distributed by elicitation condition (formal/informal, etc.).^[23] Conventionalization may also be tied to speaker communities, i.e., present in highly localized varieties. Therefore, the Kobalt essay corpus, whose speakers form a community outside the elicitation context, should be more convergent compared to the other corpora, especially the other corpus of the same format (Falko). It is also possible that speaker cohorts generally divide into smaller varieties. Then, we should find subgroups of speakers that converge less vs more, i.e., multimodal distributions of lexical overlap.

4. schema-based novelty: If the constructionist aspect of schema-based productivity dominates in lexical production, speakers should show evidence of constructional activity (e.g., blending, near-synonyms through derivational morphology), especially for the transmission of unusual or idiosyncratic meaning. Paraphrases should not outweigh novel word formation. Attempts to avoid repetition (see H1) and to signify creativity may also result in higher constructional activity.

5. situationally-dynamic variability: If situational dynamics dominate, speaker productions would lexically diverge. This can be due to differences in interpretation of the situational content, discourse paths or task engagement, and cognitive effects, such as priming/persistence, and would result in idiosyncratic lexical distributions. As word formation activity may occur abundantly as a reflection of cognitive engagement, some overlap with H4 is expected. Speaker productions would vary intra-individually about as much as inter-individually and variability would not primarily occur as register-specific.

Although this study is not of primarily contrastive or typological interest, a note on language specificity seems appropriate. Since German is a morphologically rich language that features many productive native word formation processes, productivity can be expected to be overall higher compared to English, and as high as or lower than in languages featuring even richer morphological productivity, such as Russian. Additionally, German has a well-known graphematic idiosyncrasy in the convention of not splitting compounds, which might yield lower rates of convergence where compounds are plentiful. As the witness report corpus (RUEG) allows for the comparison with parallel elicitations in English and Russian, this will be considered under the following hypothesis:

1. language-specificity: Speakers converge more in languages that allow fewer grammatical operations on lexemes (esp. word formation). English is thus most convergent between speakers, Russian is least convergent, and German is situated between the two.

4.1 Quantitative analysis

Table 2 provides an overview of type and token numbers for each corpus. Speakers are overall remarkably productive across corpora, even those featuring the most communicatively limited elicitations. For instance, German and Russian speakers in the witness report corpus (RUEG) find around 500 different verbs to describe the events of a 40-s video; and in the map task corpus (BeMaTaC), the 14 speakers presented with two maps containing 12 landmark icons each use a total of 114 different nouns to describe them.

4.1.1 Inter-individual variability

Between-speaker lexical overlap is not a very commonly used measure in corpus-linguistic research, although a few applications exist (Sinclair and Fernández 2023, Meylan and Gahl 2014, Horton and Spieler 2007).^[25] Although it abstracts from the lexemes for the comparison between sets, between-speaker overlap requires knowledge of which particular lexemes occur in each text. This makes it a more precise metric of the qualitative similarity of the lexicon between speakers compared to the measures mentioned in Section 2.2, such as between-speaker lexical diversity, sophistication, over- and underuse, etc., which can be computed from only knowing how many types, tokens, etc. of each category the corpus contains.

For the inter-individual analysis, between-speaker lexical overlap is computed for each set of three and four texts in the corpus. For example, in a corpus containing texts doc_1, doc_2 … doc_n, the first set of four texts would include documents 1–4, the second documents 1–3 and 5, the third 1–3 and 6, and so on. These three sets would then overlap in three of four texts (1–3). Lexemes are only counted if they occur in all texts of a set. This is visualized for a set of four texts from the Kobalt essay corpus in Figure 3. Only the central field of the Venn diagram is considered in the following analysis for four-text sets, and the four directly adjacent fields for the three-text sets. Only nouns and lexical verbs (no auxiliary, copula, or modal verbs) are considered in the analysis, as they form the most variable, productive, and open lexical classes, and are also most sensitive to content and situational context. Computations are performed separately for verb and noun lexemes and normalized against the total of each category ( n overlapping verb lexemes out of m verb lexemes used by a speaker – relative overlap will usually differ between speakers, as three verbs may make up 5% of one production and 10% of another; Figure 4).

Figure 3

Venn diagram representing the overlap of verb lexemes between four texts in Kobalt (DEU_002, DEU_003, DEU_004, and DEU_021). Two verb lexemes occur in all four texts, 1, 1, 2, and 0 occur in different combinations of three texts, but not all four. 23, 23, 38, and 22 occur only in one of the texts, i.e., they are not shared with either of the other texts in the set. Percentages represent the proportion of the number of lexemes relative to the total number of lexemes in the set, not each speaker’s individual set (the 23 lexemes that occur only in DEU_002 make up 17.6% of the total number of verb lexemes in the entire set of four documents). Sets of three texts are equivalent to the overlapping fields of three texts in this diagram.

Figure 4

Summary statistics of verb lexeme overlap for sets of three texts from all corpora (mean/minimum/maximum, only matching conditions).

Although the corpora are rather diverse with respect to the genre, communicative situation, and even language they elicit, the results for lexical overlap between speakers are remarkably similar and remarkably low (Figures 5 and 6. Violin plots for all distributions have been moved to the appendix for easier reading). The results can be summarized as follows (Figure 7):

1. All lexemes are shared at a low rate with medians of 0–2 shared verbs 0–6 shared nouns across conditions and languages. Verbs are overall barely shared.

