It Takes Two to Empathize: One to Seek and One to Provide

Mahshid Hosseini, Cornelia Caragea

Empathy is a cornerstone of human interaction, yet in computational linguistics, it has often been treated as a simple binary state. Research from the University of Illinois at Chicago challenges this notion, arguing that in online health communities (OHCs), empathy is a dynamic exchange between two distinct roles: the seeker and the provider.

The Directionality of Digital Support

In high-stakes environments like cancer support forums, the social function of a message depends entirely on its direction. A user expressing fear performs a different role than one offering shared experience. To capture this nuance, researchers introduced IEMPATHIZE, a dataset of 5,007 sentences from the Cancer Survivors Network (CSN), meticulously labeled for empathy direction. Each sentence was categorized as seeking-empathy, providing-empathy, or none by trained annotators, achieving a high inter-annotator agreement (83% Cohen’s kappa).

Modeling Empathy with Domain-Adaptive BERT

The study compared several modeling approaches, ranging from traditional machine learning (SVMs, Random Forests) to advanced neural architectures (CNNs, Bi-LSTMs). The clear winner was a domain-adaptive BERT model. By "warming up" bert-base-uncased on a massive unlabeled corpus of cancer-related messages before fine-tuning on the IEMPATHIZE task, the model learned the specific jargon and emotional landscape of the community. This domain-specific pre-training provided a significant performance boost, particularly in identifying the subtle linguistic cues used by those seeking empathy.

Linguistic Fingerprints of Support

The research revealed distinct emotional and grammatical patterns that differentiate seekers from providers. Seekers tend to use first-person singular pronouns ("I", "me") and express "gloomy" emotions. In contrast, providers use second-person pronouns ("you") and lean toward "optimistic" emotions like joy and trust.

Emotional distribution across seeking and providing empathy.

As shown in the emotional distribution, seekers are characterized by high levels of fear and sadness, while providers predominantly express joy, trust, and anticipation. Interestingly, the study also found that "power posters"—the most engaged community members—are significantly more likely to act as providers of empathy, whereas regular users balance both roles.

Measuring the Impact: The "Positive Shift"

The ultimate goal of identifying empathy is to measure its impact on patient well-being. By analyzing nearly 37,500 discussion threads, the authors demonstrated that empathetic intervention leads to a measurable improvement in the seeker's expressed sentiment.

Comparison of sentiment shifts in threads with and without empathetic messages.

When a seeker receives an empathetic response, they experience a "positive shift" in sentiment in 34.44% of cases—nearly four times the rate (9.23%) seen in threads lacking such support. This finding underscores the tangible value of digital compassion in patient recovery and support.

Conclusion

While current models excel at sentence-level detection, empathy is inherently contextual. Future research will focus on incorporating thread history to better understand how support evolves over time. These insights pave the way for more sophisticated empathetic AI agents capable of providing the same high-quality support as the community's most dedicated human providers.

Summary updated

Abstract

Empathy describes the capacity to feel, understand, and emotionally engage with what other people are experiencing. People have recently started to turn to online health communities to seek empathetic support when they undergo difficult situations such as suffering from a life-threatening disease, while others are there to provide empathetic support to those who need it. It is, therefore, important to detect the direction of empathy expressed in natural language. Previous studies only focus on the presence of empathy at a high-level and do not distinguish the direction of empathy that is expressed in textual messages. In this paper, we take one step further in the identification of perceived empathy from text by introducing IEMPATHIZE, a dataset of messages annotated with the direction of empathy exchanged in an online cancer network. We analyze user messages to identify the direction of empathy at a fine-grained level: seeking or providing empathy. Our dataset IEMPATHIZE serves as a challenging benchmark for studying empathy at a fine-grained level.
BibTeX
@inproceedings{hosseini2021it,
  author = {Mahshid Hosseini and Cornelia Caragea},
  title = {{It Takes Two to Empathize: One to Seek and One to Provide}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/aaai.v35i14.17539},
}
Figure