According to the idea of EM, a language model is built increasingly by collection the fractional counts of patterns (such as bigram pair) from the augmentations of all the segmentation candidates of a sentence.

• 基于em的思想，每个句子所对应的所有（或一定範围内）的分词结果构成训练集，通过这个训练集和初始的语言模型可以估计出一个新的语言模型。