Inferring Emotional Intensity from Early Dialogue Context

Most Natural Language Processing (NLP) systems for mental health conversations classify which emotion is present, but say nothing about how urgently that emotion needs a response. This matters for digital support platforms: intervention can only be timely if urgency can be judged from a conversation's opening exchanges, before it plays out in full. We investigate whether transformer-based models trained on opening exchanges can predict the final emotional intensity of support dialogues, proposing a segment-level evaluation framework with three validation settings: Early→Early, Early→Late, and Mixed→Late. Experiments with BERT, EmotionBERT, and DistilBERT across three augmentation strategies (synonym replacement, back-translation, LLM-based rewriting) show that data augmentation is the main performance driver, raising F1-macro from near-baseline (~0.08-0.10) to substantially higher results. Back-translation yields the best Early→Late result (BERT, F1-macro = 0.5606) and EmotionBERT achieves the highest overall score in Mixed→Late setting (F1-macro = 0.7210). The results indicate that data-centric augmentation strategies are more influential than encoder selection for low-resource emotion intensity prediction.