This paper investigates uses of corpora in language learning ("data-driven learning") through analysis of a 600K-word corpus of empirical research papers in the field. The corpus can tell us much--the authors and the countries the studies are conducted in, the types of publication, and so on. The corpus investigation itself starts with frequency lists of words and clusters to detect initial themes, which are then extended (via distribution plots, collocates, concordances, etc.) to look at specific items: the researchers cited, the theoretical constructs and concepts investigated and how they are treated, and so on. The paper ends by dividing the corpus into early and more recent papers to compare evolution over time. This reveals keywords that were prevalent in earlier days as a snapshot of the past, and keywords today which may give an idea of future directions. [For full proceedings, see ED565044.]