Sibylle, an assistive communication system adapting to the context and its user
Tonio Wandmacher, Jean-Yves Antoine, Franck Poirier, Jean-Paul Départe · 2008 · ACM Transactions on Accessible Computing · doi:10.1145/1361203.1361209
Summary
Sibylle is an augmentative and alternative communication (AAC) system designed for people with severe motor disabilities who cannot use conventional keyboards. The system has been in clinical use at the Kerpape rehabilitation center in Brittany, France since 2001, serving patients with conditions like locked-in syndrome, cerebral palsy, and various motor impairments. The paper focuses on the word prediction component (SibyWord), which uses a 4-gram language model with modified Kneser-Ney discounting built from newspaper corpora. The vocabulary encompasses approximately 140,000 words for each supported language (French, German, and English). What distinguishes Sibylle from other AAC systems is its dual adaptation approach: a Dynamic User Model (DUM) that learns each user's vocabulary preferences over time, and a semantic adaptation component using Latent Semantic Analysis (LSA) that adjusts predictions based on the current topic of conversation. The system also includes a letter prediction module (SibyLetter) using 5-gram models that dynamically reorganizes the virtual keyboard to minimize scanning time for single-switch users, reducing average key selection from 9 shifts to approximately 2.7-3.0 shifts.
Key findings
The baseline 4-gram model achieves keystroke saving rates (ksr) of 51.6-57.8% across languages. The Dynamic User Model provides gains of up to 9.4%, reaching a plateau after approximately 20,000 words of user input (3-6 hours of typing). LSA-based semantic adaptation adds another 1.0-1.7% improvement—five times better than traditional cache models. For German text, a partial selection method handles compound words effectively, adding 0.8-1.5% improvement. When all adaptation strategies are combined, the gains are nearly additive, with speech corpora showing the largest improvements (over 9%). Real-world deployment with more than 20 patients at Kerpape demonstrated significant benefits: teachers observed accelerated text entry, children accepted longer working sessions, and users showed reduced physical fatigue compared to previous AAC systems. Interestingly, practitioners noted decreased orthographic and grammatical errors among users—attributed to the cognitive scaffolding provided by seeing correctly spelled predictions.
Relevance
This paper demonstrates the substantial real-world impact of adaptive language technology in AAC. The seven years of clinical deployment provide rare longitudinal evidence that word prediction can meaningfully reduce fatigue and accelerate communication for motor-impaired users. The dual adaptation approach—combining long-term user learning with short-term semantic context—offers a model for personalization that remains relevant for modern AAC and assistive technology design. The observation that users often fail to select predicted words despite availability highlights an important cognitive load challenge that current AAC systems still face. For practitioners, the finding that the LSA model serves as a "thesaurus" helping users find appropriate words suggests prediction systems offer benefits beyond raw keystroke reduction.
Tags: AAC · word prediction · language modeling · user adaptation · n-gram models · latent semantic analysis · motor impairments · cerebral palsy · virtual keyboard · single-switch access