Relating Natural Language Aptitude to Individual Differences in Learning Programming Languages

Abstract

This experiment employed an individual differences approach to test the hypothesis that learning modern programming languages resembles second "natural" language learning in adulthood. Behavioral and neural (resting-state EEG) indices of language aptitude were used along with numeracy and fluid cognitive measures (e.g., fluid reasoning, working memory, inhibitory control) as predictors. Rate of learning, programming accuracy, and post-test declarative knowledge were used as outcome measures in 36 individuals who participated in ten 45-minute Python training sessions. The resulting models explained 50-72% of the variance in learning outcomes, with language aptitude measures explaining significant variance in each outcome even when the other factors competed for variance. Across outcome variables, fluid reasoning and working-memory capacity explained 34% of the variance, followed by language aptitude (17%), resting-state EEG power in beta and low-gamma bands (10%), and numeracy (2%). These results provide a novel framework for understanding programming aptitude, suggesting that the importance of numeracy may be overestimated in modern programming education environments. Computer programming has moved from being a niche skill to one that is increasingly central for functioning in modern society. Despite this shift, remarkably little research has investigated the cognitive basis of what it takes to learn programming languages. The implications of such knowledge are wide reaching, both in terms of cultural barriers to pursuing computer sciences 1 and for educational practices 2. Central to both are commonly held ideas about what it takes to be a "good" programmer, many of which have not been empirically instantiated. In fact, remarkably little research has investigated the cognitive bases of "programming aptitude" 3-5 , and, to the best of our knowledge, no research to date has investigated its neural correlates. Critically, the existing research provides inconsistent evidence about the relevance of mathematical skills for learning to program 6-8. Despite this, programming classes in college environments regularly require advanced mathematical courses as prerequisites. This gap between what we know about learning to program and the environments in which programming is taught was described by Jenkins, who argued that "If computing educators are ever to truly develop a learning environment where all the students learn to program quickly and well, it is vital that an understanding of the difficulties and complexities faced by the students is developed. At the moment, the way in which programming is taught and learned is fundamentally broken" 2. Unfortunately, little progress has been made since this call to action 15 years ago. Across the same time period, the nature of programming languages has also changed, reducing the likelihood that the original research on learning to program in Pascal 5 or COBOL 3 , for instance, will generalize to contemporary programming languages. The research described herein is motivated by a...

Computer code comprehension shares neural resources with formal logical inference in the fronto-parietal network

“…Previous studies also found that math and formal logic did not depend on classic language networks (Amalric & Dehaene, 2016;Monti et al, 2009). Lack of overlap between code and language is intriguing given the cognitive similarities between these domains (Fedorenko et al, 2019;Pandža, 2016;Peitek et al, 2018;Portnoff, 2018;Prat et al, 2020;Siegmund et al, 2014). As noted in the introduction, programming languages borrow letters and words from natural language, and both natural language and code have hierarchical, recursive grammars (Fitch et al, 2005) .…”

“…Hypotheses about how the human brain accommodates programming range widely. One recently popular view is that code comprehension recycles language processing mechanisms (Fedorenko, Ivanova, Dhamala, & Bers, 2019;Fitch, Hauser, & Chomsky, 2005;Pandža, 2016;Portnoff, 2018;Prat, Madhyastha, Mottarella, & Kuo, 2020). Computer languages borrow letters and words from natural language.…”

“…Moreover, the grammars of language and that of code share common features, including recursive structure (Fitch et al, 2005). A recent study reported that individual differences in learning a second language predict aptitude in learning to program among novices (Prat et al, 2020). One possibility then is that coding recycles neurocognitive mechanisms involved in language processing (Peitek et al, 2018;Siegmund et al, 2014).…”

Comprehension of computer code relies primarily on domain-general executive brain regions

“…The lack of the language system engagement during code comprehension adds to the body of work that demonstrates high input selectivity of these regions (Fedorenko et al, 2011;Jouravlev et al, 2019;Monti et al, 2009Monti et al, , 2012Pritchett et al, 2018). Although the language system does not appear to support code comprehension, it may play a role in learning to program (Prat et al, 2020). Studies advocating the 'coding as another language' approach (Bers, 2019(Bers, , 2018Sullivan & Bers, 2019) have found that treating coding as a meaning-making activity rather than merely a problem-solving skill had a positive impact on both teaching and learning to program in the classroom .…”

Programming experience associated with neural efficiency during figural reasoning

“…For a comprehensive description of these tasks please see the methods section. Empirical studies indicated that higher CT and programming skills come along with higher reasoning skills 14 , 19 – 21 . Training of CT skills as well as programming skills has been shown to lead to an improvement in figural reasoning tasks but not in numerical or verbal reasoning tasks 13 , 14 , 20 , 22 , 23 .…”