2. The ratio of shared lexemes per speaker (lexemes they share out of all lexemes they use) is surprisingly stable across corpora per condition (Figure 8): for the more constrained corpora, it is at median 21% for nouns and 8% of verbs, and for the essay corpora, it is at about 5% for both nouns and verbs. An exception is the diapix corpus, where verbs and nouns are shared to the same degree. This is likely due to the descriptive character of verbs therein, as a person in the picture would be described with reference to their performed activity.

3. Even for those maximally similar sets, the ratio of shared lexemes lies below 25% for the essay corpora and below 50% in most of the dialogue conditions. Overlap ratios in RUEG have longer upwards tails (some reaching higher ratios), which is largely an artifact of text length variability (higher overlap between very short and rather long productions).^[26]

4. Four-text sets have shorter tails, i.e., fewer sets with above-average overlapping lexemes, compared to three-text sets. This suggests that there is an upper bound to the qualitative similarity at three speakers even within narrowly constrained communicative situations, and very low probability of more speakers converging on the same set beyond that number. There is no evidence of a clustering of groups by similarity in a multimodal distribution. It does not appear as though speakers can be sorted into communities by lexical similarity based on choice of subdiscourse or conventionalization. Speakers in Kobalt do not converge more, but in fact less, than in Falko, in spite of uniquely forming a speaker community outside the elicitation among the corpora presented. Even in the picture description dialogue, speakers do not share more than 25% of noun lexemes per dyad. While a 100% overlap would not be expected, since it is unnecessary for both speakers in a dyad to refer to the same elements in dialogue (one speaker can expand on the previous utterance or confirm it, etc.), it is noteworthy that the ratio of shared lexemes in dialogue is barely higher compared to the ratio of shared lexemes of other, non-interlocuting speakers in the same condition.

Figure 5

Summary statistics of noun lexeme overlap for sets of three texts from all corpora (mean/minimum/maximum, only matching conditions).

Figure 6

Summary statistics of verb lexeme overlap for sets of four texts from all corpora (mean/minimum/maximum, only matching conditions).

Figure 7

Summary statistics of verb lexeme overlap for sets of four texts from all corpora (mean/min./max., only matching conditions).

Figure 8

Proportion of inter-individually shared noun and verb lexemes out of all lexemes for verbs and nouns in all analyzed corpora (only sets of three texts, only matching conditions). Please note the different scales in the top row left vs center and right.

4.1.2 Intra-individual variability

As the witness report corpus (RUEG) was elicited in four conditions for each speaker, it allows for an intra-individual analysis (Figure 9). The four contributions create a 2 × 2 matrix of formality and modality (formal/informal spoken/written). Although speakers might differentiate between register-specific vocabulary, especially between the two formal vs the two informal settings, the descriptive content remains largely identical. High within-speaker lexical overlap for lexemes that are not highly register-specific would hence be expected. Figures 10 and 11 show that this is not generally the case. Most lexemes occur as speaker-specific hapaxes. Speakers rarely reuse any verb or noun lexemes in all four settings. As predicted, the data suggest an effect of formality, which is visualized in Figure 12: if speakers use a lexeme in two settings, it is more likely that the two settings match by formality than modality or that they are unmatched.

Figure 9

Venn diagram representing overlap between the verb lexemes used in four texts by speaker DEmo50 in the German subcorpus of RUEG (fs = formal spoken, fw = formal written, is = informal spoken, iw = informal written). The speaker uses two of their verb lexemes in all four settings (kommen ‘to come’ and bremsen ‘to brake’). Eleven verb lexemes only occur in the formal spoken setting, and ten only in the formal written, etc. The speaker has four ways of using the same lexeme in a single setting (the four settings), six ways of using the same lexeme in two settings (same formality, different modality (2); same modality, different formality (2), opposite modality/formality (2)), and four ways of using the same lexeme in three settings (leaving out one setting each).

Figure 10

Reuse of verb lexemes by the same speaker in RUEG proportionally (a) and in absolute numbers (b). The x-axis describes the number of settings a lexeme is used in. For example, if the verb to run occurs in both the formal written and the formal spoken text of the same speaker, it would count towards the box visualized at 2 on the x-axis. Most verb lexemes occur in only one setting, and very few occur in all four settings for the same speaker (median = 2). The effect is stronger proportionally in English compared to the other two languages mostly because Russian and German are more productive in the verb domain due to the productivity of complex verbs. In absolute numbers, reuse is almost identical between languages.

Figure 11

Reuse of noun lexemes by the same speaker in RUEG proportionally (a) and in absolute numbers (b). The x-axis describes the number of settings a lexeme is used in. For example, if the noun ball occurs in both the formal written and the formal spoken text of the same speaker, it would count towards the box visualized at 2 on the x-axis. Most noun lexemes occur in only one setting. Proportionally, English speakers reuse as many nouns in four settings as they do in two.

Figure 12

Matching of formality or modality in verb (a) and noun (b) reuse. If speakers use the same lexeme in exactly two settings, those settings are likely to be matched by formality, but not modality. For example, if a speaker uses the verb to walk in two settings, it is more likely that these settings are both formal or both informal than it is that they are both written or both spoken, or on opposite ends, such as formal written and informal spoken.

