In the usage-based approach, language is seen as a dynamic system that is shaped by domain-general processes, such as conceptualization, analogy and (joint) attention, which are not specific to language but also used in other cognitive domains, e.g., in visual perception or (non-linguistic) memory (Bates and MacWhinney, 1989; Bybee, 2010; Ibbotson, 2020; see also Diessel, 2017). Given a particular communicative intention, speakers have to make a range of linguistic decisions in order to express the intended meaning in an utterance (Levelt, 1989), and listeners have to make similar decisions in order to interpret the elements they encounter in a sentence or phrase (MacDonald et al., 1994). Domain-general processes influence both speaking and listening, which may have long-term effects on the development of linguistic structure if speakers’ and listeners’ linguistic decisions become routinized through frequency or repetition (Diessel, 2019, p. 23–39).

Frequency of language use plays a crucial role in the emergentist and usage-based study of language (see Diessel, 2007; Diessel and Hilpert, 2016 for reviews). Linguistic elements that are frequently used to express a particular communicative intention become entrenched in memory, which does not only make these elements more easily accessible in future language use but may also alter their structure and meaning: Frequent expressions are prone to undergo phonetic reduction, semantic bleaching and chunking and may develop into lexical prefabs, grammatical markers or bound morphemes (Bybee and Hopper, 2001; Bybee, 2010).

The dynamic view of linguistic structure poses new challenges to linguistic theory. In particular, it makes it necessary to reconsider the format of linguistic representations. Traditionally, linguistic representations are derived from a small set of primitive categories and rules, or constraints, that are defined prior to the analysis of any particular structure. In this approach, grammatical categories, such as noun, case and phrase, are used as “tools” for analyzing stable and discrete representations of linguistic structure (Jackendoff, 2002, p. 75). However, if we think of language as a dynamic system, there are no primitive concepts of grammatical analysis and linguistic representations are emergent and transient. I use the term “emergent” in the sense of systems theory (Thelen and Smith, 1995) for a particular type of development whereby a complex phenomenon evolves from the interaction of many parts whose accumulated properties are not sufficient to explain the holistic properties of the phenomenon they created; and I use the term “transient” for phenomena that are in principle always changing—that never really reach a fixed state.

Over the past 25 years, linguists and psychologists have produced a large body of empirical results supporting the emergentist view of linguistic structure (e.g., MacWhinney, 1999; Tomasello, 2003; Bybee, 2010). However, while researchers agree that linguistic structure is emergent and transient, they have not yet developed an explicit theory or model to explain the various findings and to generate specific hypotheses for future research. To be sure, there are some interesting proposals as two how frequency and experience shape linguistic structure and how emergent linguistic knowledge is represented in speakers’ minds. Yet, many of these proposals are too vague and general in order to provide a structured model of grammar.

For instance, some scholars have argued that exemplar theory provides a good framework for analyzing linguistic structure (e.g., Abbot-Smith and Tomasello, 2006; Bybee, 2006; Bod, 2009). On this view, all aspects of linguistic knowledge are represented by a cluster of similar tokens that reflect a language user’s experience with particular linguistic elements. Similar tokens overlap in memory and strengthen the activation value of linguistic representations, which in turn may influence their future use.

Exemplar theory has been quite successful in modeling the emergence of speech sound categories (Johnson, 1997; Bybee, 2001; Pierrehumbert, 2001); but when it comes to grammar, exemplar theory provides nothing but a crude approximation of the effect of frequency on a speaker’s linguistic knowledge. Of course, like all other linguistic elements, grammatical categories are reinforced in memory through repetition; but this is not sufficient to explain how grammatical structure is derived from language use (see Diessel, 2016 for discussion).

Grammar is a highly complex system that involves schematic representations and different types of categories that interact with each other in intricate ways. Both abstract schemas and interacting categories are difficult to explain in a pure exemplar model. In order to analyze the emergence and interaction of grammatical categories, one needs a different approach that takes into account the full range of domain-general processes (and not just exemplar learning) and that differentiates between different aspects of linguistic knowledge (e.g., semantic vs. syntactic knowledge, schematic vs. lexical knowledge) and different types of categories (e.g., word class categories, phrasal categories, grammatical relations).

