Priming of Possessive Constructions in German: A Matter of Preference Effects?

Abstract

We investigated structural priming in adult native speakers, focusing on possessive constructions in German, where the two alternative structures involved differ in frequency. According to error-based learning approaches to priming, the less frequent structure should lead to a larger prediction error and larger priming effects than the more frequent structure. In a comparison of preferences during a pretest and preferences during priming, we did not find evidence of such an inverse preference effect. Moreover, during priming, we observed increasing production rates of the preferred structure, hence, a cumulative priming effect. In line with hybrid models of priming, we propose that two mechanisms, namely, a mechanism learning from input as well as a mechanism accumulating activation during comprehension and production, are involved in the temporal development of priming effects. Moreover, we suggest that the interaction of the two mechanisms may depend on prior experience with the alternative structures.

1. Introduction

Structural priming refers to the tendency of speakers to reuse previously heard or read constructions (Bock 1986). An error-based implicit learning mechanism has been proposed to explain the phenomenon (Chang et al. 2000). The idea that structural priming is an instance of learning comes with specific claims about the sensitivity of the learning mechanism to the given input and produced output, about the representations that are involved in learning, and about the changes that apply to these representations over time (for an overview, see, e.g., Jaeger and Snider 2013; Pickering and Ferreira 2008). The current study takes up these issues and investigates priming effects for possessive constructions in adult speakers of German.

The dual-path error-based learning model (henceforth, error-based learning model) assumes the parallel processing of thematic relations and of linear order, the latter relying on transition probabilities (Chang 2002; Chang et al. 2006, 2012). During syntactic processing of the input, speakers continuously update their knowledge about the likelihood of a sequence in a given context. In line with other predictive coding accounts (following Rao and Ballard 1999; Friston 2018), it is assumed that language users will predict what syntactic unit they are most likely to encounter next, at least in the case that the situation is sufficiently constrained to allow for such predictions (for an overview, see Huettig 2015). If the subsequent input deviates from a current prediction, language users will experience a prediction error, and this will lead to an updating of their internal model of the likelihood of a specific sequence. This mechanism constitutes the fundament of learning. Based on the new internal model, speakers will generate updated predictions for the subsequent input, and will themselves be relatively more likely to produce the structure they had not predicted to occur previously. Importantly, this process is assumed to be sensitive to the size of the prediction error, i.e., the probability difference between the predicted and the encountered input: larger prediction errors are assumed to lead to larger adaptions of the internal model. In the context of structural priming, this assumption has been claimed to be mirrored in larger priming effects for relatively rare, and hence unexpected, structures compared to alternative, more frequent structures that express the same meaning. There is evidence of such inverse preference effects (e.g., Hartsuiker and Kolk 1998; Scheepers 2003; Jaeger and Snider 2013), but they appear under a variety of conditions, and it is not always clear whether it is necessary to assume a learning mechanism to explain the effect.

This paper is structured as follows: we will first summarize what is known about inverse preference effects during structural priming and about the temporal development of priming effects. We will present explanations as to why such patterns are observed under some conditions but not all. We will then introduce the structures under investigation in the current study, our research questions, and data from a structural priming experiment. Our data provide evidence of preference, rather than inverse preference, effects. In the discussion, we will propose potential explanations for the pattern, point to limitations of our study, and outline avenues for future research.

2. Background: Factors in the Development of Structural Priming Effects

In studies on structural priming during language production, the effects of primes on the use of structural alternations are measured. The relative proportion of two alternative expressions is expected to vary as a function of the prime structure (e.g., Bock 1986). For example, priming of the dative alternation is attested for PO (prepositional object) and DO (double object) primes and responses, if the relative proportion of PO responses (PO/(PO + DO)) increases after the presentation of PO primes as compared to DO primes. There are at least two prominent accounts of structural priming that differ in the mechanisms assumed to underly these effects, a residual activation account and the implicit learning account that was briefly mentioned above. The residual activation account (Pickering and Branigan 1998) posits a lexical network, with nodes for lemmas and for the syntactic contexts they appear in. According to this model, the production of a word in a context leads to an increase in the activation of the corresponding lexical and structural nodes. This pre-activation leads to the easier activation of these nodes during subsequent encoding, giving rise to priming effects. The original proposal focuses on verb lemmas, but it was also adapted to noun lemmas (Cleland and Pickering 2003; Bernolet et al. 2013). The implicit learning account holds that priming is an instance of learning, in that a heard or read prime affects the speaker’s internal representation of the probability of the corresponding construction or sequence of words (Chang et al. 2000). Priming effects are modeled as resulting from a mismatch of the predicted and the encountered input, i.e., a prediction error (Chang et al. 2006).

Over the last decades, the strength of structural priming has been found to be modulated by several factors that are potentially informative about the underlying mechanisms. These factors relate to the linguistic similarities between primes and responses, to biases towards one of the alternative structures, and to the temporal development, as well as the longevity, of the priming effects.

Priming arises when primes and responses share linguistic properties, e.g., phrase structure (Bock 1986). Phrase structural effects are boosted when lexical materials, such as verbs in the case of the dative alternation (Pickering and Branigan 1998) or nouns in the case of possessive constructions (Bernolet et al. 2013), are repeated in primes and responses. According to the residual activation model, the lexical boost is due to the repeated activation of the link between a lexical and structural node (Pickering and Branigan 1998). The error-based learning model needs an additional mechanism, i.e., explicit memory, to deal with these lexically mediated effects, since the dual-path mechanism operates on abstract linguistic categories and produces relatively smaller effects (Chang et al. 2006).