4.2 Qualitative perspectives

This section offers qualitative perspectives on some of the phenomena that create the high degree of variability measured in the data. An exhaustive analysis of the full productivity is impractical to annotate and lies beyond the scope of this study. However, even a rough overview points towards interesting directions for future research. These concern the interface of lexical productivity and discourse granularity, referential diversity, cognitive effects, and task interpretation and interaction with the communicative situation, and will be discussed in that order below.

4.2.2 Morphological productivity and referential diversity

Divergent subdiscourses naturally entail different lexemes. The low between-speaker lexical overlap observed in the essay corpora might then simply be a cautionary tale for elicitation design, setting a higher bar for the required specificity of thematic prompts for controlling lexical parameters. However, speakers in the dialogue and witness report corpora were prompted with visual material in extremely narrow communicative settings and still only share a small proportion of their lexemes with any two other speakers. How so?

The data allows for the identification of three intertwined phenomena: (1) seemingly random choice from relatively large sets of near-synonymous lexemes, (2) many degrees of freedom in the interpretation of all details of the prompt material and task, and (3) a high degree of productivity in word formation that reflects on-the-spot situational processing and socio-pragmatic interaction. The following summary of observations is rather dense, which is unfortunately often the case with lexical data. For the purposes of this study, the many examples seem appropriate and supportive, as they highlight interesting pathways for future research as well as empirical contradictions to the theoretical models discussed in Section 2. However, to avoid unnecessary overwhelm, the reader is encouraged to freely skip ahead to the next subsection when they find their curiosity satisfied. Table 4 provides a more concise overview of examples of the different phenomena.

1. There is abundant use of near-synonyms. Aside from a small number of diatopically or diastratically marked lexemes, this includes many lexemes that do not appear to follow any particular pattern or mark register differentiation. This is particularly the case for verb semantics in the witness report corpus (RUEG), where speakers even individually employ a large set of lexemes in reference to small and seemingly less relevant aspects of the situation. For instance, one of the scenes in the video shows a woman handling groceries by her car. The German monolingual L1 corpus contains at least eight verbs describing the action of putting/stowing/packing groceries into the car: räumen, packen, laden, legen, einräumen, einpacken, einladen, verladen (‘stow/store’, ‘pack’, ‘load’, ‘lay’, ‘stow into’, ‘pack into/wrap’, ‘load into’, ‘load into/away’). At least nine more lexemes occur in another subcorpus of RUEG featuring bilingual L1 speakers of German from Germany: tun, sortieren, einsortieren, reinräumen, reinpacken, reinlegen, reintun, hineinlegen, befüllen (‘to do (put)’, ‘to sort’, ‘to sort away/into’, ‘to stow/load in’, ‘to pack into’, ‘to lay/put into’, ‘to put into’, ‘to lay into’ (variant), ‘to fill up (literally: ‘be-fill’)). Most of these verbs are directional particle verbs formed from the base lexemes to put, to sort, to lay, to load, to pack, to store, to fill), and none appear to be unquestionably marked for register. Almost all are perfectly idiomatic. The most frequent one, packen, ‘to pack’, occurs in about 10% of the productions.

2. Russian speakers use at least five lexemes to describe the placement of groceries in the trunk (клaть, гpyзить, зaгpyжaть, yпaкoвывaть, yклaдывaть – klat’, grusit’, zagruzhat’, upakovyvat’, ukladyvat’ – ‘to put’, ‘to load’, ‘to load up’, ‘to unload (refl.), ‘to pack/wrap’, ‘to lay out/put away’), but, interestingly, overall interpret the event more commonly as unloading, for which they use at least nine more lexemes: paзгpyжaть, выгpyжaтьcя, выгpyжaть, pacпaкoвывaть, yбиpaть, paзбиpaть, paзoбpaтьcя, paзлoжить, выклaдывaть (razgruzhat’, vygruzaht’sa, vygruzhat’, raspakovyvat’, ubirat’, razbirat’, razobrat’sa, razloshit’, and vykladyvat’ – ‘to unload’, ‘to unload/load out of’, ‘to load out’, ‘to unpack’, ‘to pack away’, ‘to sort out’ (2x), ‘to lay out’, and ‘to put out’).

3. The English data does not feature as many near-synonymous verbs, likely due to lower morphological richness and word formation productivity. Still, speakers use at least six different lexemes for the same event: to load, to move, to place, to sort, to pack, and to put.

4. Importantly, individual speakers frequently offer several alternatives within productions of the same condition (e.g., formal written and formal spoken), and even within the same text, as illustrated in Figure 13. Near-synonymous noun lexemes without obvious register or diastratic differentiation occur more rarely, but some examples exist, e.g., in Katzenklappe and Katzentür, ‘cat door’ or Tierhandlung and Tierfachgeschäft, ‘pet shop’ (BeDiaCo - diapix); or English RUEG junction and intersection or carriage and stroller. In the Russian subcorpus, we find competing diminutives, as in пapeнь, paren’ – ‘guy’ , пapнишкa, пapeнёк, parnishka, parenyok; or пёc, pyos – ‘(male) dog’ пcинкa, пёcик, psinka, pyosik. While the use of diminutives itself can imply register differentiation, the use of two different diminutives does not further distinguish between registers.

