TL;DR
This paper investigates how word prediction impacts communication speed for Augmentative and Alternative Communication (AAC) users. They found that both basic and advanced word prediction significantly increase communication rate, with advanced prediction offering a substantial speedup over basic prediction, even with simulated motor impairments.
What Problem Does This Paper Address?
Individuals relying on Augmentative and Alternative Communication (AAC) devices face a significant communication speed gap compared to unimpaired speech, often communicating at less than 10% of the speed. This creates a substantial barrier to social interaction and participation. While word prediction is a common feature in high-tech AAC devices, its actual effectiveness in increasing communication rate has been debated. Previous research has questioned whether the cognitive load and interface complexities of word prediction outweigh its benefits. This paper aims to clarify this by comparing the communication rates achieved with different levels of word prediction sophistication, hypothesizing that improved prediction methods will lead to greater communication rate gains.
Methods
The study employed a user study with "pseudo-impaired" participants. To simulate the motor impairments of typical AAC users, a 1.5-second delay was introduced between a key press and its registration on an on-screen keyboard (Wivik). This delay aimed to incentivize the use of word prediction, as users with fine motor control might otherwise ignore it.
Three text entry conditions were tested:
- No Word Prediction: Standard letter-by-letter typing.
- Basic Word Prediction: Utilized a recency model combined with a large word list, prioritizing the recency model for predictions.
- Advanced Word Prediction: Employed a trigram model with backoff, using a special unigram model for the first word of a sentence. If the trigram model didn't yield enough predictions, it fell back to a recency model and then a word list, similar to the basic method.
Participants typed three different excerpts from the Switchboard corpus. Keystrokes and predictions were logged to compute various metrics, including words per minute, seconds per keystroke, and keystroke savings.
Datasets
The language models for the prediction methods were trained on portions of the Switchboard corpus. For the user study, participants typed excerpts from a held-out portion of the Switchboard corpus. The task was a "copy task" to ensure controlled experimental conditions.
Key Insights and Ideas
- Word prediction does work: Despite previous skepticism, this study provides strong evidence that word prediction significantly boosts communication rate for AAC users, even when accounting for simulated motor impairments.
- Accuracy matters: The sophistication of the word prediction algorithm has a direct and substantial impact on communication rate. Advanced prediction offers a much larger speedup than basic prediction.
- Cognitive load is overcome by efficiency: While advanced prediction led to a slower per-keystroke input rate (due to scanning prediction lists), the overall communication rate (words per minute) increased significantly. This suggests that the reduction in the total number of keystrokes required is the dominant factor.
- User interface design is crucial: The authors suggest that the intuitiveness and consistency of their user interface might have contributed to the success of even the basic prediction method, potentially explaining why some users learned to utilize it effectively.
Key Results
- Communication Rate:
- Advanced prediction resulted in 59.9% more words per minute (wpm) compared to no prediction.
- Advanced prediction was 44.4% faster than basic prediction.
- Basic prediction was 10.1% faster than no prediction.
- Task Completion Time:
- Copy tasks with advanced prediction were completed in 64.5% of the time compared to no prediction.
- Input Rate (Seconds per keystroke):
- Users took longer per keystroke with advanced prediction (0.641s longer than no prediction) and basic prediction (0.345s longer than no prediction). This highlights the overhead of list scanning.
- Keystroke Savings:
- Potential keystroke savings were 55.2% for advanced prediction and 25.0% for basic prediction.
- Actual keystroke savings realized were 50.3% for advanced and 18.2% for basic.
- Prediction utilization was significantly higher for advanced prediction (90.9%) compared to basic prediction (73.3%).
Limitations and Future Work
- Simulated Impairment: The study used pseudo-impaired participants. The authors acknowledge that future work should validate these results with actual AAC users, as their specific needs and interaction patterns might differ.
- User Interface Specificity: The results are tied to the specific interface design used. The authors note that the ease of using word prediction can vary significantly across different AAC devices and interfaces.
- Future Work: The authors plan to improve word prediction algorithms and validate their findings with actual AAC users. They also suggest that further research into improving word prediction is likely to have a significant positive impact on the quality of life for AAC users.
Summary updated