Simulating Word Suggestion Usage in Mobile Typing to Guide Intelligent Text Entry Design

TL;DR

This work tackles the problem of understanding how users integrate word suggestions into mobile typing, addressing the scarcity and cost of long-term user studies. It introduces WSTypist, a reinforcement-learning-based simulator built on hierarchical control and computational rationality to model high-level decision-making about suggestion usage, incorporating efficiency, linguistic uncertainty, and personal reliance. The model is trained and evaluated on gaze- and keystroke-informed data, reproducing human-like behaviors, individual differences, and generalizing across systems, including no-ITE and auto-correction scenarios. Four design-case studies demonstrate how what-if analyses with WSTypist can guide interface and algorithmic choices, revealing actionable insights such as optimal accuracy ranges, longer-word prioritization, capitalization-aware strategies, and inline suggestion placements, all while open-sourcing the tool for broader use.

Abstract

Intelligent text entry (ITE) methods, such as word suggestions, are widely used in mobile typing, yet improving ITE systems is challenging because the cognitive mechanisms behind suggestion use remain poorly understood, and evaluating new systems often requires long-term user studies to account for behavioral adaptation. We present WSTypist, a reinforcement learning-based model that simulates how typists integrate word suggestions into typing. It builds on recent hierarchical control models of typing, but focuses on the cognitive mechanisms that underlie the high-level decision-making for effectively integrating word suggestions into manual typing: assessing efficiency gains, considering orthographic uncertainties, and including personal reliance on AI support. Our evaluations show that WSTypist simulates diverse human-like suggestion-use strategies, reproduces individual differences, and generalizes across different systems. Importantly, we demonstrate on four design cases how computational rationality models can be used to inform what-if analyses during the design process, by simulating how users might adapt to changes in the UI or in the algorithmic support, reducing the need for long-term user studies.

Figures (5)

• Figure 1: Left: The mechanisms underlying word suggestion usage during typing and the corresponding components in our model. Right: Overall structure of the WSTypist model. Similar to prior work shi2024crtypist, we use a hierarchical architecture where the supervisor agent learns to control Internal and External environment components, following computational rationality theory. We extend this architecture to simulate suggestion usage in mobile typing. In addition to including gaze and finger actions related to the suggestion list, we incorporate critical cognitive mechanisms that enable the agent to assess the utility of word suggestions. This includes assessing efficiency gains by integrating Completeness in working memory as a primary factor for suggestion usage, the influence of linguistic properties such as word length and frequency ($W_L$, $W_F$), as well as linguistic knowledge of the agent ($P_K$) affecting the error probability, and personal suggestion reliance ($P_S$) affecting the final reward.
• Figure 2: An example of how the model’s key component states transition while typing the word "return". Marked in red are the components that are newly introduced in this paper. The green rounded rectangle represents the gaze action (i.e. which UI element is looked at), and the blue rounded rectangle represents the finger action (i.e. typing the next character or selecting a suggestion). In this example, we set $P_M=0.1$, which is used to update Certainty, together with whether the agent performs proofreading. The longer the agent types without proofreading, the lower the Certainty becomes. The Completeness is computed from Certainty and the current input length (as a percentage of the target word) (a-c). When the gaze shifts to the suggestion list, Sugg becomes Yes (c) or No (b), depending on whether the target word appears there; otherwise, it remains Unknown (a). If the gaze falls on the suggestion list and the target word is detected, orthographic errors are resolved, leaving only the basic error component (c). If the target word is eventually selected, the Picking reward is Yes and is added to the final reward (d).
• Figure 3: Qualitative comparisons of typing strategies between human (observed in empirical data 2025suggestion) and the model, illustrating representative cases of word Completion, error Correction, Capitalization handling, and Select-and-Modify behavior.
• Figure 4: Effects of target word length on suggestion usage by the agent, revealing a peak in selection rates for medium-length words and a decline for longer words. This trend is attributed to reduced algorithmic accuracy for longer suggestions and closely parallels patterns observed in human behavior.
• Figure 5: Influence of suggestion accuracy on user behavior, including effects on Picked, Failed suggestions, Keystroke Savings, Gaze ratio on Suggestion List, and overall typing speed in WPM across different simulated accuracy levels. The shaded region highlights high-accuracy systems (>80%), the dot marks oracle-level performance (100%), and the dashed line projects a hypothetical trajectory toward optimal interaction.