In structural alternations, the use of structural variants may be biased, and one variant may be preferred over the other variant. For example, the PO structure may be produced more frequently in dative alternation contexts than the DO structure. Such preferences have been found to influence structural priming. The inverse preference effect refers to the finding of relatively stronger priming by less preferred structures than by more preferred structures (cf. Ferreira and Bock 2006; Hartsuiker and Kolk 1998; Scheepers 2003). The repeated use of an otherwise less preferred structure within the experiment changes the relative preference of this structure more dramatically than that of the more preferred structure. For the dative alternation referred to above, this means that a DO prime may exhibit a stronger effect on responses than a PO prime. Whereas such a pattern of results is not easily accounted for by the residual activation model, it may be explained by error-based learning: The model operates with a mechanism that predicts the next word in a structure (e.g., a prime) on the basis of transition probabilities. It learns from a prediction error that occurs whenever the predicted word (class) and the actual word (class) in the input differ. The lower the probability of the actual word is, the larger is the prediction error it evokes and the stronger is the subsequent adaptation of transition probabilities (Chang et al. 2006). Thus, the error-based learning model predicts inverse preference effects, i.e., stronger priming of the less probable word. In addition, several studies found an impact of verb-specific preferences on the strength of priming (Bernolet and Hartsuiker 2010; Jaeger and Snider 2013). This has been accounted for in terms of surprisal (an unexpected word is more informative than an expected one (Hale 2001); cf. Scheepers 2003), which resembles the error-driven model and allows for an integration of verb-specific information, without assuming an additional explicit memory mechanism (cf. Chang and Fitz 2014).

To test for inverse preference effects, preferences during priming have to be compared to preferences without priming. Priming-neutral preference data may be gained from baseline trials within a priming experiment or from a corpus. The comparison of corpus data and experimental production data may be problematic, as constraints on the data may differ. Most corpora are based on written language, whereas many priming experiments elicit spoken language, such that the degree of elaboration, complexity, and the incidence of academic language will differ. Moreover, experimenters may impose constraints on materials that increase the probability of the less preferred structure, as in the case of animacy constraints on the patient in transitive events to increase the proportion of passive responses (Bock 1986). If corpus data do not mirror these constraints, they will not be directly comparable, and it would be better to measure preferences within the experimental situation with materials similar to those presented in the priming trials.

Next, it also has to be taken into account that asymmetrical priming could be related to ceiling or floor effects. The structure that is already prevalent may have reached an upper limit (ceiling) and therefore may not gain anymore by priming. The only structure that can be primed, then, is the less preferred structure (Hartsuiker et al. 1999; Segaert et al. 2016). The assumption of ceiling effects is plausible, but it is far from clear which limit applies and whether it is reached at around 80% (Hartsuiker et al. 1999) or above 90% (Segaert et al. 2016). The counterparts of ceiling effects are floor effects for structures that are rarely produced (see Pickering et al. 2002).

Time is an additional factor in structural priming. Even though intervening trials with a structure different from the primes may reduce priming effects (Bock and Kroch 1989), structural priming has been shown to persist across trials (Bock and Griffin 2000), as well as from a priming phase to a response phase (Kaschak 2007). Effects of structural priming are attested a week (Kaschak et al. 2011) or even a month after the experimental session (Heyselaar and Segaert 2022). By contrast, the lexical boost decays quickly and cannot be traced across trials (Hartsuiker et al. 2008). This longevity of “abstract”, as opposed to lexically mediated, structural priming effects favors the implicit learning account, as well as the assumption of a different mechanism behind the lexical boost (Chang et al. 2006).

Finally, cumulative priming effects may arise as the experiment proceeds (Hartsuiker and Kolk 1998; Hartsuiker and Westenberg 2000). Structural priming does not decay completely but persists across trials (Reitter et al. 2011). Subsequent primes (Jaeger and Snider 2013) and responses (Jacobs et al. 2019) add to the overall priming potential, such that it accumulates over a repeated input or use of the critical structures. The strength of priming thus increases across experimental trials, even though additions are not necessarily linear (cf. Bernolet et al. 2013; Jaeger and Snider 2013). This pattern is predicted by residual activation models. The error-based learning model predicts the even longer persistence of priming effects, but they are suggested to develop slowly (Chang et al. 2006). Moreover, if cumulative effects are to be explained by an error-based mechanism, the less frequent structure should gain relative to the more frequent one. In line with this, Jaeger and Snider (2013) report both the cumulative priming of the less frequent structure and a decrease in verb-specific inverse preference effects across prime trials, the latter being commensurate with a decrease in surprisal by the less frequent structure (Jaeger and Snider 2013). Some approaches have combined error-based learning and activation-based processes, in order to account for temporal developments (Jacobs et al. 2019; Reitter et al. 2011), a suggestion that we will come back to in the Discussion section.

