Hex: Dynamics and Probabilistic Text Entry

John Williamson, Roderick Murray-SmithEuropean Summer School on Multi-AgentControl 2003 · DOI

TL;DR

This paper introduces "Hex," a novel gestural text entry system for mobile devices that uses continuous movements on a hexagonal grid. By dynamically adjusting the "feel" of the interface based on probabilistic language models, Hex aims to improve text entry speed and accuracy, making the system more intuitive and less effortful for users.

What Problem Does This Paper Address?

Entering text on mobile devices is notoriously challenging due to limited screen real estate and processing power. Existing methods like virtual keyboards, handwriting recognition, and gesture-based interfaces have limitations. While some gesture-based systems use dynamic layouts and probabilistic models, they can hinder user learning by constantly changing the interface. This paper addresses the need for a gestural text entry system that leverages probabilistic language models to enhance performance without sacrificing the stability crucial for user learning and developing automatic movements.

Methods

The core of the "Hex" system is a hexagonal tessellation of the input space. Users enter text by making continuous gestures across this grid. The system employs a probabilistic language model to predict likely sequences of characters. This model dynamically influences the "handling qualities" of the interface by simulating a nonlinear force-field. This force-field guides the user towards probable text sequences, making it easier to enter them, while increasing the effort required for improbable ones. The system uses a time-varying probability model to adapt these forces based on the current context, including cursor position, velocity, and acceleration. For layout optimization, the paper proposes a method to minimize a cost function based on user movement models, aiming to create more efficient layouts.

Datasets

The paper mentions using a corpus from Project Gutenberg for training the language model. No new benchmarks were introduced.

Key Insights and Ideas

  • Probabilistic Force Fields: The idea of using a probabilistic language model to dynamically alter the physical "feel" of an interface is a powerful concept. This creates a system that actively guides the user towards likely inputs, reducing cognitive load and physical effort.
  • Stability for Learning: By maintaining a stable hexagonal layout, the system supports user learning and the development of automatic, high-speed movements, which is a significant advantage over purely dynamic layouts.
  • Information Content and Effort: The system establishes a direct correspondence between the information content of a text sequence and the effort required to enter it. More probable (lower information content) sequences are easier to input.
  • Explicit Uncertainty: The system makes uncertainty explicit to the user through its dynamic feedback, which can be a valuable design principle for many gestural interfaces.

Key Results

Experimental results demonstrate that the Hex system reduces the variance in gesture trajectories for probable sequences. While specific quantitative results for throughput are mentioned as being around 10-12 words per minute for earlier versions (with 30 hours of user experience), current versions achieve around 17 wpm for free-form text entry. The paper highlights that the layout used in these tests was not optimized. Figure 5 shows that applying forces narrows the distribution of trajectories, indicating improved accuracy.

Limitations and Future Work

The authors acknowledge that further systematic user trials are needed to fully establish the effects of the system at various stages of learning. They also suggest that more complex language models could be incorporated. Future work includes extending the feedback to an audio display, which is particularly important for mobile devices where screen space is limited and visual load can be high. The authors also note that while they used a simple language model, incorporating grammar or word-level models could further improve performance.

Summary updated

BibTeX
@inproceedings{williamson2003hex,
  author = {John Williamson and Roderick Murray-Smith},
  title = {{Hex: Dynamics and Probabilistic Text Entry}},
  booktitle = {European Summer School on Multi-AgentControl},
  year = {2003},
  doi = {10.1007/978-3-540-30560-6\_15},
}
Figure