Semantic & syntactic context-aware text entry methods
Jun Gong · 2007 · Proceedings of the 9th International ACM SIGACCESS Conference on Computers and Accessibility (Assets '07) · doi:10.1145/1296843.1296903
Summary
Jun Gong's Assets '07 demonstration paper proposes a context-aware extension to the dictionary-based predictive disambiguation (DBPD) text-entry methods — the family that included the once-ubiquitous T9 on mobile phone keypads. DBPD methods let users press each multi-letter key just once and rely on a dictionary to resolve the ambiguous sequence into the intended word, which is attractive for users with motor or visual impairments who need to minimise keystrokes. The problem Gong addresses is that standard DBPD systems pick between ambiguous candidate words using raw word frequency alone, so users often have to press a 'next' key to cycle through alternatives when the most frequent word is not the one they meant. Gong's contribution is to combine three ranking signals for each candidate word: the traditional frequency score, a co-occurrence-based semantic relatedness score computed from the stems of preceding words in the current text, and a Part-of-Speech validity score obtained by running the Viterbi algorithm over likely POS tags of the prior words. The three signals are blended with tunable weights alpha, beta, gamma. A C++ implementation was evaluated in simulation against the British National Corpus across keypads ranging from 3 to 8 keys, and a PDA-based usability test with 16 participants measured real text-entry speed and error rate against a frequency-only baseline on a 3-key keypad.
Key findings
Simulation results showed the context-aware method improved both disambiguation accuracy (DA) and keystrokes-per-character (KSPC) at every keypad size tested, with the largest relative benefit on small keypads — on a 3-key layout, DA rose from 67.58% to 71.82% and KSPC dropped from 1.2124 to 1.1789, cutting about 15.8% of the extra keystrokes the user would otherwise have had to press to cycle through alternatives. As keypad size grew the gap narrowed because character ambiguity itself shrank. The sixteen-participant usability study on a 3-key PDA implementation found the context-aware method delivered 7.89 words-per-minute versus 7.01 for frequency-only (statistically significant at p = 0.011), and a small but non-significant reduction in error rate (0.076 vs 0.088 errors per word, p = 0.240). The author notes that the study was with non-disabled participants and that future work should test with actual motor- or visually-impaired users, which remains the paper's most obvious gap.
Relevance
The specific context — T9 on cellphone keypads — is largely historical, but the core idea (use semantic and syntactic context from the preceding text to rank candidate completions) has become completely mainstream: modern soft-keyboard autocorrect, GBoard and SwiftKey's language models, code-completion engines like Copilot, and AAC apps with predictive text all do a more sophisticated version of the same thing. For accessibility practitioners designing AAC boards, switch-scanning keyboards, gaze keyboards, or voice-recognition post-editing interfaces, the observation that contextual ranking yields the *largest* benefit precisely when the physical input bandwidth is most constrained (fewer keys, slower input) is directly useful — it argues for investing prediction-model effort proportional to how expensive each keystroke is for the user. Limitations in this paper are substantial: only two pages, no evaluation with actual disabled participants, dated language models (pre-neural), and the usability study is small enough that real-world generalisation is limited. It pairs well with modern work on next-word prediction for AAC and with literature on KSPC as a text-entry metric.
Tags: text entry · motor impairment · visual impairment · predictive text · T9 · natural language processing · disambiguation · mobile accessibility · alternative and augmentative communication · part-of-speech tagging