In this paper, I argue that network theory (Baronchelli et al., 2013; Barabási, 2016) provides a useful framework for the analysis of grammar in the emergentist approach (see Bates and MacWhinney, 1989 for an early network model of grammar). Network theory is based on mathematical graph theory and has been used by researchers from various disciplines to investigate a wide range of phenomena including electric power systems, economical systems, traffic systems, social relationships, the brain and the World Wide Web (Buchanan, 2002; Sporns, 2011; Barabási, 2016). Like exemplar theory, network theory can explain emergent phenomena; but the network approach is much more powerful than the standard model of exemplar theory.

The basic structure of a network model is simple. All network models consist of two basic elements: (i) nodes, also known as vertices, and (ii) connections, also known as links, arcs or relations. However, there are many different types of network models with different architectures, different mechanisms of learning and change, and different measurements for the emergence of structure (Barabási, 2016), making network theory a very powerful instrument for analyzing complex (adaptive) systems such as a person’s linguistic knowledge.

Network models are widely used by cognitive scientists to analyze the mental lexicon (see Siew et al., 2019 for a recent review) and have also been invoked by usage-based linguists to explain certain grammatical phenomena such as morphological paradigms (Bybee, 1995; Hay and Baayen, 2005) and the taxonomic organization of constructions (Goldberg, 1995; Hilpert, 2014). However, while these accounts have shed new light on some aspects of linguistic structure, grammatical categories have hardly ever been analyzed within a network model (but see Croft, 2001). In fact, although usage-based linguists agree that grammatical categories are emergent and transient, in practice, they often use them as predefined concepts, similar to the way grammatical categories are used in the “toolkit” approach (Jackendoff, 2002, p. 75).

Challenging this practice, the current paper argues that grammatical categories, such as noun, noun phrase and case, are best analyzed in the framework of a structured network model in which all grammatical concepts are defined by particular types of links or relations that specify associations between different aspects of a speaker’s linguistic knowledge. The approach is inspired by connectionism (Elman et al., 1996; Chang et al., 2006) and draws on research in morphology (Bybee, 1995; Hay, 2003) and construction grammar (Croft, 2001; Bybee, 2010; see also Diessel, 1997, 2015). However, it differs from all previous accounts in that it proposes a specific network architecture for the analysis of particular grammatical concepts. Concentrating on some of the most basic concepts of syntax, this paper considers the following phenomena:

As we will see, all of these phenomena can be analyzed as dynamic networks shaped by domain-general processes of language use. The paper builds on ideas presented in Diessel (2019), but these ideas will be discussed from a different perspective and in light of other data. We begin with one of the most basic concepts of usage-based research on grammar, i.e., the notion of construction.

In accordance with many other researchers, I assume that linguistic structure consists of constructions that combine a particular form with meaning (Goldberg, 1995, p. 5). However, contrary to what is sometimes said in the literature, constructions are not primitive units, as, for instance, suggested by Croft (2001):

I agree with Croft that syntactic categories (e.g., noun, verb) and grammatical relations (e.g., subject, object) are non-basic and derived; but I disagree with the claim that constructions are basic and primitive. It is not entirely clear what Croft means with this, but contrary to what the above quote suggests, I maintain that constructions are emergent and transient like all other aspects of linguistic structure. Specifically, I claim that one can think of constructions as networks that involve three different types of associative relations: (i) symbolic relations, connecting form and meaning, (ii) sequential relations, connecting linguistic elements in sequence, and (iii) taxonomic relations, connecting linguistic representations at different levels of abstraction (Diessel, 2019, p. 41–112; see Schmid, 2020 for a related proposal).

