documentation on cosine similarity range is wrong

[JunhaoWang]

In doc,

The concept of Answer Semantic Similarity pertains to the assessment of the semantic resemblance between the generated answer and the ground truth. This evaluation is based on the ground truth and the answer, with values falling within the range of 0 to 1.

This is wrong since cosine similarity can take on negative values

a = [1.0]
b = [-1.0]
cosine_sim_a_b = a dot_product b / (a_norm x b_norm) = -1

[omkar-334]

You're right. Cosine smiliarity can take on negative values, althought it is heavily biased towards positive values.
A few interesting experiments on this -

1. https://datascience.stackexchange.com/questions/101862/cosine-similarity-between-sentence-embeddings-is-always-positive#:~:text=This%20range%20is%20valid%20if,line%20but%20in%20opposite%20directions.
2. https://stackoverflow.com/questions/60852877/why-is-my-cosine-similarity-always-positive-fasttext

Keep in mind that vector embeddings are a result of computing the probability of the word in a given context. This means that beautiful and ugly, even though they are the opposites, they have a medium cosine score, since they're likely to appear in the same context. Completely unrelated phrases like quantum mechanics cryptography algorithms and blue kingfisher eating salmon have a negative cosine score.
However, in Dataset Generation, since all of the nodes have the same/related context, it is extremely unlikely to have a negative cosine score.