The temporal development of priming effects across experimental trials is theoretically intriguing, but it also has methodological implications: Since preferences are subject to priming, it might be less advisable to measure preferences by presenting baseline primes intermixed with experimental primes than to measure preferences in a phase before the experimental manipulation of primes (for a recent combination of both, a pre-priming baseline phase and baseline primes intermixed with critical primes, see Van Lieburg et al. 2023).

To summarize, “abstract” structural priming effects differ from lexically mediated effects, in that the former are long-lived and the latter are short-lived. Both may be subject to structural preferences. Ceiling and floor effects may hinder the development of priming effects. If not overridden by ceiling or floor effects, error-based models predict that priming effects should be stronger for less expected, e.g., less frequent, structures. Such an inverse preference effect can best be tested by comparing the impact of primes on structural preferences to preferences that have been established independent of priming. In addition, there may be cumulative effects during priming, such that structural preferences change across a priming phase and one structure gains in production rates at the expense of the other structure. If cumulative priming is due to an error-based mechanism, then it should occur in favor of the initially less frequent structure (cf. Jaeger and Snider 2013), even though the size of the error-based effect may decrease after repeated exposure and subsequent changes in syntactic representations. As for the longevity of effects, error-based models predict changes to structural preferences to be small but lasting. Activation models assume that activation may accumulate across trials, but it is not clear how long such effects can persist.

3. The Current Study

The aim of the current study is to provide further evidence on the potential impact of structural preferences on the size, the temporal development, and the persistence of priming effects. We look at constructions that have not been previously looked at, namely, possessive constructions in German (genitive construction vs. prepositional phrase construction, cf. (1) and (2) below). The structural bias associated with these structures is quite strong, which should lead to large prediction errors and thus provides an important test case for error-based models. We do not evoke a lexical boost. The relations expressed are always those between a protagonist and an object possessed by the protagonist (e.g., the artist’s rose), in a design modeled after studies by Bernolet (2008) on the priming of possessive constructions in English. We measure preferences across priming trials and in tests preceding and following the experiment (pre- and post-tests), such that we can test for:

-

An (inverse) preference effect (by comparing pretest preferences to the preferences induced during priming);

-

Cumulative priming during the priming phase (by observing whether one structure gains in production rates during the priming phase relative to the alternative structure);

-

Persistence of the effects during a post-test in the absence of primes (by comparing production rates between the pretest and post-test).

The alternative structures we focus on are a postnominal genitive construction (1) compared to a postnominal prepositional phrase attribute with von (2).

(1)	Genitive construction:
	Die	Rose	des	Künstlers-s	ist rot.
	the.NOM	rose	the.GEN	artist-GEN	is red

(2)	Prepositional phrase construction:
	Die	Rose	von	dem	Künstler	ist rot.
	the.NOM	rose	of	the.DAT	artist	is red
	The artist’s rose/the rose of the artist is red.

Note that we also conducted a second experiment, in which we compared the postnominal genitive construction to the (archaic) prenominal genitive construction. As there were no occurrences of the prenominal genitive in the pretest or post-test of that experiment, which makes it impossible to compare priming effects to a baseline or to test the persistence of effects, we aborted this experiment after having tested 16 participants. For the sake of completeness, this experiment is reported, and its results are briefly discussed, in Appendix A.

Concerning postnominal genitives and postnominal prepositional phrases with von (which we will refer to simply as “von-phrases” in the following), a recent large-scale corpus analysis (Kopf 2021) found that both are frequently used to express possessive relations (see also Kopf and Weber 2022; Smith 2003). Which structure is preferred is related to the formal properties of the attributive noun phrase; for noun phrases containing an article, there is a preference for the genitive over the von-phrase (e.g., Die Nutzung der neuen Technologien, ‘the usage the(GEN) new(GEN) technologies’, is strongly preferred over Die Nutzung von den neuen Technologien, ‘the usage of the new technologies’), while for noun phrases without an article, genitives and von-phrases are used to an equal degree, approximatively (Die Nutzung neuer Technologien, ‘The usage new(GEN) technologies’, is used to a similar degree to Die Nutzung von neuen Technologien, ‘the usage of new technologies’—example taken from Kopf and Weber 2022; see also Smith 2003). This pattern has been explained by a tendency of speakers to use genitive-marked noun phrases more often when this results in unambiguous case marking and less often when the case marking would be ambiguous. As the adjectival and noun declension paradigms are characterized by syncretism, speakers are comparatively more hesitant to use genitives for noun phrases that are case marked on nouns and/or adjectives only. For the stimuli of the current study, all noun phrases used by the confederate speaker contained a definite article agreeing with a masculine noun, which unambiguously inflected for case. This means that there were ideal conditions for the use of the genitive. It has to be noted, however, that the available corpus data are written (Kopf 2021; Smith 2003) and that von-phrases may be more frequent in informal spoken speech (cf. Bredel and Maaß 2016).

The experiment followed the confederate scripting paradigm (Branigan et al. 2000) that has been shown to produce stronger effects than other priming paradigms (Mahowald et al. 2016). We determined relative frequencies in a pretest, priming, and post-test phase to control for preferences induced by experimental materials and the spoken modality.

