FastTextKeyedVectors.word_vec(...,use_norm=True) nonsensically summing ngrams norms

[gojomo]

The logic in https://github.com/RaRe-Technologies/gensim/blob/ee6169100d13d7f684b96ac137065e302aeb7b1e/gensim/models/keyedvectors.py#L2090 regarding the summing of normed ngram-vectors, looks like nonsense to me.

A caller supplying use_norm expects (per doc-comment) a unit-normed vector back. But summing a bunch of individually unit-normed ngram vectors won't achieve that: you'd need to unit-norm the final result. And it's unlikely you'd want to use unit-normed summands: the original raw ngram vectors are what the algorithm has trained to model words-from-fragments.

Of course, matching the FB reference implementation's behavior should be a top goal, but I doubt it offers any such summing-of-normed-ngram-vectors.

If this current behavior is replaced with the more sensible "add raw vecs, unit-normalize final result", then there's no need to ever calculate/store the vectors_ngrams_norm array.

[gojomo]

Fixed in #2698.