When good-enough production is not good-enough, given at the Deutsche Gesellschaft für Kognitive Linguistik September 2024 in Osnabrück

+++++++++

Here is the PPT and the transcript of my talk at Warwick in March, 2024.

Literacy Facilitates Ultimate Native Language Attainment: More Evidence from Turkish

Abstract

Since the written form of language must be taught explicitly and is strongly influenced by prescriptivist notions, linguists often regard writing as an uninteresting add-on to spoken language. However, literacy, i.e., the availability of the written form, influences linguistic knowledge and its representation in the mind in profound ways. Literacy has been shown to affect phonological and semantic representations. However, there is surprisingly little research on how acquiring literacy influences representations of grammatical constructions. In this talk, I provide more suggestive evidence that the availability of literacy is quite important for ultimate native language attainment, especially for comprehension and production tasks that aim to elicit written language-biased constructions. Ongoing research shows that when illiterate speakers are compared against literate speakers, illiterate speakers extract fewer across-the-board generalizations — as spoken language contains fewer types-tokens of complex structures in comparison to written language, and show more individual differences in grammatical knowledge. Research on this topic has important implications for two reasons: first, it tells linguists a cautionary tale. Previously, formalist linguists asserted that adult L1 speakers converge on the same grammatical knowledge by age 3, 4 or 5. However, evidence shows that L1 speakers —both literate and illiterate alike— actually show many individual differences in their linguistic knowledge. The lack of written language results in even more individual differences, rendering the convergence hypothesis an unsound argument. Second, data from illiterate speakers provide a non-WEIRD perspective into language acquisition and help us make our language acquisition theories more inclusive. As such, more research investigating the linguistic abilities of illiterate speakers will provide more evidence to avoid arriving at incorrect overgeneralizations. 

Let’s face it. As cognitive scientists, we have paid way too much attention to data from WEIRD populations and made overgeneralizations in cognitive sciences based on such data. As Dahl (2015) explains even after controlling for Eurocentric biases in a large dataset like the WALS, languages that have been used as a source of typological comparisons are literate, with lots of speakers and official. A recent study by Blasi and colleagues also shows that linguistics has been founded on false overgeneralizations with data from only a small group of niche native speakers. From a linguistic perspective, this is not only a problem for our L1 acquisition theories, but it has profound implications for L2 acquisition studies, how we define the term native speaker; and how we think about language education policies and multilingualism. Thus, we really need to be careful with our generalizations. 

Before I start my talk, here are the ultimate take-home messages I will be focusing on in my talk: 
A) we shouldn’t over rely on conventional pieces of wisdom in our theories 

B) not all native speakers converge on the same grammatical knowledge uniformly and successfully

C) literacy facilitates ultimate native language learning

I would like to begin with some conventional knowledge in linguistics. First off, linguists regarded writing as an accessory for some time although the thing we study as language — that is speech— is almost always influenced by written language. Second, for a number of years, the generativist school (as well as others) has supported the that L1 speakers converge on the same grammar uniformly and successfully without showing much evidence for it. Here are a few quotes to begin with. As you can see, the estimated age for uniformly successful mastery of L1 goes as low as the age of 3. Here are some more quotes from other linguists that have made such assumptions. Surprisingly, there has been very little experimental evidence to support this claim. Finally, learning L1 grammar is assumed to rely on implicit learning mechanisms, which are thought to be separate from IQ scores or other variables such as linguistic aptitude or learning styles (Feldman et al., 1995; Reber et al., 1991). 

Interestingly, the few data they use to support this idea are very inconsistent and many times the experiments used are questionable with many confounding factors. Again, one of the main issues is WEIRD: such data both mostly come from highly educated literate speakers, and if you are lucky from low educated speakers if at all. Most of these highly educated speakers have at least 14-15 years of schooling, if not more. This is partially where we run into the problem of WEIRD. These highly educated literate speakers tend to be undergraduates, graduates, or university lecturers; which arguably constitute a very small percentage of the general population. But, their intuitions of language are vastly different from speakers who have less formal education for various reasons. 

Clearly, this was an apparent confounding variable, thus since the 1970s this conventional wisdom has been called into question by various linguists. That is, do all native speakers really converge on the same grammar successfully? Many studies show that this is not the case. If anything, research has shown that individual differences in linguistic knowledge persist in intriguing ways. One of the most influential researchers working on this is Ewa Dabrowska, a very dear colleague and an avid cat lover. Her work over the years has arguably revolutionized cognitive linguistics. While I cannot do justice to all of the studies on this topic, I will quickly summarize and provide a brief overview of individual differences that are borne out of literacy-related differences.

