https://www.leiphone.com/news/201706/eV8j3Nu8SMqGBnQB.html
https://blog.csdn.net/dn_mug/article/details/69852740
Neural network like a black box, which is difficult to understand the concept, this blogger personal training vector word understanding is in place:
For each word s, the training data corresponding tag is another word t, actually want to training find a relationship mapping, let s mapped to t. But obviously we do not want to find a linear function, such that s will be able to obtain a given t, we can get a desirable class of words by T s, comprising t. For each T t, due to the difference in the occurrence of frequency s context, a probability can be obtained naturally, the higher the frequency the higher the relevance of s and t.
For word vectors, or the parameter matrix W, it may be considered a word mapped to the semantic space of the bridge, s and t higher the correlation, which is considered in the semantic space closer, the closer the corresponding bridge also. If the vector is to understand the words of the vector before the smaller the angle, we use vectors to represent the information of the word, it is important to obtain semantic information. In practical applications, generating a text, we can determine the similarity between the word and the word vector, if too low, then you need to wonder whether correct.