5. Speakers exhibit high referential diversity in describing the events in all three visually prompted corpora. For instance, the grocery bags in the RUEG video are variably referred to as REWE-Tüte, Papiertüte, and Einkaufstüte (‘REWE bag’ – REWE is a local supermarket chain, ‘paper bag’, ‘plastic bag’). These uses are not synonmous – they reflect differences in the interpretation and specification of events and referents. Referential diversity is abundant with all details of the prompt material. For instance, a man playing with a ball in the RUEG video is variably referred to as father, husband, partner, friend, teenager, son, pedestrian, or passerby in the English subcorpus; the ball is identified as a ball, basketball, volleyball, or baseball in the English data, while many German and Russian speakers identify it as a football. One Russian speaker identifies the woman handling the groceries as a vendor (пpoдaвeц, prodavec). The location of the accident is variably identified as a street/road, intersection/junction or parking lot at a supermarket, grocery store, or apartment complex in the English data.

6. Some of the lexemes are byproducts of playacting, i.e., occur when speakers invent information that was not at all provided in the material. We will return to this in the next subsection. More commonly, the lexemes describe different interpretations and conceptualizations of the situation. Although objectively/externally, all speakers were shown the same material, their experience of the material and their experience of the interaction with the material differ significantly. This results in a wide range of lexical materials for the everyday referents and events they were presented with.

7. Many compounds in the dialogue corpora reflect on-the-spot semantic mapping. This includes lexical oddities as processing artifacts, for example, the incorporation of placeholders (Reparaturladendings – ‘repair shop thingamajig’), compounds with pleonastic (redundant) semantics (Zeituhr – ‘time clock’, commonly just Uhr), or unidiomatic amalgamations (Holzbretterhütte – ‘wood board hut’, commonly Holzhütte or Bretterhütte). Similarly, some referential expressions occur as almost identical word formations, such as Spaßtagbanner, ‘fun day banner’, vs Spaßtagschild, ‘fun day sign’. These cases appear to reflect linearizations of different associations into a single word (wood + board + hut; fun day + ? banner or sign?). This word formation activity is not communicatively required, as syntactic paraphrase with prepositional or relative clause modifiers would certainly be available. In a phraseological model, associative linearization of this kind with the result of idiosyncratic lexemes should be avoided, but it is abundantly found in the data.^[28]

8. More evidence for on-the-spot semantic mapping comes from a degree of specificity that seems unnecessarily high relative to the objective requirements of the communicative situation. For example, one speaker in the German subcorpus of RUEG (DEmo51M) describes the location of the accident as Parkplatz, Parkplatzfläche, Parkfläche, Parkplatzrand, and Bürgersteig (‘parking lot’, ‘parking lot area’, ‘parking area’, ‘end/edge of the parking lot’, ‘sidewalk’), all within the same production (formal written). In the map task corpus (BeMaTaC), speakers refer to an icon of a bread roll as Milchbrötchen ‘milky bread roll’ or Kaiserbrötchen ‘Kaiser bread roll’. In the absence of other types of bread rolls in the prompt material, speakers would not be required to specify the type of roll, and they could use the more general lexeme Brötchen, which would also be the more frequently occurring noun.^[29] Similarly, we find the word Metzgergeselle ‘butcher assistant’, which refers to a trained butcher who has completed a 3-year apprenticeship, but has not obtained the professional master’s qualification required to open their own butchery. It is not obvious why the depicted person – a man holding a steak – should be an assistant, and not simply a butcher (Metzger). It is unclear why this degree of specification would be required either. In the map task corpus, speakers also use highly specific spatial expressions such as Fünfundvierziggradwinkel and Fünfunddreißiggradwinkel, ‘ 4 5 ∘ angle’ and (oddly) ‘ 3 5 ∘ angle’. Communicatively, the more common schräg links, schräg nach oben would appear to suffice (‘diagonally to the left, diagonally upward’).

Table 4

Examples from BeDiaCo (diapix), BeMaTaC (map task), and RUEG (witness report corpus)

