LongEmotion: Measuring Emotional Intelligence of LLMs in Long-Context Interaction

---

Dataset Description

LongEmotion is a comprehensive benchmark designed to evaluate the Emotional Intelligence (EI) of Large Language Models (LLMs) in long-context scenarios. It includes six carefully constructed tasks that test emotion recognition, psychological knowledge application, and empathetic generation — areas crucial for emotionally coherent and human-aligned AI systems.

Key Features

• 🎯 Long-Context Evaluation: Average context length exceeds 15,000 tokens, with the longest reaching 43,588 tokens
• 📊 Comprehensive Coverage: 1,156+ samples across 6 task types covering 3 EI dimensions
• 🧠 Multi-Faceted Tasks: Emotion recognition, knowledge application, and empathetic generation
• 🌐 Real-World Data: Sourced from psychological counseling cases, academic literature, and financial documents

Dataset Statistics

Task	Type	Samples	Avg Length	Metric
EC-Emobench	Classification	200	19,345 tokens	Accuracy
EC-Finentity	Classification	200	43,588 tokens	Accuracy
Emotion Detection	Detection	136	4,592 tokens	Accuracy
Emotion QA	QA	120	-	F1 Score
Emotion Conversation	Dialogue	100 (400 turns)	-	LLM-Judge
Emotion Summary	Summarization	150	-	LLM-Judge
Emotion Expression	Generation	8 types + 1 questionnaire	-	LLM-Judge

Dataset Structure

Data Fields

Emotion Classification (Emobench)

• id: Sample identifier
• content: Long-form text content
• subject: Target entity for emotion classification
• label: Emotion label (84 emotion categories including compound emotions)
• source: Source of the text
• choices: List of emotion choices
• length: Token length of the content

Emotion Classification (Finentity)

• id: Sample identifier
• content: Long-form financial text
• subject: Target entity
• label: Sentiment label (Positive/Neutral/Negative)
• source: Source document
• token_length: Token length
• choices: Sentiment choices

Emotion Detection

• text: Text segments for comparison
• label: Labels for each segment
• length: Token length
• ground_truth: Correct answer indicating the different segment

Emotion QA

• number: Question identifier
• problem: Question text
• answer: Answer text
• source: Source academic paper
• context: Long-form context from psychology literature

Emotion Conversation

• id: Conversation identifier
• stages: List of dialogue stages
• description: Scenario description

Emotion Summary

• id: Case identifier
• case_description: Case description
• consultation_process: Consultation process
• experience_and_reflection: Therapist's reflection
• causes: Identified causes
• symptoms: Symptoms
• treatment_process: Treatment process
• treatment_effect: Treatment outcomes

Emotion Expression

• Situations: Emotion types with associated scenarios
• Questionnaires: Standardized psychological questionnaires (e.g., PANAS)

Usage

Loading the Dataset

from datasets import load_dataset

# Note: Due to schema inconsistencies, direct loading may fail
# Recommended approach: Download and load manually

from huggingface_hub import snapshot_download
import json

# Download the dataset
local_dir = snapshot_download(
    repo_id="LongEmotion/LongEmotion",
    repo_type="dataset",
    local_dir="./LongEmotion_data"
)

# Load individual task data
def load_jsonl(file_path):
    with open(file_path, 'r', encoding='utf-8') as f:
        return [json.loads(line) for line in f]

# Example: Load Emotion Classification data
ec_data = load_jsonl('./LongEmotion_data/Emotion Classification/Emotion_Classification_Emobench.jsonl')
print(f"Loaded {len(ec_data)} samples")

# Example: Access a sample
sample = ec_data[0]
print(f"Subject: {sample['subject']}")
print(f"Label: {sample['label']}")
print(f"Content length: {sample['length']} tokens")

Example Usage for Different Tasks

1. Emotion Classification

import json

# Load Emobench data
with open('./LongEmotion_data/Emotion Classification/Emotion_Classification_Emobench.jsonl', 'r') as f:
    data = [json.loads(line) for line in f]

# Process a sample
sample = data[0]
prompt = f"""
Given the following text, identify the emotion of {sample['subject']}.

Text: {sample['content']}

Choices: {', '.join(sample['choices'])}

Answer:"""

# Use your LLM to generate response
# response = your_llm(prompt)

2. Emotion QA

# Load QA data
with open('./LongEmotion_data/Emotion QA/Emotion_QA.jsonl', 'r') as f:
    qa_data = [json.loads(line) for line in f]

sample = qa_data[0]
prompt = f"""
Context: {sample['context']}

Question: {sample['problem']}

Answer:"""

# Evaluate with F1 score against sample['answer']

3. Emotion Conversation

# Load conversation data
with open('./LongEmotion_data/Emotion Conversation/Emotion_Conversations.jsonl', 'r') as f:
    conv_data = [json.loads(line) for line in f]

sample = conv_data[0]
# Multi-turn dialogue simulation
for stage in sample['stages']:
    print(f"Stage: {stage['stage']}")
    # Generate empathetic response

Evaluation Methods

LongEmotion supports multiple evaluation approaches:

1. Baseline: Direct processing of full text
2. RAG: Retrieval-Augmented Generation
3. CoEM: Collaborative Emotional Modeling (multi-agent RAG)
4. Self-RAG: Adaptive retrieval
5. Search-O1: Search-based optimization

CoEM Framework

The Collaborative Emotional Modeling (CoEM) framework integrates RAG with multi-agent emotional reasoning:

Input → Chunking → Initial Retrieval → Multi-Agent Enrichment → Re-Ranking → Generation → Output

Citation

If you use this dataset in your research, please cite:

@article{liu2025longemotion,
  title={LongEmotion: Measuring Emotional Intelligence of Large Language Models in Long-Context Interaction},
  author={Liu, Weichu and Xiong, Jing and Hu, Yuxuan and Li, Zixuan and Tan, Minghuan and Mao, Ningning and Zhao, Chenyang and Wan, Zhongwei and Tao, Chaofan and Xu, Wendong and others},
  journal={arXiv preprint arXiv:2509.07403},
  year={2025}
}

Paper

📄 ArXiv: https://arxiv.org/abs/2509.07403

Dataset Creation

Source Data

• Emotion Classification (Emobench): BookCorpus novels with fine-grained emotion annotations
• Emotion Classification (Finentity): Financial documents with entity-level sentiment
• Emotion Detection: Mixed sources for emotion anomaly detection
• Emotion QA: 30 academic papers on psychology and mental health
• Emotion Conversation: Simulated psychological counseling dialogues
• Emotion Summary: Real-world psychological counseling case reports
• Emotion Expression: Emotion generation scenarios and standardized questionnaires

Annotation Process

The dataset combines automated extraction, expert annotation, and quality validation to ensure high-quality emotion labels and psychological accuracy.

Considerations for Using the Data

Social Impact

This dataset is designed to advance research in:

• Emotionally intelligent AI systems
• Mental health support applications
• Human-AI interaction in sensitive contexts

Limitations

• English-only content
• Potential cultural biases in emotion categorization
• Long-context processing may require significant computational resources

Additional Information

Licensing

Please refer to the repository for licensing information.

Contact

For questions or feedback, please open an issue on the dataset repository or contact the authors through the paper.

Acknowledgments

We thank all researchers and data providers who contributed to building the LongEmotion dataset.

---

Dataset Version: 1.0
Last Updated: 2026-01-17