💬 NLP

Coping in Crisis: Computational Modeling of Coping Styles in Digital Crisis Discourse During the 2023 Turkiye Earthquake

This study utilizes a multi-label BERTurk classifier on over one million Turkish tweets to demonstrate that Lazarus and Folkman's coping theory can be reliably operationalized in real-time digital crisis data, revealing distinct temporal trajectories of problem-focused, emotion-focused, and meaning-making coping styles during the 2023 Turkiye earthquake.

Original authors: Şevval Çakıcı

Published 2026-06-15

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Şevval Çakıcı

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a massive, chaotic storm hits a city. In the old days, if researchers wanted to know how people were handling the disaster, they would wait weeks or months. They would sit down with survivors in a quiet room and ask, "How did you feel? How did you cope?" But memory is tricky; by the time people answer, they've already rewritten their own stories in their heads.

This paper is about a different kind of research. Instead of waiting, the researchers looked at what people were writing while the storm was still raging. They analyzed over one million tweets posted in the immediate aftermath of the February 2023 earthquake in Türkiye. Think of these tweets not as perfect windows into people's souls, but as a giant, real-time "heartbeat monitor" for a society in shock.

Here is the breakdown of what they found, using simple analogies:

1. The Three Ways People "Coped" (The Toolkit)

The researchers used a theory called "Coping Theory" to sort the tweets into three main buckets. Imagine people have a mental toolkit, and they pull out different tools depending on the stage of the crisis:

The "Fix-It" Tool (Problem-Focused): This is for when you need to solve a puzzle. In the tweets, this looked like sharing rescue coordinates, asking "Where is the nearest hospital?", or coordinating aid.
The "Feel-It" Tool (Emotion-Focused): This is for when the puzzle can't be solved yet, so you need to manage your feelings. These tweets were full of grief, prayers, "I miss you," or expressions of fear and sadness.
The "Why-It" Tool (Meaning-Making): This is for when you try to make sense of the chaos. These tweets asked, "Why did this happen?" or "Who is to blame?" They were looking for a story or a moral lesson to explain the tragedy.

2. The Computer Translator (The Robot Librarian)

To read a million tweets, the researchers couldn't do it by hand. They built a special computer program (a "BERTurk" model) trained to act like a super-fast librarian.

They taught this librarian by showing it 500 examples of tweets and telling it which "tool" (Fix-It, Feel-It, or Why-It) each one used.
Once trained, the librarian could read the remaining million tweets in seconds.
The Result: The Turkish-trained librarian was much better at understanding the local language and slang than a generic, "one-size-fits-all" international robot. It got the job done with high accuracy.

3. The Story of the Month (The Timeline)

When they watched how the "tools" were used over 30 days, they saw a very clear, predictable pattern, like the changing seasons after a disaster:

Days 1–3 (The Rush): The "Fix-It" tool was used the most. People were frantic, sharing locations and asking for help. It was pure survival mode.
Days 4–7 (The Shift): As the immediate rescue urgency faded, the "Fix-It" tweets dropped sharply. The "Feel-It" tweets started to rise. People began to process the grief.
Weeks 2–4 (The Plateau and the Search): The "Feel-It" tweets stabilized (people were still sad and grieving). But the "Why-It" tool started climbing steadily. People began asking hard questions about why the buildings fell and who was responsible.

4. The Anger Connection

One of the most interesting findings was about anger.
The researchers found that when people were angry, they rarely used the "Fix-It" tool. Instead, anger was strongly linked to the "Why-It" tool.

The Analogy: Think of anger not as a fuel for practical action (like digging through rubble), but as a spotlight. When people were angry, they used that energy to shine a light on blame and accountability. They weren't just asking "How do we help?" they were asking "Who broke the rules that let this happen?"

5. Why This Matters (The Practical Takeaway)

The paper suggests that if we can read this "heartbeat" in real-time, we can help people better.

If a population is mostly using the "Fix-It" tool, they need information, maps, and coordination.
If they are mostly using the "Feel-It" tool, they need space to grieve, acknowledgment of their pain, and emotional support.
If they are shifting to the "Why-It" tool, they are looking for transparency and answers about responsibility.

The Bottom Line:
This study proves that we can use computers to understand how a society processes trauma in real-time. It shows that even in a chaotic, polarized environment, human responses to disaster follow a predictable rhythm. By understanding which "tool" a community is using at any given moment, aid workers and leaders can stop guessing and start responding exactly to what people actually need in that moment.