First of all, individual differences (IDs) in language learning are differences in how speakers diverge quantitatively or qualitatively from one another in learning various levels of linguistic units (i.e., phonetics, grammar, lexis, pragmatics). In a smart study, Ewa Dabrowska (2018) measured individual differences among L1 English speakers and demonstrated two very important things about grammatical knowledge in L1 speakers. First, print exposure seems to modulate comprehending complex syntax. Second, when linguistic experience is not enough to comprehend these structures nonverbal IQ acts as a compensatory mechanism to support comprehension. I would like to start with the first point and explain why print exposure should matter for complex syntax. Written language contains more types and tokens than spoken language. So written language provides both more quantity and quality in comparison to spoken language. As usage-based models would also predict, variation in relevant ambient linguistic input should result in variation in linguistic knowledge. Quite simply, a plethora of research studies show that if you read more, you simply attain a more schematic representation of a linguistic structure. Second, nonverbal IQ is practically problem-solving skills: if you don’t have experience with a construction, you are more likely to use problem-solving skills to establish relationships between constituents, suffixes and the general structure. 

Let’s look at the grammatical knowledge measure: the picture selection task. The experimenter reads the stimuli and asks the participant to match the sentence they heard with either of the pictures. The participant could have the sentence repeated many times. So it is not exactly a cognitively demanding task. This has been used with speakers as young as 2 year olds. In this study, she tested subject/object relatives, quantifiers, passives, and clefts. 

Now let’s take a look at the descriptives and regression results from Dabrowska’s study. Overall accuracy is 79%. This is quite odd: if all speakers converged on the same grammar successfully, why do we have individual differences in such a simple task? As you can see, grammar is explained by both nonverbal IQ and reading as measured by the ART among other variables such as education and language analysis skills. Interestingly, IQ explained roughly 12% of grammar, whereas reading did about 3% of variance in grammar. So clearly, IQ must be important in mapping words to pictures.

Now you might wonder why print exposure modulates comprehending complex syntax. Recall that written language contains more complex language than spoken language. And that is quite simply because speakers who read more move away from lexically-specific representations of language to a more abstract representation; they move towards what some foreign language teachers and linguists might call “rules” (you can see the example on the screen). In the same vein, Ewa Dabrowska and James Street demonstrate this in a study in 2014 conducted with English native speakers. Their study shows that speakers that read more could comprehend faster and more accurately not only lexically-specific passive constructions (stimuli that use verbs prototypically occurring in the English passive) but also others: stimuli that use non-prototypical verbs in the English passive. Thus, we claim that speakers with more reading experience extract a more fine-tuned representation of language than speakers who do not read as much or at all. 

OK, we established that reading and cognition matter for grammatical knowledge. But if reading influences linguistic knowledge specifically grammar in profound ways, how is grammar affected by the absence of literacy? After all, even within literate languages, not every speaker is literate. Surprisingly there is very little research investigating this. The existing literature does show that the acquisition of a writing system seems to facilitate the acquisition of complex grammar, that is, structures that are less frequent in spoken language and structures that require more semantic mapping such as passives, object relatives, or subordination. Furthermore, education improves IQ about 3 to 5 points on average each year, thus we see stark differences in IQ between literate and illiterate speakers.

Ewa Dabrowska calls the facilitatory effects of writing the training wheels hypothesis. And this is for three reasons. 1) written language provides more types and tokens (that is quality and quantity) than spoken language, fostering the extraction of more across-the-board-generalizations (or highly schematic constructions); 2) writing acts as memory crutches for long and complex sentences, such that speakers can read them at their own pace and re-read them if necessary for a deeper analysis; and finally 3) writing improves metalinguistic awareness, that is, the extent to which speakers can analyze the internal structure of linguistic constructions. 

I would like to turn to 2 pioneering studies that have investigated the effect of acquiring literacy on complex grammar comprehension. These studies are also the reason why I wanted to pursue my research.  Both studies were conducted by Ewa Dabrowska and her colleagues in 2022 and 2023 respectively. Both studies were conducted with Spanish L1 illiterate and literate speakers with 3 groups: semi-literate; late-literate; high-literate. Semi-literates were speakers who could only do very basic reading. Late-literate were speakers with a little bit more reading experience, who could read longer texts with more independence; and finally high-literates were speakers who went to school at an early stage in life and had attained at least a BA degree. 

