Empathy Practices in Social Media Discourse: A Multidimensional and Relational Perspective

Paper Title

Empathy Practices in Social Media Discourse: A Multidimensional and Relational Perspective

Publication Info

• Topic area: Empathy practices in online social media interactions
• Keywords: Empathy, online communities, social media, Reddit, Stack Exchange, language models, multidimensional analysis, relational dynamics, fine-tuning, online discourse

Background and Problem

• Problem / challenge: Existing research often treats empathy in online interactions as a binary or unidimensional construct, overlooking its multidimensional and relational nature. This limits the understanding of how empathy is solicited and expressed in different contexts.
• Significance: Understanding empathy practices in online communities is crucial for improving platform design, fostering supportive interactions, and addressing the cognitive and emotional costs of empathy.
• Motivation and related work: Prior studies have explored empathy in specific contexts, such as mental health forums, but often lack granularity, relational perspectives, and contextual awareness. This paper builds on psychological theories and computational approaches to address these gaps.

Solution

• Proposed approach: A multidimensional and relational framework for studying empathy practices in online communities, combined with fine-tuned language models (LMs) to detect these practices at scale.
• Novelty:
  1. Development of a comprehensive framework capturing distinct empathy practices in posts and replies.
  2. Creation of a high-quality annotated dataset of Reddit and Stack Exchange interactions.
  3. Fine-tuning of LMs to identify multidimensional empathy practices, achieving higher granularity and contextual sensitivity than prior models.
  4. Empirical analysis of platform- and community-specific patterns in empathy practices.
• Procedure and key techniques:
  • Thematic analysis to define six empathy-seeking and seven empathy-giving practices.
  • Annotation of 1,302 post–reply pairs for these practices.
  • Fine-tuning of LMs (Phi-3-Medium-Instruct, Llama-3.1-8B-Instruct, Mistral-Nemo-Instruct) using 4-bit quantization.
  • Application of the best-performing model to a large-scale dataset (21,432 posts and 97,071 replies) for statistical analysis.

Results

• Concrete findings:
  • Fine-tuned LMs achieved macro-F1 scores of 0.72–0.90 across empathy themes, outperforming few-shot baselines.
  • Empathy practices in posts and replies were systematically identified, with some themes (e.g., "asking for advice") easier to detect than others (e.g., "expressing validation").
  • Platform differences: Reddit showed more emotional empathy, while Stack Exchange emphasized informational and task-oriented empathy.
  • Community differences: Parenting and social skills communities exhibited higher emotional and cognitive empathy than work-related communities.
• Advantage over baselines:
  • Fine-tuned models consistently outperformed random, majority-class, and few-shot baselines, demonstrating the value of task-specific fine-tuning.
• Experiments / evaluation:
  • Dataset: 21,432 posts and 97,071 replies from six communities (three each on Reddit and Stack Exchange).
  • Metrics: Macro-F1 scores for classification tasks.
  • Analysis of correlations between empathy-seeking and empathy-giving practices, as well as their effects on engagement (e.g., reply counts, OP responses).
• Limitations and future work:
  • Limited representation of some themes (e.g., "seeking emotional support").
  • Focus on explicit empathy practices; implicit signals remain underexplored.
  • Need for longitudinal studies to assess the long-term effects of empathy on community health and user retention.

Summary

This study introduces a multidimensional and relational framework for analyzing empathy practices in online communities, supported by fine-tuned language models. The framework captures six empathy-seeking and seven empathy-giving practices, enabling nuanced analysis of Reddit and Stack Exchange interactions. Fine-tuned LMs demonstrated strong performance in detecting these practices, revealing platform- and community-specific patterns. The findings challenge the assumption that empathy is uniformly positive, highlighting its contextual variability. This work provides actionable insights for platform design, community support, and AI-driven tools, while paving the way for future research on implicit empathy signals and long-term impacts.