1.1.1 The Component Processes for Speaking

In this section, we review theories of grammatical encoding for speech production, focusing on the proposed relationship between words and syntax. We begin, however, with setting the process of grammatical encoding in context. All cognitive models of speech production are heavily influenced by Levelt’s classic blueprint for the speaker (Reference LeveltLevelt, 1989), which in turn built on the seminal work of Reference GarrettGarrett (1975). The proposal is that utterances are produced in a number of more-or-less successive processes, and there is also agreement on the broad structure of the processes involved (see Figure 1). The starting point is message generation, which involves the construction of a conceptual representation that details the information that the speaker wants to convey. This representation is usually known as the message (Reference LeveltLevelt, 1989). The current view is that messages are non-linear and must at least contain conceptual category information. Messages can be very short (e.g., mapping onto utterances like ‘Hi’ or ‘Look there!’) or much longer, including a thematic structure which assigns concepts to thematic roles such as agent or patient (e.g., mapping onto utterances like ‘The politician was amazed by the volume of fan mail’; see Reference Konopka, Brown-Schmidt, Goldrick, Ferreira and MiozzoKonopka & Brown-Schmidt, 2014, for a review). In addition, messages should contain information that is required to generate a grammatical sentence including time, mood and focus, as well as any language-specific information required by a language for obligatory syntactic or morphological markers (see Reference LeveltLevelt, 1989, chapter 3, for a detailed discussion).

Figure 1

A representation of the key processing stages of spoken sentence production.

The message triggers grammatical encoding processes, which include selecting the appropriate lexical items, assigning grammatical roles and generating a syntactic structure to fix their linear order. The phonological structure of the utterance is constructed in the subsequent phase, where an abstract prosodic representation is generated which forms the input to phonetic and articulatory processes. Grammatical encoding processes therefore form the link between the conceptual structure to be conveyed and the sound structure of the utterance that will convey it. The component processes are lexical retrieval and syntactic structure building.

1.1.2 Lexical Retrieval Processes

Lexical retrieval refers to the activation and retrieval of words from the mental lexicon. During production, activation at the conceptual level triggers a lexical search. Psycholinguistic models largely agree that lexical representations exist independently of semantics, at the lemma and lexeme levels (Reference Kempen and HuijbersKempen & Huijbers, 1983). Lemmas are abstract, modality-general and language-specific lexical entries that are activated by information at the conceptual level. In turn, lemma selection activates lexemes, that is, representations that include word-form information (see Reference Caramazza and MiozzoCaramazza & Miozzo, 1997, vs. Reference Roelofs, Meyer and LeveltRoelofs, Meyer & Levelt, 1998), and then phonological encoding processes. For example, a speaker wishing to convey information about one person (a woman) transferring something (a book) to another person (a man) will generate a message-level representation consisting of conceptual nodes that correspond to the nominal concepts woman, man and book, as well as the action of transferring X to Y, and this information may activate the lemma nodes for the nouns ‘woman’, ‘man’, ‘book’ and the verbs ‘give’ and ‘donate’. Lemmas include item-specific syntactic information, such as grammatical gender for nouns and restrictions on syntactic alternations for verbs (e.g., the verb ‘give’ can be used with both prepositional-object [PO] and double-object [DO] syntax, while the verb ‘donate’ can only be used with PO syntax).

The majority of models describing lexical access focus on retrieval of individual words – most often nouns (e.g., ‘woman’, ‘man’, ‘book’) – or production of short sequences of words in simple or complex noun phrases (NPs) (e.g., ‘the woman’, ‘the woman and the man’). The likelihood of selecting a lemma and the speed of selecting one noun lemma over another vary as a function of (a) word-specific variables (e.g., lexical frequency, age of acquisition, name agreement), (b) properties of the words’ lexical neighbours (e.g., neighbourhood density, recent activation of neighbouring lexical nodes, the degree to which relationships between words are taxonomic or thematic), and (c) the proposed architecture of the production system (e.g., the direction of information flow between the conceptual, lexical and phonological levels). Two classes of models, Levelt and colleagues’ serial model (Reference Levelt, Roelofs and MeyerLevelt et al., 1999; also see Reference RoelofsRoelofs, 1992) and Dell and colleagues’ interactive models of lexical access (Reference DellDell, 1986; Reference Dell, Schwartz, Martin, Saffran and GagnonDell, Schwartz, Martin, Saffran & Gagnon, 1997), have led the theorising in the field. In both models, the concepts or lexical nodes that are most strongly activated are selected for production, but the models differ in the degree to which they allow activation from lower levels to influence selection: serial models assume a feedforward flow of activation from concepts to lemmas and to phonological encoding, while interactive models allow for feedback from lower levels.

Lexical retrieval models also differ in their assumptions about the selection process at the lemma level, specifically the degree to which lemmas do or do not compete for selection (Reference Levelt, Roelofs and MeyerLevelt et al., 1999 vs. Reference Mahon, Costa, Peterson, Vargas and CaramazzaMahon, Costa, Peterson, Vargas & Caramazza, 2007; see Reference Abdel-Rahman and MelingerAbdel-Rahman & Melinger, 2009, for a review). The predictions of these models are often tested with the picture–word interference paradigm, where speakers name individual pictured objects while ignoring superimposed printed words. Retrieval times normally increase in the presence of semantic competitors, such as when trying to name the picture of a cat while seeing the printed word ‘dog’, and decrease in the presence of phonological neighbours, such as when trying to name the picture of a cat while seeing the printed word ‘cap’. Debates concerning the size and direction of these effects often hinge on determining the joint effects of multiple individual processes: conceptual priming (semantically related words prime each other), lexical interference (taxonomically related words compete against each other for selection), lexical facilitation (thematically related word prime each other) and phonological facilitation (phonologically related words prime each other). Production of a sequence of words, either in phrases (e.g., ‘the cat and the dog’) or without a phrasal context (‘cat dog’), naturally multiplies the number of processes to be completed and adds an additional parameter: retrieval of each word (word n) can be influenced by anticipatory activation of word n+1, and likewise, retrieval of word n+1 is influenced by production of word n. As in most picture-word interference paradigms, retrieval of word n is slower when word n+1 is a semantic competitor, but retrieval of word n+1 is also slower when word n is a semantic competitor (an effect known as cumulative semantic interference).

In a recent meta-analysis, Reference Bürki, Elbuy, Madec and VasishthBürki, Elbuy, Madec and Vasishth (2020) concluded that existing research does not adjudicate between models assuming competitive and non-competitive lexical access. Reference Oppenheim and NozariOppenheim and Nozari (2021) also showed that behavioural indexes such as the presence of semantic interference or facilitation cannot be used to conclusively distinguish between competitive and non-competitive lexical access, as competitive and non-competitive selection rules can produce similar behavioural outcomes. A more promising approach is to track context-specific changes in retrieval speed in order to model experience-driven changes in activation levels and connections between the conceptual level and word level (see Reference Dell, Jacobs, Hickok and SmallDell & Jacobs, 2016; Reference Dell, Nozari, Oppenheim, Goldrick, Ferreira and MiozzoDell, Nozari & Oppenheim, 2014; Reference Oppenheim, Dell and SchwartzOppenheim, Dell & Schwartz, 2010, and Reference Oppenheim and NozariOppenheim & Nozari, 2021, for more detail with supporting empirical evidence and simulations). For example, the degree to which both taxonomically and thematically related distractors interfere with production of a target word depends on the way these relationships are represented in the model, rather than depending on selection rules.

The models of lexical retrieval reviewed in the preceding text are concerned with the nature of lexical representations for content words (mostly nouns) and thus do not make explicit claims about processes responsible for integrating sequences of lexical items into longer utterances. In the rest of the Element, we focus primarily on a different long-standing debate in psycholinguistics – namely, the contribution of the lexicon to grammatical encoding (see Reference BockBock, 1982, Reference Bock and Ellis1987, for early reviews). This area of research focuses on production of longer, multi-word utterances with complex syntactic structures and, critically, utterances requiring retrieval of verbs.

1.1.3 The Need for Syntax

Producing grammatically correct multi-word utterances requires that words be produced in a specific order, that is, that they be sequenced according to language-specific word-order rules. This sequencing is referred to as linearisation. Interestingly, while it is clear that linguistic utterances are structured, the nature of the structural representations generated to output grammatically correct word sequences is debated. This puzzle concerns the degree to which the lexicon is involved in the generation of sentence structure.

Broadly speaking, the generation of sentence structure has been described, in different accounts, as a by-product of lexical retrieval processes or as the outcome of processes operating outside of the lexicon (e.g., see Reference Bock and EllisBock, 1987, for a review). Lexicalist (or functional) accounts propose that there is no strict separation between the lexicon and grammar: speakers retrieve lexical items as required by the preverbal message they want to communicate, and it is the lexical retrieval process that initiates the building of a syntactic structure. In other words, the building of a linguistic structure is dependent on lexical activation. By implication then, syntax is largely epiphenomenal. However, the linearisation of a longer, complex message that requires activation of multiple content words poses a problem for this account, as lemma activation can be responsible for the activation of ‘local’ syntactic information but is less likely to be responsible for the building of larger syntactic frames (also see Section 3 for a discussion of planning scope in multi-word utterances). Abstract structural accounts are better able to account for linearisation in longer utterances, as they propose that larger structures (or frames) are built by abstract syntactic procedures independently of the lexical items that will be slotted into them. These procedures are sensitive to word-specific syntactic requirements, but they are not, crucially, triggered by activation of individual lemmas.

