Embarking on the journey of learning a new language is an enriching experience, but it is not without challenges. In the face of rapid speech rates, transient linguistic signals, and inherent limitations in the human cognitive system to perceive and process auditory information, all language users, including both first language (L1) speakers and second language (L2) learners, are fundamentally constrained by the Now-or-Never bottleneck in real-time language processing. Bypassing the necessity to learn a new language exclusively by extracting statistical patterns from fleeting, continuous, and incomprehensible linguistic input, L2 learners and speakers tend to rely on existing conceptual, linguistic, and metalinguistic knowledge and over-segment L2 speech in favor of L1 transference at the lexical level, underusing the mechanism of chunking for effective real-time processing. This tendency is not only associated with less efficient online processing but also limits the integration of linguistic and conceptual information in learning. This dissertation investigates chunking as a unifying mechanism that underlies sentence processing, language proficiency, and knowledge acquisition in a second language across three empirical studies. Chapter 2 examines the relationship between auditory chunk recall and online sentence processing among L2 learners, showing that higher L2 chunk recall was associated with faster L2 processing across sentence types and that processing efficiency depends on chunking and representing chunks as coherent linguistic units rather than as flexible combinations of individual words. Chapter 3 evaluates the efficacy and applicability of utterance recall, paired with fine-grained automatic scoring, as a measure of real-time L2 proficiency. The task taps into chunking, real-time processing, as well as long-term L2 knowledge while simultaneously assessing comprehension and production skills. Chapter 4 extends the investigation into educational contexts, linking chunk-based language skills with memory outcomes for lecture content. Lecture recall performance is predicted by chunking sensitivity and utterance recall across both L1 and L2 speaking students, and when individual differences in these skills are accounted for, the group-level difference in lecture recall performance largely disappeared. Together, these findings highlight a central role of chunking across domains and levels of language use, providing empirical support for integrating chunk-based approaches in language education.