Corpus-based pedagogy, in which learners consult a corpus and generate patterns from the corpus, has been shown to be effective in collocation teaching and learning. The present research investigated how 140 English learners from Taiwan corrected different types of collocation errors by consulting the Corpus of Contemporary American English (COCA) accompanied by other consultation resources. Three aspects, including COCA performance and improvement, COCA use, and reflection toward COCA use, have been explored together to provide comprehensive understanding. Students were required to take two tests: a paper-based test and a COCA-based test. The paper-based test required them to correct eight collocation errors of different types (verb+noun collocations vs. adjective+noun collocations; easy collocations vs. difficult collocations; L1 congruent collocations vs. L1 incongruent collocations) adapted from their essay writing, by providing three answers to replace each collocation error.

In the COCA-based test, they were required to correct the same eight collocation errors by consulting COCA and other consultation resources. Thirty out of 140 students who volunteered to have their corpus consultation videotaped were interviewed later. A mixed-methods approach that included quantitative statistics and qualitative analysis was used. The results show that in COCA performance and improvement, learners successfully induced patterns from COCA consultation and corrected collocation errors. They improved approximately 30% in both collocation accuracy and collocation richness after having access. Additionally, even though learners of higher English proficiency performed better, learners across different levels of English proficiency received a similar amount of assistance from COCA.

As for COCA use, different learners coordinated the corpus use and other consultation resources differently, and they gained benefits and encountered some difficulties when consulting COCA and other consultation resources. Nevertheless, the number of times they consulted COCA was similar among learners across different levels of English proficiency, but lower proficiency learners utilized more of the external consultation resources. Lastly, the COCA Reflection section shows that learners generally found COCA to be helpful and easy to use, but higher proficiency learners had higher evaluations of COCA use. Nonetheless, they did not necessarily think that COCA was easier to use, compared to lower proficiency learners.

It is concluded that, for learners of all levels of English proficiency, corpus use accompanied by other consultation resources greatly assisted learners in finding collocations, even though data-driven learning (DDL) itself still caused difficulties for learners in the process. Nonetheless, the advantages outweigh the difficulties. Because the use of external consultation resources helps learners to check unfamiliar vocabulary items encountered in COCA use, the key point is to encourage learners to embrace DDL, as this is different from their previous learning styles. Teachers should inform learners of the benefits of learning collocation through corpus and should guide learners in how to coordinate effectively the use of corpus and other consultation resources.