Category	Corpus	Examples	Translation
Regional variation	BeDiaCo	Pantoffel vs Hausschuh	Slipper (regional vs general)
	BeMaTaC	Broiler, Brathähnchen, Hühnchen, Hähnchen	Fried chicken (regional variants)
		Möhre, Karotte	Carrot (regional variants)
Near-synonyms with register difference	RUEG-DE	Fahrzeug, Kleinwagen, Auto	Vehicle, car
Word formation with placeholder	BeDiaCo BeMaTaC	Ovalding, Ziehding Reparaturladendings	Oval thing, pull thing repair shop Thingamajig
High specificity	BeMaTaC	Milchbrötchen, Kaiserbrötchen, Metzgergeselle	Milky bread role, Kaiser bread roll, butcher assistant
High specificity with amalgamation	BeDiaCo	Holzbretterhütte	Wood board hut (more common: Holzhütte, Bretterhütte)
Conceptual blending within agreed-upon common ground	BeMaTaC	Toasterladen, Motorradladen, Hafenbar	Toaster shop, motor bike shop, port bar
Near-synonymous	BeDiaCo	Spaßtagbanner, Spaßtagschild	Fun day banner, fun day sign
Word formation		Katzentür, Katzenklappe	Cat door (both variants common)
		Tierhandlung, Tierfachgeschäft	Pet shop (both variants common)
		Stadtfassade, Stadtsilhouette, Stadtkulisse	City facade (uncommon), city silhouette (skyline, rare), city backdrop
Rare lexemes	BeDiaCo	Sackhüpfrennen, Stadtsilhouette	Sack race, city skyline (Skyline more common in German)
Referential diversity	RUEG-DE	Typ, Mann, Kerl, Dude, Junge, Herr	Type, man, dude, dude (English), boy, Sir
(differences in interpretation/identification of referents)		Couple, couple (diminutive), family, man and woman, sir and madam, mother and two sons	Paar, Pärchen, Familie, Mann und Frau, Herr und Dame, Mutter und zwei Söhne
		VW, Renault, Van, Kleinwagen	Volkswagen, Renault, van, compact car
Referential diversity, possibly unintended	BeDiaCo	Quadrat, Quadrant, Quader	square, quadrant, cuboid
High specificity for spatial orientation	BeMaTaC	Fünfunddreißiggradwinkel, Fünfundvierziggradwinkel, Stückchen	3 5 ∘ angle, 4 5 ∘ angle, tiny bit
Meronymy, hyperonymy, hyponymy	BeDiaCo	Treppenstufe, Treppe	Step, staircase
Plenoastic	BeMaTaC	Wohnwagenbild, Nagelbild, Kornähre	Caravan picture, nail picture, grain oar
(redundant semantics)	BeDiaCo	Zeituhr	Time clock
Slip of the tongue	BeDiaCo	Radsteinfeger	? (Rad = wheel, Schornsteinfeger = chimney sweep)
		Schmettersching	Butterfly (should be Schmetterling)

Figure 13

Verbs used with the GROCERIES subevent by three speakers in RUEG-DE. Verbs sometimes occur as the simplex, i.e., just the stem (räumen, ‘to load, to stow’, packen, ‘to pack’), sometimes as complex verbs (einräumen, einpacken, ‘to load/pack into, to stow away’) within the same text; the groceries can be described as being dropped (fallen lassen) or as falling (fallen) by the same speaker.

The cases discussed in this section demonstrate that speakers do not generally pick the most frequent, most general, or most phraseologically predictable word in spontaneous reference or opt for syntactic paraphrase. It is possible that the high degree of specificity is found more in the visually prompted corpora because speakers in these elicitations need to (a) first map the visual input to their own semantic space and (b) formulate references that are recognizable, i.e., can be related to the same material by the interlocutor or understood in a way that the same images would emerge in the imagination of the interlocutor case of the witness report. This may invite a higher degree of specificity compared to the types of argumentative texts elicited in Falko and Kobalt, which are fundamentally intertextual, i.e., relate to more general, external, and abstract concepts and discourses. Moreover, word formation can also be used to signify socio-pragmatic meaning. For instance, compounds such as Ovalding ‘oval thing’ and Reparaturladendings ‘repair shop thingamajig’ may signify social and status information such as a relaxed attitude or a willingness to cooperate and remain within the common ground; or semiotic status, such as referential uncertainty in an ambiguous space (e.g., ‘I am aware and willing to warn you that this is not easy to recognize: the thing I want to refer to is somewhat oval’). The fact that these types of word formations are not abundantly found in the essay corpora suggests that different types of elicitation invite different types of path dependence and cognitive engagement.

4.3 Summary

This section summarizes the results relative to the hypotheses presented in Section 3.2 (- hypothesis not met, + met, ++ met and exceeded). The analysis has yielded no evidence to support high between-speaker convergence in the sense of a limitation to a single or a small set of linguistic forms in context, phraseological or otherwise ( − H 1 ). A 1:1 mapping of individual or groupwise preferences or distributional niches for particular expressions in context, e.g., for common activities such as loading a car with groceries, is not supported by the data. No evidence of an observable set of qualitatively self-similar distributions was found. Thus, the stronger interpretation of transitivity through distributional niches of the constructicon cannot be confirmed at least at the level of individual contextualized productions ( − H 2 ). Rather, speakers make use of what appears to be the full range of near-synonyms in descriptive tasks, and only overlap in very generic thematic vocabulary in the essay productions, especially for verbs. Maximal variability between speakers and between productions of the same speakers appears to be the baseline across elicitation formats, and there is abundant evidence of idiosyncratic, path-dependent speaker engagement with the situation (++H5). This is partially realized through the use of constructional blends for idiosyncratic and socio-pragmatic meaning. However, constructional activity is not limited to this area, but overall abundant in descriptive tasks (++H4). Intra-individual distributions do not primarily vary by elicitation context, and near-synonyms do not divide into groups by elicitation factors. There are no identifiable subgroups of more convergent speakers, and speakers who form a community outside of the elicitation do not converge more than other speakers in the essay corpora. However, as it appears that different types of elicitation foster different types of path dependence, distributional/variationist perspectives might still explain some of the variability on a meta-linguistic level ( + ∕ − H 3 ). Different communicative contexts might in fact work as ecologies for different types of cognitive strategies and occur with predictable diversity in linguistic expression. This might result in the use of more or fewer exemplars of particular types of word formation, without predicting particular identical word forms. Hypothesis 6 regarding language-specificity is partially confirmed (+/-H6). The English subcorpus of RUEG shows overall lower productivity as expected, and English speakers converge more with one another compared to Russian speakers. At the same time, the German subcorpus is the most convergent in RUEG. More research including parallel elicitations of the other formats in other languages is needed to fully understand those effects.

