Language Analytic Ability, Print Exposure, Memory and Comprehension of Complex Syntax by Adult Native Speakers

1.1 Implicit and Explicit Memory

Learning a language involves storing vast amounts of information: words and their grammatical properties, idioms and collocations, and a large number of grammatical constructions (some lexically specific and some more abstract). One might therefore expect individual differences in memory to be relevant for acquisition, and indeed, there is a considerable amount of research investigating the relationship between memory and linguistic attainment (see below). A large part of this research has focused on the distinction between declarative memory, or memory for facts and events (“knowledge that”), on the one hand, and procedural memory, which contains information about how to carry out sequences of operations (“knowledge how”). The declarative memory system is supported by the hippocampus and the medial temporal lobe. Since the information stored in this system is normally accessible to consciousness, the latter is often referred to as “explicit memory”. The contents of procedural memory, in contrast, are not consciously accessible but are expressed through better performance in skilled tasks (e.g. faster reaction times); this system is therefore a type of implicit memory. Neuronanatomically, procedural memory depends on the basal ganglia and the frontal premotor cortex.¹

The prevailing view is that procedural/implicit memory supports the acquisition of grammar, while declarative/explicit memory supports lexical learning (see Hamrick et al., 2018; Ullman, 2015; 2016). However, while the latter part of this claim is uncontroversial, the relationship between implicit memory and grammar has not gone unchallenged (cf. West et al., 2017). Furthermore, although a number of studies have found a relationship between implicit learning and grammar, a recent meta-analysis (Oliveira et al., 2022) found that the average correlation across a large number of studies was close to zero (r = .04) and not significant.

Furthermore, there is some evidence to suggest that explicit memory may also be relevant for first language (L1) grammar. For example, West et al. (2017) found a significant relationship between grammatical comprehension (assessed using TROG-2) and performance on declarative memory tasks in seven- to eight-year-old English-speaking children, with correlation coefficients ranging from .25 to .52. Conti-Ramsden et al. (2015), in contrast, found a positive but non-significant relationship (r = .25, p = .096) between declarative memory and TROG in their sample of slightly older children (ages from 8;6 to 11;5). It should be noted, however, that their sample was considerably smaller (45 as opposed to 101 children for West et al.). Moreover, since the main focus of the study was to compare SLI and typically developing (TD) children, they excluded TD children whose receptive language skills were 1 SD or more below the group mean. While this move was motivated by the design of the study, it obviously resulted in less variation and could be responsible for the non-significant result.

Two other studies (Lum & Kidd, 2012 and Kidd & Kirjavainen, 2011) examined the relationship between explicit/declarative memory and children’s production of past tense verb forms in English and Finnish respectively. Lum and Kidd found no direct relationship between declarative (or procedural) memory and performance on the language task; however, they found that declarative memory predicted vocabulary, which in turn predicted morphological knowledge. Kidd and Kirjavainen found very weak, marginally significant effects of declarative memory for some classes of irregular verbs, but not for regulars, and no significant effects of procedural memory.

Finally, Llompart and Dąbrowska (2020) examined the relationship between memory and L1 grammatical proficiency (assessed using an acceptability judgement task) and memory in adult speakers. Their study revealed a robust relationship between grammar and explicit memory (r = .64); the correlation between grammar and implicit memory, in contrast, was close to zero.

Thus, the results of previous research on the relationship between grammatical abilities on the one hand and the two memory systems are mixed, with some studies reporting a relationship with implicit memory, some with explicit memory, and some failing to find any significant relationships.

1.2 Working Memory

While declarative and procedural memory involve the storage of knowledge for long periods of time, working memory is often seen as the mental “workspace” in which information is temporarily stored and manipulated during processing – similar to the random access memory of a computer. It seems reasonable to assume that working memory (WM) is relevant for language. During language comprehension, for instance, listeners or readers must recognize words, identify the contextually relevant meanings and use grammatical cues to integrate them into a coherent structure; working memory is thought to provide a space in which information is temporarily held while being processed. And indeed, many previous studies have demonstrated a relationship between WM capacity – traditionally assessed using complex span tasks in which participants must store and manipulate information simultaneously (cf. Conway et al., 2005) -- and performance on a variety of language comprehension and production tasks. Daneman & Merikle’s (1996) meta-analysis reports an average correlation of .52 between measures of verbal WM and “specific” language tasks (e.g. assigning pronominal reference, making inferences, detecting ambiguity) and a somewhat weaker correlation (.48) between measures of non-verbal WM (e.g. operation span, see below) and language tasks. These results appear to support the capacity theory of comprehension, which maintains that WM capacity constrains language comprehension: in other words, individuals with smaller capacity are unable to process more complex structures (Just & Carpenter, 1992; see also King & Just, 1991; Miyake et al., 1994; Just, Carpenter & Keller, 1996).

