Learners normally acquire their first languages in context. While this poses
problems – e.g., referential ambiguity – it might also provide benefits, such as
ensuring that appropriate patterns of relations between words are acquired.
Another benefit that could arise from learning in context is that while obvious
similarities among referents might be learnable in isolation, some subtle
differences could be better discerned in contexts that allow referents to be
aligned (Markman & Gentner, 1997). Since second language learning often
occurs in sparse contexts, this may explain why L2 learners struggle to acquire
aspects of grammar encoding differences. To explore the impact of learning in
and out of context, we trained participants on an artificial language containing a
grammatical gender system – which L2 learners typically struggle to learn – that
systematically matched a set of morphological markers to differences among the
features of its nouns. A between-participants experiment contrasted a traditional
"presentation training paradigm" in which referential phrases and corresponding
images were introduced individually, with a novel "alignment error paradigm"
where referential phrases were presented alongside two images, and learners
had to guess the correct match. Our results show that although noun-phrases
were learned well in both paradigms, morphological generalization only occurred
when training allowed for contextual alignment, suggesting that this may be
critical for learning many aspects of grammatical structure.

Markman, A. B., & Gentner, D. (1997). The effects of alignability on memory.
Psychological Science, 363-367.