The increasing use of learner corpora in second language acquisition (SLA) research has led to a growing body of evidence that supports the usage-based construction grammar approach to language learning. This approach suggests that language development is strongly influenced by statistical regularities in learners’ input, including the frequency, dispersion, and contingency of linguistic items. However, limitations to our understanding remain. A particularly notable gap relates to the fact that while numerous studies corroborate the effect of statistical properties in learners’ input on their acquisition and use of less schematic constructions, there is a dearth of evidence relating to more complex constructions, particularly ones which are non-adjacent, i.e., where the collocating elements are separated by an intervening string of words. As a result of this limitation, some influential scholars remain sceptical and argue that more research is needed to understand the role of statistical learning on second language (L2) acquisition. 

In light of these issues the present study aimed to investigate L2 learners’ development and use of a type of complex non-adjacent verb-argument construction in Mandarin Chinese - the ‘duì-construction’. It adopted a corpus linguistics methodology, utilising data from three corpora – an L2 learner corpus, a general Chinese reference corpus and a teaching Chinese as a foreign language textbook corpus. The reference and textbook corpora served as proxies (albeit imperfect ones) for the input that learners receive inside and outside the classroom. To facilitate its analysis, the study used a number of statistical models and machine learning techniques, including non-parametric tests, various types of generalised linear models, and a random forest model. Key aims were to establish L2 learners’ duì-construction development trajectory and then to explore the factors (both distributional and linguistic) that influence their verb usage in duì-constructions relative to native speakers and textbooks. 

The study produced three main sets of findings. Firstly in terms of L2 learners’ development trajectory, it revealed that as their proficiency increased, their dui-construction usage became more productive, complex and sophisticated. However, the development was non-linear and exhibited instances of regression and plateau. In terms of accuracy a more complicated picture emerged. Specifically, while advanced learners exhibited reduced surface-level errors, collocation errors increased. Secondly, the study shows that advanced L2 learners’ usage of dui-constructions generally align with the patterns evident in the general reference and textbook corpora but with certain important nuances and inconsistencies across different figures. Finally, the findings reveal that while verb usage frequencies in the general reference and textbook corpora are strong correlating factors with L2 learners’ verb usage in duì-constructions, other factors such as input accessibility, reliability, and L2 speakers’ preferences for certain functions and semantic classes, can override the frequency effect. However, the impact of learner’s first language, task type, and proficiency were not as influential as expected. 

The findings of the study have important theoretical, methodological and pedagogical implications. These apply both to usage-based language learning in general and to the specific field of Chinese as foreign language.