Technical Summary: Coping in Crisis: Computational Modeling of Coping Styles in Digital Crisis Discourse During the 2023 Türkiye Earthquake

Problem Statement
Traditional research on psychological coping during disasters relies on retrospective self-reports (e.g., surveys, interviews) administered weeks after an event. This approach suffers from memory reconstruction biases and fails to capture real-time psychological dynamics under actual stress. While social media offers a massive, timestamped record of public response, existing computational approaches have predominantly focused on English-language data and have not systematically operationalized established psychological theories, such as Lazarus and Folkman's (1984) coping theory, for non-English, low-resource contexts. Furthermore, the specific interplay between collective trauma and political polarization in crisis discourse remains under-explored in computational social science.

Methodology
The study utilizes a corpus of 1,037,663 original Turkish-language tweets posted between February 6 and March 6, 2023, following the Türkiye earthquake. The data was sourced from the Politus corpus, filtered for keywords related to the disaster, and divided into four theoretically motivated crisis phases: Urgency (Days 1–3), Blame (Days 4–7), Grief (Weeks 2–3), and Meaning-Making (Week 4+).

The core methodological contribution is the development of a multi-label classification pipeline using BERTurk (a BERT-base model pre-trained on a large Turkish corpus).

Annotation: A stratified random sample of 500 tweets was annotated for three coping styles derived from Lazarus and Folkman (1984): problem-focused, emotion-focused, and meaning-making. A fourth category, avoidance, was excluded due to near-zero inter-annotator agreement ( $\kappa = 0.021$ ).
Training: The 500-tweet dataset was split into training (70%), development (15%), and test (15%) sets. The BERTurk model was fine-tuned for multi-label sequence classification using sigmoid activation and binary cross-entropy loss.
Baseline: Performance was compared against a zero-shot baseline using mDeBERTa (a multilingual model), where coping categories were defined via natural language hypotheses.
Analysis: The trained classifier was applied to the full corpus to track temporal dynamics. Spearman rank correlations were computed between coping styles and engagement metrics (likes, retweets) and pre-computed emotion scores (anger, fear, sadness, etc.) from the Politus model.

Key Results

Model Performance: The fine-tuned BERTurk model achieved a macro F1 score of 0.693, substantially outperforming the zero-shot mDeBERTa baseline (macro F1 = 0.324). The improvement was most significant for emotion-focused coping ( $\Delta F1 = +0.542$ ), highlighting the necessity of language-specific models for capturing Turkish emotional nuances.
Temporal Dynamics: The analysis revealed a clear, non-random trajectory in collective coping:
- Problem-focused coping dominated the Urgency phase (56.2%) and declined monotonically to 19.4% by Phase 4.
- Emotion-focused coping rose from 32.0% in Phase 1 to stabilize at approximately 55% in Phases 3 and 4.
- Meaning-making increased steadily across all phases, rising from 12.4% to 37.3%.
Demographic and Emotional Correlates:
- Demographics: Female users exhibited higher rates of problem-focused coping, while male users showed higher rates of emotion-focused and meaning-making coping. Organizational accounts displayed significantly lower meaning-making rates than personal accounts.
- Emotion-Coping Link: Anger showed the strongest correlation with meaning-making ( $r = 0.387$ ) and a negative correlation with problem-focused coping ( $r = -0.285$ ). This suggests that in this polarized context, anger functioned as a mobilizing force for blame attribution and political interpretation rather than practical action. Sadness correlated most strongly with emotion-focused coping ( $r = 0.184$ ).

Key Contributions

Empirical Validation: This study provides the first large-scale empirical test of coping theory in real-world digital crisis discourse, extending a literature previously confined to laboratory settings and self-reports.
Methodological Feasibility: It demonstrates that coping style detection is feasible in non-English, low-resource contexts using language-specific transformer models, bridging a gap in computational social science.
Substantive Insights: The findings offer a granular view of how collective coping evolves across crisis phases, revealing that coping dynamics are not merely psychological but are deeply entangled with the political context (e.g., the proximity to a national election and issues of state accountability).

Significance and Claims
The paper claims that coping theory can be reliably operationalized in digital data to reveal how societies process collective trauma. The authors argue that these findings are not merely descriptive but have practical utility for humanitarian organizations and crisis communicators. By detecting the dominant coping style of a population in real-time, responders can tailor their interventions:

Problem-focused dominance signals a need for information, coordination, and resource access.
Emotion-focused dominance signals a need for acknowledgment, presence, and space for grief.
Meaning-making dominance signals a need for institutional transparency and accountability mechanisms.

The study concludes that while tweets are public textual traces rather than transparent windows into individual psychology, the aggregate patterns provide a valid signal of collective psychological states. The authors emphasize that such tools should be used to understand and respond to population needs rather than for surveillance, noting the ethical responsibilities inherent in analyzing vulnerable populations in crisis.