To optimize the proposed prediction model in the future
work, we need to intensively select the input features and the
learning methods. In the case of features, the current number
of features is 12. Only the most significant subset of these 12
features might be optimal for a particular learning method. A
suitable way to select a subset of significant features such as a
minimum classification error (MCE) algorithm [13] can be
applied. Apart from the current features, one of interesting
features is the window size of POS sequences as it has been
proven to highly affect the performance of phrase break
prediction for English [1]. If the window size of POS
sequences is suitable, the learning method can better gather
crucial information from the input text.