It is worth recalling that, as outlined in the introduction, if the prevalence of genitive structures over von-phrases can be confirmed in the pretest, error-based learning models would expect the subsequent priming effects to be stronger for the von-phrase than for the genitive phrase. This should be reflected in an inverse preference effect (e.g., in the trials where the von-phrase is used as a prime during the priming phase, it should cause a larger change in the relative frequency of the von-phrase and genitive than when the genitive is used as prime). Moreover, if error-based models are applied to cumulative priming, they would also predict that the relatively less frequent structure should continue to cause a stronger prediction error than the relatively more frequent structure, as long as it remains the less frequent one overall. While the size of the prediction error could decrease, this should nevertheless lead to an overall increase in the comparatively less frequent structure over time (cf. Jaeger and Snider 2013). Finally, according to error-based models, if an advantage for the initially less frequent structure can be observed, this effect should persist during the post-test phase.

4. Method

4.1. Participants

Forty native speakers of German (23 female, 16 male, 1 non-binary; mean age: 25, range: 19–34) participated in the experiment. All of them were students at TU Dortmund University or at Heidelberg University, and none of them were bilingual as a child.

4.2. Materials

We created two sets of pictures, a description set and a verification set. An item was defined as the combination of a picture from the verification set, together with the description of this picture given by the confederate speaker (for experimental items in the priming phase: the prime), and the description picture meant to elicit the description by the participant (the target). For the pre- and post-test phase, there were 12 experimental items and 24 filler items. For the priming phase, there were 24 experimental items and 48 filler items. Figure 1a,b illustrate a combination of a verification picture and a following description picture from a pretest trial without priming (Figure 1a), and a priming trial of the priming phase (Figure 1b). The participant had to first verify the confederate speaker’s description of the verification picture (1) and then turn to the description picture (2) to describe it.

Items were created, closely following the procedure described in Bernolet (2008) (see also Bernolet et al. 2013), except that we used a larger variety and greater number of fillers and a higher number of different characters, to reduce the likelihood of different experimental trials influencing each other. Different items were created for the pre- and post-test phase than for the priming phase.

Concerning the priming phase, for the experimental items, both the verification picture and the description picture consisted of drawings of two characters with an identical object in their hand (identical to the set-up in Bernolet 2008, see Figure 1b). The object held by one of the characters was colored in blue, green, yellow or red, while the rest of the picture was black and white. Pictures were created based on twelve different characters and 48 objects. All characters could be described by masculine, disyllabic nouns in German, which fell in frequency classes 9–12 according to the Projekt Deutscher Wortschatz (mean: 10.83; Leipzig Corpora Collection 2011; cf. Goldhahn et al. 2012), and all ended with the suffix –er and formed the genitive form by adding the suffix –s. The objects were taken from the Snodgrass and Vanderwart (1980) picture database. These descriptions consisted of 48 nouns, 16 of each of the three German gender classes. A total of 26 of the nouns were monosyllabic and 22 were disyllabic. The nouns fell into frequency classes 8–15 according to the Projekt Deutscher Wortschatz (mean for the verification set: 11.75, mean for the description set: 11.92). For the description and verification set of each item, the objects of the verification and the description set were matched for gender and number of syllables. Moreover, for each item, the four characters appearing on the elicitation and description picture were always different from each other, and the color of the object was always identical in the two pictures. Each character appeared twice in each of the following positions: on the verification picture holding the colored object, on the verification picture not holding the colored object, on the elicitation picture holding the colored object, and on the elicitation picture not holding the colored object. The character holding the colored object occurred on the left or right side of the screen for half of the pictures, respectively, and each of the four colors appeared equally often on the pictures and in the confederate’s descriptions. The position of the color was kept constant, such that all red and blue objects always appeared to the left and all yellow and green objects to the right.

Regarding the filler items of the priming phase, all 48 filler items consisted of pairs of the same characters that were also used for the experimental items but without any objects, as in Bernolet (2008). In half of the cases, only one of the two characters was colored (as in Bernolet 2008, again), while in the other half of the cases, both characters were completely colored, always in the same color. Whether one or both characters were colored was crossed across the verification and the elicitation set, such that all four possible combinations (one character colored in the verification set, one in the elicitation set, one in the verification set, two in the elicitation set, and so on) occurred equally often. All colors appeared equally often on the pictures and in the confederate’s descriptions. The position of the colors was kept constant, such that single yellow characters always appeared on the left of the screen, single blue characters on the right, and that in pictures where both characters were colored, these were always either both red or both green.

When describing the verification picture, the confederate used the following three types of structures:

(3)	For filler items with one colored character:
	Der Kellner ist blau.
	‘The waiter is blue’.

(4)	For filler items with two colored characters:
	Der Kellner und der Radler sind blau.
	‘The waiter and the cyclist are blue.’

(5)	For experimental items (genitive or von-phrase constructions):
	Die Glocke des Kellners ist rot.
	Die Glocke von dem Kellner ist rot.
	‘The waiter’s bell/the bell of the waiter is red.’

Two experimental lists were created, such that in each list, the confederate used each of the two genitive structures 12 times and for different items in the two lists. Items appeared in a pseudo-randomized order, in which there were always between one and three filler trials between two experimental trials. To control for any effects of the specific order, about half of the participants received the items in reverse order.

