Chapter 5 Sentence Production
Written by Laurel Brehm
Edited by Vitória Piai
Imagine that instead of simply saying “dog!” like was discussed in Section 2, you wanted to describe the scene in Figure 5.1. We might describe the message behind this sentence as something like <🐕, 🏃,🐈> or as <dog, chase, cat>– in either case, this is a shorthand for non-verbalized ideas.
![A simple scene that elicits syntactic alternations. Cat and dog both from Wikimedia Commons. https://commons.wikimedia.org/wiki/File:Running_Cat_01.JPG and https://commons.wikimedia.org/wiki/File:Gleich_hab_ich%27s!_(6865180445).jpg]](img/dog-chasing-cat.png)
Figure 5.1: A simple scene that elicits syntactic alternations. Cat and dog both from Wikimedia Commons. https://commons.wikimedia.org/wiki/File:Running_Cat_01.JPG and https://commons.wikimedia.org/wiki/File:Gleich_hab_ich%27s!_(6865180445).jpg]
.jpg){kind=link}
If you are speaking in English, you might say that in this picture, ‘The dog is chasing the cat.’ But you can also say ‘The cat is running from the dog’ or ‘The cat is being chased by the dog’ to convey similar ideas.
The fact that there are different ways to describe the same event highlights the decisions we have to make for sentence production. We have to decide what the objects are called, decide which of them is the agent of the action, and decide which of them comes first in the sentence. These decisions are all parts of the sentence production process, and each has been studied extensively in psycholinguistics.
5.1 The Bock and Levelt model of sentence production
K. Bock and Levelt (1994) is a foundational model of sentence production (Figure 5.2). This model does what was described in the last paragraph. Deciding what objects are called (is it ‘cat’ or ‘kitten?’) is lexical selection. Deciding which is the agent of the action (the dog is the ‘do-er’ of the chasing) is function assignment. Finally, deciding which content word comes first (cat or dog) is constituent assembly. This model presumes that lexical selection and functional assignment happen in parallel in a first stage of planning called ‘functional processing.’ In a second stage of planning called ‘positional processing,’ constituent assembly happens along with adding inflections to the lemmas (marking cat, is, and dog all as singular).
One crucial piece of evidence for Bock and Levelt’s model is that when we select the wrong word (i.e., picking ‘kitten’ instead of the intended ‘cat’ for the object noun), it often gets selected with the right morphology and case (see Stemberger 1982). This is easier to see in languages other than English which have a richer inflectional paradigm: for example, in Dutch, the gender on ‘het poesje’ (the kitten) is different than ‘de kat’ (the cat), which makes errors like these more obvious. (See Berg 1987 for examples in German and other Germanic languages). This pattern of dissociation means that functional processing must happen before positional processing: even if functional processing goes wrong, positional processing can (and often does) still go right.
![Bock and Levelt model of sentence production [after @bock1994language]](img/bLmodel.png)
Figure 5.2: Bock and Levelt model of sentence production (after K. Bock and Levelt 1994)
5.2 Syntactic Priming
Another line of evidence for two separate stages of sentence planning is syntactic priming, also known as structural priming or syntactic persistence. As described above, there are often several ways to describe the same event, with nouns in different syntactic positions. These are called alternations; common ones in English include active/passive (‘the dog chased the cat’/ ‘the cat was chased by the dog’), double-object/prepositional-object datives (‘the dog gave the cat the ball’/ ‘the dog gave the ball to the cat’), and the spray/load construction (‘the dog sprayed the wall with mud’/ ‘the dog sprayed mud onto the wall’). All of these can be primed. J. K. Bock (1986) showed that people are much more likely than chance to re-use the structure of a previous sentence. That is, a sentence like ‘the mouse was eaten by the dinosaur’ will prime the production of ‘the cat was chased by the dog’ to describe Figure 5.1. Syntactic priming is evidence that function assignment has psychological reality. However, syntactic priming is also evidence that function assignment and lexical selection might happen in the same stage. The lexical boost phenomenon is where priming effects are increased by repeating the same verb: ‘the mouse was chased by the dinosaur’ primes ‘the cat was chased by the dog’ at an especially high rate (e.g. Pickering and Branigan 1998). This pattern suggests that choices of structures and words are not independent.
5.3 Adding Inflections
Many languages add marking to nouns, pronouns, verbs, or articles for properties like number, gender, person, or case. This happens during positional encoding. A type of speech error called agreement attraction shows how this process goes awry. In English, when there are two nouns in a sentence (‘The key to the cabinets’), the verb is supposed to be controlled by the first (head) noun, but people sometimes produce a verb that agrees with the second (local) noun […*were shiny; K. Bock and Miller (1991)]. This is evidence that adding inflections happens late in sentence planning. Agreement attraction is pretty universal: it happens in lots of kinds of inflections, in lots of languages, including Russian (Lorimor et al. 2008), Spanish (Vigliocco, Butterworth, and Garrett 1996); Hebrew (Deutsch and Dank 2009); French (Vigliocco et al. 1996), Dutch (Vigliocco et al. 1996), and Italian (Vigliocco, Butterworth, and Semenza 1995), and in all languages it is influenced more by grammatical form than meaning (see for more K. Bock and Miller 1991; Vigliocco et al. 1996).