This conclusion, however, has been challenged by Caplan and Waters (1999). The authors distinguish between interpretive and post-interpretive processes, which call on different WM (sub)systems, and therefore different resource pools. According to them, interpretive processes, which include word recognition, lexical access, and syntactic analysis (including thematic role assignment) occur first. These processes are online (word by word, or even sound by sound) and obligatory. Post-interpretive processes occur later and involve accessing the meaning of the sentence in order to perform other cognitive activities (e.g. answering a comprehension question). For Caplan and Waters, what is measured by complex span tasks is only relevant for post-interpretive processes. Interpretive processes, on the other hand, are supported by a specialised WM independent of the general-purpose system.

An alternative explanation for the observed correlations between verbal working memory and language tasks is offered by constraint-based models (see e.g. MacDonald & Christiansen, 2002; MacDonald & Seidenberg, 2006), which maintain that language processing involves the rapid combination of different partially informative information sources. In this approach, linguistic knowledge is stored in a large interactive network in long term memory. This network is “co-opted” by verbal working memory tasks; thus, such tasks do not measure a separate WM capacity but “a person’s skill in encoding and maintaining verbal information” (Schwering & MacDonald, 2020: 11). Constraint-based models, therefore, predict a strong relationship between language tasks and verbal (but not non-verbal) working memory tasks.

1.3 Language Analytic Ability

There is ample evidence that the individual differences in grammatical skills in second language (L2) learners are partly due to differences in language aptitude, and specifically the learners’ grammatical sensitivity, as measured, for example, by the Words in Sentences subtest of the Modern Language Aptitude Test (MLAT), as well as their inductive language learning ability, which is assessed by tests such as the Language Analysis subtest of the Pimsleur Language Aptitude Battery (PLAB). Two recent meta-analyses of this research can be found in Li (2015) and Li (2016).

Grammatical sensitivity and inductive learning ability share a number of similarities. Both are most relevant for grammar learning; and the tests that measure them are strongly metalinguistic and require conscious effort; not surprisingly, performance on such tests correlates strongly with intelligence (Sasaki, 1996; Wesche et al., 1982). Therefore, these two aspects of language aptitude are sometimes grouped under the term language analytic ability (Li, 2016; Skehan, 2002).

It is generally assumed that language analytic abilities are only relevant for late L2 acquisition; L1 and early L2 grammatical development, in contrast, are supposed to depend (almost) entirely on implicit processes. If correct, this would imply a “fundamental difference” (Bley-Vroman, 1989; 1990) between L1 and L2 learning. However, several recent studies (e.g. Dąbrowska, 2018; Llompart & Dąbrowska, 2023; Prela et al., 2022) have found that language analytic ability is also a reliable predictor of L1 grammatical proficiency, thus challenging the fundamental difference hypothesis.

1.4 Print Exposure

There is considerable evidence that individual differences in grammatical knowledge and processing in adult native speakers are influenced by the amount of exposure to printed text. People who read more produce more tokens of some complex syntactic structures – such as passive relatives – which are more frequent in written texts (Montag & MacDonald, 2015). They also read faster and show different patterns of eye-fixations and word-by-word reading times, including larger wrap-up effects (Ng et al., 2020; Payne et al., 2020), engage more in predictive processing (Favier et al., 2021; Ng et al., 2018), perform slightly better on grammaticality judgement tasks (Favier and Huettig 2021) and are more likely to rely on syntactic rather than semantic cues when interpreting ambiguous pronouns (Langlois & Arnold, 2020).

More relevant for the purposes of the current study is research examining the relationship between print exposure and grammatical comprehension. Here the results are somewhat inconsistent. Dąbrowska (2019) and Street & Dąbrowska (2010) report a significant positive correlation between these variables; Acheson et al. (2008) found a small positive correlation which, however, was not significant; and Misyak and Christiansen (2012) found a significant correlation between ART and one of the three sentence sets they used; however, the relationship was not significant in the regression analysis. Finally, a training study by Wells et al. (2009) found that additional experience with written object relatives resulted in shorter reading times but had no effect on accuracy.

Stronger effects of literacy were observed in two studies which compared the grammatical abilities of adult native speakers of Spanish who are just beginning to learn to read with those of late literates (who learned to read as adults) and early literates (who learned to read as children and thus have a lifetime of reading experience behind them). The first study (Dąbrowska, Pascual & Gómez-Estern, 2022) examined the ability to comprehend a relatively difficult syntactic structure, namely object relatives, and the second (Dąbrowska, Pascual, Gómez-Estern & Llompart, 2023) the ability to provide past tense forms of novel verbs. Both studies observed very large group differences favouring the more literate participants.

