Author Identifier
Sharjeel Tahir: http://orcid.org/0009-0008-9012-0490
Date of Award
2026
Document Type
Thesis
Publisher
Edith Cowan University
Degree Name
Doctor of Philosophy
School
School of Science
First Supervisor
Syed Afaq Shah
Second Supervisor
Jumana Abu-Khalaf
Abstract
Empathic communication demands more than recognizing sentiment from words. It requires inferring when language, prosody, and facial cues diverge; how they evolve over time; and adapting to each person’s distinctive way of expressing disparate emotions. Contemporary conversational AI (chatbots), despite their linguistic fluency, remain brittle on the aforementioned hindrances. This creates a threefold gap: a measurement gap (what to optimize for empathic understanding), a data gap (where such instances are represented), and a control gap (how systems adapt to users’ mood and behavior during interaction).
This thesis investigates computational foundations for empathic conversational AI that address these gaps through five contributions. (i) A structured review that identifies the gaps in the field of artificial empathy- with respect to the existing modeling and evaluating techniques as well as the datasets and how to bridge them. (ii) An empirical testing of LLMs in therapeutic dialogue generation. (iii) A novel E-THER dataset that introduces annotation of verbal–visual incongruence (mismatch) across 789 dialogue pairs; 20% of expressed emotions are misaligned, with recurring patterns (e.g., minimisation of negative emotions). (iv) A feedback-aware retrieval framework with temporal decay (SCIRAG), personalizes responses while maintaining communication appropriateness, yielding a 15% gain over static baselines and 85.1% semantic accuracy. (v)Finally, person-centred evaluation metrics (attunement, responsive engagement, concision) that correlate with psychology expert judgments (r = 0.74), whereas BLEU/perplexity show near-zero correlation with empathic quality.
Analyses further reveal wide individual variability in incongruence (some people may habitually conceal their visual emotions more than others) and greater misalignment as conversations deepen (15% average incongruence at beginning of sessions to 23% incongruence in the later stages of conversations), motivating user-calibrated models. Our goal is augmentation, not substitution, of human connection. The datasets, training methods, and evaluation frameworks introduced in this thesis provide foundations for responsible progress in support-oriented AI that is attentive to what is said- and what remains unspoken.
Access Note
Access to this thesis is embargoed until 14th February 2027
DOI
10.25958/fa5y-dr15
Recommended Citation
Tahir, S. (2026). Novel computational frameworks for empathy modeling and evaluation in artificial agents. Edith Cowan University. https://doi.org/10.25958/fa5y-dr15