a common technique in natural language processing is to operationalize certain semantic concepts (e.g., "synonym") in terms of syntactic structure (two words that tend to occur nearby in a sentence are more likely to be synonyms, etc). This is what word2vec does.

Traditional word embedding algorithms assign a distinct vector to each word. This makes them unable to account for polysemy. In a recent work, Upadhyay et al. (2017) provided an innovative way to address this deficit. The authors leveraged multilingual parallel data to learn multi-sense word embeddings.

This is very important as training embeddings from scratch requires large amount of time and resource. Mikolov et al. (2013) tried to address this issue by proposing negative sampling which is nothing but frequency-based sampling of negative terms while training the word2vec model.

A general caveat for word embeddings is that they are highly dependent on the applications in which it is used. Labutov and Lipson (2013) proposed task specific embeddings which retrain the word embeddings to align them in the current task space.

One solution to this problem, as explored by Mikolov et al. (2013), is to identify such phrases based on word co-occurrence and train embeddings for them separately. More recent methods have explored directly learning n-gram embeddings from unlabeled data (Johnson and Zhang, 2015).

The context words are assumed to be located symmetrically to the target words within a distance equal to the window size in both directions.