[gensim:11892] Re: doc2vec+elmo/bert

[Gordon Mohr]

Gensim's `Doc2Vec` generally doesn't use outside word-vectors as a separate
input. Rather, it trains doc-vectors in a style similar to the original
word2vec's word-vectors, a process that sometimes also involves
simultaneously-training word-vectors.

So, there's no standard or well-supported way to incorporate either another
style of training word-vectors, or the results of some outside word-vector
training, into a `Doc2Vec` model.

There was an experimental method (`intersect_word2vec_format()`), that used
to work for merging overlapping words into an existing `Doc2Vec` model,
which might have some follow-on effects (good or bad) on subsequent
`Doc2Vec` training in some modes. (That method is no longer available on
`Doc2Vec`.) Others have at times patched the source to support similar
operations. So if you have a strategy in mind you can review the code and
perform whatever kind of word-vector-replacement you think might help.

- Gordon

please tell me how can I incorporate bert or elmo these kind of
contextuallized word embedding into gensim‘s doc2vce models