Class imbalance problem afflicts medical data classification due to the inherent bias towards the over-represented class. This results in improper prediction of the minority class, which is generally the class of interest. In this study, imbalanced data is handled for text-based depression detection using popular oversampling techniques. Depression is a critical health condition with serious consequences if not handled early. However, the lack of substantial data for depressed individuals in comparison with that of healthy individuals leads to low-performing decision models. This study is an exhaustive investigation of the combinations of various text vectorization methods and popular oversampling techniques over varied classifiers. Experiments exhibit best performance with the combination of CountVectorizer and adaptive synthetic sampling using the k-nearest neighbour classifier with F1 depressed score and F1 non-depressed score as 0.80 and 0.93, respectively. Taking the severity levels of depression into consideration, a context-aware severity-based adaptive oversampling technique for depression (AdaDepSyn) is proposed to generate a normally distributed synthetic data. The proposed technique shows improved performance with F1 depressed score and F1 non-depressed score as 0.81 and 0.90 respectively and is statistically significant for depression detection.