The descriptions of the verification picture were wrong in half of the cases and correct in the other half. For the experimental items, wrong descriptions always consisted of the confederate naming the wrong color. For the filler items where one character was colored, the wrong description always consisted of the confederate naming the wrong character (e.g., “The biker is blue”, for a picture with a blue hunter and a black and white biker). This was done to ensure that participants could not predict whether the color of the object or the object itself was wrongly identified in the wrong descriptions, which presumably should enhance the motivation to use precise descriptions, and hence, possessive structures for the experimental items. Finally, for the filler items where the elicitation picture consisted in two colored characters, the wrong descriptions always consisted of wrongly identifying the color, as was the case in the experimental items. Moreover, the confederate answered with “yes” to half of the descriptions of the participant and with “no” to the other half, with both types of answers being equally likely both for filler and experimental items.

In the pretest and the post-test phase, for the experimental items, we used six different characters and 12 different objects. As in the priming phase, the six characters chosen for the pre- and post-test could be described by masculine, disyllabic nouns in German. These nouns fell in frequency classes 9–12 according to the Projekt Deutscher Wortschatz, had the same mean frequency (10.83), and all ended with the suffix –er and formed the genitive form by adding the suffix –s. The objects could be labeled by 12 different nouns, 4 of them belonging to each of the three genders. Six of these nouns were disyllabic and six were monosyllabic. The pictures of the description set were constructed following the same principles as were applied for the construction of the pictures presented during the priming phase. Each character appeared twice on the pictures holding the colored object and twice not holding the colored object, and an equal number of times at the left and at the right side of the screen. As no priming was intended during this phase, the verification pictures were constructed as were the filler pictures. That is, they always displayed two characters without any objects, and in half of the cases, one character was completely colored, while in the other half, both were colored. All characters appeared equally often in these verification pictures, and the four characters of one experimental item (verification and elicitation picture combined) were always different from each other. All four colors appeared an equal number of times in the verification pictures and the description pictures, and half of the confederate’s descriptions were correct, while the other half were wrong. Each character appeared twice in the description picture holding the colored object and twice not holding the colored object. For the filler items, we used the same six characters and combined them such that there was an equal number of filler pictures with one colored character and both characters being colored, and all four possible combinations of elicitation and verification picture (one person colored/both persons colored, etc.) appeared equally often. All materials, as well as the data to be reported below and statistical analysis scripts, can be accessed at the study’s OSF page.

4.3. Procedure

All experimental sessions were conducted with an experimenter, a naïve participant, and a confederate participant. The naïve participant was not aware of the confederate’s role until after the completion of the experiment.

The naïve participant and the confederate were sitting opposite to each other, such that each of them could see his or her own screen but only the back of the opposite screen, and it was thus impossible for the participant to see that the confederate was reading the descriptions and judgments, instead of making them spontaneously. The experimenter sat such that she could observe both participant and confederate. Sessions were recorded for the post-experimental coding of responses.

At the beginning of the experimental session, the participants gave written consent to participating. Next, the training phase began. The two participants each saw an identical PowerPoint presentation with the six characters and 12 objects used during the pretest and post-test phase, with a written sentence introducing the label for the respective character or object (“This is the…”). The experimenter read the sentences out loud. The naïve and confederate participants were instructed to look at each of the characters and objects carefully and to try to remember the label, but were also assured that they could use other labels during the experiment if they could not remember the one presented during the training phase. They were then informed that the purpose of the experiment was to investigate the descriptions of the pictures, and that they were going to describe the pictures in turn and judge whether the pictures they saw matched the description just given by the other participant or not. They were told to describe their pictures as quickly and accurately as possible, and to make the judgment whether their picture matched the description by saying ja (‘yes’) or nein (‘no’), and to proceed to the following picture after each description or judgment by pressing the space bar. The pictures that the participant saw were presented in a PowerPoint presentation on the laptop screen, while the confederate speaker saw the same pictures with the expected descriptions, as well as the ‘yes’ or ‘no’ judgments, from a PowerPoint presentation on her laptop.

After the pretest phase, there was a new training session for the priming phase, where participants were presented with the 12 characters and the 48 objects used during the priming phase in the same way as described for the pretest phase, before completing the priming session.

They could then look once more at the characters and objects for the post-test phase (that were identical to the pretest), before completing this phase.

Following the priming experiment, all participants completed a background questionnaire asking for their age, their gender, and their language-learning history. Finally, they were paid for their participation and were fully informed about the purpose of the experiment.

5. Results

All target answers produced by the participants were coded as structures containing postnominal genitives (“genitive”), von-phrases (“von”), or other responses (“other”). Other structures often included cases where the possessive relation was expressed as a predicative, as in Der Künstler hat eine rote Rose. (‘the artist has a red rose’).

5.1. Descriptive Results

Table 1. Total numbers of other responses, genitive responses, and von-phrase responses in the three phases of the experiment, as well as percentages of genitives and von-phrases with other responses included or excluded in the total count.

	Total Other	Total Gen.	Total von	% gen. (Others Incl.)	% von (Others Incl.)	% gen. (Others Excl.)	% von (Others Excl.)
Pretest	336	106	38	22.08	7.92	73.61	26.39
Priming (gen. primes)	89	369	22	76.88	4.58	94.37	5.63
Priming (von primes)	93	235	152	48.96	31.67	60.72	39.28
Post-test	127	320	33	66.67	6.88	90.66	9.35

displays the absolute numbers of genitives, von-phrases, and other responses produced in the pretest, in the two conditions of the priming phase, and in the post-test. Note that there is a large number of other responses, in particular in the pretest. While the inferential statistics will be conducted on a dataset from which the other responses were excluded, to give a complete picture, Table 1 displays percentages for the two critical structures, both when the other responses are excluded and when they are not excluded.

