💬 NLP

Dealing with Controversy: An Emotion and Coping Strategy Corpus Based on Role Playing

This paper addresses the gap between psychological and computational emotion research by introducing a role-playing-based corpus and a coping identification task to investigate how emotions function as behavioral strategies in text, revealing the challenges and potential for models to better capture these mechanisms.

Original authors: Enrica Troiano, Sofie Labat, Marco Antonio Stranisci, Viviana Patti, Rossana Damiano, Roman Klinger

Published 2026-02-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Enrica Troiano, Sofie Labat, Marco Antonio Stranisci, Viviana Patti, Rossana Damiano, Roman Klinger

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

The Big Idea: Emotions Are Like Action Plans

Imagine you are watching a play. Usually, when we analyze a character's feelings, we just put a label on them: "He is angry," or "She is sad." It's like sticking a name tag on a box.

But this paper argues that emotions are more like action plans. When you feel a specific emotion, your brain immediately comes up with a strategy to deal with the situation. The authors call these strategies "Coping Strategies."

They use a framework (based on the work of psychologist Ira Roseman) that says every emotional reaction falls into one of four buckets:

Attack: You want to fight the problem head-on. (Like a lion roaring at a threat).
Contact: You want to get closer and connect. (Like a puppy wagging its tail to say hello).
Distance: You want to back away and create space. (Like a turtle retreating into its shell).
Reject: You want to completely push the problem away or deny it exists. (Like slamming a door in someone's face).

The paper asks a simple question: Can we see these "action plans" hidden inside the words people write?

The Problem: You Can't Just Ask People

The researchers wanted to build a database (a corpus) of text where people clearly used these strategies. But there was a catch: you can't just go to Twitter or Facebook and find people saying, "I am currently using the 'Distance' strategy." Real life is messy, and people don't label their own coping mechanisms.

The Solution: The "Method Actor" Experiment

To solve this, the researchers didn't just ask people to write about their feelings. Instead, they used Role-Playing, similar to an improv theater class.

Think of it like this:

The Setup: They created four distinct "characters," each representing one of the four strategies.
- Character A (Attack) is a fiery, aggressive person who never backs down.
- Character B (Contact) is a calm, open-minded mediator.
- Character C (Distance) is someone who needs space and avoids drama.
- Character D (Reject) is a stubborn person who refuses to engage with negativity.
The Script: They used AI to write controversial scenarios (like arguments about immigration or racism) where one person says something upsetting.
The Performance: Real people were hired to "become" one of these characters. They had to read the scenario and write a reply as if they were that character.

It's like asking an actor, "You are playing a grumpy lion. Someone just stepped on your tail. Write a sentence reacting to it." The goal was to see if the actor could successfully channel the "lion's" strategy into their words.

What They Found

1. The Actors Got the Feelings Right
When the researchers asked the participants, "How does your character feel?" the answers matched the theory perfectly. When playing the "Attack" character, people felt angry and frustrated. When playing the "Contact" character, they felt hopeful and loving. The actors successfully understood the emotional "script."

2. The Text Was Hard to Decode
Here is the tricky part. Even though the actors understood the strategy, it was very hard for other people (and computers) to guess which strategy was being used just by reading the text.

When the researchers asked humans to read the replies and guess the strategy, they only got it right about 34% of the time (which is barely better than random guessing).
The "Contact" strategy was the easiest to spot (people used polite words like "I understand").
The "Reject" strategy was the hardest (it was often confused with other negative emotions).

3. Computers Struggled Too
The researchers trained computer models (AI) to do the same guessing game. The AI models also struggled, getting scores below 55% accuracy.

Interestingly, the AI did better when it read descriptions of the character's body language (e.g., "clenched fists," "looking away") than when it just read the words they spoke.
This suggests that while the words might be subtle, the behavior associated with the strategy is a stronger signal.

The Conclusion

The paper concludes that while emotions are deeply tied to how we plan to act (our coping strategies), these strategies are not always obvious in text.

The Good News: We can successfully create data about these strategies using role-playing.
The Bad News: It is much harder for computers (and humans) to spot these strategies in a sentence than we thought. The "action plan" is often hidden or subtle.

The authors say this opens up a new path for research: instead of just labeling "Anger," we should try to understand what the person is trying to do with that anger. But to do that, we need better tools to read between the lines.

Technical Summary: Dealing with Controversy: An Emotion and Coping Strategy Corpus Based on Role Playing

Problem Statement

There exists a significant mismatch between psychological and computational studies on emotions. While psychological research aims to explain the internal mechanisms of emotional phenomena, computational work in Natural Language Processing (NLP) often simplifies emotions into discrete labels or vectors defined by dimensions such as valence, arousal, and dominance. Consequently, fundamental aspects of how emotions develop and how individuals cope with them remain under-explored in NLP. Specifically, the complex mechanisms governing the emergence of emotions and their subsequent effect on behavior—conceptualized in psychology as "coping processes"—are rarely addressed. Current models struggle to capture the link between emotions and the behavioral strategies (e.g., attacking, contacting, distancing, or rejecting a stimulus) that people employ in response to emotion-eliciting events.