Traditionally, argument structure is determined by verbs (Levin and Rappaport Hovav, 2005), but in construction grammar, argument structure is not just a matter of verbs but also of constructions (Goldberg, 1995). Verbs select a set of participant roles and argument-structure constructions provide slots for certain semantic types of participants. If a verb and a construction specify the same participant roles, they are semantically compatible with each other and may fuse. This is, in a nutshell, Goldberg’s Semantic Coherence Principle (Goldberg, 1995, p. 50), which has been very influential in the constructivist approach to the analysis of argument structure. However, this principle is not without problems. As I see it, there are two general problems that can be easily resolved if we think of argument structure as a network.

The first problem is that there are many idiosyncrasies. In Goldberg’s theory, fusion is a matter of semantic compatibility, but very often fusion is not semantically motivated. Take, for instance, the double-object construction (She gave her friend a present), which denotes an act of transfer and typically occurs with transfer verbs, e.g., give, send, offer, bring. Most of these verbs also appear in the to-dative construction (She gave a present to her friend), but there are various idiosyncrasies. Donate and say, for instance, designate transfer—physical or communicative transfer—like give and tell; yet, unlike give and tell, donate and say occur only in the to-dative construction (She donated some money to the Red Cross; He said no to her) but not in the double object construction (^∗She donated the Red Cross some money; ^∗He said her no). Conversely, there are verbs such as forgive and envy that may occur in the double-object construction (She forgave him his faults; I envy you your car), although these verbs do not denote any obvious sense of transfer (Goldberg, 1995, p. 130).

Goldberg is aware of these idiosyncrasies and considers them “exceptions” (Goldberg, 1995, p. 129–132); but since lexical inconsistencies of this type are very common, some scholars have questioned the importance of high-level schemas for the analysis of argument structure. In particular, Boas (2003, 2008) has argued that argument-structure constructions are organized around particular verbs, or narrow verb classes, and that fully schematic constructions are only of minor importance to the analysis of argument structure (see also Faulhaber, 2011).

A related problem is that current theories of argument structure do not account for the statistical asymmetries in the distribution of individual verbs. As many corpus linguists have pointed out, verbs and constructions are skewed in their distribution. Give, for instance, is more frequent in the double-object construction than statistically expected and less frequent than expected in the to-dative construction; but for bring it is the other way around (Gries and Stefanowitsch, 2004).

Lexical idiosyncrasies and asymmetries have also been noted with regard to many other types of argument-structure constructions. Consider, for instance, the active-passive alternation. Most transitive verbs can appear in both active and passive voice, but in some languages, the active-passive alternation is not fully productive. German, for example, has a number of transitive verbs (i.e., verbs selecting an accusative object) that do not occur in passive voice, e.g., kennen “to know,” wissen “to know,” besitzen “to own,” kosten “to cost,” bekommen “to get” (Eisenberg, 2004, p. 128–130). In English, most transitive verbs can be passivized (a notable exception is the main-verb use of have, see below); but there are statistical biases in the distribution of individual verbs. For example, the verbs get, want and do occur with a higher frequency ratio of active/passive uses than one would expect if the co-occurrence of verbs and constructions was random; but for the verbs use, involve and publish it is the other way around: They are biased to appear in passive voice (Gries and Stefanowitsch, 2004, p. 109).

Both the item-specific constraints on the occurrence of individual verbs and the distributional asymmetries in the co-occurrence of particular verbs and argument-structure constructions are motivated by general conceptual and discourse-pragmatic factors (e.g., Pinker, 1989; Goldberg, 1995). Nevertheless, they are not strictly predictable from these factors. There are, for instance, no obvious semantic or pragmatic reasons why the main-verb use of have, meaning “to own” or “to possess,” cannot be passivized given that the verbs own and possess are frequently used in passive voice (e.g., The farm was owned by a wealthy family; He was possessed by a devil); and there is also no obvious semantic or pragmatic reason why the English verb know can appear in passive voice while its German counterparts wissen and kennen are banned from the passive construction.