These data show that, while both genitives and von-phrases are initially less frequent than other responses, the genitive structure is used relatively more often than the von-phrase structure at pretest (at a ratio of about 74 to 26%). During the priming phase, the production rate of genitives increases after genitive primes (from about 74% to about 93%, when other responses are excluded), while it decreases after von-phrase primes (from about 74% to about 61%, when other responses are excluded). During the post-test, the genitive is the prevalent structure, both when other responses are included in the count (about 67% genitives) and when they are excluded (about 91% genitives). This means that the genitive structure gains in production rates during the experiment, both in comparison to the alternative von-phrase structure and in comparison to other responses, which are produced at marginal rates at post-test. As these numbers give no insights into the temporal development of production rates during the priming phase, Figure 2 plots the percentage of genitives across the trials of the priming phase.1 Other responses have been excluded from this count. The regression line indicates a gradual increase in genitive responses and implies a complementary decrease in the rate of the alternative von-phrase.

5.2. Inferential Analyses

We restricted the inferential analyses to the two alternative structures that we had aimed at eliciting, e.g., genitives and von-phrases; that is, we excluded all structures that were coded as “other” from the data analyses that are reported in the following.

In a first step, we tested whether there was a significant priming effect during the priming phase. We computed a logistic mixed-effects model with “Target” (genitive, coded as “1” vs. von-phrase, coded as “0”) as the dependent variable and “Condition” (genitive vs. von-phrase) as the independent variable. In addition, we included the factor “Order” (referring to the order of experimental trials during the priming phase), which was centered and scaled, as well as its interaction with “Condition”. All analyses were conducted using the Lme4 (Bates et al. 2015) and lmerTest (Kuznetsova et al. 2017) packages in R (R Core Team 2022). “Condition” was sum-coded using the contr.sum function.

Table 2. Output of generalized mixed-effects model for priming phase.

	Estimate	Std. Error	z Value	Pr (>|z|)
(Intercept)	3.10	0.57	5.40	<0.001 ***
Condition	−2.24	0.24	−9.33	<0.001 ***
Order	0.50	0.19	2.66	<0.01 **
Condition: Order	0.36	0.19	1.90	0.06.

Note: Formula: glm(target~condition ∗ order + (1|participant)). Note that the model does not include the random intercept of item, because models with both intercepts or with only the random intercept of item yielded singular fits. ** indicates a p-value of <0.01, *** indicates a p-value of <0.001.

summarizes the model output.

There was a main effect of Condition, indicating that the use of each of the target structures increased in the respective priming conditions. There also was a main effect of Order, which indicated the cumulative priming of one structure (genitives) at the expense of the other (von-phrases). This main effect was qualified by a marginal interaction between Condition and Order, however. As it seems premature to interpret the main effect of Order without exploring how it may be qualified by the interaction with Condition, Figure 3 displays the development of production rates separately for the “genitive” and the “von-phrase” priming conditions. Again, for better readability, the graph only plots the percentage of genitive structures, which is complementary to the percentage of von-structures.

These data show that, in the genitive priming condition, the target structures are produced at extreme ranges, with virtually no changes over time. It seems that genitive responses reach the ceiling from the start, leaving no possibility of a further increase relative to the alternative structure. The picture for the von-phrase priming condition is different. In this condition, genitive target responses start with a relatively lower production rate than the von-phrases and become increasingly frequent over the course of the experiment.

In a second step, we aimed at testing effects over time, by comparing the relative amount of genitive and von-phrase structures at pretest to the relative amounts in the two priming conditions during the priming phase, and to the relative amount at post-test. We combined the datasets from pretest, priming phase, and post-test, and defined a new independent variable “Phase” with the four levels “Pretest”, “Priming phase (genitive primes)”, “Priming phase (von-phrase primes)”, and “Post-test”. “Pretest” was set as the reference level to which the three other levels were compared, applying treatment coding. The dependent variable was the target structure, with von-phrases being coded as 0 and genitives as 1. Importantly, defining each of the two priming levels as a separate level that can be compared to the pretest allows for exploring (inverse) preference effects. The comparison between the “pretest” and the “priming phase (genitive primes)” levels directly compares the change in the relative production rate of genitives and von-phrases from the pretest to this priming condition, and hence, assesses the effect of genitives as primes. The comparison between the “pretest” and the “priming phase (von-phrase primes)” levels assesses the effect of von-phrases as primes. Given that genitives are more frequent than von-phrases in the pretest, an inverse preference effect would mean that there should be priming by von-phrase primes but no, or lesser, priming by genitive primes. The results are summarized in

Table 3. Output of generalized mixed-effects model for pretest, two priming conditions, and post-test (outcome: genitive vs. von-phrases).