The viability of the lexical account, and thus the origins of the debate between lexical and abstract accounts, has historical roots. Language research has been largely skewed in favour of comprehension rather than production, and comprehension studies show strong reliance on the lexicon during parsing. In comprehension, listeners receive a linguistic signal that comes in word by word over time and they must integrate this information to decode the speaker’s message. Naturally, given that listeners process incoming information as soon as it becomes available, the processor may give more weight to new lexical information (which can be quickly integrated with those parts of the utterance that have already been heard) than to structural information (as the structural representation of a spoken utterance is built up or inferred from a string of words rather than from individual words). Listeners do generate predictions about upcoming words, but evidence of prediction based on the semantic or lexical content of a sentence (be it coarse-grained, i.e., involving entire words, or finer-grained, i.e., involving sublexical units) is currently more plentiful than evidence of prediction of structure based on grammatical markers or parts of speech (see Reference Huettig, Rommers and MeyerHuettig, Rommers & Meyer, 2011, for a review). Thus, the demands of comprehension for structural processing may be less stringent than in production and may effectively ‘hide’ potential effects of abstract structural processes. Levels of engagement during comprehension can also vary, such that ‘good enough’ processing (i.e., the build-up of underspecified representations) may be sufficient for successful comprehension in many contexts (Reference Karimi and FerreiraKarimi & Ferreira, 2016). Indeed, finding evidence of the involvement of abstract structural processes in comprehension requires development of more sensitive measurement tools or ensuring greater engagement on the listener’s part (see Reference Tooley and BockTooley & Bock, 2014).

In contrast, the distinction between lexical sources of structure and abstract structural processes is more salient and thus more relevant in production. The processing demands of language production on the speaker are arguably higher than the demands of comprehension on the listener. To produce an utterance, speakers must first decide what they want to say (albeit not necessarily in large, sentence-sized chunks) and must then begin generating the linguistic material they will need to communicate their message from scratch. This involves both structural and lexical processing, so stronger reliance on lexical than structural information may not be as viable in production as it is in comprehension: producing a sequence of words cannot bypass structural processing and rely exclusively on lexically specific syntactic information. An empirical challenge in the field of language production is therefore the need to delineate the boundary between lexically driven and lexically free influences on word order, and to explain when and how these processes interact.

1.1.4 Models of Grammatical Encoding: The Relationship between Words and Syntax

Models of grammatical encoding differ in the relationship they propose between words and structure. There are different claims about which level of representation encodes links between lexical and structural information, with some models encoding explicit links between lexical concepts and thematic roles at the conceptual level (e.g., Reference ChangChang, 2002; Reference Chang, Dell and BockChang et al., 2006), and others in grammatical representations between lemmas and syntactic information, allowing lexical retrieval and structure building to interact during grammatical encoding (e.g., Reference Bock, Levelt and GernsbacherBock & Levelt, 1994; Reference Cleland and PickeringCleland & Pickering, 2003, Reference Cleland and Pickering2006; Reference Ferreira and WheeldonFerreira, 2000; Reference Ferreira, Morgan, Slevc, Rueschemeyer and GaskellFerreira, Morgan & Slevc, 2018; Reference LeveltLevelt, 1989; Reference Levelt, Roelofs and MeyerLevelt et al., 1999; Reference MommaMomma, 2021; Reference Pickering and BraniganPickering & Branigan, 1998). Models also diverge in the degree to which lexical or structural information guide grammatical encoding.

The earliest models of grammatical encoding were lexically driven and accorded a central role to lemma representations, which comprised semantic and syntactic-lexical information (e.g., Reference Bock, Levelt and GernsbacherBock & Levelt, 1994; for reviews, see Reference Bock, Ferreira, Goldrick, Ferreira and MiozzoBock & Ferreira, 2014; Reference Ferreira, Slevc and GaskellFerreira & Slevc, 2007; Reference Ferreira, Morgan, Slevc, Rueschemeyer and GaskellFerreira et al., 2018). Later versions of this approach limited lemmas to encoding aspects of lexical syntax, including grammatical category (e.g., noun, verb, adjective) as well as syntactic features (e.g., tense, number, grammatical gender; e.g., Reference Levelt, Roelofs and MeyerLevelt et al., 1999, see also Reference Roelofs, Ferreira and HagoortRoelofs & Ferreira, 2019). These models also assume a discrete flow of information, with lemma selection occurring during grammatical encoding prior to the activation of phonological form (see Section 1.1.2). Two distinct stages are proposed for structure building. In the initial stage, termed functional encoding, the lemmas which best match the conceptual representation in the message are retrieved and assigned to grammatical functions appropriate for the thematic structure (e.g., agent → subject, patient → object, for a transitive active sentence such as ‘Anne saw Bill’). Following function assignment, an appropriate phrase structure is generated to which the lemmas are attached. The process for generating phrase structure was elaborated in a model proposed by Reference Pickering and BraniganPickering and Branigan (1998), which also incorporated links from lemma representations to nodes specifying the possible phrase structures in which they can occur. These ‘combinatorial nodes’ were initially linked only to verbs and encoded subcategorisation information. Later versions of the model extended the approach to nouns (Reference Cleland and PickeringCleland & Pickering, 2003, Reference Cleland and Pickering2006). Following function assignment, the selection of phrase structures in the model is driven by activation spreading from the lemmas with the most highly activated combinatorial node being selected (constituent assembly). Due to the direct links between lemmas and syntactic structures, this approach provides a clear mechanism through which lexical and syntactic representations can interact to determine the structure of the sentence produced.

Another approach which encodes explicit links between lemmas and syntactic structures employs tree-adjoining grammar (TAG; Reference Ferreira and WheeldonFerreira, 2000; Reference FrankFrank, 2002; Reference MommaMomma, 2021, Reference Momma2022). Reference MommaMomma (2021, Reference Momma2022) proposes a TAG-based grammatical encoding model in which the syntactic structure for an utterance is constructed based on elementary trees. Elementary trees are complex structures headed by clause-taking verbs comprising a hierarchical syntactic structure with open nodes for constituents. For example, the elementary tree for a transitive verb like ‘chase’ would have two determiner phrase nodes for the sentence subject and object. More complex structures are created in the model by combining elementary trees either by a process of substitution or adjoining. In substitution, open nodes in an elementary tree are filled by appropriate tree structures; for example, an open determiner phrase node could be filled by a determiner phrase tree ‘the girl’. The process of adjoining fills nodes with auxiliary trees containing recursive elements like adverbs and adjectives. In this model, lemmas are represented at a sub-tree level and are connected to the appropriate nodes of an elementary tree. The sub-tree level also contains nodes representing functional heads for structural options, such as DO and PO datives, which can be activated by thematic representations. Inhibitory links between sub-trees allow for a competitive lemma selection process. In contrast, elementary trees do not compete for selection. Elementary trees are stored in long-term memory and activated by the conceptual structure, either directly or via the conceptual activation of their sub-tree representations.

Reference MommaMomma (2021) proposed the model to explain the grammatical encoding of long-distance syntactic dependencies, such as the cross-clausal filler-gap dependency in the sentence ‘Who do you think that the girl likes?’. This sentence has a syntactic dependency between the words ‘who’ (the filler) and ‘likes’ (the gap – i.e., the missing object for the verb). According to the model, speakers plan the structural dependency between such elements prior to planning the intervening material. Critically, elementary trees must encode all syntactic dependencies, including long-distance dependencies, within a phrase. Therefore, a minimal elementary tree for the sentence above must represent the cross-clausal filler-gap dependency. This elementary tree is abstract in that it represents critical grammatical information about the syntactic nature of the gap and the clause structure in which it occurs. However, it does not represent the material intervening between the filler and the gap. This is represented in a separate tree, and the process of tree adjoining enables this material to be inserted into the elementary tree at a later point during grammatical encoding. This model therefore encodes explicit links between lexical and syntactic representations, allowing them to interact during grammatical encoding.

In contrast, interactions between lemmas and syntactic structures are not a feature of a series of computational learning models of grammatical encoding (Reference ChangChang, 2002; Reference Chang, Dell and BockChang et al., 2006; Reference Dell and ChangDell & Chang, 2014). The Dual Path approach adopted in these models hinges on a strict separation between lexical retrieval and structure building. Similar to the lexically driven models described in Section 1.1.3, grammatical encoding is initiated following the construction of a conceptual message in which lexical concepts are bound to thematic roles and appropriate lemmas are activated by these lexical concepts. The models diverge, however, in that there is no process of function assignment in the Dual Path approach. Instead, the order of activation of lemmas is determined by the activation level of their lexical concepts, which is in turn determined by the weighting of their associated thematic roles. For example, during the production of an active sentence, the agent role would be most highly activated, while for passive sentences the patient role would have the highest activation level. Importantly, the activation of a lemma would be blind to the thematic role assignment of the associated lexical concept.

Syntactic structure is built by a sequencing system modelled as a simple recurrent network (SRN). This network has access only to the event semantics in the thematic structure and is blind to the lexical concepts. It learns syntactic categories and relationships between words through trial and error by predicting word order during training. The SRN stores the links between thematic structures and word orders via a layer of hidden units. During grammatical encoding, the most appropriate word order to convey the information encoded in a message is determined by the weighted thematic structure and the learned syntactic relationships in the SRN. The Dual Path model therefore explains grammatical encoding in terms of a predictive learning process operating as we comprehend speech, actively predicting the next word we will hear and learning from our mistakes (error-driven learning). It therefore provides an explicit mechanism for the acquisition of syntax (e.g., Reference Fitz and ChangFitz & Chang, 2017). Critically this approach allows no interaction between lexical information and structure building during grammatical encoding.

