FU-TU-DFKI@eRisk 2025: A Linguistically Informed but Overdiagnosing Approach to Early Depression Detection
This paper describes the participation of the FU-TU-DFKI team in the eRisk 2025 Task 2, Contextualized Early Detection of Depression. We propose a hybrid approach that combines transformer-based modelling with linguistic and meta feature analysis. While our model achieved high recall, it exhibited low precision, resulting in an overall F1-score of 0.29 in the official evaluation. We interpret this cautious behaviour as a tendency toward overdiagnosis. Beyond the technical system, we investigated the linguistic characteristics of user messages via corpus-linguistic methods, including Collostructional Analysis– a method for identifying statistically significant associations between words and grammatical constructions. Additionally, we examine the ethical implications of automated depression detection, and highlight the reductionist interpretation of complex affective utterances in such systems. Our submission emphasizes the importance of interpretability and caution in high-stakes, health-related NLP tasks, particularly when system performance remains limited.
