Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session B: Wednesday, August 13, 1:00 – 4:00 pm, de Brug & E‑Hall

Toward Affective Empathy in AI: Encoding Internal Representations of "Artificial Pain"

Angeline Wang¹, Iran R Roman¹; ¹Queen Mary, University of London

Presenter: Angeline Wang

Current chatbots excel at demonstrating cognitive empathy through language analysis but they lack mechanisms to internalize emotional intensity, a hallmark of human affective empathy mediated by neural substrates like the anterior cingulate cortex (ACC). We propose a framework inspired by ACC-mediated "Artificial Pain" encoding, integrating emotion classification with intensity regression. Using the Emotional Support Conversations (ESConv) dataset, we carry out transfer learning using MentalBERT, MentalRoBERTa, and ModernBERT in a multi-task setup that jointly models emotion categories and corresponding intensity levels on a 1–5 scale. We then evaluate these models to assess their capacity for emotion understanding and graded affective representation.MentalRoBERTa achieves state-of-the-art performance in single-task classification (F1=0.59) and multi-task settings (F1=0.63), with intensity regression showing significant correlations to the human-annotated ground-truth, but with relatively high estimation error. While multi-task learning improves emotion classification through shared intensity signals, predicting the intensity of emotions remains challenging, highlighting the need for model training with larger datasets. This work establishes a benchmark for emotion intensity-aware affective AI, bridging natural language processing methods with neuroscientific principles. Future implications include the advancement of affective empathy in human-agent interactions.

Topic Area: Language & Communication

Extended Abstract: Full Text PDF