Figure 6

Training the W2V BoW model. A sliding window with the size of five words is moved over the text, word by word. The input layer nodes of the network corresponding to the words in the context window of the word w₃ are activated. The error in the output layer prediction and the expected prediction for the focus word w₃ is back-propagated through the network. When the training is completed, the context vector $\overrightarrow{\mathsf{\text{C}}{w}_3}$ constitutes the set of weights connecting the input layer node for w₃ and the hidden layer.