5.4 Ordering constraints
In the K. Bock and Levelt (1994) model, constituent assembly– choosing what order to put the parts of the sentence in– also happens during positional encoding. The same underlying syntactic structure can be assembled in different ways: for example both ‘the cat was chased by the dog’ and ‘the dog chased the cat’ have the same agent/patient assignments (cat = patient, dog = agent), but a different constituent order. At constituent assembly, non-syntactic factors related to accessibility (i.e., how quickly nouns can be retrieved) impact production. Two of these factors that are important are givenness (whether something has been previously discussed) and animacy (the degree to which an item is human-like). Psycholinguists think that the bias where given items appear before new items comes about during the constituent assembly stage in languages such as English (e.g. J. K. Bock and Irwin 1980), Japanese (e.g. Ferreira and Yoshita 2003), and Ukrainian and Russian (e.g. Mykhaylyk, Rodina, and Anderssen 2013). However, psycholinguists are less sure where the animacy-first bias in English (K. Bock, Loebell, and Morey 1992) and Japanese (Tanaka et al. 2011) occurs at function assignment or constituent assembly; likely, it has an effect at both stages.
Not all languages may be the same in terms of constituent assembly. Languages with flexible word ordering, such as the Australian languages of Pitjantjatjara (Kidd et al. 2025) and Murinthpapa (Nordlinger, Rodriguez, and Kidd 2022) also show a preference for both agent-first and human-first orders. Tracking people’s eyes while they are planning to speak shows that in both languages (which are quite different from each other), people split their attention between the agent and the patient of the sentence for about 500 ms before beginning to speak, as if trying to decide the structure and constituent order simultaneously. This is very different from English, where people seem to quickly direct their attention to the first-named referent in the sentence (Griffin and Bock 2000). This suggests that Pitjantjatjara and Murinthpapa have more hierarchical planning than English does. This means that it is possible that these and other free word order languages do not require a separate constituent assembly phase of planning, and all of their decisions about word order happen at function assignment.
5.5 Evidence for interactivity between levels
The data from free word order languages put a wrench in the separation between functional and positional processing in K. Bock and Levelt (1994). Some psycholinguists now think that these two stages are also not entirely independent in languages like English and Dutch. Konopka and Kuchinsky (2015) found that similar meaning verbs (‘duwt,’ pushes, and ‘tackelt,’ trips) have especially strong syntactic priming effects in Dutch, but only in the typical SVO verb order and not in the less frequent Dutch VSO order. This means that constituent assembly influences the amount of interplay between the message and functional processing (see also Fukumura and Yang 2024).
Neurolinguistic evidence also suggests that syntax and lexical items are probably not stored or used separately, providing more support for interactivity in sentence production. A recent meta-analysis finds that while there are separate brain areas that tend to be activated the most for syntactic processing and lexicalization in production, both also tend to activate lots of the same areas (Yeaton 2025). So, instead of having lexical access and function assignment occurring separately and in parallel, like in K. Bock and Levelt (1994), it might be that lemmas are stored with syntactic frames or treelets that then are composed in preparation for production. This composition might occur in a single operation called Unification or Merge (Hagoort 2017) or in two separate syntactic operations relating to hierarchical planning and linearization (Matchin and Hickok 2020).
Take-home messages
Formulating sentences involves two stages of syntactic planning.
- Functional processing, during which words are selected (lexical selection) and assigned to thematic relations like agent and patient (function assignment)
- Positional processing, during which structures are linearized and mapped to grammatical constitutents (constituent assembly) and inflections are added (inflection)
Exercise 5.1 How many different ways can you describe Figure 5.1? List them all out. Underline the agent (‘do-er’) of each sentence and the grammatical subject (controller of the verb). Notice how the two processes can be dissociated.
Exercise 5.2 As you go about your day, listen for errors in number/gender inflection or in case marking. In the K. Bock and Levelt (1994) model, what processes went right, and what processes went wrong?
Suggestions for further reading
Two other models of sentence production with slightly different focuses are:
MacDonald (2013), on the relationship between production and commprehension, over time.
Eberhard, Cutting, and Bock (2005), on how meaning and number marking both influence number agreement.
There’s also lots more to read on syntactic priming, which is perhaps the best studied phenomenon in sentence production. Some further reading shows that it occurs even at long delays (e.g. K. Bock and Griffin 2000), across a bilingual’s two languages (e.g. Robert J. Hartsuiker et al. 2008), in small children (Savage et al. 2003), and in people with memory loss (Ferreira et al. 2008); see Mahowald et al. (2016) for a meta-analysis of syntactic priming. Despite all this research, there is some debate on what cognitive mechanism creates syntactic priming. It might occur because of implicit learning (e.g. F. Chang, Dell, and Bock 2006), short-term activation changes (e.g. Pickering and Branigan 1998), or a combination of both (e.g. Ferreira and Bock 2006).