	Estimate	Std. Error	z Value	Pr (>|z|)
(Intercept)	1.26	0.59	2.13	<0.05 *
Phase (pretest vs. priming-gen)	3.95	0.48	8.17	<0.001 ***
Phase (pretest vs. priming-von)	−0.35	0.33	−1.07	0.29
Phase (pretest vs. post-test)	2.51	0.38	6.62	<0.001 ***

Note: Formula: glm(target~phase + (1|participant)). Note that the model does not include the random intercept of item, because a model with both intercepts and a model with the random intercept of item only yielded singular fits. * indicates a p-value of <0.05, *** indicates a p-value of <0.001.

.

Importantly, the increase in the rate of genitive production relative to the pretest that is caused by the genitive prime condition is statistically significant, while the increase in the rate of von-phrases caused by the von-phrase priming condition (reflected in the negative estimate) is numerically weaker and not significant. This is not compatible with an inverse preference effect, which would have predicted stronger priming of the less frequent structure. In addition, the advantage of the more frequent structure, the genitive, persists in the post-test, as reflected in the significant difference in the rate of genitives between pretest and post-test.

6. Discussion

To sum up our results, our pretest data show that, as expected, postnominal genitives were more frequent than von-phrases, which is in line with the corpus data (Kopf 2021; Kopf and Weber 2022; Smith 2003). Crucially, the incidence of genitive responses was boosted by the priming manipulation. Genitives were the prevalent structure in the priming phase, and their high production rate was maintained during the post-test. We interpret the difference between the rates at the pretest and rates during priming as evidence of a preference effect. This observation complements the previous findings of inverse preference effects (e.g., Jaeger and Snider 2013; Kaschak 2007; Scheepers 2003), suggesting that both patterns can occur and that further factors will need to be taken into account, to understand whether the lesser or the more frequent structure will lead to relatively stronger priming, as discussed in more detail at the end of the Discussion session.

Second, a closer inspection of responses in the priming phase reveals how production rates changed over time. An effect of the priming manipulation is visible from the very first trial onward. Across the trials, the proportion of genitive responses increased relative to the von-phrase. This pattern of results is indicative of a cumulative priming of the prevalent construction. To better understand this effect, it is informative to look at it separately by condition. For genitive primes, genitive responses were immediately boosted to ceiling, with no changes over the trials and therefore no chance of cumulative priming. For von-primes, the picture is more complex. In the first trials, von-responses were descriptively more frequent than the generally preferred genitive structures, but the rate of von-responses soon decreased below that of genitive responses. Since the primes and accumulating responses differed in structure, this effect cannot be attributed to priming by the immediately preceding input. It seems plausible to assume that cumulative priming of the genitive construction was involved in the dynamics of the von-prime trials, most probably by participants priming themselves by the repeated production of the prevalent genitive structure throughout the priming phase (cf. Jacobs et al. 2019). Since the structures are complementary, it is not easy to discern potentially conflicting effects, and we cannot say whether we found evidence for error-based learning of the less preferred structure or not. But we may assume that we found cumulative priming of the preferred structure. A likely explanation of the cumulative priming of the initially preferred structure seems to be that the use of a dominant structure reinforces itself, because, presumably, participants are not only primed by the utterances produced by the confederate but also by the utterances that they themselves produce. This explains why the dominance of the preferred structure increases throughout the priming phase and does not fall back to pretest levels at post-test. It leaves open the question why other studies have found inverse preference effects and have often not found cumulative priming in favor of the more frequent structure, to which we turn in the following.

The mechanism underlying inverse preference effects is assumed to operate on comprehended input in general (cf. Jaeger and Snider 2013) and, more specifically, on primes presented in an experiment. It may be envisaged as has been modeled in error-based learning approaches (cf. Chang et al. 2006; Dell and Chang 2014) or in expectation adaptation models based on surprisal (cf. Jaeger 2010; Jaeger and Snider 2013). In the following, we stick to a simplified version of the error-based learning approach. During comprehension, the mechanisms recur to transition probabilities between words that are based on previous input. These are accessed by the production system to generate predictions on the unfolding structure. A mismatch between word-wise predictions and the structure at hand causes a prediction error that is fed back to tune structural representations. This purely input-based mechanism accountss for implicit learning and for inverse preference effects in structural priming. The initial increase in von-phrases in the von-priming condition might be an incidence of error-based learning, but if there is such an effect, it is clearly weaker than the priming effect boosting genitives. Moreover, genitives continue to be boosted throughout the priming phase, with a complementary decrease in von-phrases. This cannot be explained by the error-based mechanism, because within the error-based model, there is no additional mechanism that might account for cumulative preference effects. The short-lived lexical boost in structural priming is ascribed to the explicit memorization of word forms (Chang et al. 2006), irrespective of their phrasal status (heads or no heads). As the lexical boost for attributed noun phrases has been shown to hinge on the repetition of nouns (Cleland and Pickering 2003) and nouns were not repeated between primes and targets in our experiment, an explicit memory account of the attested cumulative preference effects is not plausible either, which is why we turn to hybrid models.

