A Versatile Architecture for a Multilingual Word Prediction System

At the 10th International Conference on Computers Helping People with Special Needs (ICCHP) on July 12-14 at the University of Linz, Austria, the staff of Department of Electronics, Polytechnic School, University of Alcala, presented their paper on the development of high-tech aids for people with physical disability and how assistive technology can improve their quality of life.

A Word Prediction application is an assistive technology solution that is highly beneficial for people with physical disabilities. It tries to predict which word a user is typing or going to type before the user types it completely. The user can select the desired word from the list of guessed words and include it in the text. It reduces the time and effort while typing a text because the user does not have to type of the entire word. It is useful for slow typists and for people with special needs who are unable to use a conventional keyboard.

Profet, PredWin, Fazly, Fasty, Aurora Prediction and WordQ are some of the assistive technology applications based on word prediction. Some of these applications are speech prostheses that help non-vocal users to communicate. One of the tools is speech synthesis in which a user enters a word, and it is read aloud. Word prediction increases the communication rate of non-vocal users with a physical disability. Usually, these applications include prediction in English, or automatically trained lexicons for other languages like Portuguese, Swedish and Spanish. However, this application supports automatically trained lexicons and prediction methods. In the case of Spanish, the high-quality prediction system with specially designed grammatical information, rules and lexicon improve the results considerably.

The general architecture of this word prediction system is modular, which has a high degree of flexibility and is divided into upper and lower parts. The modules related to the software application from the user’s application can be seen in the upper part and the modules related to the automatic training of all the information sources used in the prediction process are located in the lower part. The system design and portability to other languages is easier as the prediction methods and lexicons are independent. Lexicons, Word Prediction Methods and Heuristics are some of the most important modules. Personal and subject lexicon provide the best results for all languages and overcome the limitations of fixed lexicons. The prediction system is common for all the languages. The databases and the information sources vary from one to the others.

Depending on the application, the evaluation of a word prediction system as a writing aid is measured in two different ways. For people with physical problems, the aim is to reduce the number of keystrokes needed to write the text and for people with linguistic problems, the focus is on ‘prediction coverage,’ which is the number of words that are correctly predicted. The most efficient word prediction system is the Spanish prediction system, since a large amount of information has been specifically designed, trained and adapted to optimize the prediction process. The word prediction for Spanish is presently included in PredWin, a widely used text editor for people with severe disabilities. The percentage of words predicted before the user types a word is usually up to 90-95% for all languages, lexicons and methods.

Assistive technology products, such as the Word Prediction System, provide better accessibility to high-tech writing aids to people with physical disabilities, linguistic problems, slow learners, children and people learning a second language.