5.1 Revisiting models of lexical production

Sections 2.2 and 2.3 have pointed out that the principle of no synonymy, the idiom principle, and the stochastic principles of entrenchment predict stable lexical distributions by speaker or situation. Let us now revisit this claim with empirical data. In the RUEG corpus,^[32] speakers use 17 different lexemes to refer to the activity of loading groceries into the trunk of a car they witness in the same video: räumen, packen, laden, legen, einräumen, einpacken, einladen, verladen, tun, sortieren, einsortieren, reinräumen, reinpacken, reinlegen, reintun, hineinlegen, befüllen (‘to stow’, ‘to pack’, ‘to load’, ‘to lay’, ‘to stow in’, ‘to pack into/wrap’, ‘to load up’, ‘to load away’, ‘to do (to put)’, ‘to sort’, ‘to sort into’, ‘to stow into’, ‘to pack into’, ‘to lay into’, ‘to do (put) into’, ‘to lay into’, ‘to fill (literally to be-fill)’). This does not say much about synonymity: the verbs are not formally synonymous and would likely occur in complementary distributions in large corpora. Even where the same speakers produce several verbs to refer to the same activity in the same production, we cannot know whether they perceive these words as interchangeable or whether they aim to provide new details with every lexical realization. Then, in a lexicographic sense, the observation may not be striking.

However, if we consider the procedural and cumulative effects of this variability, logical challenges arise for the model of dynamic entrenchment of finely distributed semantic niches. In the RUEG example, the most frequent lexemes packen and einpacken (‘to pack’; ‘to pack/to wrap’) are plausible semantic prototypes in the given situation, but only occur in about 10 and 6% of productions, respectively. This means that well over 80% of productions do not rely on prototypes, or that the most frequent lexemes would not occur 8–9 out of 10 times a similar situation is encountered. In the meantime, regardless of whether or not speakers perceive semantic differences between near-synonyms, their minds would be entrenched with associations between the lexemes of the set and with the situational meaning, as the model expects to ‘wire together what fires together’ (Sprenger et al. 2024). These mutual associations should also make the lexemes more synonymous over time, unless entrenchment is a priori semantically selective.

This is all the more powerful since interlocutors, like corpus linguists, have no way of knowing whether a speaker used a particular verb to relate the same meaning that they would have associated with a different verb, or whether the other person was highlighting another, subtly different aspect of the situation. Did they interpret the situation in a way that is better expressed with einladen than reinräumen, or are the two semantically indistinguishable in their map? Unless speakers engage in meta-communication to reflect on all sensory, perceptual, and interpretational aspects of a situation (which they rarely do), situational ambiguity will always prevail. This may seem trivial, but it has a direct effect on the model by imposing an upper boundary for the potential sharpness of the constructicon: under high lexical variability, neither form, nor meaning would be sharpened by naturalistic interactions. Entropy should then increase with every interaction.

Moreover, to successfully navigate such a space, i.e., to avoid perpetual cognitive dissonance and confusion from slightly off descriptions, speakers must accept and anticipate imprecise mappings between words and situational meaning. Then, even if they started out with an a priori ideal 1:1 map, it would quickly blur into an approximate n : m mapping – referential diversity in a passive entrenchment environment entails semiotic ambiguity. This means that the data presented is inconsistent with an interpretation of a unique and bidirectional (1:1) mapping of the constructicon as far as language production is concerned. The constructicon must either be of a different composition, not divided into stochastically precise semantic niches, or not be applied in a precise manner in production.^[33]

Similar conclusions can be drawn with respect to the role of stochastic learning of those mappings. In the presence of many acceptable slot fillers, the precise acquisition of fine-grained lexicographic distributions from stochastics alone is challenged by limitations of time and space. To remain with the previous example: all of the verbs used by the RUEG speakers seem acceptable for the expressed meaning. All pertain to the set of basic German vocabulary and would not generally be considered as marked for formality. Although loading groceries into the trunk of a car is a frequent activity for many speakers, it would still only occur a few times a week. Let us simplify and maximize exposure by assuming that speakers handle groceries twice a week and talk about it everytime. Let us further assume that they do not repeat lexemes until the entire set is instantiated – this is only for simplicity of illustration; of course, speakers are not expected to actualize the distribution in such orderly ways in reality. Even in this unlikely case, it would take 17 car fillings, or 8.5 weeks, to go through each of the verbs once. Most of the verbs in the set are formed from the directional particles ein, rein, and hinein (all meaning ‘in’/‘ìnto’) and various bases denoting variants of ‘packing’. Speakers in the RUEG elicitation did not actualize the full productive potential of this pattern: they do not use reinladen, for instance, although it is perfectly acceptable and would probably occur in another elicitation.^[34] Iterating through the complete list of plausible verbs would then require more instantiations, i.e., take more time. At this point, no distributional information at all would have been transmitted yet, as each verb would only have been realized once. How many months would it take to instantiate a stable distribution that accurately reflects the different frequencies of each verb as they are found in large corpora even for this everyday activity? A structural interpretation of the constructicon that requires constant reinforcement of the distributional properties of all lexemes does not seem plausible in a field as lexically variable as it is observed in the data presented, especially at scale for the entire lexicon.

