AdaptiFont: Increasing Individuals' Reading Speed with a Generative Font Model and Bayesian Optimization

Title of the Paper

AdaptiFont: Increasing Individuals’ Reading Speed with a Generative Font Model and Bayesian Optimization

Paper Information

• Domain: Human-Computer Interaction, Font Generation and Optimization, Reading Speed Enhancement
• Keywords: Reading Speed, Adaptive Fonts, Generative Model, Bayesian Optimization, Font Design

Research Background and Problem

• Identified Problem or Challenge: Digital text has become the primary medium for knowledge exchange, but existing research on the relationship between fonts and reading speed or comprehension yields inconsistent and contradictory results. Solutions for individualized font optimization to enhance reading speed remain underexplored.
• Significance: Improving the readability of text on digital screens can enhance information processing efficiency, especially in an era of information overload. Optimizing fonts for individuals can address diverse user needs in terms of visual perception, cognitive processing, and device display conditions.
• Motivation and Related Work:
  • Traditional font design focuses on aesthetics and artistry, but its functional optimization has not been sufficiently data-driven.
  • Font characteristics and reading readability have been widely studied, but individual differences have received insufficient attention.
  • The proliferation of digital devices has driven user interfaces toward dynamic and customizable designs. While previous studies have explored parametric fonts and adaptive rendering technologies, closed-loop systems focusing on personalized optimization are scarce.

Solution

• Proposed Method or Solution:
  • Develop a closed-loop human-computer interaction system, “AdaptiFont,” which dynamically generates and optimizes fonts to enhance individual reading speed using a generative font space model and Bayesian optimization algorithm.
  • Employ Non-Negative Matrix Factorization (NMF) to learn a continuous font space from 25 classic fonts, enabling the generation of new fonts and iterative optimization of reading speed.
  • Use Bayesian optimization to efficiently explore the font space, forming a closed-loop interaction between font generation and user feedback.
• Innovations:
  • Generative Model Innovation: Use NMF to create a continuous font space and enable real-time vector and bitmap conversion during font generation.
  • Optimization Mechanism Innovation: Apply Bayesian optimization to balance exploration and exploitation, dynamically adjusting font generation based on user reading speed.
  • Experimental Design Innovation: Guide participants through word classification tasks during reading, evaluating reading speed and comprehension using both subjective and objective metrics.
• Implementation Steps and Key Techniques:
  1. Construct font space: Extract three-dimensional feature vectors from classic fonts using NMF.
  2. Dynamic font generation: Generate new font vectors through linear combinations and produce TrueType files.
  3. Bayesian optimization: Establish a Gaussian process to optimize the individual reading speed function within the font space.
  4. User experiments: Validate the effectiveness of the optimization system through reading tasks involving 95 texts.

Research Outcomes

• Specific Results:
  • Experiments demonstrated that AdaptiFont identifies regions in the font space associated with high reading speeds and significantly improves participants’ reading speeds.
  • Significant individual differences in font preferences were observed, underscoring the importance of personalized font design.
• Advantages Compared to Existing Solutions:
  • Compared to traditional fonts, AdaptiFont-generated fonts significantly enhance reading speed, maximizing interaction efficiency between text and users.
  • Dynamic font generation overcomes the limitations of predefined fonts in classical design and integrates personalized adaptability to font parameters.
• Experimental or Evaluation Results:
  • Average reading speed in AdaptiFont experiments was significantly higher than in traditional font experiments, with Kolmogorov-Smirnov tests confirming statistical significance.
  • Bayesian clustering algorithms identified font clusters associated with high reading speeds, validating the existence of locally efficient regions within the font space.
• Limitations and Future Directions:
  • Limitations:
    1. Did not incorporate other optimization objectives such as text comprehension or memory.
    2. Did not consider the dynamic relationship between reading speed and contextual or device display variations.
    3. Fonts generated using NMF may violate typographic rules, requiring further improvements to constrain practical font design.
  • Future Directions:
    1. Explore more complex objective functions, such as aesthetic satisfaction.
    2. Apply Generative Adversarial Networks (GANs) to improve font generation quality.
    3. Further investigate the long-term adaptability of optimized fonts in diverse contexts.

Code and Sharing

The research team has open-sourced the code and related resources, including generated TrueType font files and scripts for NMF components. Link: AdaptiFont GitHub