Testing the limits of data-driven learning: language proficiency and training

The potential for corpora in language learning has attracted a significant amount of attentionin recent years, including in the form of data-driven learning (DDL). Careful not to appear toover-promote the field, enthusiasts have urged caution in its application, in particular withregard to lower-level learners, and have argued that extensive learner-training in corpustechniques is an essential condition for DDL to be successful. Such limits seem eminentlyreasonable, but there is a notable dearth of empirical studies to support them. This paperdescribes a simple experiment to see how lower-level learners cope with corpus data with noprior training.The language focus here is on linking adverbials in English, which are renowned to bedifficult to teach using traditional methods. The subjects are 132 first-year students at anengineering college in France of roughly intermediate and lower levels of English. They weredivided into random groups to compare their ability to deal with the target items usingtraditional sources (extracts from a bilingual dictionary or a grammar/usage manual) orcorpus data (short contexts or truncated concordances). Performance was tested prior to theexperiment, subsequently to check ability to use the different information sources as a reference,and later to test recall.No evidence was found that traditional sources promote better recall, and corpus dataseemed to be more effective for reference purposes. While the results of any single experimentmust be treated with caution, these findings suggest the need for more empirical studies tocomplement the theoretical arguments and qualitative data which currently dominate thediscussions of DDL