If you rely on an on-screen keyboard to communicate, anything that improves typing speed and reduces effort makes a big difference. Word prediction is one technology that can make such a difference, as long as it takes less effort to locate the right suggestion than it takes to type the next character. A word prediction system needs to feel as if it is reading your mind. This was the objective we had while developing our PolyPredix word prediction engine.
There are two basic types of word prediction: systems based on linguistic rules and systems based on statistics. Linguistic prediction systems have the benefit and drawback that they guide the user into writing certain types of sentences. Statistical prediction systems allow the user more freedom in writing style but do not guarantee grammatically correct sentences. For PolyPredix, we chose the statistical approach so that we could combine great prediction with support for many different languages.
In the world of word prediction, there are three different methods for predicting words: word completion, next-word prediction and multi-word prediction. Word completion systems suggest an ending to the word currently being typed but are unable to suggest the most likely next word. Next-word prediction systems combine word completion with the ability to suggest the next word after a space is typed. The most advanced method is multi-word prediction, which builds upon word completion and next word prediction by including multi-word suggestions, such as "want to" after the user has typed "I wa". PolyPredix is such a multi-word prediction system.
The PolyPredix engine takes a wide range of factors into account to provide the best possible suggestions. Using highly optimized prediction dictionaries containing 90,000 common words and word combinations, the system analyzes how frequently and how recently a word is used. In addition, when predicting a word, PolyPredix takes into account the preceding words as many sentences follow similar patterns.
The PolyPredix engine is self-learning. As the user types, the system adjusts frequencies to the individual user and learns new words and word combinations. It is even capable of learning a new language from what the user types. For users who make frequent typing errors, it is possible to turn learning off or avoid learning words unknown to the spellchecker. Small errors, such as errors in capitalization or missing accents in words, are often automatically corrected by the prediction engine.
Efficient word prediction saves effort and helps maintain typing speed over a longer period of time. Seeing more suggestions increases the likelihood of the right suggestion appearing. However, more suggestions also leads to increased visual effort while scanning the suggestions. In the figure below, you can see the impact the number of suggestions has on the number of taps saved. Out-of-the-box, with as few as three suggestions displayed, the number of taps can be reduced by over 50% when typing in English with the PolyPredix 3 engine included in Proloquo4Text. Effort reduction varies per language depending on the complexity of the language. Once the system becomes accustomed to the user's vocabulary, effort reduction rises to over 75% for English. These huge reductions in effort can significantly speed up typing for slower typists, making our PolyPredix engine one of the most effective prediction engines on the market.
Figure: Out-of-the-box effort reduction for 5 languages for 0 to 15 displayed suggestions.