Machine Learning Approaches applied to Brazilian Portuguese Word Prediction

People with physical disabilities may have serious problems to use computer keyboards to write. For this reason, they may use specific tools that include systems to assist the writing process, such us word prediction, in order to reduce the number of keystrokes needed to write the text. Word prediction may be based on different sources of information: statistical, grammatical, specific of the subject or/and the user, etc. In this paper we increase the quality of the word prediction in Brazilian Portuguese by improving the prediction of the part of speech (POS) of the predicted word. We propose the following methods to predict the POS: artificial neural networks, support vector machines, regularized logistic models and a naive Bayes classifier. When included in the word prediction system, they save between 32.55 % and 34,58% of the keystrokes needed to write the text.