Functional arguments towards predictability seem to fail under high local variability as well: using a verb should prime the speaker towards its reuse locally, and add to its global frequency. In a frequency-optimized model, it should then be preferred for the course of the production and in the future as well. The same should go for phraseological models that center conventionalization through convergence on form, or limitation of choice in context. By being used in context, a lexeme should gain phraseological weight and be prioritized in the future for easy access. This should be most valid for the most common and recurrent situations, as to provide the greatest cumulative advantage. But that is not the case – speakers barely reuse any lexemes and actively provide many variants even for the most basic referents and events.

The simplest way to coherently interpret that data is this: each subset of near-synonymous lexemes may reflect a lexically interconnected semantic cluster or network from which lexemes are realized more or less randomly. Higher frequency in production reflects (a) prototypicality in context and (b) genericity/low semantic weight, not idiosyncratic probability features. This view not only explains, but expects higher referential diversity in basic situations:^[35] everytime a type of situation is experienced and expressed, a speaker has a chance to find a new way of perceiving and relating it, which can then be integrated into the long-term association of the frame. We would then find less conventionalization, more diversity, and more innovation in more frequent situations, as well as more lexical diversity over a lifetime.

5.2 Population vs individual and system vs behavior

The degree of variability in the presented corpora suggests that population averages may not scale down to speaker behavior for lexical frequency in a way that optimizes predictability. The most frequent lexemes only occur in a small number of productions, and most lexemes occur at very low rates. This is at odds with the established practice of connecting corpus frequencies to adult speaker behavior via population statistics, highlighting the need for more research into the distributions of individual behavior and their relationship to models of linguistic abstraction. In fact, the high inter- and intra-individual variability and the high degree of local interaction suggests general non-ergodicity and local non-stationarity even within limited semantic fields, which raises more general concerns with respect to the statistical paradigm in lexicology. Even if statistical patterns of pooled averages were stable,^[36] speakers would only have access to the distributions they encounter, which do not seem to converge. There is no guarantee that speakers (a) could gather enough frequential evidence to reconstruct the underlying distributions, (b) would notice differences between the distributions they encounter, would then (c) average out and internalize those averages in order to (d) generalize from these patterns to their own usage. As words are discrete – a slot can only be occupied with a particular lexeme, not a distribution of lexemes or relative frequency of a lexeme – the loss in resolution for the representation of the distribution grows proportionately to the size of the set of variants.

Let it be noted that these conclusions are not incompatible with lexical association experiments that correlate corpus data with response times or collocational preferences based on corpus frequency. These associations may still reflect prototypes/basic level categories, i.e., semantically salient, connotationally most neutral, or featurewise representative members of a category. Their frequency might be elevated due to their position in the ontological space, not vice versa. They may also be neutral association patterns in the absence of contextual alternatives. If the stochastic space changes in context (as would be predicted by the CDS perspective), global associations may only occur as one of many factors in local production. The epistemological problem arises when the two are equated and their presumed equivalence loaded with functional or causal explanations. Similarly, the observation that category learning is influenced by category distributions (e.g., Emberson et al. 2019, Casenhiser and Goldberg 2005) is untouched by this, as a skewed prototype distribution does not require fine-grained stochastic details, but can as well be understood in qualitative terms of feature similarity and abstraction.

More questions arise to the extrapolation from speaker productions to speaker systems. First, what appears as inter-individual variability in the presented corpora may well reflect intra-individual variability, i.e., participants might behave like each other in other elicitations or longer productions. Corpus-linguistic research, where it aspires to cognitive plausibility, is thus in need of repeated elicitations from the same speakers, i.e., longitudinal L1 data beyond acquisition contexts, to disentangle situational dynamics from stable individual and population effects. A better understanding of the relationship between behavior and system requires a clear demarcation of the factors that influence behavior – qualitatively, quantitatively, and from actual production data.

If there is no simple connection between population statistics and individual behavior, models of the relationship between individual behavior and individual system may also require more specification. Is language production an actualization of the existing system (a direct mapping of what already exists in the mental lexicon), or is it an active application of the productive/generative potential of the system?^[37] The degree to which this matters depends on the proportion of such applications we expect to find in corpus data. How much of a speaker’s situated lexical production is a reflection of their stable lexicographic system, and how much is a creative application under local uncertainty, and what does that say about the organization of words? The data presented suggests that productivity in context is actually abundant. This would be relevant to psychologically or cognitively oriented subfields that connect lexicographic evidence with theories of behavior, e.g., language acquisition and language dynamics research; but also to the descriptive adequacy of lexicographic models. For instance, when a speaker produces both ‘parking lot area’ and ‘edge of the parking lot’ (Parkplatzfläche, Parkplatzrand) in the same production, this is likely not a simple actualization of an existing lexicographic mapping to referents in space. In fact, the speaker may not have had a concept or word to speak of ‘parking lot edges’ before this interaction – while not impossible, it seems at least unusual. It may be spontaneously mapped because a) the speaker may be socio-pragmatically expressing their willingness to cooperate and be helpful by providing high resolution details, which may exceed the usual specificity of the parking lot semantics in their system; or b) the two realizations may be traces of a preliminary, and ongoing, categorization of an a priori undelineated amount of sensory input, and not of primarily communicative purpose. Such an explication of inner monologue, an otherwise internally and silently ongoing meaning-making process, may not be representative of the speaker’s lexical system in a lexicographic sense, but an active application thereof.^[38]

