DeepFM-driven personalised recommendation algorithm integrating multimodal multimedia for college English listening: an empirical study with controlled experiment

To address limitations in college English listening instruction, including limited resource diversity and insufficient personalisation, this study developed an adaptive learning platform that integrated multimodal resources with a personalised recommendation algorithm.A deep learning (DL)-based feature extraction framework was constructed to capture key representations from audio, video, and text modalities, which were then fused using the ResNet50 model.An intelligent listening platform was subsequently designed and implemented, and a DL-based recommendation algorithm, DeepFM.A total of 120 non-English major sophomores from a local undergraduate university in Sichuan were recruited.The students were randomly assigned to an experimental group and a control group (n = 60 each) and participated in an 8-week instructional intervention.The results showed that DeepFM achieved a Precision@10 of 0.823 and an NDCG@10 of 0.864, outperforming baseline models.The experimental group's listening scores increased by 31.3%, compared to 11.7% in the control group, indicating a significant instructional effect.Unlike traditional platforms that merely aggregate multimedia, this study innovatively introduces an attention mechanism within the DeepFM framework, dynamically adjusting the modal weights of audio, video, and text based on learners' specific skill gaps, thereby achieving more precise resource matching.