Incorrect CBOW implementation in Gensim leads to inferior performance

[piskvorky]

Problem description

According to this article https://aclanthology.org/2021.insights-1.1.pdf:

Steps/code/corpus to reproduce

I haven't tried to verify / reproduce. Gensim's goal is to follow the original C implementation faithfully, which it does. So this is not a bug per se, more a question of "how whether / how much we want to deviate from the reference implementation". I'm in favour if the result is unambiguous better (more accurate, faster, no downsides).

Versions

All versions since the beginning of word2vec in Gensim.

[piskvorky]

@gojomo @mpenkov WDYT?

[gojomo]

I remain as unconvinced as when this was raised earlier this year. I'll add more comments there. I suggest this be closed as dupe.

[piskvorky]

Thanks! I remember that discussion now.

Closing as duplicate of #1873.