Text this: Predicting student mental health through entropy-based features and interpretable cross-attention transformer networks