The effects of literacy on grammatical comprehension can be explained in several ways. First, written texts tend to be syntactically more complex than spoken texts (Cameron-Faulkner & Noble, 2013; Karlsson, 2009; O’Donnell, 1974 – but see also Beaman, 1984). Thus, individuals who read more have more experience with complex structures, and therefore more opportunities to practise them, which could result in stronger representations and hence better performance. Secondly, when reading or writing, language users can process sentences at their own pace and backtrack or edit as necessary, which is not possible during speaking or listening. Thus, written representations can act as a kind of “processing crutch” which enables language users to understand and produce structures that they might not be able to process in the spoken modality. Once these structures have been practised in writing, their representations might become strong enough to be accessed under time pressure. In this way, reading and writing may act as “training wheels” for processing spoken language (cf. Dąbrowska, 2021). Finally, exposure to written representations could also facilitate the development of grammatical skills in more indirect ways, for example by promoting the development of metalinguistic skills and/or verbal working memory.

1.5 This Study

The study we present here investigates the extent to which individual differences in the ability to use grammatical information in comprehension can be explained by the five factors discussed above (implicit and explicit memory, working memory, language analytic ability and print exposure). To ensure that any observed effects are relatively specific and not simply due to familiarity with the test-taking situation or engagement with the experimental tasks, we also include two control measures which could be expected to correlate with general test taking abilities and engagement, but not with grammatical understanding. These include mental calculation (a secondary measure used in our working memory task) and a measure of ability to stay focused on a task (sustained attention).

2.1 Participants

We recruited 79 UK participants via Prolific (https://app.prolific.co/). The participants (28 males and 51 females) were monolingual native speakers of English living in the UK. The mean age was 45.1 (range 23–73) and the participants had spent had spent on average 15.5 years in full-time education (range 10–24). Forty-three either had or were studying for a university degree (28 B.A., 12 M.A. and three Ph.D); of the remaining 36 participants, five had a Higher National Diploma, six had A levels or equivalent and five had a BTEC or equivalent; thus, the sample is relatively highly educated. Thirteen participants were unemployed, 14 had relatively low-skill jobs (e.g. healthcare worker, retail assistant, cleaning technician), and 41 held professional positions (e.g. radiographer, civil engineer, accountant). The remaining participants were retired (eight participants) or students (three participants).

All participants received 15 pounds of financial compensation for taking part in the experiment, regardless of the time they took to complete the study. On average, the study took 80 minutes to complete.

2.2 Tasks

Implicit memory for sequences: This task was based on the Visual Statistical Learning (VSL) task developed by Siegelman, Bogaerts and Frost (2017). The task tested participants’ ability to identify triplets (i.e. sequences of three complex shapes that always follow each other in the same order), without having been told that the shapes they observed were organised into triplets. The task consisted of a familiarisation phase and a test phase. In the familiarisation phase, participants were presented with a series of complex black shapes. Each shape appeared in the centre of the screen for 1000 ms. The shapes were distributed across 12 blocks, each block containing the same eight triplets, repeated twice, in a pseudo-random order. In the test phase, knowledge of the triplets was tested by means of a forced-choice task in which participants had to discriminate between familiar and foil sequences. The score reflects the proportion of correct answers in the test phase.

Explicit memory for sequences: Like the preceding task, this task involved a familiarisation phase in which participants were presented with triplets of (new) shapes and a test phase assessing learning. However, the instructions and the mode of presentation were different, in order to encourage explicit learning. Participants were explicitly told that the shapes came in triplets and asked to try to memorise the sequences within triplets; furthermore, a fixation cross was presented at the beginning of each triplet, thus marking triplet boundaries. There were eight triplets, each presented seven times. The score was the proportion of correct responses in the test phase.

Sustained attention: Sustained attention was assessed using a modified version of the Continuous Performance Task-X (CPT-X) originally developed by Rosvold et al. (1956). Participants saw letters appear one by one in the centre of the screen in rapid succession; their task was to press the space key when the letter was an X. Individual performance was assessed by means of d’ scores (Huang & Ferreira, 2020), which take into account hit rates (hit = when participants correctly identified the target letter) and false alarm rates (false alarm = when participants responded with a button press to a non-target letter).

