Effects of using data-driven and generative AI-assisted instructions on learning multiword constructions in written academic English

With a pretest/posttest/delayed posttest design, this quasi-experimental study compares the effectiveness of corpus use versus generative AI (GenAI) use in helping undergraduate students learn multiword constructions (MWCs) in written academic English and examines the moderating effects of learners’ learning styles. Six intact writing classes with a total of 210 students were assigned to one of three instructional groups with each group including two intact classes: Experimental Group 1 (using data-driven learning [DDL] only), Experimental Group 2 (employing GenAI-assisted learning combined with corpus analysis of teacher-generated concordance lines), and Control Group. MWC learning was measured by a MWC production test and a writing test that assessed both the overall writing quality and the frequency and number of types of MWCs used. Statistical analyses revealed that while both experimental groups made significantly more improvements in their MWC use than the control group across all four learning measures in both posttests, Experimental Group 2 (GenAI combined with corpus analysis) significantly outperformed Experimental Group 1 (DDL only) on nearly all measures in both posttests. Survey results also show that while these two instructional methods were well-received by most participants, the combined instruction better accommodated diverse learning styles than the DDL-only instruction did. Together, these findings indicate that combining DDL and GenAI can significantly enhance the effectiveness of MWC instruction in academic writing.