Add support for Sentence Transformers

[alanakbik]

The sentence transformers library has great pre-trained models to produce embeddings for entire sentences.

We should add a new DocumentEmbeddings class to support these embeddings in Flair.

[shameelct]

This commit does not cover contextual word embedding as in other transformer models support in Flair right? Any plans to add that?

[alanakbik]

Not sure if I understand - can you elaborate?

[shameelct]

The following is how a BERT embedding can be used in Flair to obtain contextual word embedding as far as I understand:

from flair.data import Sentence
bert_embedding = TransformerWordEmbeddings('bert-base-multilingual-cased')
sentence = Sentence('The grass is green .')
bert_embedding.embed(sentence)
for token in sentence:
print(token)
print(token.embedding)

Is there a way to obtain contextual word embedding from Sentence Transformer using Flair, like we did for BERT like above code snippet.

[alanakbik]

Sentence transformers are trained to give an embedding for the whole sentence, so getting single word embeddings is currently not possble through Flair. For words, regular transformer embeddings will probably be better.

[shameelct]

Cool thanks