Leveraging Conceptual Similarities to Enhance Modeling of Factors Affecting Adolescents’ Well-Being

While large language models consistently outperform their smaller transformer-based counterparts, there are constraints on their deployment. Model size becomes a critical limiting factor in cases involving sensitive data, particularly when the imperative is to execute inference on edge devices such as smartphones. We explore the possibility of detecting common positive and negative influence factors that impact adolescents’ well-being in instant messenger communication using a newly annotated dataset. We show that by leveraging the similarities between the concepts, we can produce classifiers with a small ELECTRA-based model with 14M parameters that can run on resource-limited edge devices. Our findings can be used to advance intervention and parental control software, creating a safer digital environment for children and adolescents.