In the 2022 study, Ewa and her colleagues used the picture selection task and tested the comprehension of Spanish subject and object relatives. As mentioned earlier, object relatives are structures that tend to appear more frequently in written language — although there are exceptions to this in different languages. Almost all speakers performed at ceiling on the subject relatives, a structure that is relatively frequent in spoken language. However, as you can see, accuracy on object relatives was in descending order: high literates performed almost at ceiling; then late literates who outperformed semi literate speakers. Group significantly predicted performance. Interestingly, nonverbal IQ was also an important predictor of performance. 

In the 2023 study, Ewa and her colleagues used a nonce-verb inflection task in Spanish to examine morphosyntactic productivity. As some of you might know, Spanish verbs come in different flavors, such as -AR/-IR/ and -ER forms. And these forms have different conjugation patterns. They observed vast individual differences in productivity with inflectional endings. High-literates consistently provided the appropriate form more often than late-literates, who in turn were better than semi-literate participants. Crucially, group interacted with person, number, and conjugation, such that the between-group differences were larger for the less frequent cells in the paradigm, indicating that literacy-related differences are not merely a consequence of the high-literacy group being more engaged or test-wise. This suggests that the availability of written representations may facilitate the acquisition of certain aspects of grammar. These results add to the growing body of research which challenges the assumption that all native speakers converge on the same grammar early in development.  

This now connects me to my own studies where I expand on the relationship between literacy and linguistic knowledge from both a comprehension and a production perspective. I used the picture selection task. For production, I used the picture elicitation method to elicit subject, and object relatives and passives in Turkish. In the picture selection task, I was interested in subject, object relatives, passives, quantifiers, and nominalizations. Previous research shows that speakers tend to struggle with these constructions, either due to structural complexities or due to asymmetries in frequency across modalities. All of these constructions show a frequency asymmetry in Turkish across modalities. They are more frequent in the written modality ranging anywhere from 2.5 to 3 times. I recruited 30 illiterate and 27 literate speakers. Unfortunately, I did not have the gradient grouping of literacy levels Ewa and her colleagues did in their studies. However, I developed a 1-minute word reading task: where speakers have to read as many real words as they can in 1 minute. This they repeated this with nonce words, timed under 1 minute. The composite score of the two would be our reading measure. This would act as a proxy for reading abilities. 

In our regression models, we preferred to use group over reading for three reasons although reading is a continuous variable and should provide better results: 1) reading and group are strongly correlated at .91 (which results in multicollinearity), 2) the reading measure measures two different things in these groups: in illiterates potentially the speed at which orthographic encoding/decoding happens and reading fluency in literates, and 3) determining real effects of only writing requires speakers to have been exposed to written language for many years, whereas our illiterate participants had been learning literacy for roughly 7 months on average. However, when we swap group with reading, we get highly comparable results. Therefore, we are confident that group acts as a proxy measure for literacy.

As you can see, illiterate speakers have lower performance on the reading task, and lower nonverbal IQ than the literate group. This is quite expected as most of the illiterate participants were not allowed to go to school when they were younger— due to patriarchal reasons. It shouldn’t be surprising that all of our illiterate speakers were also female and quite young: 45 years old on average. This is a commonly observed pattern in Turkey and around the world. Unfortunately, illiteracy is a gendered issue. 

Here are our results for the picture selection task, overall and per construction. As you can see that literate speakers outperform illiterates. As can be seen in this figure, literate speakers perform almost always at ceiling, while illiterate speakers show vast individual differences. 

But is there anything that predicts performance on the picture selection task? Yes. Group and nonverbal IQ significantly predict performance on the picture selection task and account for about 33% of the variance each. Clearly, being literate as well as having above average IQ helps with grammar comprehension in your L1. 

Our production data is a bit more complicated. You can also see the overall performance on producing the three constructions. There is a decreasing performance in both groups, with SRC > ORC > passives. However, illiterate speakers seem to show much less variation: illiterates showed the most individual differences in the subject relative production condition, whereas literate speakers were practically at ceiling. In the object relative condition, individual differences in illiterate speakers decrease and they perform close to floor, whereas literate speakers now show more individual differences. And in the passive condition, illiterate speakers perform virtually at floor, whereas literate speakers are all over the place. Our regression analyses revealed an interaction between group and nonverbal IQ. As you can see in this figure here, it is not difficult to see why there is an interaction. For illiterate speakers, no predictors were statistically significant. Interestingly, nonverbal IQ was the only significant predictor for literate speakers, explaining a whopping 23% of the variance in production. 

