GloVe dot product optimized for non-comutative data whilst the operation itself being commutative
To my current knowledge, GloVe word vectors dot product are optimized to be the
w_i ⋅ w_j = log(P(ⅈ|j))
The probability being computed from a cooccurance matrix. However, dot product is a commutative operation, whilst the log probablity isn't. Is this issue being adressed in GloVe? Am I missing something?
Topic mathematics stanford-nlp word2vec word-embeddings nlp
Category Data Science