Skip-Gram Training Pairs

Generate the (center, context) training pairs that word2vec's skip-gram model learns from.

Skip-gram learns a word's vector by asking it to predict the words around it. For every position i in a token list, the center word tokens[i] is paired with each word at most window positions away on either side, but not with itself. Words that appear in similar contexts end up with similar vectors.

Emit the pairs in this order: centers from left to right, and for each center, context positions from left to right.

Function signature

def skipgram_pairs(tokens, window):

Example

Input

skipgram_pairs(["the", "cat", "sat", "down"], window=1)

Output

[("the", "cat"), ("cat", "the"), ("cat", "sat"),
 ("sat", "cat"), ("sat", "down"), ("down", "sat")]

Reasoning

With a window of 1, each word pairs with its direct neighbours. The first and last words have only one neighbour each, so they give one pair. The middle words give two.

Constraints and notes

• The window is cut at the ends of the list.
• Pair by position, not by value: a word that appears twice next to itself still forms a pair with the other copy.
• Return a list of tuples.