Dataset Overview

This directory contains four different datasets across three files:

• openai_english.json includes Dataset 1 and Dataset 2, which are interactions generated by OpenAI GPT-4o in explicit and implicit versions, respectively.
• openai_implicit_translated.json contains Dataset 3, which includes the implicit interactions from Dataset 2, translated into Italian by GPT-4o.
• deepseek_implicit_english.json contains Dataset 4, which features implicit interactions generated by DeepSeek.

Each dataset consists of user interactions categorized into nine different predicates representing various mental states. Each predicate is further divided into explicit and implicit levels, capturing different ways users express themselves. The datasets provide structured conversations between a child and a robot across different contexts.

---

Dataset Structure

The dataset are stored in 3 different json files with the following structure:

{
  "predicate": {
    "level": [
      {
        "id": integer,
        "context": "string",
        "phrases": [
          "string",
          "string",
          ...
        ]
      }
    ]
  }
}

Key Components:

1. predicate: Represents the emotional or behavioral state of the user. There are nine predicates in total:

   • Hard
   • Easy
   • Bored
   • Tired
   • Hungry
   • Succeed
   • Fussy
   • Curious
   • Uncomfortable

2. level: Specifies whether the interaction is explicit or implicit:

   • Explicit: The user directly states their feelings (e.g., "This is too hard!").
   • Implicit: The user indirectly expresses their feelings (e.g., "I’m not sure how to do this.").

3. id: A unique identifier for each interaction.

4. context: The scenario in which the interaction occurs (e.g., "Math Homework," "Playing a Game").

5. phrases: A list of dialogue exchanges between a child and a robot.

   • Each interaction contains 8 phrases, but they can be easily sliced into 2, 4, or 6 at a time for different analytical approaches.

---

Usage Notes

• The dataset can be filtered by predicate, level, or context to extract specific interactions.
• The phrase sets can be sliced into smaller groups (e.g., 2, 4, or 6 phrases) to analyze different interaction dynamics.
• The dataset can be converted into different formats (e.g., CSV, structured text) for further analysis.

---

This dataset provides a structured and versatile resource for understanding human-computer interactions and emotional states in various learning and engagement scenarios.