Methodology

To address this gap, the authors introduce COPING, the first corpus focused on coping information, constructed via a role-playing methodology. The study is grounded in Roseman's (2013) theory, which maps 16 specific emotions to four generic coping strategies: Attack, Contact, Distance, and Reject.

Data Creation

Given the difficulty of finding naturally occurring text annotated with coping strategies, the authors employed a crowdsourced role-playing approach:

Character Definitions: Four fictional characters (X) were defined, each embodying one of the four coping strategies. These definitions were refined through iterative pre-testing rounds to ensure annotators could correctly identify the intended strategy.
Scenario Generation: Using ChatGPT-4, the authors generated controversial scenarios across five socially relevant topics (abortion, immigration, racism, LGBTQ+ rights, and drugs). In each scenario, a fictional interlocutor (Y) makes a contentious statement threatening the identity of character X.
Role-Playing Task: Crowdworkers were assigned to impersonate character X. They were required to:
- Generate a verbal reply opposing Y's statement.
- Describe X's non-verbal behavior.
- Rate X's emotional responses based on Roseman's taxonomy.
- Self-reflect on whether they would react similarly to X.
Demographic Targeting: To elicit stronger and more authentic coping expressions, participants were selected based on specific demographics relevant to the topics (e.g., LGBTQ+ individuals for LGBTQ+ scenarios).

The resulting COPING corpus contains 1,200 data points, balanced across strategies, topics, and scenarios. Each entry includes the generated text, behavioral descriptions, and emotional ratings.

Experimental Setup

The paper investigates two primary Research Questions (RQs):

RQ1: Can coping strategies be elicited in language and annotated?
RQ2: Can coping strategies be identified automatically?

To answer RQ2, the authors evaluated the feasibility of automatic detection using:

Supervised Learning: Fine-tuning three Small Language Models (SLMs)—BERT, DistilBERT, and RoBERTa—on the corpus.
Zero-Shot Prompting: Utilizing four Large Language Models (LLMs)—phi-2, mistral-2, nous-hermes-2, and dolphin-2—in a zero-shot setting.
Input Variations: Models were tested on three input types: the verbal answer only, the behavioral description only, and a concatenation of both.

Key Contributions

Theoretical Integration: The paper takes an initial step in studying the four coping categories (Attack, Contact, Distance, Reject) that underlie standard emotion labels, linking them to specific behavioral functions and emotion groups as defined by Roseman (2013).
COPING Corpus: The release of a highly-structured corpus built with a novel role-playing strategy for collecting emotion-centered information. This addresses the scarcity of data annotated with coping mechanisms.
Task Benchmarking: The introduction and benchmarking of the coping strategy identification task. The study provides the first empirical evidence regarding the feasibility of detecting these strategies in text.

Results

Human Annotation and Understanding

Strategy Elicitation: Quantitative and qualitative analyses confirm that coping strategies can be elicited in language. Crowdworkers demonstrated a robust understanding of the strategies, with high accuracy in mapping character descriptions to the correct strategies (e.g., 100% accuracy for Attack and Contact in pre-testing).
Emotional Alignment: The emotional profiles generated by the role-players largely aligned with Roseman's theoretical framework (e.g., frustration correlated with Attack, hope with Contact). However, some deviations were noted, such as the difficulty in distinguishing "Reject" from other strategies.
Subjectivity: Inter-annotator agreement on the generated texts was low ( $\kappa = .34$ ), indicating that the perception of coping strategies in text is highly subjective. "Contact" was the most recognizable strategy, while "Reject" was the least.

Automatic Classification

Performance Limits: Automatic identification of coping strategies proved challenging. All models, both fine-tuned SLMs and zero-shot LLMs, achieved F1 scores below 0.55.
Input Sensitivity:
- DistilBERT performed best on verbal answers (F1 = 0.523).
- RoBERTa performed best on behavioral descriptions (F1 = 0.537).
- Combining answers and behaviors did not consistently improve performance over single inputs.
- Zero-shot LLMs performed on par with fine-tuned models, with behavior-based inputs generally yielding better results than answer-only inputs.
Strategy Difficulty: Consistent with human annotation, Contact was the easiest strategy to classify (F1 $\approx$ 0.64), while Reject remained the most difficult (F1 $\approx$ 0.41).

Significance and Claims

The authors claim that their work opens a promising research direction for enhancing the capability of models to capture emotion mechanisms from text. By shifting the focus from static emotion labels to dynamic coping strategies, the study bridges the gap between psychological theories of emotion-behavior interaction and computational NLP.

The paper modestly concludes that while the emergence of coping from text is less straightforward than initially hypothesized—particularly for strategies involving disengagement like "Distance" and "Reject"—the corpus and the observed links between strategies and emotions validate the potential of this approach. The authors emphasize that their setup, while artificial, successfully demonstrated that data on this psychological concept can be obtained via crowdsourcing and that these strategies are theoretically grounded in text, even if their automatic detection remains a significant challenge. This work suggests that integrating coping strategies could improve downstream tasks such as argumentation analysis and event-centered emotion analysis.