Hybrid models, such as the one proposed by Reitter et al. (2011), propose a second mechanism, next to underlying error-based learning, to account for short-lived priming effects. This mechanism tends to be envisaged in terms of the residual activation model proposed by Pickering and Branigan (1998) (cf. Jaeger and Snider 2013; Reitter et al. 2011). Every time a structure is processed, its activation and therefore its probability of being produced again increases. This mechanism accounts for the lexical boost in structural priming, but it is also capable of accounting for effects of abstract structural priming. Contrary to the error-based mechanism, it is not only sensitive to priming by the input but also to priming by previously produced output, thus bringing conservativity (as opposed to learning) into the system. The mechanism captures alignment with the interlocutor (Branigan et al. 2000) and self-priming (Jacobs et al. 2019). We propose that assuming such an activation-based mechanism has the potential to account for preference effects, as well as for the cumulative priming of an initially preferred structure. If structures are produced repeatedly, the respective representations accumulate activation and their opportunity to decay decreases (cf. Reitter et al. 2011). Importantly, according to this model, the representations linked to the less frequent structure of course receive activation as well, every time it is produced, but as the more frequent structure is produced more frequently, effects for this structure should reinforce themselves to a stronger degree than for the less frequent structure. As for the temporal persistence of preference effects till a post-test, whether these are to be expected is, to our knowledge, not explicitly addressed by the existing hybrid models. We found an effect that survived till a post-test that immediately followed the priming phase. Recent evidence suggests that structural priming may affect preferences over longer time intervals such as one week (Kaschak et al. 2011), and that inverse preference effects persist even one month after the experimental manipulation (Heyselaar and Segaert 2022). Future research will show whether preference effects survive longer time periods without priming or whether they are dependent on conditions where they can continuously renew themselves.

The assumption of two competing mechanisms, error-based learning in addition to an activation-based mechanism, calls for an account of their interaction. The relative weight of the two mechanisms must be different in different situations, as evidenced by conflicting findings of both preference and inverse preference effects in the literature. Which mechanism is prevalent in any specific situation may depend on several factors. Given that, compared to the biases looked at in most previous studies, the bias for genitives we focused on was a rather strong one, we suspect that the relative frequency of the structures to be produced should be considered as one important issue. Both accumulating activation and prediction errors are assumed to be sensitive to the initial frequency distribution, but the way in which they change depending on frequencies need not be similar, and their respective changes could follow different, potentially non-linear, trajectories. This makes it hard to predict a tipping point at which either the accumulation of activation—which works in favor of the more frequent structure—or prediction error—which works in favor of the less frequent structure—outweighs the other process. More evidence is needed concerning distributional effects, therefore it will be important to vary the constructions presented in future priming experiments, both within and across languages.

As a further, more speculative avenue of future research, it may also seem worthwhile to explore why structures differ in frequency in the first place. If a structure is much more frequent than an alternative structure, this may indicate that the two structures are in fact no true alternatives. The von-phrase is a marginal structure for the specific type of attributive noun phrase that we looked at (appellative noun phrases containing a case-marked article). Speakers may perceive the less frequent structure as not appropriate in this morphosyntactic context, and this may lower production rates. To take this one step further, it is also possible that whether a structure is perceived as a viable alternative depends on the social context. One reason why the von-structure was rare in our experimental setting could be that it is often assumed to be more frequently used in informal settings. It is possible that this may have contributed to the gradual decrease in its use during the experiment, as participants may have increasingly adapted to the formal setting of an experiment2. These ideas suggest a more complex interplay between knowledge about possible constraints, built up by long-term experience, and short-term experiences in an experiment, than one may expect when taking models of syntactic priming at face value. In particular, it is possible that explicit processes related to language attitudes (such as which type of structure is expected in which setting) interact with underlying implicit processes, be they based on activation or prediction error. This possibility could be further explored in future studies that could attempt to systematically study effects of register.

7. Prospects

While we have observed some important patterns, we do not think that our evidence allows for firm conclusions about different models of priming, but they help us to ask precise questions in order to gather more conclusive evidence. In particular, the conditions under which preference and inverse preference effects arise need further investigation: Is it possible to identify frequency ranges that constrain the development of cumulative priming? To which extent is lexical information involved in the rise in preference and inverse preference effects? Are the units involved in preference and inverse preference effects identical or do they differ? Might differences in cumulative structural priming between younger and older speakers (Heyselaar and Segaert 2022) be accounted for by an input-based mechanism ceasing to interfere with the residual activation mechanism across the lifespan (cf. Chang et al. 2006)?

To understand language learning, it would be beneficial to look more closely at the relations between the influence of prior experience (e.g., what is the baseline frequency of a given structure) and the experience within an experiment (cf. Jaeger and Snider 2013). With regard to possessive constructions, one interesting avenue for future research would be to look at whether priming effects for the structures tested here play out differently for speakers who may have less knowledge of the subtle factors governing the choice between the genitive and the von-phrase, such as children who are in the process of acquiring German and might have had less exposure to complex noun phrases such as the ones we looked at here, as well as second-language learners of German.

More generally, the mechanisms underlying inverse preference and preference effects have the potential to account for individual differences in learning and fossilization during second-language acquisition. The combination of a priming phase with a pretest and a post-test phase, used in cumulative priming studies, corresponds to a standard design in intervention research (cf. Ionin and Montrul 2023), paving the way for a more extensive use of the paradigm in second-language research. Taking advantage of frequency ranges that constrain preference and inverse preference effects, intervention studies thus might help to identify an optimally skewed input for implicit learning (cf. Goldberg and Casenhiser 2008; McDonough and Trofimovich 2013).