Predicting Final Emotional Intensity from Early Support Dialogue: Effects of Data Augmentation and Multi-Task Learning

Emotional intensity is an important but underexplored aspect of affective language understanding, particularly in conversational settings where early intervention may be beneficial. This study investigates whether emotional intensity can be predicted from the opening exchanges of emotional support conversations. Using the ESConv corpus, each dialogue was segmented into early and late phases, enabling a controlled evaluation of cross-phase prediction. We compare single-task and multi-task transformer models based on BERT and DistilBERT, and systematically evaluate three augmentation strategies: synonym replacement, back translation, and LLM-based rewriting. Across 120 experimental configurations, the strongest single-task result on the target Early→Late setting was obtained by BERT with back-translation at 50% augmentation (macro-F1 = 0.561). All augmentation strategies substantially outperformed the non-augmented baseline, with LLM-based rewriting providing the most stable overall gains. In contrast, multi-task learning with auxiliary emotion classification consistently underperformed the single-task baseline, indicating negative transfer rather than beneficial shared supervision. The findings show that early dialogue context contains useful intensity-related signal, but that performance depends critically on augmentation and task formulation.