TL;DR
This paper introduces Functional Concept Graphs (FCGs) to combat "cognitive fixation" in innovation by representing functional relationships between problems and solutions. Their MUSE engine uses these graphs to generate creative inspirations, which a user study showed significantly increased participants' ability to generate novel and feasible ideas.
What Problem Does This Paper Address?
Innovators often get stuck on familiar solutions, a phenomenon known as "cognitive fixation," which hinders the exploration of truly novel ideas. This limits the effective navigation of the vast "design space" – the abstract realm of all possible solutions. Existing methods, like keyword searches, often fail to uncover deeper functional analogies crucial for creative leaps. The authors aim to provide a structured way to explore this design space and break free from fixation by representing the core functional essence of ideas.
Methods
The core of this work is the construction of Functional Concept Graphs (FCGs). These are directed graphs where nodes represent functional concepts (problems/purposes and solutions/mechanisms), and edges link them.
- FCG Construction:
- Data Extraction: They leverage patent data, extracting titles, abstracts, and Cooperative Patent Classification (CPC) tags.
- Problem & Solution Node Creation: * Mechanisms (Solutions): CPC tags are used to extract mechanism information, refined by a trained classifier.
- Purposes (Problems): GPT-3 is used for few-shot learning to extract purpose tags from patent descriptions.
- Clustering: Agglomerative clustering (with a two-stage K-means and then refined clustering) is used to group similar purpose tags into "problem nodes" and induce "solution nodes" based on co-occurrence with problem nodes.
- Edge Creation: * Problem-Solution: Edges connect a problem node to a solution node if a patent links them.
- Problem-Abstraction: A Natural Language Inference (NLI) model is used to identify hierarchical relationships between problem nodes (e.g., "protecting an organism" is an abstraction of "protecting plants").
- Enhancing Connectivity: Virtual "LLM-based" and "verb-based" nodes are introduced to capture more abstract or far-off analogies that the NLI model might miss.
- MUSE Engine: This algorithm takes an FCG and a target problem, encodes the problem, finds the closest node in the FCG, and then samples inspirations by traversing specific types of paths (e.g., "v-structure" paths involving abstraction and concretion steps). They use MMR (Maximal Marginal Relevance) to ensure diversity in sampled inspirations.
Datasets
- Patent Data: The primary dataset used is from Patentsview.org, specifically 500,000 patents sampled from the US patent corpus (originally ~3 million patents) across three top-level CPC categories (Human Necessities, Operations and Transport, Mechanical Engineering).
- No new benchmarks were introduced.
Key Insights and Ideas
- FCGs as a Scaffold: The core idea is that FCGs can act as a powerful "cognitive scaffold," enabling designers to reframe problems and discover analogies from seemingly unrelated domains.
- Abstraction is Key: Explicitly encoding abstraction relations between problem nodes is crucial for moving beyond simple semantic similarity and finding deeper functional analogies.
- LLMs for Abstraction: Leveraging LLMs (like GPT-3 and Llama3.1) is essential for both extracting purposes and creating more abstract connections within the FCG, overcoming limitations of previous, more manual or less scalable approaches.
- MUSE's Impact on Creativity: The user study demonstrates that providing structured inspirations derived from FCGs can significantly boost human creativity, particularly in generating novel and feasible solutions.
- Sentence-Level Inspirations are Powerful: Presenting inspirations as full sentences, rather than just purpose/mechanism pairs, yielded the best results in the user study, suggesting that context and narrative help users integrate the inspirations.
Key Results
- FCG Quality: The proposed method constructs richer, better-connected, and less noisy FCGs compared to prior work, with edges explicitly encoding abstraction relations.
- User Study - Creativity Boost:
- Participants using MUSE inspirations generated up to 19% more creative solutions in absolute numbers compared to those without inspirations.
- The ratio of creative solutions was significantly higher: 75% for participants receiving sentence-based inspirations versus 49% for those without.
- Inspiration-based conditions generally produced a higher ratio of feasible and creative ideas than the "empty" (no inspiration) condition.
- Trajectory Performance: LLM-based and NLI-based inspirations were superior to verb-based inspirations in producing feasible and creative solutions.
- Solution Generation Over Time: While participants without inspirations started faster, inspiration-based groups eventually produced more feasible and creative ideas, suggesting that inspiration integration takes time.
Limitations and Future Work
- Patent Data Bias: The reliance on US patents written in English might miss relevant products and ideas from different cultures or languages, potentially introducing bias.
- CPC Tag Reliance: Using CPC tags for mechanism extraction might require alternative methods when generalizing to new domains.
- User Study Limitations: The user study participants were native English speakers, and the tool's effectiveness for non-native speakers was not tested.
- Future Work:
- Explore more sophisticated sampling methods for inspirations, including "far-off" analogies.
- Investigate how to enhance creativity in state-of-the-art LLMs by enriching their training or inference with FCG-like representations.
- Further research into automatically finding structured representations for navigating design spaces and finding analogies in large repositories.
Summary updated