A competing alternative is Reference Reitter, Keller and MooreReitter, Keller and Moore’s (2011) lexical model of priming. This model based on Adaptive Control of Thought—Rational (ACT-R) accounts for both short-term and long-term structural priming, and views both as a consequence of lexical priming. The model implements Combinatory Categorial Grammar (Reference SteedmanSteedman, 2000), where words are bound to subcategorisation information. It includes a declarative memory element with chunks of lexical information connected to chunks of syntactic information and a procedural memory element with if-then rules. Production occurs by sequential activation and retrieval of individual lexical and syntactic chunks from declarative memory; priming occurs because words are kept in a short-term buffer (the ‘working memory’ of the model) and are more likely to be reactivated, with their associated syntactic details, if they have been used recently. The spread of activation in the model explains lexically supported short-term priming as well as cumulative priming.

Finally, other approaches allow both lexical and structural representations to guide grammatical encoding (or, more specifically, linearisation). Theories of sentence planning are particularly well-suited to addressing questions about the degree of lexical and structural control of production because they make specific predictions about the type of information that starts or triggers production of a sentence. A well-known property of the language system is that it allows production to unfold incrementally: speakers plan utterances in small increments rather than in proposition-like units. So, when producing a novel utterance, what information do speakers tend to encode first? The precise nature of the incremental build-up of an utterance is described by two accounts – Linear Incrementality and Hierarchical Incrementality – which roughly follow from the assumptions of accounts proposing lexical and abstract syntactic control of production (see Figure 2). Linearly incremental planning assumes that utterances can be built up in word-like chunks. For example, the generation of a sentence can begin with the retrieval of a single word (corresponding to the concept that is mentioned first; e.g., either ‘cowboy’ or ‘bull’ in a sentence that will eventually be articulated as ‘The cowboy caught the bull’ or ‘The bull was caught by the cowboy’). Activation of either noun commits the speaker to selection of either active or passive syntax (at least in English), which will result in the ‘projection’ of a syntactic structure to subsequently guide encoding of the remaining content words. The various word-sized planning units are joined together by basic sequencing rules (which are underspecified in this account). The attachment or joining of each new word-like unit to the previous one results in the emergence of a specific syntactic structure. In contrast, Hierarchical Incrementality assumes that control of sentence planning and word sequencing lies in the hands of abstract structural processes responsible for generating a relational structure at the conceptual level and a corresponding syntactic structure at the linguistic level. Importantly, this structural framework is generated without lexical support: it precedes lexical retrieval and controls its timing, rather than being projected from a sequence of lexical retrieval operations (see e.g., Reference Bock, Ferreira, Goldrick, Ferreira and MiozzoBock & Ferreira, 2014, for a review). In other words, speakers activate lexical items in the order in which the sentence syntax calls for them.

Figure 2

Schematic illustrating the relationship between message-level and sentence-level planning under the strong versions of

(a)

hierarchical incrementality and

(b)

lexical incrementality in the production of a transitive sentence with three content words in a sentence like ‘The cowboy caught the bull’ in four time steps (see Figure 3 for a paradigm eliciting such sentences).

Hierarchical Incrementality is consistent with an elegant solution to the linearisation problem in production proposed by Reference Dell, Oppenheim and KittredgeDell and colleagues (2008; also see Reference Momma, Buffinton, Slevc and PhillipsMomma, Buffinton, Slevc & Phillips, 2020). Dell and colleagues describe a model with a strict division of labour between semantics and syntax. Determining word order is the job of a syntactic ‘traffic cop’, or a mechanism consisting of a series of weights between syntactic sequential states and lexical items that enables speakers to ‘say the right word at the right time’. The traffic cop tracks syntactic sequential states to ensure that only words of specific classes (e.g., determiners vs. nouns vs. verbs) are activated at specific points in time in the order required by the syntax of the developing sentence. Lexical retrieval is thus managed in an efficient manner by sequentially activating and deactivating lexical items that are semantically relevant and syntactically appropriate during production. For example, when producing the sentence ‘The cowboy is catching the bull’, the words ‘cowboy’ and ‘catch’ will not compete for the subject slot (but a suitable alternative word for referring to the agent, like ‘man’ or ‘rancher’, can). The division of labour is instantiated in the weights between syntax and the lexicon: content words have stronger links to semantics than to syntax, while function words have stronger links to syntax than to semantics. Reference Momma, Buffinton, Slevc and PhillipsMomma and colleagues (2020) provided confirmatory experimental evidence that prime verbs and nouns with the same lexical form (e.g., ‘is singing’ vs. ‘her singing’) have different effects on production of semantically related verbs (e.g., ‘whistling’) in target sentences: prime verbs (‘is singing’) delay production of target verbs but prime nouns (‘her singing’) do not. Likewise, prime nouns delay production of target nouns, but prime verbs do not.

The strong versions of Linear and Hierarchical Incrementality described previously provide a useful reframing of the lexicalist versus abstract syntax debate. It is useful not only because it generates explicit testable predictions about control of production (as do other frameworks) but primarily because it does so at a level that provides fine-grained insight into production: adjudicating between these accounts requires tracking the time-course of production from message encoding until articulation, which allows for analysis of the coordination of multiple production processes. Importantly, it also provides a way of reconciling lexicalist and abstract syntax accounts. Determining the degree to which the production system supports one type of incremental planning over another by default – that is, the degree to which lexicalist and abstract syntax accounts provide a better description of the data – can lead to somewhat of a theoretical impasse. Instead, recent work in this area has suggested that there may be no default form of incrementality that is strictly determined by the architecture of the production system (e.g., Reference Ferreira and SwetsFerreira & Swets, 2002); instead, the production system may be flexible in supporting linear planning or hierarchical planning under different conditions (e.g., Reference Konopka, Meyer and ForestKonopka et al., 2018). Outlining the conditions under which planning is more likely to proceed in a linearly or hierarchically incremental fashion may ultimately lead to a more precise characterisation of grammatical encoding as being primarily lexically driven or primarily under the control of abstract syntax, but crucially, at present, this approach provides a new framework for explaining the complex processes through which language processes change and adapt to experience

The following sections of this Element describe empirical investigations into lexical and abstract syntactic control of grammatical encoding, first from the perspective of structural priming paradigms, which allow us to test what factors influence speakers’ binary structure choices (Section 2), and then from the perspective of utterance planning paradigms, which allow us to test what factors influence grammatical encoding with finer-grained continuous temporal measures (Section 3).

2.2.1 The Lexical Boost in Structural Priming

Repetition of meaning normally entails repetition of lexical items. Thus, as with repetition of event semantics, one might expect to see evidence for the involvement of the lexicon in speakers’ structural choices in the form of a lexical boost in structural priming. Indeed, using a sentence completion task, Reference Pickering and BraniganPickering and Branigan (1998) showed that speakers were more likely to repeat the syntax of a prime PO (or DO sentence) in a new target sentence when primes and targets used the same verb (8a, 8b) than when they did not (8c, 8d).

1. 1. a.  The racing driver showed the torn overall … (prime sentence biased towards PO completion)

   2. b.  The patient showed … (to-be-completed target sentence)

   3. c.  The racing driver gave the torn overall … (prime sentence biased towards PO completion)

   4. d.  The patient showed … (to-be-completed target sentence)

Since then, this lexical boost (or lexical enhancement of priming) has been replicated in numerous studies with a range of sentence elicitation paradigms and in corpora (Reference GriesGries, 2005), and with repetition of both verbs and nouns (Reference Cleland and PickeringCleland & Pickering, 2003). Reference Pickering and BraniganPickering and Branigan’s (1998) proposed representation basis of lexical influences on structure choice focuses on activation at the lemma level. On their account, lemma nodes are linked to combinatorial nodes, which specify what structures a verb can be used in (e.g., give can be used in both prepositional-object and double-object dative structures, while donate can only be used with propositional-object syntax). This information is activated during processing of a prime sentence. Residual activation of combinatorial nodes can bias structure selection in a subsequent target trial in favour of the recently activated structure. Importantly, the link between a combinatorial node and the verb lemma also remains temporarily activated, so repetition of the same verb from prime to target increases the likelihood of selecting a recently used structure beyond the level supported by activation of combinatorial nodes alone. The magnitude of this lexical enhancement of priming is noteworthy: the odds of repetition of structure double in sentences with than without lexical overlap (Reference Mahowald, James, Futrell and GibsonMahowald et al., 2016).

It is not clear from Pickering and Branigan’s account (Reference Pickering and Branigan1998) whether the lexical boost is a conflation of two effects: a semantic boost driven by repetition of meaning plus a lexical boost due only to repeated activation of the same lexical information. Reference Santesteban, Pickering and McLeanSantesteban, Pickering and McLean (2010) provided evidence that distinguished between these possibilities by testing whether lexical similarity without semantic overlap can modulate structural priming. Their experiments compared structural priming of NPs (e.g., ‘the red bat’ vs. ‘the bat that’s red’) with non-homophonous and homophonous nouns. Priming was equally strong from primes with nouns that repeated meaning and sound information (e.g., the animal bat in primes and targets) and homophonous nouns that repeated sound information alone (the animal bat and a cricket bat in primes and targets, respectively; but see Reference Cleland and PickeringCleland & Pickering, 2003). This homophone boost suggests that overlap at the word-form level is sufficient to observe enhanced priming (also see Reference Konopka and BockKonopka & Bock, 2009, for evidence of a lexical boost in sentences with idiomatic verbs that repeat word forms but not meaning).

A number of findings regarding the involvement of the lexicon in structure building are important for keeping these effects in perspective. The first observation is that structural priming is boosted by – but not determined by – lexical repetition: the lexical boost is an enhancement of structural repetition rather than a precondition for structural repetition to occur. Observing lexically unsupported priming is indeed treated as the golden standard for classifying repetition of structure as ‘syntactic’ in nature (e.g., Reference Fehér, Wonnacott and SmithFehér et al., 2016). This definition places key constraints on lexical accounts of syntax and requires clarification of how potential links between lexical items and syntax modulate structure choice.

