PhraseFlow: Designs and Empirical Studies of Phrase-Level Input

Title of the Paper

PhraseFlow: Designs and Empirical Studies of Phrase-Level Input

Paper Information

• Domain: Human-Computer Interaction, Text Input Technology
• Keywords: Text input, autocorrection, phrase-level input, keyboard, visual feedback, user study

Research Background and Problem

• Identified Problems or Challenges:
  • Current autocorrection technologies primarily rely on word-level decoding, which struggles with error correction and spatial error handling.
  • In word-level decoding, keyboards cannot leverage contextual information from subsequent input to improve correction accuracy.
• Significance:
  • Enhancing text input accuracy and user experience on mobile keyboards is critical for improving interaction efficiency on mobile devices.
  • Phrase-level input has the potential to address the limitations of word-level decoding.
• Motivation and Related Work:
  • Existing studies (e.g., VelociTap) suggest that phrase-level decoding may lead to higher text input accuracy, but there is limited research on user behavior and interaction design.
  • Additionally, phrase-level input involves cognitive load issues, necessitating optimized design.

Solution

• Proposed Approach:
  • Developed a keyboard prototype called PhraseFlow, which supports phrase-level decoding-based input.
  • The keyboard utilizes contextual information from subsequent input to provide more accurate corrections for previously entered text.
• Innovations:
  • Buffer Commit Method: Gradually commits correction results after multi-word phrases are entered, instead of committing all at once.
  • Real-Time Correction Feedback: Provides real-time text correction results as users continue typing, reducing attention disruption caused by uncertainty over uncorrected text.
  • Optimized Visual Feedback Design: Uses underlining to mark active text and flashing cues to highlight correction changes, meeting users' cognitive needs for system status awareness.
• Implementation Steps and Key Techniques:
  1. Modified Gboard's existing Finite State Transducer (FST) decoder to support phrase-level decoding.
  2. Applied various design schemes to optimize commit methods, suggestion display, and correction effects.
  3. Conducted multiple simulations and experiments to validate the impact of the improved design on correction accuracy and user experience.

Research Outcomes

• Specific Results:
  • Testing showed that PhraseFlow's phrase-level decoding improved Word Error Rate (WER) by 16.6% compared to word-level decoding.
  • The Buffer Commit Method significantly reduced users' cognitive load, enabling PhraseFlow to achieve input speed and accuracy comparable to commercial keyboards.
• Comparison with Existing Solutions:
  • PhraseFlow demonstrated significant advantages in handling spacebar-induced errors and leveraging subsequent context for error correction.
  • Compared to existing word-level decoders, PhraseFlow resulted in fewer correction errors and higher user acceptance.
• Experiment and Evaluation Results:
  • Experiment 1 (Design Optimization): Identified effective visual feedback using background flashing and underlining.
  • Experiment 2 (Simulation Testing): Confirmed that the phrase-level decoding method reduces error rates.
  • Experiment 3 (Lab Testing V1): Found that the initial design imposed high cognitive load, requiring optimization.
  • Experiment 4 (Lab Testing V2): The V2 prototype with buffer commit and real-time correction feedback achieved speed and error rates comparable to Gboard, reducing error rates by 19%.
  • Experiment 5 (Real-World Usage): During six days of use by 42 participants, approximately 78.6% expressed willingness to use phrase-level input in future keyboards, while only 7.1% expressed dislike.
• Limitations and Future Directions:
  • Limitations:
    • The current language model is relatively small and does not support multilingual or gesture input.
    • Some users lack trust in corrections, and the error reversal process is complex.
  • Future Directions:
    • Develop larger-scale, deep learning-based language models.
    • Optimize correction methods for more efficient error handling.
    • Explore the applicability of phrase-level gesture input and multilingual environments.