Understanding Empathy in Online Support Forums

Empathy is a powerful psychological tool for supporting people in both therapeutic and daily settings. With the growing demand for peer support in online communities, particularly within mental health and social support forums, understanding how empathy is communicated through text has become increasingly important. Most prior computational work has focused on modeling empathy through emotional signals, such as identifying affective states or sentiment in text. While these approaches capture what emotions are expressed, they provide limited insight into how empathy is conveyed through language. As a result, the linguistic mechanisms that shape empathetic communication, such as the use of figurative language, remain largely underexplored. This dissertation addresses this gap by advancing the modeling and interpretation of empathy in text-based online support forums, with an emphasis on domain-specific peer-support settings that allow for more detailed analysis of empathetic expressions. As an initial step, we introduce AcnEmpathize, a publicly available dataset capturing empathetic expressions in acne support forums, a mental health–relevant domain that has received little attention in prior work. Using this dataset, we establish strong baselines for automated empathy detection and demonstrate that empathy can be reliably modeled even in specialized, underexplored settings. Building on this foundation, we investigate the role of figurative language in empathetic communication through a series of detection and modeling studies. Through empirical analysis and participation in the WASSA shared task on empathy and emotion prediction in conversations, we demonstrate that incorporating figurative language signals–such as metaphors, idioms, and hyperbole–consistently improves performance in empathy and emotion prediction in both static and conversational settings. These findings suggest that figurative language is a meaningful component of empathetic expression, not a mere stylistic feature. We then move from empathy detection to generation, with the goal of producing contextually aligned and linguistically expressive replies that more effectively support individuals in online peer-support forums. In this study, we manually annotate empathy-evoking sentences in the seeker’s post and incorporate them into the generation process with figurative language labels. Results from both automatic metrics and human evaluations reveal that combining these contextual and linguistic elements improves overall response quality, including fluency, lexical diversity, and perceived empathy. Finally, we introduce Empathy–Metaphor, the first corpus to explicitly annotate metaphorical spans in empathetic replies. This dataset builds on AcnEmpathize and focuses on metaphors due to their strong emotional resonance, frequency, and diversity in empathetic communication. By annotating metaphors at the span level, we enable a more detailed examination of how they function within empathetic replies. Through quantitative and qualitative analyses, we show that metaphors are frequent, diverse, and strategically positioned, often framing experiences of shared struggle and perseverance. Benchmark experiments support that metaphorical spans can be reliably identified using transformer-based models, providing a concrete resource for future research on figurative language and empathy. In this dissertation, we study empathetic language from multiple perspectives, moving beyond emotion-related signals. By leveraging figurative language as a central mechanism, we advance empathy modeling through more interpretable and contextually grounded approaches. All studies rely on publicly available online support forum data and received a Not Human Research Determination from the University of Illinois Chicago Institutional Review Board. This work provides a foundation for future research on empathy and its application in computational tools for online peer-support communities.

Author dc:creator

• Gyeongeun Lee (22940746)

• Computer Science
• Artificial Intelligence
• Natural Language Processing

Statement dc:rights

• In Copyright
• Open Access after 2028-05-01

DOI dc:identifier

https://doi.org/10.25417/uic.32995079.v1

OAI identifier oai:identifier

oai:figshare.com:article/32995079