Directed Diversity: Leveraging Language Embedding Distances for Collective Creativity in Crowd Ideation

Title of the Paper

Directed Diversity: Leveraging Language Embedding Distances for Collective Creativity in Crowd Ideation

Paper Information

• Research Domain: Human-Computer Interaction, Collective Innovation, Natural Language Processing
• Keywords: Diversity, Collective Creativity, Crowdsourcing, Creative Ideas, Incentive Information, Collective Intelligence, Creativity Support Tools

Research Background and Problem

• What problems or challenges did the authors identify?

  • Crowdsourcing tasks often result in numerous redundant ideas during the process of collecting creativity from individual participants.
  • Current methods to reduce redundancy (e.g., displaying peer ideas or peer-generated prompts) require significant manual coordination, making them inefficient and lacking scalability.

• Why is this problem important?

  • The goal of crowdsourcing tasks is to efficiently collect highly diverse creative ideas to support practical applications in areas such as design, decision-making, and behavioral motivation.
  • Helping groups generate more novel ideas within limited timeframes holds broad practical significance.

• Research Motivation and Related Work:

  • Existing studies emphasize that diversity fosters collective creativity, but there is a lack of automated methods to support scalable diversity prompting.
  • Artificial intelligence (e.g., language models) enhances the potential for processing natural language and modeling diversity, with proven effectiveness in fields like recommendation systems and ecology.

Solution

• Method or Solution: The authors propose a method called "Directed Diversity," which leverages language model embedding distances to automatically select the most diverse prompts, guiding participants to generate novel ideas.

• What is innovative about this solution?

  • Utilizes pre-trained language models (e.g., Universal Sentence Encoder) embeddings to quantify prompt diversity.
  • Introduces a new Diversity Prompting Evaluation Framework, combining diversity metrics from various disciplines to systematically analyze the propagation of creativity through diverse prompts.
  • Employs a Minimum Spanning Tree (MST) approach to select prompts that maximize semantic distance, balancing novelty and extensibility of content.

• Implementation Steps and Key Technologies:

  1. Prompt Extraction: Extract phrases from domain-relevant documents, followed by cleaning and syntactic analysis.
  2. Phrase Embedding: Obtain vector representations of phrases using the Universal Sentence Encoder.
  3. Phrase Selection: Apply the Minimum Spanning Tree algorithm to select semantically distant phrases, maximizing diversity.
  4. Evaluation Framework: Develop a set of diversity and creativity metrics, and validate the framework's effectiveness through experiments.

Research Outcomes

• What specific outcomes were achieved?

  • The proposed Directed Diversity method effectively increased prompt diversity and reduced redundancy.
  • Prompt diversity directly enhanced the content and format diversity of ideas generated by participants.
  • Experiments demonstrated that ideas collected using this method were not only more flexible and original but also covered a broader range of topics.

• What advantages does it have compared to existing solutions?

  • Unlike manual methods, this solution is fully automated, significantly scaling up the scope and efficiency of creative idea generation.
  • More sensitively detects and quantifies the process of diversity propagation while substantially reducing the generation of repetitive ideas.

• What were the experimental or evaluation results?

  • Across multiple experimental evaluations, the Directed Diversity prompting method achieved significantly higher creativity diversity scores compared to random selection or baseline methods without prompts.
  • After implementing prompt diversity, participants' ideas scored higher in flexibility and originality.

• Limitations and Future Directions:

  • Limitations:

    • Phrases selected using Directed Diversity may sometimes be difficult to understand, requiring greater effort from participants.
    • The method is currently focused on text-based tasks, and its adaptability to other types of creative generation (e.g., images, designs) needs further validation.

  • Future Directions:

    • Enhance the quality and semantic relevance of phrase prompts using domain-specific embedding models.
    • Extend the mechanism to creative generation tasks beyond text, such as product design or concept modeling.
    • Explore ways to reduce participants' cognitive load while maintaining the diversity and effectiveness of prompts.