I am using GloVe in a context where the order of the tokens and distances between tokens in the context window don't have the same meaning as in most language modeling tasks. Just counting the unweighted cooccurrences leads to more meaningful vectors in some contexts. In the proposed change I am adding an argument in cooccur.c to take a flag to ignore distance between words when counting coocurrences.
The default behavior remains the same as before, so an explicit flag needs to be passed to switch on the new behavior.
I am using GloVe in a context where the order of the tokens and distances between tokens in the context window don't have the same meaning as in most language modeling tasks. Just counting the unweighted cooccurrences leads to more meaningful vectors in some contexts. In the proposed change I am adding an argument in cooccur.c to take a flag to ignore distance between words when counting coocurrences.
The default behavior remains the same as before, so an explicit flag needs to be passed to switch on the new behavior.