PopBlends: Strategies for Conceptual Blending with Large Language Models

Title of the Paper

PopBlends: Strategies for Conceptual Blending with Large Language Models

Paper Information

• Subject Area: Creativity Support Systems, Natural Language Processing, Applications of Language Models
• Keywords: Creativity Support Tools, Applications of Large Language Models, Conceptual Blending, Human-Computer Interaction, Artificial Intelligence

Research Background and Problem

• What problems or challenges did the authors identify?

  • On social media, "conceptual blending," which combines products or services with popular culture, is an engaging way to spread creative ideas. This requires identifying connections between two domains, posing a cognitive challenge.
  • Traditional conceptual blending demands significant user effort, involving extensive searches for associative information and identifying inspiration from a vast combination of data.
  • While computers can perform large-scale searches, they often lack cultural knowledge and commonsense reasoning capabilities.

• Why is this problem important?

  • Achieving efficient "conceptual blending" can enhance the creative quality of social media content and support amateur creators in generating content related to brand promotion.
  • Effectively combining popular culture with product domains can help brands establish deeper emotional connections with consumers.

• Research Motivation and Related Work

  • Conceptual blending has been proven to be important in the creative design process, but achieving this creativity remains a cognitive and technical challenge.
  • Previous research and tools focused on using knowledge graphs, databases, and small-scale natural language processing methods for blending, but these approaches often lack comprehensive coverage or cultural knowledge.
  • The emergence of large-scale pre-trained language models (e.g., GPT-3) has the potential to transform technology applications in creative design support.

Solution

• What methods or solutions did the authors propose?

  • The authors proposed a system called PopBlends, which utilizes a two-phase divergent and convergent process to match conceptual blends between input products or services and popular culture.
  • They employed three strategies (No-GPT, Half-GPT, Full-GPT) that combine traditional natural language processing techniques with large language models in a step-by-step process.

• What are the innovative aspects of this solution?

  • The system integrates traditional knowledge processing methods with the associative capabilities of modern large language models.
  • It explores varying levels of dependency on GPT-3 to generate "connecting concepts," enabling diversity in the generated results.
  • The system provides automated retrieval of popular culture and product-related scene images, supporting users in quickly generating blended content.

• What are the implementation steps? What key technologies were used?

  1. Step 1: Define the input (popular culture domain and product/service).
  2. Step 2: Use three strategies—No-GPT, Half-GPT, and Full-GPT—to identify connecting concepts:
     • No-GPT: Use traditional natural language processing methods to extract related concepts from Wikipedia plot summaries of popular culture.
     • Half-GPT: After extracting popular culture entities, use GPT-3 to expand these entities with activities, adjectives, and iconic phrases.
     • Full-GPT: Directly use GPT-3 to generate associative vocabulary between popular culture and the product.
  3. Step 3: Retrieve corresponding scenes and images based on the connecting concepts.
     • Popular culture scenes are retrieved based on Wikipedia plot summaries.
     • Product scenes are generated using GPT-3 to describe related activities and scenes, which are then matched.

Research Outcomes

• What specific results were achieved?

  • Experiments showed that PopBlends can generate connecting concepts and efficiently support users in generating creative ideas.
  • Users of the system generated twice as many blended ideas compared to relying solely on internet searches, while reducing cognitive load by 50%.
  • The accuracy of the three strategies was similar, but the styles of the generated connecting concepts varied. All strategies produced at least one valid connecting concept in 90% of cases.

• What advantages does it have compared to existing solutions?

  • The system's structure clearly supports the divergent and convergent process.
  • By combining GPT-3 with knowledge databases, it achieves broader coverage while providing nuanced cultural associations.
  • Designed for amateur creators, the system significantly lowers the barrier to generating creative ideas.

• What were the experimental or evaluation results?

  • Annotation studies showed that 56.7% of the generated connecting concepts were both relevant to popular culture and the product.
  • User studies indicated that compared to traditional internet searches, PopBlends was significantly more effective in helping users generate creative ideas.

• Limitations and Future Directions

  • Limitations:

    • The current system focuses only on the popular culture domain of movies and TV shows, and cannot directly apply to music, toys, or other fields.
    • GPT-3-generated results still face the risk of "hallucinations" (e.g., incorrect scenes or information).
    • The system does not fully automate the generation of blended images, requiring users to manually complete the final image design.

  • Future Directions:

    • Explore how to incorporate other popular culture domains, such as songs and political events, into PopBlends.
    • Optimize GPT-3 prompt engineering to reduce errors or irrelevant content generation.
    • Test image generation technologies like DALL-E and integrate them into the system to achieve fully automated blended designs.
    • Release the tool to a broader user base to evaluate its effectiveness and requirements in real-world applications.

The research on PopBlends demonstrates the potential of combining large language models with knowledge graphs for creativity support tools, providing valuable insights into system design, technology, and user evaluation.