For example, not all lexical repetition results in a lexical boost. Repetition is mediated by the syntactic role of the repeated words: a lexical boost has been observed with repetition of open-class words that are syntactic heads (Reference Carminati, van Gompel and WakefordCarminati, van Gompel & Wakeford, 2019; Reference Ivanova, Branigan, McLean, Costa and PickeringIvanova et al., 2017; Reference van Gompel, Wakeford and Kantolavan Gompel, Wakeford & Kantola, 2022) but not syntactic non-heads. Repetition of closed-class words such as to and for in dative sentences (Reference BockBock, 1989), or by in locatives and passives (Reference Bock and LoebellBock & Loebell, 1990) does not create a lexical boost (but see Reference Ziegler and SnedekerZiegler et al., 2018). Further, lexical similarity alone is also not sufficient for repetition. For example, Reference FerreiraFerreira (2003) used a sentence-recall production task to demonstrate that the use of the optional complementiser ‘that’ in sentences such as 9d could only be primed by sentences that included a complementiser ‘that’ (i.e., the same lexical item playing a similar syntactic role, 9a) but not by a determiner ‘that’ (i.e., the same lexical item playing a different syntactic role, 9b) or a noun-complement ‘that’ (i.e., the same lexical item in a different syntactic context, 9c). Importantly, both the inclusion and exclusion of the complementiser ‘that’ could be primed – a finding consistent with a structural rather than lexical locus for priming effect.

1. 1. a.  The company ensured that the farm was covered for two million dollars

   2. b.  The company insured that farm for two million dollars

   3. c.  The theory that penguins built the igloo was completely false

   4. d.  The mechanic mentioned (that) the antique car could use a tune-up

Reference MommaMomma (2022) built on these findings by demonstrating that the priming of complementiser ‘that’ can be lexically boosted by the repetition of verbs biased for its use (e.g., Reference Bernolet and HartsuikerBernolet & Hartsuiker, 2010). This boost was observed when both prime and target either did or did not feature cross-clausal filler-gap dependencies (10a and 10b, respectively). However, when the prime sentence (but not the target) contained a cross-clausal filler-gap dependency (e.g., 10a), the lexical boost disappeared.

1. 1. a.  Who did the manager imply (that) he would promote?

   2. b.  The manager implied (that) he would promote the employee

This result is problematic due to the explicit memory account of the lexical boost, which would predict that a boost should be observed in all conditions. Reference MommaMomma (2022) proposed that lexically independent structural priming is the result of an enhanced link between a concept and a node of an elementary tree, whereas lexical boost effects are due to the residual activation of elementary trees (similar to the priming mechanisms proposed by Reference Pickering and BraniganPickering & Branigan, 1998). Critically, according to the TAG model, elementary trees must contain all syntactic dependencies as well as complementiser features. Therefore, the sentences in 10 will be represented by different elementary trees which are headed by the same verb.

Second, repetition of structure across sentences with and without overlap in content words has a different time-course. Lexical and abstract syntactic accounts make different predictions about the duration of structural priming effects, following directly from differences in their mechanistic explanations of priming. Activation-based accounts, such as Reference Pickering and BraniganPickering and Branigan’s (1998) lexical account, predict short-lived priming: by definition, activation dissipates quickly, so any activation of links between verb nodes and combinatorial nodes that produces a lexical boost in the short term may fail to produce a lexical boost in the long term. Accounts that emphasise a syntactic locus of structural repetition predict persistence of lexically unsupported priming.

Consistent with abstract accounts, structural priming has been observed across different lags within the same experiment (Reference Bernolet, Collina and HartsuikerBernolet, Collina & Hartsuiker, 2016; Reference Bock and GriffinBock & Griffin, 2000; Reference Bock, Dell, Chang and OnishiBock, Dell, Chang & Onishi, 2007; Reference Hartsuiker, Bernolet, Schoonbaert, Speybroeck and VanderelstHartsuiker et al., 2008; Reference Kaschak and BorreggineKaschak & Borreggine, 2008; Reference Kaschak, Loney and BorreggineKaschak, Loney & Borreggine, 2006), across different sessions of the same experiment (Reference Kaschak, Kutta and SchatschneiderKaschak, Kutta & Schatschneider, 2011), and in natural speech (Reference GriesGries, 2005) in sentences without lexical overlap (see Figure 3 for an example of Lag 0 and Lag 2 priming). Priming effects are also obtained regardless of the cover task given to participants (i.e., regardless of whether participants’ attention is directed to the sentence form or not; Reference Bock, Loebell and MoreyBock et al., 1992) and remarkably even in participants with compromised episodic memory (anterograde amnesics; Reference Ferreira, Bock, Wilson and CohenFerreira, Bock, Wilson & Cohen, 2008; Reference Heyselaar, Segaert, Walvoort, Kessels and HagoortHeyselaar, Segaert, Walvoort, Kessels & Hagoort, 2017). In contrast, the lexical boost in syntactic priming declines when primes and targets are separated by as few as two intervening sentences (Reference Hartsuiker, Bernolet, Schoonbaert, Speybroeck and VanderelstHartsuiker, Bernolet, Schoonbaert, Speybroeck & Vanderelst, 2008; Reference Kaschak and BorreggineKaschak & Borreggine, 2008; Reference Konopka and BockKonopka & Bock, 2005). For example, Reference Hartsuiker, Bernolet, Schoonbaert, Speybroeck and VanderelstHartsuiker and colleagues (2008) compared the magnitude of dative PO/DO priming at lags 0, 2, and 6, and showed a lexical boost only at lag 0. The sharp decline in lexically supported priming suggests that lexical contributions to structure repetition are short-lived.

Figure 3

Schematic illustrating a prime-target structural priming paradigm in

(a)

lag 0 with adjacent prime and target trials (white cells), and in

(b)

lag 2 with the prime and target separated by two intervening filler trials (grey cells). Recorded prime sentences with active or passive syntax: ‘The man is lifting the bench’ / ‘The bench is being lifted by the man’. Target sentences eliciting active or passive sentences: ‘The cowboy is catching the bull’ / ‘The bull is being caught by the cowboy’.

A complementary approach showed that the magnitude of priming effects can vary with the extent of repeated exposure to alternative structures but not to repeated exposure with specific verbs. Reference Kaschak, Loney and BorreggineKaschak and colleagues (2006) tested whether participants’ sensitivity to a PO-DO priming manipulation was modulated by recent and repeated experience with the primed structures (i.e., a form of cumulative priming). In Experiment 1, after an exposure phase where the ratio of PO and DO sentences was manipulated (50:50 in the same block vs. 50:50 but in separate blocks vs. 100:0), priming was only observed in participants who had been exposed to both structures at the beginning of the study. Participants who had been exposed to only PO or only DO sentences (i.e., a 100:0 ratio of PO:DO structures or DO:PO structures) did not respond to the priming manipulation. In Experiment 2, sensitivity to the priming manipulation also varied in a graded fashion, in line with the strength of the structural bias introduced in the exposure phase (50:50 vs. 75:25 vs. 100:0). This effect was not replicated by Reference KaschakKaschak (2007), who proposed that the exposure phase shifts base rates for individual structures (with stronger shifts for the dispreferred PO structure) but not the magnitude of priming effects. Subsequently, Reference Kaschak and BorreggineKaschak and Borreggine (2008) tested whether these biases are affected by verb repetition in the exposure phase and again showed structural repetition only in participants who had been exposed to both structural alternatives prior to the priming task. Crucially, these effects were not modulated by verb repetition or by the presentation of the verb in one or both structures. In other words, what mattered for structure choice was the frequency of use of individual structures rather than the frequency of use of individual structures with specific verbs.

Thus, testing the longevity of abstract syntactic priming and lexically supported priming shows a critical dissociation. On balance, the short-lived nature of the lexical boost – arguably the strongest evidence for involvement of the lexicon in structure building – and the persistence of abstract priming suggests that the two effects have a different source. These results favour a multi-factorial account of priming, one where the binding of words to structures responsible for the lexical boost is dynamic and short-lived, while the persistence of structure choice long after the lexical boost has decayed arises from longer-term learning of structure-building procedures in an abstract syntactic system. Reference Chang, Dell and BockChang and colleagues (2006) account for this finding by proposing an implicit learning mechanism that predicts both short-term and longer-term repetition of structure and that is resistant to episodic forgetting, and speculate that a separate mechanism, relying on explicit memory, may explain lexical influences on structure choice (see Section 1.1.4; Reference Chang, Janciauskas and FitzChang, Janciauskas & Fitz, 2012).

The proposal of explicit memory retrieval explaining the lexical boost is important for clarifying the coordination of lexical and syntactic processes during grammatical encoding. If repetition of structure is driven by memory for the prime sentences, this makes the priming paradigm less suitable for addressing questions about the production architecture, such as the nature of the links between individual lexical items and structural information. For example, Reference Scheepers, Raffray and MyachykovScheepers, Raffray and Myachykov (2017) showed that repetition of any content word from PO/DO-primes to PO/DO targets can produce a lexical boost (agents, verbs, recipients, themes). In fact, the magnitude of priming increased with increasing overlap in content words, producing a cumulative lexical boost effect. Scheepers and colleagues proposed that repetition of lexical items serves as a powerful retrieval cue: quite simply, the more lexical repetition, the stronger the memory cues and the higher the likelihood of structural repetition. Reference Bernolet, Collina and HartsuikerBernolet and colleagues (2016) also suggested that explicit memory may contribute to structural repetition even when primes and target do not share lexical items, based on the finding that both structural priming and explicit memory for the prime sentences declined over lags (from Lag 0 to Lag 6) in their experiments. There was, however, evidence of cumulative lexically unsupported priming: speakers’ production of target structures increased as a function of the number of these structures produced within an experimental session. In a more direct test of the memory hypothesis, Reference Zhang, Bernolet and HartsuikerZhang, Bernolet and Hartsuiker (2020) showed that adding a cognitive load to the production task reduced structural priming effects in adjacent prime and target trials, both in sentences with and without lexical overlap (but see Reference Yan, Martin and SlevcYan, Martin & Slevc, 2018, for a different view). Further insight into the coordination of lexical and structural processes is provided by studies assessing verb biases and cumulative priming effects (Section 2.2.2).

