Mavdol/NPC-Valence-Arousal

NPC Valence Arousal Dataset

Dataset Summary

This dataset contains text-emotion annotations designed for training AI models to understand and predict emotional states of NPCs (Non-Player Characters) in video games. Each text entry is annotated with valence and arousal scores based on Russell's Circumplex Model of Affect.

The dataset is specifically optimized for gaming applications where NPCs need to interpret player inputs and respond with appropriate emotional reactions.

Dataset Visualization

🎮 Interactive Visualization: Explore the emotional distribution of this dataset using our interactive web tool: Valence-Arousal Visualizer

This visualization tool allows you to:

• Plot your data points on Russell's Circumplex Model
• Explore emotional clusters and distributions
• Understand the dataset's emotional coverage
• Validate annotation consistency across quadrants

Dataset Structure

Data Fields

• text (string): The input text (dialogue, action, or phrase)
• valence (float): Emotional valence score from -1.00 to +1.00
  • -1.00 = extremely negative (hatred, despair, rage)
  • 0.00 = neutral
  • +1.00 = extremely positive (joy, love, triumph)
• arousal (float): Emotional arousal/activation score from -1.00 to +1.00
  • -1.00 = extremely low energy (depression, lethargy)
  • 0.00 = neutral energy level
  • +1.00 = extremely high energy (excitement, panic, rage)

Theoretical Foundation

This dataset is based on Russell's Circumplex Model of Affect (Russell, 1980), which represents emotions in a two-dimensional space defined by valence (pleasure-displeasure) and arousal (activation-deactivation). This model is particularly suitable for NPC applications as it:

• Captures emotional nuances beyond discrete categories
• Enables smooth emotional transitions in character behavior
• Provides continuous values for fine-tuned NPC responses
• Maps well to gameplay mechanics (animations, dialogue choices, behavioral states)

Emotional Quadrants

• Q1: High Arousal + Positive Valence: Excitement, joy, triumph, euphoria
• Q2: High Arousal + Negative Valence: Fear, anger, panic, rage
• Q3: Low Arousal + Negative Valence: Sadness, depression, melancholy
• Q4: Low Arousal + Positive Valence: Contentment, peace, satisfaction

Data Sources

The dataset includes diverse text types representative of video game contexts:

• RPG Dialogues: Conversations from games like Skyrim, Baldur's Gate 3, Dragon Age, Cyberpunk
• Action Descriptors: Single words/phrases describing player or NPC actions
• Emotional Expressions: Exclamations, reactions, and emotional statements
• Interactive Phrases: Commands, requests, and social interactions

Text Variety

The dataset contains both:

• Complete sentences: Complex dialogue with contextual emotion
• Simple words/actions: Basic emotional triggers (e.g., "attacked", "helped")
• Various punctuation: Capturing intensity through formatting (!, ?, ...)
• Case variations: Different emotional intensities through capitalization

Annotation Process

1. Initial Manual Annotation: ~100 carefully balanced examples across all emotional quadrants
2. LLM-Assisted Expansion: Large Language Models used for consistent annotation scaling
3. Gaming Context Optimization: Emotions slightly amplified for clear NPC behavioral triggers
4. Quality Control: Validation against manual annotations and inter-annotator agreement checks

Intended Use

Primary Applications

• NPC Emotion Recognition: Training models to understand player input emotions
• Dynamic Dialogue Systems: Emotionally-aware conversation generation
• Behavioral State Machines: Triggering appropriate NPC reactions and animations
• Player Experience Enhancement: Creating more immersive and responsive game worlds

Model Compatibility

Optimized for lightweight models suitable for real-time gaming applications:

• DistilBERT
• MiniLM

Limitations

• Gaming Domain Focus: Optimized for video game contexts
• English Only: Currently limited to English text
• Amplified Emotions: Values slightly amplified for gaming clarity, may not reflect subtle real-world emotions
• Cultural Context: Annotations based on Western gaming culture and emotional expressions

Citation

@dataset{npc_valence_arousal_prediction,
  title={NPC Emotion Dataset: Valence and Arousal Annotations for Interactive Characters},
  author={Mavdol},
  year={2025},
  url={https://huggingface.co/datasets/your-username/video-game-npc-emotions},
  note={Based on Russell's Circumplex Model of Affect for NPC emotion recognition}
}