Generating context-based features

We use Word2Vec [22] to generate word embedding. Specifically, we use a skip-gram model which aims to find word representations that are useful for predicting the surrounding words in a given sentence or a document consisting of sequence of words; w₁,w₂,...,w_K. The objective is to maximize the average log probability using the following formula:

where word vectors V(w) are computed by averaging over the number of words K and c is the size of the training context. We generated the word embedding by using the default parameter settings of the Word2Vec gensim implementation: vector size (dimensionality) of 100, window size 5, minimum occurrence count of 5, and we use a skip-gram (sg) model.