2.2.2 Verb Bias and Structural Priming

Stronger evidence for the involvement of the lexicon in structure building comes from studies of verb bias. Among verbs that can appear in alternative syntactic structures (e.g., PO and DO datives), some demonstrate a strong bias for one structure over the other, while others show weaker biases or are considered to be equi-biased. These biases illustrate one of the main premises of lexicalist accounts of syntax, that is, the claim that lexical activation is a key driver of structure-building procedures, or that, put differently, syntax is ‘projected’ from the lexicon. For such biases to arise, the production system must keep track of individual verb-structure pairings over a speaker’s lifetime (a form of cumulative priming) and store verb-specific frequency information that reflects the statistics of the input. This information is then activated when a known verb is used on a new occasion and can bias selection of the most frequent structural alternative for that verb.

In support of this lexical view, Reference Melinger and DobelMelinger and Dobel (2005) showed that speakers’ structure choices can be influenced by the presentation of a single verb. In prime trials in their study, participants read non-alternating dative verbs (i.e., verbs that were strongly biased towards either PO or DO dative structures), and then saw pictures meant to elicit dative descriptions in target trials. Target descriptions were consistent with the bias of prime verbs: speakers produced more PO descriptions after PO-biased verbs and more DO descriptions after DO-biased verbs. Thus, activation of a verb with strong structural biases out of a sentence context was sufficient to influence structure choice in line with these biases.

However, the mere existence of verb biases is seemingly in direct contrast with the observation that the lexical boost in priming declines over time (discussed in the previous section). How do results such as Reference Melinger and DobelMelinger and Dobel’s (2005) as well as Reference Kaschak and BorreggineKaschak and Borreggine’s (2008) square with the observation of a short-lived lexical boost in priming? Conversely, if verb-structure pairings are dynamic, as assumed by implicit accounts of priming, how do verb biases come about in the first place? The answer arguably lies in the difference between short-term priming effects and long-term (or cumulative) priming.

Studies assessing structural priming with biased verbs in the short term provide an explanation that is more compatible with implicit learning accounts than with lexicalist accounts of syntax. Reference Bernolet and HartsuikerBernolet and Hartsuiker (2010) tested how existing verb biases modulate the magnitude of dative PO/DO structural priming in a full-sentence prime-target paradigm and showed an important difference between baseline structural preferences and sensitivity to priming of different verbs. Production of target sentences in a baseline condition showed strong effects of verb bias: participants generated more sentences with PO syntax than DO syntax, consistent with the biases of the verbs used in these sentences. At the same time, structure choice was also modulated by the priming manipulation: more PO/DO sentences were produced after PO/DO-primes. Importantly, participants’ productions showed an inverse priming effect: priming from DO-primes (i.e., primes with dispreferred DO syntax) was stronger than priming from PO-primes (i.e., primes with preferred PO syntax; see Reference Segaert, Weber, Cladder-Micus and HagoortSegaert, Weber, Cladder-Micus & Hagoort, 2014, for a similar effect with active and passive structures). This effect was further modulated by the individual ‘bias scores’ of both prime and target verbs: DO-primes had the strongest effect on target sentences when they featured PO-biased verbs, and target sentences with PO-biased verbs showed the strongest effects of DO-primes. In other words, the verb-structure combinations that are encountered less frequently produced the strongest priming.

The results are consistent with a key prediction of the implicit learning account of structural priming (Reference Bock and GriffinBock & Griffin, 2000; Reference Chang, Dell and BockChang et al., 2006; Reference Jaeger and SniderJaeger & Snider, 2013), namely the fact that encountering a surprising structure in a prime trial results in stronger error-based learning and thus increases the likelihood of producing a dispreferred structure in a target trial. This account also explains why dispreferred structures continue to exist in speakers’ linguistic repertoire: these structures receive a large boost with each use, which effectively ensures that they do not prime themselves out of existence (Reference Ferreira and BockFerreira & Bock, 2006).

In contrast, studies assessing structural priming with biased verbs in the long term show that verb biases do persist. Reference Coyle and KaschakCoyle and Kaschak (2008) proposed that repeated exposure to verbs in the same structures (i.e., using a manipulation meant to simulate specific verb biases) can produce longer-lasting, lexically enhanced structural priming, and that these effects can be observed in production tasks with a minor modification. Namely, Coyle and Kaschak hypothesised that any longer-lasting effects induced in the exposure phase of their experiment may not be visible in the priming phase because prime trials exert a strong and immediate influence on structure choice on target trials, but should be observable in the same target trials when not preceded by primes. Indeed, examining structure choice in target trials after an exposure phase meant to induce specific verb biases showed the persistence of these biases. Thus, the rapid decay of the lexical boost in priming studies need not imply that lexical contributions to structure repetition are short-lived: verb biases may come about due to repeated exposure to specific verb-structure bindings on a time scale that exceeds that of most priming studies in the lab.

2.2.3 Structural Priming in Bilinguals

As a field, psycholinguistics initially favoured research on language processing in monolingual populations. Studies on bilingual language processing, however, are now plentiful. Much like cross-linguistic research, bilingualism provides new opportunities for establishing the nature of processing similarities across languages or constraints on processing (also see Reference Blasi, Henrich, Adamou, Kemmerer and MajidBlasi, Henrich, Adamou, Kemmerer & Majid, 2022, for a discussion of the need to broaden the field’s scope of research to languages other than English). In particular, by testing whether syntactic structures can generalise from one language to another, bilingual studies contribute valuable evidence to the debate about the balance of lexical and structural influences in grammatical encoding (Reference Hartsuiker and PickeringHartsuiker & Pickering, 2008).

Early research in this area showed lexically unsupported repetition of structure in bilinguals that closely mirrored findings from monolingual speakers (Reference Loebell and BockLoebell & Bock, 2003): priming occurred between English and German dative sentences (both languages allow PO and DO syntax) but not between English and German transitive sentences (English and German passive syntax differs). Later studies showed reliable between-language priming even across languages with different word orders (e.g., Reference Bernolet, Hartsuiker and PickeringBernolet et al., 2009; Reference Hwang and ShinHwang & Shin, 2019; also see Reference Khoe, Tsoukala, Kootstra and FrankKhoe, Tsoukala, Kootstra & Frank, 2021, for a model). These results are strongly supportive of abstract syntactic accounts in demonstrating that syntactic structure can persist in the absence of any lexical similarity between prime and target sentences, as long as there was syntactic similarity in the prime and target sentences (Reference Bernolet, Hartsuiker and PickeringBernolet et al., 2007). Based on similar evidence with English and Spanish transitive sentences, Reference Hartsuiker, Pickering and VeltkampHartsuiker, Pickering and Veltkamp (2004) proposed a shared-syntax model that extended Reference Pickering and BraniganPickering and Branigan’s (1998) lexical account of priming by adding combinatorial nodes linked to lemmas in two languages rather than only one language. However, bilinguals and language learners do show sensitivity to structure frequencies (i.e., stronger priming for less frequent structures; Reference Hwang and ShinHwang & Shin, 2019; Reference Kaan and ChunKaan & Chun, 2018; Reference Kootstra and DoedensKootstra & Doedens, 2016), which is more consistent with implicit learning accounts.

As in research on monolingual production, further studies considered the extent of lexical involvement in structural processing and showed a high degree of similarity in within-language and between-language priming. Most strikingly, Reference Salamoura and WilliamsSalamoura and Williams (2006) found evidence of single-verb priming comparable to Reference Melinger and DobelMelinger and Dobel (2005) in bilingual speakers: presenting PO-only and DO-only verb primes in one language (L1, Dutch) increased production of PO and DO target sentences, respectively, in a second language (L2, English). Reference Schoonbaert, Hartsuiker and PickeringSchoonbaert, Hartsuiker and Pickering (2007) showed that priming PO and DO syntax from L1 to L2 was enhanced when primes and targets shared the same verb (irrespective of the verb’s cognate status), although the cross-linguistic translation-equivalent boost was smaller than the within-language lexical boost. An analogous translation-equivalent boost was not observed from L2 to L1, which can be explained by differences in the ease with which L1 and L2 verbs activate each other (L1 targets may not reactivate L2 primes as strongly as L2 targets reactivate L1 primes). Cross-linguistic priming of NP structure (e.g., the girl’s apple vs. the apple of the girl; Reference Bernolet, Hartsuiker and PickeringBernolet, Hartsuiker & Pickering, 2012) did show a cognate effect, suggesting a possible role for feedback from phonology to the lemma level in line with interactive models of lexical access (but see Reference Cai, Pickering and BraniganCai et al., 2012, for a different argument). However, the longevity of the translation-equivalent boost – that is, the key parameter supporting interpretations of structural repetition in terms of implicit learning of abstract structures rather than in terms of lexicalist accounts in monolingual speakers – remains to be determined (see Reference van Gompel and Araivan Gompel & Arai, 2018, for a review).