Language analytic ability: Language analytic ability was measured using the Sentence Pairs task (Llompart & Dąbrowska, 2023) which is an adaptation of the Words in Sentences subtest from MLAT (Carroll & Sapon, 1959). This is a forced choice task in which participants were asked to identify words in two sentences that had the same function. In each trial, participants were presented with two sentences, as in (1):

• (1) a. Bethany SANG a nice song in front of her friends.

•       b. Every Saturday, John reads the newspaper at home while having a coffee.

Participants were then asked to choose which of the words in bold in (1b) was most similar to the word in capitals in (1a), the correct answer being “reads”. The test consists of 32 items; the score is the proportion of correct responses.

Working memory: Most studies of language processing use the reading or listening span task as a measure of working memory capacity. These tasks involve reading or listening to a series of sentences and answering comprehension questions about them while at the same time trying to remember the last word in each sentence for later recall. Thus, these tasks involve both a processing component and a memory component, both of which must be executed simultaneously.

Note that reading and listening span tasks require the participant to process sentences for comprehension. Since our dependent variable is also a measure of comprehension, using a verbal span task is potentially circular in that establishing a correlation between the span task and a comprehension measure would simply show that comprehension correlates with comprehension. In order to avoid this circularity, we used an operation span task (Ospan), in which participants were asked to solve simple mathematical problems while retaining a sequence of letters in memory.

Our task was based on von der Malsburg’s (2015) implementation of the Operation span task described in Conway et al. (2005). The task began with two training phases. In the first of these, participants were presented with equations (e.g. “( 6 + 5 ) × 4 = 44”) and had to press the c key if the equation was correct and the i key if the equation was incorrect. After they responded, participants were given feedback (a green tick if the response was correct and a red cross if it was incorrect). There were four practice items in this phase. In the second training phase, participants saw an equation displayed on the screen, had to press the c or i key, received feedback, and then a letter was displayed on the screen for 1000ms. After a sequence of three equations and three letters, the participants were asked to recall the letters in the correct order. This series of three operations followed by a recall probe was repeated four times.

The test phase was similar to the second training phase, except that participants no longer received feedback on the equations: they simply saw an equation on the screen and had to indicate whether it was correct. After this, a letter was displayed for 1000ms, followed by the next equation, and so on. At the end of the sequence of equations and letters, the recall screen appeared. The recall screen prompted participants to type the letters in the correct order. The expected number of letters was indicated and participants were asked to type a question mark (?) in the correct position in the sequence if they were unable to recall a particular letter. The recall screen also provided participants with information about the percentage of correct responses on the arithmetic task. Participants were told that they should be at least 85% correct on this task and were asked to monitor their performance. The purpose of this was to ensure that they did not simply focus on the recall task. We excluded the results from participants who were below a threshold of 80% correct.

The sequences of equations and letters differed in length (from three to seven equation+letter combinations). There were three sequences for each length (three sequences of three equations and three letters, three sequences of four equations and four letters, and so on), for a total of 15 sequences. The 15 sequences were presented in random order.

The score for this task is the total number of letters that participants remembered correctly and in the right position over the whole test phase. As argued by Friedman and Miyake (2005), this is a more fine-grained measure of working memory than the traditional measure, i.e. working memory span (the longest sequence recalled correctly on at least two-thirds of all trials). In addition, we also use the proportion of correct responses on the arithmetic problems as a measure of participant engagement with the task.

Print exposure was assessed using a computerised adaptation of the Author Recognition Test (ART; Acheson Wells & MacDonald, 2008). In this task, participants are presented with a list of 130 names. Half of these are names of well-known authors; the other half are foils. Participants are asked to indicate which names belong to authors. In our adaptation of the task, the names were presented one at a time and participants responded by clicking either a “yes” or a “no” button. Since this was an unsupervised online experiment, participants had only five seconds to make their choice -- enough time to make a decision, but too short to look up the answer. We computed d’ scores by participant, taking into account hit rates (hit = when participants answered “yes” for a real author) and false alarm rates (false alarm = when participants answered “yes” to a foil). Computing d’ scores instead of total scores leads to better reliability for the computerized version of the ART (Wright et al., in prep.).

Grammatical comprehension task: Our dependent variable in this experiment was participants’ ability to use syntactic cues to interpret four different types of sentences (see Table 1). Three of these (Complex NPs, X-Is-Difficult-to-Answer constructions, and Reduced Relatives) are relatively complex structures which involve two levels of embedding and long-distance dependencies:

• Sentences with Complex NPs contained a noun phrase of the form [the fact that VERBing NP is ADJ] in the subject position of an embedded clause (e.g. Linda complained that the fact that cycling in the main square is forbidden annoys tourists.).

