Machine Learning Methods for Word Prediction in Brasilian Portuguese

Objective Computers have become essential in the life of many people with disabilities. A common activity that can be computer assisted is text generation. People who cannot accurately control their extremities (due to cerebral palsy, etc.) may use computers as writing tools, and if they have problems to speak, they may use a computer to communicate. In both cases, the generation of text is a necessary activity that can be physically demanding and extremely slow. Word prediction methods are commonly used to assist in this task. The objective of our work is to improve the quality of a word prediction system for Brazilian Portuguese, in order to reduce the effort and time needed to write texts. Main content The selection of the predicted words is partly based on a two steps process. Firstly, the possible parts-of-speech (POS) of the next word are predicted from the POS of the previous words. Secondly, the list of predicted words is generated from these predicted POS and the information contained in the lexicons. In this paper we present prediction algorithms based on machine learning methods adapted to POS prediction (the first step of the process). Specifically, this work describes the use of artificial neural networks, support vector machines and regularized logistic models to predict word POS in Brazilian Portuguese, based on the POS of the 1, 2, 3 or 4 previous words. We also briefly describe a meta-learning strategy for algorithm selection and a fusion algorithm to combine them. Results These methods increase the word prediction quality, saving a maximum of 38.26% of the keystrokes needed to write the text (a relative improvement of 9.85% with respect to the unigram method), and correctly predicting 79.95% of the words in the experiments performed (with a maximum of 28.5% of hit rate). Conclusions Besides presenting evidences that such methods can be adapted to predict word POS, it is also shown that they are robust, consistent and easy to incorporate into a general word prediction system. In future works, the Portuguese prediction system will be included in PredWin, a freely available text editor and communicator already working for Spanish. The whole word prediction system will also be adapted, trained and evaluated in other languages, such as English, and included in PredWin in case good results were obtained.