Taken together, these findings suggest that speakers “know” how individual verbs are used across argument-structure constructions, on top of any semantic or pragmatic factors that may motivate their use in a particular construction. Considering these findings, I suggest that argument structure is best analyzed in the framework of a dynamic network model in which verbs and constructions are related by filler-slot associations that are determined by two general factors: (i) the semantic fit between lexemes and constructions (i.e., Goldberg’s Semantic Coherence Principle), and (ii) language users’ experience with particular co-occurrence patterns (cf. 5) (see Diessel, 2019, p. 121–141 for a more detailed account).(5)

Good evidence for this hypothesis comes from psycholinguistic research on sentence processing. For instance, Trueswell (1996) showed that the processing difficulty of (reduced) passive relatives varies with the frequency with which individual verbs occur in passive voice. Since a verb such as consider is much more frequent in the passive than a verb such as want (Gries and Stefanowitsch, 2004, p. 109), passive relatives including consider cause significantly fewer processing problems in comprehension experiments than passive relatives including want (cf. 6a–b).

Similar effects have been observed in psycholinguistic research with other types of constructions and other verbs (e.g., MacDonald et al., 1994; Spivey-Knowlton and Sedivy, 1995; Garnsey et al., 1997), supporting the hypothesis that speakers’ knowledge of argument-structure constructions includes filler-slot associations between individual verbs and the verb slots of particular constructions.

The same network approach can be applied to grammatical word classes and phrase structure (Diessel, 2019, p. 143–171, 191–195). Traditionally, word class categories are seen as properties of lexical items (e.g., tree is a “noun”), but one can also think of word classes as slots of constructional schemas. Consider, for instance, the contrast between nouns and verbs in English. There are morphological, phrasal and clausal constructions including noun slots and verb slots, to which I refer as N/V-schemas (cf. 7a–c).(7)

Following Croft (1991, p. 36–148) and Langacker (1991, p. 59–100), I assume that word class schemas give rise to particular conceptualizations of lexical expressions in order to use these expressions for particular speech act functions. N-schemas conceptualize the content of a lexeme as a non-relational and a-temporal entity that is used to perform an act of reference; whereas V-schemas conceptualize the content of a lexeme as a relational and temporal entity that is used to perform an act of predication. The lexeme fax, for instance, refers to an entity if it occurs in an N-schema (cf. 8a), and it designates a process if it occurs in a V-schema (cf. 8b).

N/V-schemas attract particular semantic types of lexical items: items that designate an entity such as the word table typically occur in N-schemas; whereas items that designate an action, such as the word drink, tend to occur in V-schemas (see Croft, 1991, p. 87–93 for quantitative corpus data from several languages supporting this analysis). However, crucially, while the co-occurrence of lexemes and word class schemas is semantically motivated, this is not just a matter of semantics but also of experience. Speakers “know,” for example, that a word such as crime is exclusively used in N-schemas despite the fact that crime designates an action, and they also “know” that table and drink appear in both N-schemas and V-schemas despite the fact that table (in its basic use) designates an entity and drink an action. In other words, speakers associate particular lexemes with specific word class schemas and the strength of these associations is again determined by two factors: the semantic fit between lexemes and schemas, and language users’ experience with particular co-occurrence patterns (cf. 9).(9)

The network approach to nouns and verbs can be extended to other word classes and subclasses (Diessel, 2019, p. 157–171). Count nouns and mass nouns, for instance, are expressed by different types of N-schemas. In English, count noun schemas construe an item as a bounded entity (e.g., That’s a cake), whereas mass noun schemas construe an item as an unbounded substance (e.g., I like cake) (Talmy, 2000, p. 50–55). Both schemas are associated with alternating and non-alternating lexemes (cf. 10). Cake, for instance, is an alternating lexeme, whereas cat is non-alternating (e.g., That’s a cat vs. ^∗I love cat).(10)

The associations are semantically motivated and entrenched by frequency of language use, but speakers can create novel connections, as in the oft-cited example There was cat all over the driveway, which nicely illustrates that the English mass noun schema evokes a particular conceptualization if it is applied to a new item (Langacker, 2008, p. 128–132).