• X-Is-Difficult-to-Answer constructions (cf. Herbst and Hoffmann, under review) had the form [NP1 will be ADJ to V1 NP2 to V2] (e.g. James will be easy to persuade Walter to help.), where the missing object of the nonfinite verb in the sentence final position (help) is understood to be coreferential with the subject of the main clause (James). Previous research (e.g. Chipere, 2001; 2003; Dąbrowska, 1997) has shown that sentences with Complex NPs and X-Is-Difficult-to-Answer constructions are difficult for many adults, so we expected considerable individual differences in performance on these items.

• The third structure were sentences containing a reduced relative clause inside another reduced relative clause inside the subject NP (e.g. A child [staring at a dog [chasing a postman]] was afraid.). Scholes and Willis (1987) showed that children between 8 and 10 as well as semi-literate adults exhibited very high error rates (from 35% to 60%) when interpreting such sentences. Since structures which are acquired late tend to show much more variation in adults than early-learned structures (Brooks & Sekerina, 2005; Dąbrowska, 2018; Dąbrowska, Pascual & Gómez-Estern, 2022; Indefrey, 2002), we expected this kind of item to pose difficulties for at least some adult speakers.

• The final structure were Ditransitives in which the word immediately after the verb could function either as a determiner or a pronoun (e.g. her) followed by two nouns and a determiner (e.g. Mr Peters showed [her baby] [the pictures] and Mr Peters showed [her] [the baby pictures]). In order to interpret such sentences, the comprehender must be able to correctly identify the constituents and their roles: in the example just given, they must determine whether the indirect object is her or her baby. The only available cue is the position of the determiner the. Thus, unlike the other three sentence types, the Ditransitive items were relatively simple syntactically but required the comprehender to make use of a relatively subtle grammatical cue. Scholes et al. (1976) have shown that children continue to misinterpret such sentences throughout middle and late childhood; the rationale for their inclusion was thus the same as that for the Reduced Relatives items.

There were eight items for each sentence type, giving a total of 32 test items. For the Ditransitives, half of the items had a structure analogous to example (5) in Table 1 and the other half were like (6).

Comprehension was assessed by means of a two-alternative forced choice task. Complex NP and X-Is-Difficult-to-Answer items had three experimental questions each and the Reduced Relatives and Ditransitives had one question. The Ditransitives and Reduced Relatives items also included control questions which probed participants’ ability to identify the subject and object respectively. In both cases, the subject or object was a simple NP which occurred in the canonical position immediately before or after the verb, so we did not expect them to pose any difficulty for adult native speakers. These questions were included to enable us to determine whether participants had engaged with the task.

To avoid fatigue, the experimental items were divided into two blocks, each block containing half of the items for each type of construction. The items within blocks and the questions for each item were presented in a pseudo-random order with the constraint that items belonging to the same condition should not occur immediately next to each other. Since we were interested in whether participants were able to extract the relevant information from the stimuli and not in how good they were in remembering them, the stimuli sentences remained on the screen while participants answered the comprehension questions.

We employed two scoring methods: proportional scores, which were computed by dividing the number of correct answers by the total number of questions for a given item (excluding control questions) and an “all-or-nothing” score where an item was considered as correct (1) if the participant answered all the experimental questions correctly and as incorrect (0) otherwise. The two scoring methods were very highly correlated (r =.96), so “all-or-nothing” scores were used in all subsequent analyses.

2.3 Procedure

Before starting the experiment, participants were invited to fill out a background questionnaire in which they were asked to provide basic demographic information (age, gender, occupation, years spent in full-time education, highest educational attainment and native language). The experiment consisted of seven tasks (one of which was divided into two blocks), with three breaks during which the participants were invited to do some guided gentle gymnastics movements. Figure 1 shows the order in which the tasks were presented to participants.

Figure 1

The study was conducted via the online experimental platform Gorilla (gorilla.sc; Anwyl-Irvine et al. 2020).

3.1 Descriptive Statistics

Table 2 reports the descriptive statistics and reliability estimates for all the measures in our experiment. Split-half reliability estimates were computed averaging the results of 5000 random splits using the splithalf package (version 0.8.2; Parsons, 2021) in R, except for the d’ scores, where the reliability estimate was computed by averaging the results of 1000 random splits using the splithalfr package (version 2.2.0; Pronk 2021). Cronbach’s coefficient alpha was computed using the psych package (version 2.3.9; Revelle 2023). Reliability estimates were acceptable for all tasks. They were also relatively high for individual structures tested in the grammatical comprehension task, with the exception of Ditransitives.