Given that our results cannot be explained with non-linguistic factors (such as not being engaged or cognitive overload), one question arises: Do these individual differences reflect differences in underlying linguistic knowledge? Providing an unequivocal answer requires additional assumptions about the nature of language learning as well as the cognitive machinery responsible for it. For instance, for usage-based approaches the answer must be positive, since such approaches readily hypothesize that differences in cognitive machinery and linguistic experience will result individual differences in linguistic competence. For generativist approaches, giving an unequivocal answer is more difficult because of all the extra assumptions needed in how we situate experience in the universal grammar hypothesis and the language acquisition device.

We now turn to a discussion of some plausible accounts to explain the group differences in the comprehension and production of complex syntax. 

Frequency might be the first account. Many complex structures occur more frequently in written language than in spoken language. From a usage-based perspective, we would expect speakers with more exposure to written texts to have stronger representations of complex structures. However mere frequency cannot account for all these differences because for instance object relatives occur more frequently in spoken language, at least in English (Roland et al. 2007), Spanish (Llompart & Dabrowska 2021), and in child-directed Turkish. Again, the frequencies of the structures in my study show about 2-3 times difference in frequency between modalities. So we can’t exactly claim that frequency is the mere reason behind group differences. 

However, if we expand the account from quantity to quality, a frequency approach might be more plausible. Spoken language involves fewer type-token counts of various constructions, and there is evidence, at least for English, that written relative clauses include more types (i.e., more varied; cf. Montag & MacDonald 2015). From a usage-based perspective (e.g., Bybee 2010), speakers that do not have much written language exposure might extract fairly specific/low-level templates. Usage-based studies suggest that more variation of instantiations of a construction helps with acquiring a more general/higher-level template (e.g., Street & Dabrowska 2014). Thus, it is possible that the overall higher performance of literate speakers could be attributable to the fact that they experienced more different types of various constructions. 

Another possible explanation is that the quality of exposure may reveal distinctions that are difficult to discern without the necessary linguistic experience. Speakers with more written language experience may have experienced more type-tokens of such structures, leading to stronger representations of these form-meaning pairings. We could argue that the structures tested in this study are complex because the suffixes that occur in these constructions have multiple form-meaning mappings and they are non-canonical structures. With more print exposure, it could be argued that speakers gain more metalinguistic awareness, leading to an improved grasp of structures that appear in multiple form-meaning pairings.

A third possibility the availability of literacy might help with fine-tuned grammatical representations and ease processing them. This is known as the training wheels hypothesis of Dabrowska (2020). This hypothesis proposes that written representations eases the cognitive load on working memory, which enables skilled-readers to process complex structures more efficiently than they would be able to in spoken modality. With prolonged exposure to written language, speakers will have experienced more types and tokens of various structures, potentially entrenching these structures with stronger representations, allowing for more efficient accessibility and processing. 

Another important note is that, in an ideal world, these accounts are not mutually exclusive. It is probable that a combination of all these factors contribute to heightened online comprehension and production of complex syntax. Recall that education improves IQ, which improves problem solving skills and potentially the acquisition of other complex structures and literacy: in other words cognition. Based on previous studies, it is highly likely that the causal relations that we postulate here are entangled with each other (cf. Cunningham & Stanovich, 1997, 1998; Stanovich, 1986; Stanovich & Cunningham, 1992). We know reading improves cognition, which in turn would influence linguistic abilities, which affects the speed at which reading and cognition develop, which improves language and so on.  

Similarly, as for production, schooling improves IQ and teaches the native language back to its native speakers, making them literate. Highly literate societies and speakers are influenced by written conventions when producing speech (i.e., Montag & MacDonald 2015). Thus, according to our findings, producing complex syntax is tied to both years of education, literacy and above average nonverbal IQ. Again, in an ideal world, they are all entangled with one another. 

This brings me to my conclusion. The extent of individual differences in L1 speakers is a lot greater than previously postulated. Does acquiring literacy modulate L1 grammatical knowledge? Based on our findings and accounts here, we suggest that the answer is yes, especially in light of previous studies. However, there is still a long way to go to confidently claim this. In any case, literacy does one thing right (in languages with a writing system). And I could not explain it any better than women who have already experienced it, who have emancipated themselves from the patriarchy one letter at a time.  

Before I finish the talk, I will repeat the take-away messages and finish the talk to welcome any questions or comments:

A) we shouldn’t over rely on conventional pieces of wisdom in our theories 

B) not all native speakers converge on the same grammatical knowledge uniformly and successfully

C) literacy facilitates ultimate native language learning

Thank you very much for your attention.