These results are broadly compatible with accounts assuming shared syntactic representations across languages, as well as some degree of dependence on lexical information. Importantly, the fact that proficiency levels vary across speakers as well as within speakers over time offers a unique means of determining how the balance between lexically unsupported and lexically supported syntactic processing might shift with linguistic experience (Reference Hartsuiker and BernoletHartsuiker & Bernolet, 2017; see Reference JacksonJackson, 2018, for a review). For example, the magnitude of cross-linguistic priming without lexical repetition was found to increase with speakers’ proficiency in the target (L2) language but the magnitude of priming with lexical support was larger in speakers with lower L2 proficiency (Reference Bernolet, Hartsuiker and PickeringBernolet, Hartsuiker & Pickering, 2013; also see Reference Jackson and RufJackson & Ruf, 2017). This suggests greater overlap in L1 and L2 structural processes in more proficient speakers but more reliance on the lexicon (i.e., less abstraction of syntax) in less proficient speakers. Such results are consistent with early descriptions of developmental changes in children (see e.g., Reference TomaselloTomasello, 2000, for arguments about lexical dependence; Reference FisherFisher, 2002, for arguments about abstract syntax; Reference Rowland, Chang, Ambridge, Pine and LievenRowland et al., 2012 and Reference Peter, Chang, Pine, Blything and RowlandPeter, Chang, Pine, Blything & Rowland, 2015, for an updated account).

2.2.4 Structural Priming in Dialogue

Going beyond lexicalist and abstract syntactic accounts, psycholinguistic theories have also suggested that repetition of structure, together with repetition of a number of linguistic properties, may serve a communicative function in dialogue (Reference Ferreira and BockFerreira & Bock, 2006; Reference Pickering and GarrodPickering & Garrod, 2004). Interlocutors normally establish common ground and align representations over the course of a conversation. Alignment at the semantic and lexical level in conversational settings is well-established: for example, speakers repeat lexical items previously used in the discourse (Reference Levelt and KelterLevelt & Kelter, 1982) and begin to refer to known referents with reduced lexical descriptions (‘the dancer’ instead of a longer description for a tangram resembling a cartoon person in motion; Reference Clark and Wilkes-GibbsClark & Wilkes-Gibbs, 1986). They also reliably produce target sentences that repeat structures recently used by conversational partners in prime sentences (datives and NPs in collaborative card-matching tasks; Reference Branigan, Pickering and ClelandBranigan, Pickering & Cleland, 2000; Reference Branigan, Pickering, McLean and ClelandBranigan, Pickering, McLean & Cleland, 2007; Reference Cleland and PickeringCleland & Pickering, 2003), both within and across languages (Reference Hartsuiker and PickeringHartsuiker & Pickering, 2008). Such findings have a number of implications.

First, the occurrence of structural repetition in dialogue shows that priming can occur from comprehension to production as readily as from production to production. Early priming studies that used non-interactive paradigms involved a production task in both prime and target trials: participants heard a prime sentence which they had to repeat out loud and then generated a new sentence in target trials (e.g., Reference BockBock, 1986). Repetition of a prime sentence implies that, in principle, any changes observed in target trials could be attributed to the engagement of production processes in prime trials, rather than showing generalisation directly from comprehension to production. More recent research has ruled out this explanation: in comprehension-to-production priming studies, participants are exposed to sentences with a given structure in a prime trial without immediately repeating them out loud, and production is then monitored in a subsequent target trial. Reference Bock, Dell, Chang and OnishiBock and colleagues (2007) reported structural priming effects of similar magnitude – and similar persistence across lags – to tasks requiring production of both primes and targets, suggesting that generalisation of structure is not compromised by changes in modality (repetition of prime sentences, however, may serve a more supportive function in second language production; see Reference Jackson and RufJackson & Ruf, 2018).

By analogy to prime-target paradigms used in the lab, utterances produced by one conversational partner in a dialogue serve as ‘comprehension primes’ and utterances produced by the other conversational partner are ‘production targets’. But does structural repetition in dialogue occur automatically or is dialogue ‘special’? Communicative success is often described as a joint effort, prompting questions about modulation of priming by the social nature of the production setting: one might expect repetition effects to be larger in dialogue than in single-speaker, non-interactive settings (or rather, given that language use in interactive settings is the norm rather than the exception, one might expect repetition effects in non-interactive settings to underestimate the magnitude of repetition effects in everyday language use). This is indeed often the case. For example, Reference Branigan, Pickering and ClelandBranigan and colleagues (2000, Reference Branigan, Pickering, McLean and Cleland2007) and Reference Cleland and PickeringCleland and Pickering (2003) reported priming effects that were larger than in most non-interactive studies. Reference Schoot, Hagoort and SegaertSchoot, Hagoort and Segaert (2019) confirmed these differences in a between-participant comparison of priming in an interlocutor-present and interlocutor-absent condition, suggesting that repetition of structure may not simply occur automatically but rather that it may be additionally influenced by speakers’ communicative goals. However, Reference Ivanova, Horton, Swets, Kleinman and FerreiraIvanova, Horton, Swets, Kleinman and Ferreira (2020) obtained priming effects of similar magnitude in interlocutor-present and interlocutor-absent conditions, both with between-participant and within-participant manipulations, and argued that potential differences between interactive and non-interactive priming effects in earlier studies may be due to differences in participants’ attention and engagement in the two types of settings. Consistent with this hypothesis is the observation of stronger priming in corpus data from a goal-driven task than in spontaneous conversation (Reference Reitter and MooreReitter & Moore, 2014).

If dialogue is special, one might also expect the magnitude of repetition effects to vary between two-party and multi-party settings together with participants’ roles in the conversational exchanges. Speakers are indeed more likely to reuse recently heard structures when they are addressed directly by an interlocutor than when are were not addressed directly (i.e., when the prime sentence is addressed to another speaker in the conversational setting; Reference Branigan, Pickering, McLean and ClelandBranigan et al., 2007), consistent with the hypothesis that repetition of structure may be mediated by heightened attention and task engagement. The magnitude of structural repetition can also vary with the identity and social evaluations of one’s conversational partner (e.g., humans vs. computer-like avatars in Reference Heyselaar, Hagoort and SegaertHeyselaar, Hagoort & Segaert, 2017; likeable vs. less likeable confederates in Reference Balcetis and DaleBalcetis & Dale, 2005; teachers evaluated more vs. less positively in Reference Hwang and ChunHwang & Chun, 2018) as well as evaluations of the likelihood of communicative success (e.g., humans vs. computers in Reference Branigan, Pickering, Pearson and McLeanBranigan, Pickering, Pearson & McLean, 2010), although ‘social’ effects can also be observed in non-interactive paradigms (Reference Weatherholtz, Campbell-Kibler and JaegerWeatherholtz, Campbell-Kibler & Jaeger, 2014). Interestingly, the magnitude of repetition does not seem to vary with the degree to which conversational partners align with the participants’ productions (Reference Schoot, Hagoort and SegaertSchoot et al., 2019).

The larger question of whether structural alignment systematically supports communicative success, however, is still open. While communicative pressures can shape language structure (e.g., Reference Christensen, Fusaroli and TylénChristensen et al., 2016; Reference Fehér, Wonnacott and SmithFehér et al., 2016), the evidence linking repetition of structure to communicative benefits is mixed. Structural alignment can reduce processing times in production and comprehension (e.g., Reference Ferreira and BockFerreira & Bock, 2006; Reference Pickering and GarrodPickering & Garrod, 2004), but structural repetition effects need not be partner specific (e.g., Reference Ferreira, Kleinman, Kraljic and SiuFerreira, Kleinman, Kraljic & Siu, 2012) and are not directly correlated with task success (e.g., Reference Branigan, Pickering, McLean and ClelandBranigan et al., 2007; Reference Ivanova, Horton, Swets, Kleinman and FerreiraIvanova et al., 2020). The scarcity of supporting evidence may be due to a number of factors. It is possible that repetition of structure alone is less clearly linked to communicative success than repetition of lexical items (i.e., lexical alignment, which is an explicit and strategic choice made by the speaker; e.g., Reference Suffill, Kutasi, Pickering and BraniganSuffill, Kutasi, Pickering & Branigan, 2021), or that alignment at multiple levels is needed to boost communicative success. It may also be the case that laboratory tasks elicit relatively unchallenging conversational exchanges and thus fail to uncover possible benefits of alignment. Interestingly, Reference Reitter and MooreReitter and Moore (2014) found evidence supporting the alignment hypothesis in rich and unsupervised task-driven dialogues between conversational partners completing the Map Task, that is, a task where one participant gives instructions for drawing a route on a map to another participant and thus where alignment of situation models is critical for task success. Participants who showed more long-term (but not short-term) structural alignment in this task had more similar paths. Drawing causal inferences about the role of abstract syntax from these data is complicated, but the relationship between long-term linguistic adaptation and task performance suggests that alignment needs to be tracked over longer time intervals than is typically done in the lab. Designing suitable tasks to detect such effects is a methodological challenge that we return to in Section 4.1.