Note that while word class categories are defined by semantically motivated filler-slot relations, they are also influenced by formal considerations. For example, speakers of English associate particular verb forms with particular past tense schemas based on their phonetic properties (Bybee and Slobin, 1982), which is readily explained by filler-slot relations. To illustrate, the vast majority of English verbs form the past tense by adding the -ed suffix (e.g., walk → walked). However, given a nonce verb such as spling, speakers may create the past tense form splang based on the phonetic similarity between spling and certain “irregular” verbs such as sing that form the past tense by changing the vowel [I] to [æ] (cf. 11) (Bybee and Modor, 1983).(11)

The formation of the English past tense has been a showcase for the power of the network approach in early research in connectionism (Rumelhart and McClelland, 1986). However, if we think of nouns and verbs in terms of networks (as in 11), the same approach could also be used to model the emergence of grammatical categories, if the input nodes and output nodes of a (neural) network are specified for certain conceptualizations and speech act functions.

Finally, the network approach sheds new light on cross-linguistic aspects of word classes. Most European languages have roots that are categorically linked to N-schemas or V-schemas (English is unusually flexible in this regard). However, there are other languages like Nootka (Jakobsen, 1979) and Mundari (Evans and Osada, 2005) in which lexical roots are linked to both N-schemas and V-schemas with almost no restrictions (cf. 12).(12)

This has led some researchers to argue that languages like Nootka and Mundari do not distinguish between nouns and verbs (e.g., Jelinek, 1995), but this claim is potentially misleading as it restricts the analysis of grammatical word classes to “lexical nouns and verbs.” While lexical roots are categorically unspecified in Nootka and Mundari (with some minor restrictions; Jakobsen, 1979), there is no doubt that these languages have formally distinct N/V-schemas in which lexical roots are used as nouns and verbs for reference and prediction (Croft, 2001). Recent research in typology has questioned the existence of language universals, including the existence of universal word classes (Evans and Levinson, 2009). However, the distinction between N-schemas and V-schemas appears to be a universal trait of language that is foundational to the cognitive and linguistic organization of grammar (see Diessel, 2019, p. 152–161 for discussion).

Like word classes, constituent structure involves filler-slot relations. The best evidence for the traditional toolkit approach comes from the analysis of syntactic constituents (Jackendoff, 2002). In generative grammar, syntactic constituents are discrete building blocks that are combined to larger structures by a set of phrase structure rules (in older versions of generative grammar) or a single syntactic operation called “merge” (in recent versions of generative grammar). The resulting structures are commonly represented in phrase structure graphs consisting of nodes and arcs that could be interpreted as some kind of network (Diessel, 2019, p. 172–173).

However, while phrase structure graphs bear some resemblance with network models, the traditional approach to the study of constituent structure is not consistent with the emergentist view of grammar in the usage-based approach. If we think of grammar as an emergent phenomenon, we need a more dynamic model of grammar that explains how constituent structure is derived from language use.

In what follows, I argue that traditional phrase structure graphs can be re-analyzed as dynamic networks of interrelated constructions. In order to understand the dynamics of these networks, one has to consider both the processes that give rise to syntactic constituents and the processes that explain how the various phrasal constituents are related.

In the three previous sections, we have been concerned with filler-slot relations. In the remainder of this paper, we will consider constructional relations, which specify associations between constructions at the same level of abstraction. Constructional relations have long been ignored in usage-based construction grammar, but a number of recent studies have argued that constructional relations, also known as lateral or horizontal relations, are key to understand grammatical phenomena (e.g., Diessel and Tomasello, 2005; van Trijp, 2010; Van de Velde, 2014; Norde and Morris, 2018).

Constructional relations can be divided into two basic types: (i) relations of similarity, which constitute construction families, and (ii) relations of contrast, which constitute paradigmatic alternatives of grammatical categories such as voice, case and number (Diessel, 2019, p. 199–248). We begin with the latter.

