For DyslexiWrite, when we talk about "word prediction," we're talking about a writing tool that helps you write by constantly trying to guess what word you are typing. If you're not typing anything, the word predictor will try to guess what will be your next word based on the content around existing text. As you type, if you see the word you want among the guesses, you can simply click it to insert it. Once you get the hang of it (and it doesn't take long), you will be able to write faster and more accurately than you ever have before.

Other forms of language technology, such as grammatical analysis tools or the language models used in speech recognition have other goals and limitations. However, they often use similar technology and algorithms.

There are two fundamental models for word prediction: statistical and grammatical. Roughly speaking, statistical word prediction works from a database of frequently used words and expressions taken from a large collection of documents. Grammatical word prediction, on the other hand, is more like a kind of tic-tac-toe game the computer plays with the text, in which it uses logic to deduce the most likely outcomes. Both methods have their own advantages and disadvantages.

At first glance, the grammatical method may seem like the most sensible. After all, basic grammar is something we all learn in grade school, so it can't be too difficult for a computer, right? And if the computer has been analyzing the text as you write it, it should be able to guess intelligently as to which words should come next. At the very least, for example, it should know whether the next logical word is a noun or a verb.

Unfortunately, it's not always that simple. In fact, there are many reasons that the grammatical method is not especially suited to effective and robust word prediction:

• First, very few people write grammatically correct all the time. The less perfect a sentence is, the more difficult it becomes for the computer to guess the next words. Unlike humans, computers have a hard time thinking for themselves. They can only follow the rules they know, which is a very serious limitation.
• Second, it is very difficult for a computer to make a grammatical analysis of a sentence in progress. No sentence is grammatically perfect while it is still in the process of being typed.
• Third, in many cases it is difficult to make correct grammatical analyses of grammatically perfect sentences-even for people.

Therefore, the vast majority of effective word predictors are based primarily on statistical methods-and that includes DyslexiWrite. A statistical word predictor's quality is first and foremost dependent on the size of its vocabulary and the quantity and type of text it was trained on. For example, say a statistical model was designed from newspaper articles, which are typically written in 3rd-person ("The announcement was made this morning…," or, "Employees at Johnson's Pillow Factory have declared a strike..."), and with language that sounds a bit "dry." Such a model won't be very good at predicting words for a document written in a more conversational tone, like emails ("Hi honey!", "and then she said...").

In addition, the majority of statistical models are very local in their focus. Typically they rely only on the previous one or two words to formulate their predictions, and often make no attempt to ensure agreement in case. The result is that it is sometimes possible to "cheat" a statistical word predictor into providing suggestions that do not correspond to the preceding text. For a statistical word predictor, the sentence "My father had his birthday yesterday, and we have her a new carpenter's bench" will, in all likelihood, seem perfectly acceptable, despite the disagreement between "father" and "her"-the two words are so distant from each other, they fall outside the model's horizon.

Statistical word prediction is, in other words, no miracle solution. However, it is a robust method for providing smart word suggestions, even (or better yet, especially) when the writer doesn't always keep to proper English.