3.2.1 Cognitive Load in Linguistic Processing

Cognitive load can vary within the language planning system due to differences in the ease with which utterance increments can be planned or retrieved at different levels of representation. A number of studies suggest that varying cognitive load can affect the scope of grammatical planning processes. Message-level representations that are easier to construct result in increased planning scopes (e.g., Reference Konopka and MeyerKonopka & Meyer, 2014; Reference Kuchinsky and BockKuchinsky & Bock 2010; Reference Skalickyvan de Velde et al., 2014). Reference Konopka and MeyerKonopka and Meyer (2014) used a spontaneous production task in which speakers described pictures of transitive events eliciting active and passive sentences. Analyses of eye movements during production of active sentences (e.g., ‘The dog is chasing the postman’) showed that speakers allocated more attention to both characters shortly after picture onset when the gist of the event was easy to encode and to express linguistically (suggesting planning of a larger message, consistent with Hierarchical Incrementality) but quickly directed their attention to the character they would mention first when the gist of the event was more difficult to encode and to express (suggesting planning of a small, one-character increment at the outset of the planning process, consistent with Linear Incrementality). Further, facilitating generation of an active sentence via structural priming also resulted in a shift towards planning of a larger message shortly after picture onset. Such differences across items and shifts in planning strategies due to structural priming suggest that speakers can change planning strategies flexibly and dynamically. Specifically, speakers appear to prioritise processes that can be completed quickly at the outset of planning, so planning may proceed in larger increments or in small increments for different sentences.

Effects of cognitive load on lexical retrieval have also been observed. Word retrieval is a notoriously costly process, prone to retrieval failures (e.g., tip-of-the-tongue states; Reference Brown and McNeillBrown & McNeil, 1966; Reference Meyer and BockMeyer & Bock, 1992) which increase with the learning of addition languages (e.g., Reference Gollan and AcenasGollan & Acenas, 2004) and in older age (Reference Segaert, Lucas and BurleySegaert et al., 2018). As mentioned above, in picture-naming tasks, there is evidence of parallel activation of upcoming picture names (e.g., Reference Schotter, Ferreira and RaynerSchotter, et al., 2013), but also evidence that the pre-activation of upcoming pictures is influenced by the ease of name retrieval (e.g., Reference KonopkaKonopka, 2012; Reference Malpass and MeyerMalpass & Meyer, 2010; Reference Morgan and MeyerMorgan & Meyer, 2005; Reference Wheeldon and KonopkaWheeldon et al., 2011). It has further been shown that syntactic processing load can affect lexical planning scope. Lexical processing scope is smaller when the sentence structures to be produced vary from trial to trial (e.g., Reference Wheeldon and KonopkaWagner et al., 2010). Conversely, lexical planning scope increases when the syntactic processing load is reduced using structural priming (Reference KonopkaKonopka, 2012, Reference Konopka and KuchinskyKonopka & Kuchinsky, 2015, Reference Konopka and MeyerKonopka & Meyer, 2014; see Reference Wheeldon, Konopka, Rueschemeyer and GaskellWheeldon & Konopka, 2018, for a review). For example, Reference KonopkaKonopka (2012) used a structural priming methodology to facilitate the production of sentences beginning with complex NPs that included semantically related or unrelated nouns (e.g., ‘The axe and the cup are above the book’ vs. ‘The axe and the saw are above the book’). Lexical planning scope was measured by comparing sentence onsets for the two types of NPs in structurally primed and unprimed conditions, that is, after producing prime sentences beginning with complex NPs or simple NPs (see Reference Smith and WheeldonSmith & Wheeldon, 2001; Reference SkalickyWheeldon & Smith, 2003). The results showed earlier retrieval of the second noun when the complex NP structure was primed, as evidenced by the presence of semantic interference delaying sentence onsets, but not when the structure was unprimed. However, increases in lexical planning scope were not observed beyond the initial phrase.

A related question is the extent to which lexical availability can affect syntactic planning scope. Retrieving and buffering words in memory is cognitively demanding, so can speakers extend their syntactic processing scope to encompass available lexical material? Reference Wheeldon, Ohlson, Ashby and GatorWheeldon, Ohlson, Ashby and Gator (2013) tested this by manipulating both lexical availability and initial phrase structure at the same time. Speakers saw a previewed picture followed by an array of four moving pictures, which again elicited sentences beginning with a coordinate or a simple NP. In the critical sentences, the previewed picture was always the second picture to be named. The position of the previewed picture was either unpredictable (i.e., filler trials were used to vary the position) or predictable (always occurring in second position in experimental and filler sentences). Unpredictable preview replicated the benefit for pictures falling within the initial phrase but not beyond it, as observed by Reference Allum and WheeldonAllum and Wheeldon (2009). When preview was predictable, a significant benefit was observed for pictures beyond the first phrase, as well as a significant effect of initial phrase length. This pattern of results is consistent with speakers extending their planning to include some processing of the second picture in a display when it is previewed but is not consistent with the picture’s name being retrieved and incorporated into the grammatical structure prior to speech onset. Interestingly, the effect of picture preview beyond the first phrase was shown to be inhibitory rather than facilitatory in older adults (Reference Hardy, Wheeldon and SegaertHardy et al., 2020), suggesting age-related differences in the ability to hold on to upcoming lexical information while planning sentence-initial phrases. Finally, Reference Wheeldon, Ohlson, Ashby and GatorWheeldon and colleagues (2013) observed no preview benefit for the predictable preview of pictures occurring in the final position of three of a three-noun coordinate phrase, although the effect of initial phrase length (between a two-noun and three-noun coordinate) was significant. Along with similar findings reviewed above (e.g., Reference KonopkaKonopka, 2012), this demonstrates that initial phrase structure does not necessarily determine the upper limit of lexical access, confirming that phrasal and lexical processing scope do not necessarily coincide (see Reference Roeser, Torrance and BaguleyRoeser, Torrance & Baguley, 2019, for similar findings in the planning of written phrases).

Language experience and proficiency also affect the cognitive demands of speaking, and studies of bilingual language planning have shown that planning scope can differ in a bilingual’s dominant and non-dominant language (e.g., Reference Gilbert, Cousineau-Perusse and TitoneGilbert, Cousineau-Perusse, & Titone, 2020; Reference Konopka, Meyer and ForestKonopka et al., 2018). For example, Reference Konopka, Meyer and ForestKonopka and colleagues (2018) compared planning of active SVO sentences (e.g., ‘The dog is chasing the postman’) by Dutch speakers with high proficiency in English. Analyses of eye movements before speech onset showed that the speakers began fixating and thus linguistically encoding the sentence-initial character (the agent: ‘The dog …’) earlier when generating a description in Dutch than in English. Early encoding of the sentence-initial noun implies that speakers allocated fewer resources to the advance planning of the information to be produced next (‘… is chasing the postman’), that is, that they engaged in more linearly incremental planning rather than in hierarchical planning. Thus, speakers were more likely to adopt an opportunistic (or risky, on-the-fly) planning strategy when using their native language, possibly because they would be able to plan subsequent conceptual and linguistic increments (verbs and the patient names) quickly or would be able to correct any errors if they ran into problems from their chosen starting point (‘The dog …’). By comparison, production is more effortful in a second language, so a highly incremental, risky planning strategy is not optimal: when preparing English sentences, speakers allocated more resources to encoding information about the whole event before beginning linguistic encoding of the sentence-initial character, consistent with a hierarchically incremental planning strategy. Reference Konopka, Meyer and ForestKonopka and colleagues (2018) verified that this effect was indeed due to speakers’ preference to encode relational (verb-related) information early in the planning process rather than due only to a delay in retrieving the sentence-initial referent name.

3.2.2 Individual Differences in Cognitive Abilities

The evidence reviewed in the previous section demonstrates that cognitive load can influence grammatical planning scope within the language system. These findings raise the question of whether individual differences in cognitive abilities can have similar effects. There is some evidence for effects on language planning of individual differences in cognitive abilities related to attention and memory. Arguably, cognitive limitations might necessitate the adoption of smaller, less demanding planning increments (e.g., Reference Christiansen and ChaterChristiansen & Chater, 2016). For example, there is evidence of a relationship between working memory (WM) capacity and planning scope (e.g., Reference Martin, Miller and VuMartin et al., 2004; Reference Martin, Slevc, Goldrick, Ferreria and MiozzoMartin & Slevc, 2014). A number of studies have demonstrated that higher performance in WM tasks is related to planning of larger increments (e.g., Reference Petrone, Fuchs and KrivovokapićPetrone, Fuchs & Krivovokapić, 2011; Reference SkalickySwets, Desmet, Hambrick, & Ferreira, 2007; Reference Wheeldon and KonopkaSwets, Fuchs, Krivovokapić & Petrone, 2021; Reference Swets, Jacovina and GerrigSwets et al., 2014). For example, Reference Petrone, Fuchs and KrivovokapićPetrone and colleagues (2011) demonstrated a relationship between WM capacity and phrase-initial fundamental frequency (F0, the acoustic measure related to pitch). Phrase-initial F0 was higher for speakers with high WM than for those with low WM. There is a decline in F0 across an utterance, and longer phrases are initiated with a higher F0 than shorter phrases, suggesting that phrase-initial F0 is a measure of planning scope (see also Reference Fuchs, Petrone, Krivokapić and HooleFuchs, Petrone, Krivokapić & Hoole, 2013). In a different approach, Reference Swets, Jacovina and GerrigSwets and colleagues (2014) related individual differences in WM to speakers’ performance in an interactive speech production task in which participants directed a listener to move pictured objects in grids on a screen. Speakers with better WM performance showed evidence of a broader planning scope in this task. They were more likely to look at the third object in a scene (e.g., a wheel) prior to initiating a sentence such as ‘The cat moves below the train and the wheel moves above the train’. They were also more likely to produce disambiguating modifications to the first NP such as ‘The four-legged cat moves below the train … .’. Onset latencies to contrasting sentence structures did not vary for low- and high-WM speakers; however, these data suggest that high-WM speakers were able to plan further ahead than low-WM speakers within the same time frame. In a follow-up study, Reference Wheeldon and KonopkaSwets and colleagues (2021) investigated whether differences in language requirements interact with individual differences in cognitive factors, such as WM and processing speed. Using a similar methodology to Reference Swets, Jacovina and GerrigSwets and colleagues (2014), they tested speakers of English, French and German. English and German are Germanic languages that allow modifiers in NPs to occur before or after the noun (e.g., ‘the four-legged cat’, ‘the cat with four legs’). In contrast, modifiers in French are almost always post-nominal (e.g., ‘le chat à quatre pattes’). Previous research has shown faster latencies for the planning of post-nominal modification (Reference Brown-Schmidt and KonopkaBrown-Schmidt & Konopka, 2008; Reference Myachykov, Scheepers, Garrod, Thompson and FedorovaMyachykov et al., 2013), suggesting more incremental planning of such phrases. Reference Wheeldon and KonopkaSwets and colleagues (2021) also found evidence for more incremental planning in French speakers than in English and German speakers. French speakers also showed a relationship between speech latency and individual differences in processing speed, which was not observed in the Germanic language speakers. However, the data patterns were not robust, and the relationship between WM and planning scope observed by Reference Swets, Jacovina and GerrigSwets and colleagues (2014) for English speakers did not replicate. Nevertheless, the study is the first to address the possibility that cognitive capabilities may differ in the extent to which they predict planning scope in speakers of different languages.