Paradigmatic alternatives are related constructions, such as active and passive sentences or singular and plural nouns, that are commonly seen as members of particular grammatical categories such as voice and number. In formal syntax, paradigmatic alternatives have been analyzed in terms of syntactic or morphological derivations. Construction grammar has abandoned the idea that linguistic structures are derived from one another or from underlying representations. Nevertheless, like any other grammatical theory, construction grammar must account for alternating categories such as active and passive voice.

If we think of grammar as a network, paradigmatic alternatives constitute pairs of horizontally related constructions. Crucially, one of the alternating categories typically serves as the default. For instance, in the case of voice, the active construction functions as the default: active sentences are more frequently used than passive sentences (Biber, 2006) and occur within a wider range of contexts (Weiner and Labov, 1983). Moreover, the linguistic encoding of active and passive sentences is asymmetrical. As it turns out, across languages, passive constructions are often marked by an extra morpheme, as illustrated by example (18b) from Sre (Mon-Khmer, Vietnam), in which the passive verb is marked by a particular passive prefix. Note, also, that in addition to the passive prefix, the agent of a passive sentence in Sre is marked by a preposition that does not occur in the corresponding active construction (cf. Engl. This letter was written by John).

Encoding asymmetries of this type also occur with many other grammatical categories including number (car vs. car-s), tense (walk vs. walk-ed), aspect (go vs. is go-ing), case (car vs. car’s), degree (beautiful vs. more beautiful) and polarity (He is lazy vs. He is not lazy). Linguistic typologists refer to these asymmetries as structural markedness (Croft, 2003; see also Greenberg, 1966). Markedness is an important concept of grammar that is readily explained within a network model (Diessel, 2019, p. 223–248).

Since the occurrence of an extra morpheme correlates with frequency of language use, it has been argued that the encoding asymmetries of grammatical categories are shaped by domain-general processes of language use (Haspelmath, 2008; Haspelmath et al., 2014). Specifically, we may hypothesize that frequency of language use gives rise to particular linguistic expectations. To simplify, all else being equal, listeners expect speakers to use the more frequent member of an alternating pair of constructions. Yet, if, for whatever reason, the less frequent member is used, speakers may find it necessary to indicate their choice of construction by an extra morpheme (Kurumada and Jaeger, 2015). The best example for this is perhaps the alternation of polarity constructions. Since the majority of sentences are affirmative, negative sentences usually include a negative marker (cf. 19).(19)

This strategy of morphological flagging is arguably the driving force behind the emergence of structural markedness (Diessel, 2019, p. 223–248). The default construction is often “zero-coded” (Haspelmath, 2006, p. 30), whereas the less frequent member takes an extra morpheme (cf. 20).(20)

This does not only hold for syntactic constructions, such as active and passive sentences, but also for morphological constructions including inflectional categories such as number and case. Consider, for instance, the following forms of the noun pa. t. ti meaning “dog” in (21) from Malayalam (Dravidian, India).

In Malayalam, nouns are inflected for number and case, which is usually described as a morphological paradigm consisting of a lexical root and a set of inflectional affixes. However, in construction grammar, each word form constitutes a construction in which the root is stored and processed together with a sequentially related affix (or string of affixes). The various word forms constitute a network that reflects language users’ experience with individual members of the paradigm (cf. 22).(22)

As can be seen, the various word forms differ in terms of frequency (as indicted by the strength of the boxes), which correlates with the occurrence of grammatical markers. The most frequent word form is nominative singular, which is formally unmarked, as it functions as the default. All other word forms carry at least one extra marker (for number or case), and plural nouns in accusative case take two markers (for both number and case), as they are the least frequent and least expected member of the paradigm.

What this example shows is that every construction has a particular “ecological location” in the grammar network that is defined by its relationship to other constructions in the system (Diessel, 2019, p. 223–248). This does not only concern paradigmatic alternatives of grammatical categories such as voice, number and case, but also groups of similar constructions, to which I refer as “construction families.”

The term construction family is used in analogy to the notion of lexical family in the study of the mental lexicon, which is commonly characterized as an association network (Anderson, 1983; Dell, 1986). In order to explore the structure of this network, psycholinguists investigate how lexemes are accessed in online language use (Collins and Loftus, 1975; Anderson, 1983; Schreuder and Baayen, 1997).