As anticipated, participants were at ceiling on the control questions, indicating that they had understood the task, were cooperative etc. The scores for reduced relative clauses and Ditransitives were also relatively high, but not quite at ceiling. The other two structures (Complex NPs and X-Is-Difficult-to-Answer) were considerably more difficult. The mean overall proportion of correct responses on the grammatical comprehension task was .57, with an interquartile range from .45 to .69, indicating considerable variation in individual performance.

3.2 Correlational Analysis

Table 3 shows the correlations between all measures used in our study. As we can see, grammatical comprehension shows relatively robust correlations with language analytic ability and print exposure, and a much weaker correlation with implicit learning and calculation. It is worth noting that explicit and implicit sequence learning are relatively strongly correlated, which is most likely due to the fact that the two tasks were very similar. Language analytic ability is strongly correlated with print exposure and the calculation score.

The correlation between reaction times and accuracy comprehension task was close to 0 (r = –.07), indicating that there was no speed-accuracy tradeoff.

Table 4 shows the correlations between the different sentence types in the comprehension task. While all the correlations are positive, they range from moderately strong to very weak. It is noteworthy that the correlation between the two most difficult structures, namely X-Is-Difficult-to-Answer constructions and Complex NPs is the highest. The relatively low correlations between the easier structures are most likely due to the fact that performance on these two structures is close to ceiling (see above).

3.3 Regression Analyses

To examine which factors contribute to differences in comprehension, we fitted a mixed-effects model with a logistic linking function using the lmerTest package (version 3.1-3; Kuznetsova et al., 2017) in R. The data and the R code used in the analysis are provided in the supplementary materials. The dependent variable was the all-or-nothing score for each item in the comprehension task.² We included the following predictors: working memory (measured by the score on the Ospan task), print exposure (measured by the d’ score on the ART), implicit memory for sequences, explicit memory for sequences, language analytic ability and sustained attention. We added random intercepts for participants and items.³

The model (summarized in Table 5) shows three significant main effects: language analytic ability, print exposure and implicit learning. These factors have positive coefficients, indicating that higher scores for these abilities are associated with better performance on the grammatical comprehension task. The estimate for language analytic ability is the highest, indicating a somewhat larger effect, while the estimate for implicit memory is the lowest, indicating that this effect is smaller than the other two.

All other predictors (i.e. calculation, explicit memory, working memory and attention) had lower estimates. They were not significant predictors in our model. Even though many predictors correlate with each other, no multicollinearity could be detected.

We also attempted to fit a model with sentence type and its interactions with the other predictors to determine if they affect performance on all constructions in the same way, but it did not converge. We therefore ran four different models, one for each sentence type. Appendix 7.3 provides summaries of these models. Language analytic ability was a significant predictor for all sentence types. Print exposure was a significant predictor for X-Is-difficult-to-Answer and Ditransitives, but only marginally significant for the other sentence types. Implicit memory was not significant in any of these four additional models, which could be due to of a lack of power. Surprisingly, working memory was a significant predictor for Ditransitives. Since this was the syntactically simplest construction in the test, this could be a type 1 error.

4.1 Language Analytic Ability

Our results revealed a robust effect of language analytic ability. This finding is in line with recent studies by Dąbrowska (2018), Prela et al. (2022) and Llompart and Dąbrowska (2023), which also found this aspect of “foreign” language aptitude to be a significant predictor of performance on tasks tapping grammatical knowledge in adult native speakers. It should be noted that these earlier studies used a variety of measures of aptitude and grammatical ability. Thus, Dąbrowska (2018) assessed aptitude using the Language Analysis subtest of the PLAB, while Prela et al. and Llompart and Dąbrowska’s study 2 employed the Sentence Pairs task used here; and Llompart and Dąbrowska’s study 1 employed both measures. Grammatical proficiency was assessed by means of picture selection in Dąbrowska (2018) and Llompart & Dąbrowska’s study 2 and by means of a grammaticality judgement task in their study 2 and in Prela et al. (2022). Dąbrowska (2018) and the two studies in Llompart and Dąbrowska (2023) report remarkably similar correlations between performance on the aptitude test and the grammar task: .46 in Dąbrowska’s study; .48 and .45 for Sentence Pairs and Language Analysis respectively in Llompart and Dąbrowska study 1; and .43 in Llompart & Dąbrowska’s second study. The correlation reported by Prela et al. is somewhat lower (.37), which is most likely due to ceiling effects (the participants were 93% accurate on the native language grammar task). In the present study the correlation was higher (.68); this is most likely because our measure of grammatical proficiency was considerably more difficult than those used in the previous studies, and hence the differences in performance between speakers were much larger. The fact that all these studies obtained very similar results in spite of the fact that they used different measures provides strong convergent evidence that the relationship we report here is real.