Finally, attention is a multifaceted ability (Reference Miyake, Friedman and EmersonMiyake et al., 2000) and individual differences in some components of attention predict picture-naming performance (Reference Piai and RoelofsPiai & Roelofs, 2013; Reference Wheeldon and KonopkaShao, Roelofs & Meyer, 2012). In phrase production, the ability to sustain attention has been shown to affect production latencies (e.g., Reference Jongman, Meyer and RoelofsJongman, Meyer & Roelofs 2015; Reference Jongman, Roelofs and MeyerJongman, Roelofs & Meyer, 2015). For example, Reference Jongman, Meyer and RoelofsJongman, Meyer and Roelofs (2015) measured individual differences in sustained attention using a continuous processing task (CPT) involving the monitoring of a series of digits for the target digit 0. Single digits were presented for 100 ms and participants made a button-press response to the target digit. Participants with lower CPT performance also showed an increase in their number of slow responses in the production of conjoined NPs in L1 Dutch (e.g., ‘de wortel en de emmer’, i.e., the carrot and the bucket), consistent with the effect of lapses of attention. A similar correlation was shown when single-picture naming was followed by a non-linguistic arrow-categorisation task. The findings suggest that speakers need to maintain attention when coordinating the production of NPs, either with another NP or a non-linguistic task. While these results do not speak directly to effects of sustained attention on processing scope, they highlight the need to account for individual differences in cognition in language planning research.

3.2.3 Cognitive Load in Dialogue

Finally, most sentences are spoken in conversational contexts which demand much more than the planning of one’s own utterances. An interlocutor must comprehend the speech of other speakers, keep track of what they have said, and constantly update their representation of the unfolding discourse. These processes are demanding of both attention and memory (e.g., Reference Barthel and SauppeBarthel & Sauppe, 2019; Reference Fairs, Bögels and MeyerFairs, Bögels & Meyer, 2018; Reference Fargier and LaganaroFargier & Laganaro, 2016). Moreover, timing in conversational turn-taking is very tight, with a new speaker often taking the floor about 200 ms from the offset of the previous speaker’s utterance (e.g., Reference Levinson and TorreiraLevinson & Torreira, 2015). This raises the possibility that speakers may reduce their processing scope in order to speed up utterance onset under conversational time pressure. However, the literature on turn-taking focuses on the degree to which utterance planning occurs in parallel with listening to the current speaker, rather than on changes in planning scope (e.g., Reference Barthel, Sauppe, Levinson and MeyerBarthel, Sauppe, Levinson & Meyer, 2016; Reference Lindsay, Gambi and RabagliatiLindsay, Gambi & Rabagliati, 2019; Reference Sjerps and MeyerSjerps & Meyer, 2015) and there is evidence that the speed of turn-taking might be overestimated (e.g., Reference Corps, Knudsen and MeyerCorps, Knudsen & Meyer, 2022). Some studies have shown that pause length in turn-taking is sensitive to measures of the length of the utterance to be produced (e.g., Reference Roberts, Torreira and LevinsonRoberts, Torreira & Levinson, 2015; Reference Torreira, Bögels and LevinsonTorreira, Bögels & Levinson, 2015), but the effect of conversational constraints on the scope of grammatical encoding remains poorly understood. Nevertheless, it remains true that, in order to keep the floor in a conversation, it is important to avoid long planning pauses between one’s own utterances, and there is some experimental evidence that planning scope can be reduced under conditions of increased time pressure (e.g., Reference Ferreira and SwetsFerreira & Swets, 2002).

4.1.1 Paradigms

Picture-naming paradigms that elicit simple sentences consisting of NPs (e.g., ‘The A and the B’) meet the criteria outlined above as these sentences are relatively easy to elicit without training. Studies employing such paradigms have provided much of the initial evidence about vertical information flow in the production system (i.e., top-down information flow as well as feedback from lower levels to higher levels in the lexicon) and horizontal information flow (i.e., estimates of the amount of information that can be planned in parallel at the message level and sentence level). However, the production of more complex sentences is needed to address theoretical issues such as planning of long-distance dependencies. To elicit complex structures, picture-naming studies present participants with hand-drawn pictures or photographs showing transitive or dative events (Reference Griffin and BockGriffin & Bock, 2000; Reference Konopka, Meyer and ForestKonopka et al., 2018; Reference SkalickySegaert, Wheeldon & Hagoort, 2016). Often, picture-naming studies include training blocks where participants are taught to use specific words prior to starting the main production task, or participants are shown printed words to be used in their picture descriptions on a trial-by-trial basis (e.g., Hartsuiker and colleagues; Reference Ziegler and SnedekerZiegler & Snedeker, 2018). Studies eliciting sentences that communicate messages that are hard to depict employ sentence memory tasks: sentences are presented to participants either in full or word by word, and the task for participants is to then repeat these sentences back from memory after a short interval (e.g., Reference FerreiraFerreira, 2003; Reference Konopka and BockKonopka & Bock, 2009; Reference MommaMomma, 2022). Despite evidence that speakers will reproduce these sentences from a gist representation (rather than repeating sentences verbatim from information stored in working memory), the prior processing of the full structure may still have consequences for planning scope that differ from more spontaneous sentence production methodologies.

These paradigms are far from exhaustive in terms of capturing the complexity of language produced outside of the lab. Nevertheless, they have provided data that, so far, can be only partially accounted for by existing computational models. A promising approach for future research is to adapt such paradigms for use in dialogue settings to be able to test effects of interactivity and conversational history.

4.1.2 Dependent Measures

Studies using priming paradigms, such as the structural priming paradigm, have typically relied on two dependent measures: binary outcomes in each trial (such as repetition of the primed structure or production of the unprimed structure), which are aggregated to compute the magnitude of the priming effect across conditions, and onset latencies in the sentences with primed and unprimed structures, which are aggregated to compute facilitation effects across conditions. The former provides a measure of changes in the degree to which the production system is biased to select one structure over another (see Section 2: Reference BockBock, 1986; Reference Chang, Dell and BockChang et al., 2006; Reference Pickering and BraniganPickering & Branigan, 1998), and is typically used to make claims about learning of a structural alteration. The latter provides a continuous measure of how quickly a particular structural procedure can be implemented (Reference KonopkaKonopka, 2012; Reference Smith and WheeldonSmith & Wheeldon, 1999), regardless of whether speakers select the primed or unprimed structure, and provide key insight into the planning process (Section 3).

Comparisons of sentence latencies across priming conditions are also particularly useful for assessing the effects of structural primes on sentences with preferred structures, that is, sentences that are hard to prime. For example, speakers overwhelmingly prefer to use active syntax than passive syntax. Given that selection of active syntax is often at ceiling, structural primes rarely increase selection of active syntax further (frequent structural alternatives also show less priming; Reference Jaeger and SniderJaeger & Snider, 2013). However, priming in active sentences is observable with a different measure: sentences that repeat active syntax have shorter onset latencies (Reference Segaert, Weber, Cladder-Micus and HagoortSegaert et al., 2014). This suggests that repetition of structure can have facilitatory effects on production in the absence of changes in structure choice. At the same time, sentence onsets are highly sensitive to a large number of variables that influence the speed with which speakers produce linguistic output under any conditions (e.g., lexical constraints such as the ease of retrieving individual content words) as well as changes specific to structural repetition itself, such as structure-driven changes in planning scope (Reference KonopkaKonopka, 2012; Reference Konopka and MeyerKonopka & Meyer, 2014).

A particularly informative approach in sentence planning is the use of eye tracking to track spontaneous production of sentences that can be elicited with visual stimuli. Sentence onsets vary across studies but speakers rarely begin speaking before 1000 ms. Tracking participants’ eye movements provides a rich implicit record of what information participants began encoding at different points in time and how easy this information was to encode before speech onset. This is especially relevant for making inferences about incremental planning (see Reference Norcliffe, Konopka, Mishra, Srinivasan and HuettigNorcliffe & Konopka, 2015). The richness of the eye-movement record in spontaneous sentence production also presents a challenge: eye movements can capture effects of low-level non-linguistic variables (such as perceptual salience) as well as higher-level linguistic variables (such as message-level encoding difficulty). The time windows in which these variables are likely to influence production in theoretically interesting ways, as well as the way that eye movements reflect an influence of these variables, are still debated (see e.g., Reference Griffin and DavisonGriffin & Davison, 2011; Reference KonopkaKonopka, 2019).