Lexical access is a competition process that is determined by several factors. First, all else being equal, frequent items are more easily accessed or activated than infrequent ones (Forster and Chambers, 1973 among many others). Second, lexical access is facilitated by priming: if the target word is preceded by a lexical prime, it is more easily activated (Dell, 1986). And third, lexical access is crucially influenced by neighborhood density, which refers to the number of items that are phonetically and/or semantically similar to the target word. The word cat, for instance, has many phonetic neighbors, e.g., rat, hat, vat, pat, mat, bat and at, whereas cup has only a few, e.g., cut, up. Neighborhood density can slow down lexical access in word recognition tasks (Luce and Pisoni, 1998), but has facilitatory effects on the activation of lexemes in speech production (e.g., Dąbrowska, 2008) and word learning (e.g., Storkel, 2004). Taken together, these findings have led psychologists to characterize the mental lexicon as an activation network in which lexemes are grouped together into families of semantically and/or formally similar expressions (cf. 23).(23)

Like lexemes, constructions are organized in families of semantically or structurally similar grammatical patterns that influence each other in processing and acquisition (Diessel, 2019, p. 199–222). Construction families share some properties with paradigmatic alternatives such as active and passive sentences (see above). Yet, in contrast to the latter, the members of a construction family are only loosely associated with each other. They do not form tightly organized paradigms of grammatical categories such as voice and number, but are open-ended groups of constructions that do not (usually) exhibit the encoding asymmetries to which typologists refer as markedness. Consider, for instance, the following examples of the English resultative construction (cf. 24).

Resultative constructions designate an action that puts an NP argument into a particular state (Boas, 2003). Like many other argument-structure constructions, resultative constructions vary along several parameters. They generally include a resultative element, but this element can be an adjective or a prepositional phrase (24a–b). The verb is usually transitive, but there are also intransitive resultative constructions (24c). If the verb is transitive, the direct object may or may not be selected by the verb (24a–b vs. 24d). If the verb is intransitive, the construction either lacks a direct object (24c) or includes a “fake object,” usually a reflexive pronoun (24e) (Boas, 2003, p. 4–8). Considering this variation, Goldberg and Jackendoff (2004, p. 535) argued that resultatives do NOT form a “unified phenomenon” but “a sort of family of constructions.”

Like resultatives, copular clauses constitute a family of semantically and formally related constructions (e.g., Hengeveld, 1992 and Stassen, 1997). In English, for example, the copula be may be accompanied by a nominal, an adjective or a prepositional phrase (25a–c). If be is followed by a nominal, the copular clause may express identity (25a) or existence (25d); and if be is followed by an adjective, the copular clause expresses either a permanent state (25b) or a transitory event (25e) (which in some languages are formally distinguished by the use of different copular verbs, e.g., Spanish ser vs. estar).

Crucially, while the members of a construction family may be subsumed under a constructional schema, they are also horizontally related to one another. One piece of evidence for this comes from structural priming. Like lexemes, constructions prime each other (see Pickering and Ferreira, 2008 for a review). In the simplest case, the priming effect is caused by the prior use of the same construction. For instance, as Bock (1986) demonstrated in a pioneering study, speakers’ choice between the double-object construction (e.g., Give me the money) and the to-dative construction (e.g., Give the money to me) is crucially influenced by the prior use of these constructions. If the previous discourse includes a double-object construction, speakers tend to describe a scene depicting an act of transfer by a double-object construction, but if the previous discourse includes a to-dative construction, they are likely to describe the same scene by a the to-dative (cf. Bock and Griffin, 2000; Gries, 2005).

Crucially, structural priming does not only occur when prime and target have the same structure; it also occurs with distinct but similar constructions, suggesting that these constructions are related in speakers’ linguistic memory. For example, Bock and Loebell (1990) showed that sentences including a directional prepositional phrase prime the to-dative construction (cf. 26a–b), and Hare and Goldberg (2000) showed that sentences including a verb such as provide (with) prime the double-object construction (cf. 27a–b). In the first case, prime and target have similar structures but different meanings, and in the second case, they have similar meanings but different structures.