It is also worth noting that all the correlations between language analytic ability and grammatical knowledge observed in these five studies are all higher than the correlations typically reported in L2 studies: the average correlation coefficient in Li’s (2015) meta-analysis of L2 research of the effects of language analytic ability on grammatical proficiency is .31 (95% CI = .25–.36). Language aptitude tests such as the MLAT and PLAB were originally developed to predict foreign language achievement in instructional contexts, particularly in the early stages of language learning. However, they have also proven to be good predictors of second language learning in naturalistic settings, even at later stages. In fact, some L2 researchers (e.g. McLaughlin, 1990; DeKeyser, 2000) have proposed that the relationship between aptitude and achievement may be even stronger in naturalistic contexts, because naturalistic learners have to discover the rules of the language for themselves, without the help of a teacher. The results reported here and the earlier studies we reviewed suggest that this may also be true of L1 learners.

4.2 Print Exposure

While there is considerable evidence that print exposure is related to vocabulary size and reading comprehension (for a meta-analysis see Mol & Bus, 2011), relatively few studies have examined the relationship between print exposure and grammatical comprehension in adults, and, as noted in the introduction, they have produced somewhat inconsistent results. The effect observed in the current study (r = .55) is larger than that observed in most previous studies. This could be due to the fact that our grammatical comprehension task was more demanding than those used in previous research.

4.3 Implicit Memory for Sequences

As noted in the introduction, it is widely assumed that the acquisition of native language grammar depends largely on implicit learning abilities, and a number of studies report significant correlations between procedural/implicit learning abilities and grammatical abilities. However, not all studies found such an effect; and in fact, a recent meta-analysis (Oliveira et al., 2022) found no significant relationship between these two variables. It should be noted that Oliveira et al.’s meta-analysis only included studies which used the serial reaction time task as a measure of implicit learning. As the authors note, such tasks are notoriously unreliable, which could explain the lack of a relationship.

The visual statistical learning task used in the current study has much better reliability (.83 in the original study by Siegelman et al., and .75 according to our estimates). With this improved measure, we did find a significant relationship. It should be noted, however, that the observed effect was rather small (r = .26, see Table 3).

4.4 Working Memory

As discussed above, many studies have demonstrated strong links between working memory and language comprehension. However, it is not clear whether this relationship is due to the fact that working memory capacity imposes constraints on individuals’ ability to process language (as argued for example by Just & Carpenter, 1992), whether these effects appear at the post-interpretive stage (a hypothesis put forward by Caplan & Waters, 1999), or whether verbal working memory tasks effectively measure linguistic knowledge (as proposed by Schwering and MacDonald, 2020).

Our grammatical comprehension task differed from the tasks used in most previous studies in that it alleviated the demands on working memory by leaving the sentence on the screen while participants answered the comprehension question and by allowing participants to process the stimulus sentences at their own pace (i.e. without time constraints). Under these conditions, we observed no significant effect of individual differences in operation span (with the possible exception of the Ditransitive condition), despite the fact that the sentences were very difficult, as evidenced by high error rates.

These findings seem to argue against Just and Carpenter’s capacity theory of comprehension and are consistent with Caplan and Waters’ theory. They could also be accommodated in constraint-based theories of comprehension, which predict a strong relationship between comprehension and verbal working memory tasks, but not necessarily between language comprehension and operation span tasks.

4.5 Explicit Memory for Sequences, Sustained Attention and Mental Calculation

None of the other measures (sustained attention, mental calculation or explicit memory for sequences) were significant predictors of grammatical comprehension.

The lack of effect for explicit memory contrasts with the findings of Llompart and Dąbrowska (2020), who observed robust correlations between performance on a grammaticality judgement task and two explicit memory measures (digit span and paired associates) in adult native speakers. This could be due to the fact that the explicit memory task used in this study was entirely non-verbal – although it should be noted that we did find significant effects of implicit memory, which was also assessed using a non-verbal task.

Be that as it may, the lack of any effect of explicit memory for sequences, selective attention or accuracy on the computational part of the WM task indicates that individual differences in scores on the comprehension task cannot be explained by appealing to factors such as engagement with the task, familiarity with testing, or other linguistically irrelevant performance factors: in other words, they reflect genuine differences in language knowledge.

