trying to add pipes to en_coreference_web_trf from other model and all tokens are NOUNs

[afriedman412]

I'm working something that I uses coreferencing and the basic spacy pipes. I was using en_coreference_web_trf and en_core_web_sm separately and parsing the text twice, but that seemed inefficient. I figured I would combine the two.

It looks like this:

nlp = spacy.load("en_coreference_web_trf")
for p in ['lemmatizer', 'attribute_ruler', 'parser', 'tagger']:
    try:
        nlp.add_pipe(p, source=spacy.load("en_core_web_sm"), first=True)
    except ValueError:
        continue

But when I run nlp(t) (where t is some text), it tags every single token as a NOUN.

help?

[danieldk]

The lemmatizer, parser and tagger pipe use the tok2vec pipe to get the contextualized word representations. So you have to add the tok2vec pipe as well and ensure that it runs before the other pipes. E.g. the following does work correctly:

nlp = spacy.load("en_coreference_web_trf")
for p in ['tok2vec', 'tagger', 'parser', 'lemmatizer', 'attribute_ruler']:
    try:
        nlp.add_pipe(p, source=spacy.load("en_core_web_sm"))
    except ValueError:
        continue

[afriedman412]

lol so I was just forgetting one of the pipes.

thanks -- it works now!!