Taken together, this research suggests that argument-structure constructions are organized in construction families with overlapping structural and/or semantic properties (cf. 28) similar to lexical expressions in the mental lexicon (see 23 above).(28)

To conclude, there is a large body of empirical results supporting the usage-based view of linguistic structure as a dynamic and emergent phenomenon. However, there is no consensus in the usage-based literature as to how the many results can be explained and integrated into a coherent model. In particular, the analysis of syntactic phenomena is unclear in this approach.

In this paper, I have argued that linguistic structure is best analyzed within a dynamic network model of grammar. The general idea has been expressed in previous studies. In fact, usage-based linguists seem to agree that grammar constitutes some kind of network (Langacker, 2008; Bybee, 2010). However, while the network view of grammar is frequently invoked in the usage-based literature, it has not yet been developed into an explicit theory or model.

In this paper, I have proposed network accounts for several core concepts of syntax including the notion of construction, grammatical word classes and constituent structure, which are commonly treated as primitive concepts of syntactic analysis. However, as we have seen, all of these concepts can be analyzed as emergent phenomena if they are construed as networks.

At the heart of the proposed analyses is a set of associative relations that concern different aspects of a speaker’s linguistic knowledge and that are shaped by various cognitive processes. Specifically, I have proposed the following set of relations:

Taken together, the proposed relations provide a framework for the analysis of a wide range of grammatical phenomena as emergent concepts. In Diessel (2019), I have proposed additional network analyses for other grammatical phenomena and have discussed some of the topics of the current paper in more detail. Let me conclude with some general remarks on future research. There are many open questions, but here are three general points which, I believe, are of particular importance.

First, the various associative relations have different properties. For example, while one might assume that symbolic relations involve bidirectional associations between form and meaning, sequential relations are unidirectional in that sequential relations have an inherent forward direction, as evidenced by the fact that the language users anticipate upcoming elements in the speech stream. Each relation is influenced by particular cognitive processes and has specific properties that have to be investigated in more detail. This requires both experimental research and computational modeling. There are various computational frameworks using network models, but the conceptual and computational tools of Network Science appear to be particularly useful (Barabási, 2016). These tools have been used in psycholinguistic research on the mental lexicon (e.g., Vitevitch, 2008), but have not yet been used in research on grammar. Second, constructions and lexemes are the basic units of the grammar network. Construction grammar has emphasized the parallels between lexemes and constructions—both are commonly described as signs or symbols. Yet, in this paper, I have argued that constructions are best analyzed as networks and that the symbolic associations of constructions and lexemes have different properties. In my view, the notion of construction has to be revised in the context of a dynamic network model, but this needs careful consideration.

And third, the grammar network has been devised to account for dynamic processes in both language change and language acquisition. The latter comprises L1 acquisition and the acquisition of a second language. There are conspicuous parallels between language change and language acquisition (Diessel, 2011, 2012), but there are also differences between them. For instance, language change is influenced by social factors, such as prestige, which is of no or little importance to (early) L1 acquisition, but may have an impact on second language learning. Moreover, early L1 acquisition is a bottom-up process whereby children extract linguistic schemas from the ambient language (Gómez and Gerken, 1999); whereas language change typically involves the extension and modification of existing schemas, rather than the creation of entirely new ones; and L2 acquisition is influenced by interference from a learner’s native language (Diessel, 2019, p. 37–39). These differences raise questions about the general architecture of the grammar network and its status. Can we model language acquisition and language change within the same network or do we need two separate models to account for acquisition and change? Does L2 acquisition involve two separate networks or just one? And how do we account for language attrition in the context of a grammar network model?

All datasets generated for this study are included in the article/supplementary material, further inquiries can be directed to the corresponding author.

The author confirms being the sole contributor of this work and has approved it for publication.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.