A knowledge-driven and data-driven integration of the Ko-rean auto-pronunciation method

In the data-driven method, the decision tree algorithm is easy to understand, and the mechanism is highly interpretable. The resulting model is easy to visualize, and non experts are also easy to understand. Compared with other algorithms, the decision tree algorithm has a relatively small amount of computation and is easy to incorporate knowledge rules, and can handle continuous fields. Other technologies often require data generalization, such as removing redundant or blank attributes, while the decision tree algorithm does not need complex data preprocessing, such as normalization or standardization, and can handle a large number of data in a relatively short time with good results. In the initial stage of building the decision tree, the Gini index of all features is calculated, and the feature with the smallest Gini index is selected as the root node. According to the optimal cut-off point of the root node, it is divided into two forks and the corresponding data set, and then continue to calculate the Gini index of other features along the leftmost branch, and then select an optimal feature as the node of the fork, and so on. When all the calculations of this branch node are completed or cannot continue to grow, it will return to the previous unfinished branch to continue the growth of the tree, Until all forks have completed node calculation or the tree cannot continue to grow downward.