Tailoring Adaptive-Zero-Shot Retrieval and Probabilistic Modelling for Psychometric Data

Likert scales are widely used tools in psychology, employed to quantify individuals' feelings, attitudes, and perceptions through structured questionnaires. However, administering these questionnaires can be time-consuming and resource-intensive, limiting their practicality in fast-paced mental health screening scenarios. This study presents a novel approach to predict Beck Depression Inventory (BDI-II) scores using social media posts. Our method introduces two key innovations: an adaptive strategy for identifying relevant social media content according to each survey question (aka item) and a probabilistic extension of BERT to predict item-specific scores. The results show that our implemented approach is particularly accurate in correctly predicting responses to BDI-II questionnaire items compared to the considered benchmarks.