Obviously, lexical productivity is a well-known phenomenon. However, in corpus linguistics, it has been treated as the exception, an addition to a largely complete lexicographic map. This is partially due to the fact that it is often impossible to know whether a word is productively formed or lexicalized for the speaker from corpus data alone. However, this problem is a correlate of the global lexicographic approach with lexical productivity: viewing all unique words in a large corpus as a bag of productively formed lexemes does not allow for behavioral analysis in context. The data presented in this article shows that productivity is not only omnipresent in description tasks, but also easy to spot due to many idiosyncrasies and divergences between speakers. In finding words like ‘ 3 5 ∘ angle’ (Fünfunddreißiggradwinkel) and ‘parking lot edge’, we can observe a process of situational meaning making and deduce, with some caution, that these exemplars are not plain realizations of generally used vocabulary (even though they may not be ‘truly novel’), but local applications – especially if none of the other speakers use it. The contrastive analysis of the commonalities and idiosyncrasies in lexical realizations between many different speakers in the same situation and multiple elicitations from the same speakers can thus illuminate aspects of the relationship between speaker system and speaker behavior. Even the analysis of the verbalization of minor aspects of a visual prompt, like the same micro-subevent of loading a car with groceries, has shown to be surprisingly rich in this regard (see also Shadrova et al. in prep.).

5.3 Perceptions of identity

One fascinating aspect to working with lexical corpus data is that readers of the material tend to vastly underestimate the degree of variability. For some reason, text seems to appear more lexically similar than it is – even to linguistic experts: in an informal survey of researchers working closely with the witness report corpus (RUEG), 14 out of 17 participants estimated an intra-individual lexical overlap of 40% between the four elicitation conditions (median estimation = 65%). In reality, speakers on average reuse about 12% of their lexemes in all four conditions in the German and Russian data, and less than 20% in the English data. Survey participants named 13 different verbs as candidates for the most frequent to occur with groceries, four of which do not occur at all. Fourteen participants estimated the prevalence of those verb lexemes at 40% and above (median = 189 productions or 71%). In reality, the most frequent verb occurs in only 10% of texts and was named only by one participant (packen, ‘to pack’).^[39] Not only were participants closely familiar with the entire corpus – every production – but they were also informed on previous occasions that lexical overlap is lower than they might expect. Even so, their estimations exceed observations by half a magnitude.

A similar phenomenon on a meta-scale can be observed in the high estimations of the phraseological extent of everyday speech presented by lexicologists in the past (50% to over 80%, see Section 2.2). Recurrent linguistic structures per se are not evidence of situationally conventional ways of speaking, as situational conventionality can only be measured against chance of occurrence: to find many instances of a combination is only surprising if those numbers are higher than expected relative to the total number of realizations of the meaning in this or other ways. However, the fact that this view seems so natural and has gained so much traction might speak to an as-of-yet undescribed linguistic phenomenon: perhaps in perception, near-synonyms are actually summarized into a common semantic space, i.e., a mapping of n words to one fuzzy meaning complex (e.g., a semantic density field). Like an ‘inner phraseologist’, there might be a function that parses meaning from objectively diverse lexemes into one convenient semantic handle. Since recent research shows that conceptualizations even of common nouns do not align almost at all between speakers (Marti et al. 2023, Wulff et al. 2022), a tendency towards generally approximating, rather than well-defined and precise, semantic interaction seems to be plausible.

The constraints to verbatim production seem to further be defined by the limits of perception and memory. In fact, adult speakers are generally not particularly good at memorizing language verbatim. While humans possess the general ability to memorize long text, for instance, in oral culture or theater, for most, this is also a laborious process that tends to require much focused attention and support through memorization techniques. This view is further supported by Roberts et al. (2023), who found that symbols are remembered better than words of the same meaning, e.g., $ vs dollar.^[40]

The debate continues on whether stochastic learning is indeed a key process in child lexical acquisition (Unger et al. 2023, Kray et al. 2024). In either case, the verbatim precision of this learning may change over lifetime as a physiological effect of nervous system maturation or a structural effect of lexical system connectivity. Distributional learning may be helpful for initialization, but not for the maintenance of the associative meaning structure of many exemplars. Indeed, a recent study by Benitez and Li (2023) shows that children prefer unique labeling and avoid variable multiplex ( n : m ) mappings, while adults do not. However, even children will reformulate sentences in sentence repetition tasks to only include structures and lexemes that they have acquired (passive to active voice, Polišenská et al. 2015; related lexemes like ‘aunt’ to ‘uncle’, Torregrossa et al. 2023), i.e., repeat from language parsing, not from sound mimicking. Along with the well-known fact that the phonological processing of unfamiliar languages is limited to the degree that phonemic differentiation is acquired, these findings challenge the assumption that, even if we had access to a speaker’s entire input corpus, they would generally process the same content that we can observe with corpus-linguistic methods.

Perceived lexical identity of what is objectively a variable lexical space may then be indexical of a parsing activity in which meaning is not triggered (read from a dictionary of constructions in a discrete and sharp 1:1 map), but (inter-)actively reconstructed from all available contextual information, resulting in a complex and multidimensional, but approximate semantic object that is related to many lexemes with partially overlapping semantics. Any of those lexemes might then support the recreation of that object. If stochastics plays a role in this process, it would seem unlikely for it to take scope over the entire lexicon for each localized interaction.