I'm not surprised that industry prefers W2V over Transformers, given how heavy-duty the latter can be at inference time. It's been a few years since I looked a

I'm not surprised that industry prefers W2V over Transformers, given how heavy-duty the latter can be at inference time. It's been a few years since I looked at it, but IIRC fastText is basically just w2v with subwords, so it's also possible this negative sampling fix applies to w2v and fastText equally.