4.6 Directionality of Causation

Our main interest in this study is in the reasons for individual differences in native speakers’ grammatical knowledge. Obviously, since our data is correlational, we cannot make strong inferences about causation. It is, nevertheless, useful to reflect on the plausibility of various causation scenarios involving our variables. Given a correlation between two different abilities (A and B), four scenarios are possible:

1. A causes B;

2. B causes A;

3. reciprocal causation (A affects performance on B which in turn affects performance on A and so on); and

4. the correlation is due to a third variable, C, which causes both A and B.

In this section, we discuss the plausibility of these four scenarios with regard to the three variables which emerged as significant predictors of performance on the grammar task, namely implicit learning, print exposure and language aptitude.

4.6.3 Language analytic ability

A number of studies have shown robust relationships between language aptitude and foreign language achievement (see Li, 2015; 2016), and it is generally assumed that the relationship is causal: that is to say, high-aptitude learners are more successful because the cognitive abilities assessed by aptitude tests play an important role in adult second learning. Given that the relationship between aptitude and achievement appears to be even stronger for the native language (at least with regard to grammatical sensitivity and grammatical comprehension), the question arises whether language aptitude could also play a role in first language acquisition.

It is not difficult to envisage how language aptitude (and grammatical sensitivity in particular) might help with learning grammar. In order to be able to produce and understand novel sentences, speakers need to generalise beyond the sentences they have previously encountered. According to usage-based models, generalisation involves analogical reasoning (Behrens, 2017; Goldwater, 2017; Tomasello, 2003). This, we argue, is true for L1 as well as L2 acquisition. Consider a learner who wants to ask about the location of a pteranodon but does not have a constructional template for location questions or a lexically specific unit which corresponds to their semantic intention. However, our learner has acquired the lexically specific unit Where’s the dog? as a means of asking about the location of a dog and knows the noun pteranodon. To produce the target question, the learner needs to recognize the overall similarity between the meaning of the lexically specific unit and their semantic intention (they are both questions about the location of an object), perform an analogical mapping between the relevant parts (the ?LOCATION substructure in their semantic intention corresponds to ?LOCATION in the semantic representation of the familiar expression; PTERANODON corresponds to DOG) and, finally, match these semantic substructures with the appropriate phonological forms (where corresponds to ?LOCATION and PTERANODON corresponds to pteranodon).⁴ In other words, the learner must in effect solve a proportional analogy problem, as illustrated in Figure 2.

Figure 2

However, generalisation by analogy is not without dangers (e.g. Chomsky, 1966; Searchinger, 1994). To solve the proportional analogy problem correctly, the learner needs to understand which chunk of meaning corresponds to which chunk of form and identify structural parts which play the same role in different utterances – and this is exactly what the Sentence Pairs task measures. One could even speculate that the relationship between language analytic abilities and grammar is even stronger for L1 than for L2 because L1 learners, in contrast to most adults learning a second language, have to discover the rules of grammar for themselves, without the help of a teacher or a textbook.

Thus, a causal link from grammatical sensitivity to grammatical knowledge makes sense theoretically if we assume a usage-based account of acquisition. A causal link in the opposite direction, in contrast, is implausible, given that once grammatical knowledge becomes automatized, it is not accessible to consciousness and hence presumably cannot influence performance on a highly metalinguistic task such as Sentence Pairs (see Llompart & Dąbrowska, 2023 for further discussion).

This leaves the final scenario, namely, that individual differences in performance on grammatical sensitivity as well as grammatical comprehension tasks are caused by a third, unobserved variable. The obvious candidate for this third variable would be general intelligence, which is known to correlate strongly with aptitude (Li, 2016). This is not surprising, as there is overlap in the tasks used in standard IQ and language aptitude tests (for example, both often include vocabulary and memory measures). In fact, some researchers appear to equate language aptitude with (verbal) intelligence: for instance, DeKeyser et al. (2010) assessed aptitude using a test designed to select candidates suitable for university study which contained tasks traditionally included in IQ tests (e.g. definitions, analogies and verbal reasoning).

There are, however, reasons for thinking language aptitude and IQ are distinct, though related, concepts. For second language learners, aptitude appears to be a better predictor of language achievement than intelligence (Gardner & Lambert, 1965; Li, 2019). Furthermore, the findings reported by Dąbrowska (2018) suggest that both variables contribute independently to individual differences in native speakers.

Because of the partial overlap between language aptitude and intelligence, it is difficult to distinguish between the direct causation and third variable scenarios. Perhaps the most plausible conclusion is that language analytic ability is something that emerges when general intelligence is applied to language data, or possibly to written linguistic representations. The latter formulation would explain the links that many researchers have observed between literacy and meta-linguistic awareness (Karanth et al., 1995; Nagy & Anderson, 1995; Nunes et al., 2006).