Help: Can't reproduce demo about neuralcoref

[thomwolf]

Hmm the site is running an older version of NeuralCoref (v1 in fact) with spaCy 1 so there are indeed so differences.
Two things:

• have you tried with en_coref_lg?
• can you give me an example of sentence and coref that work on the demo and not with your version?

[sightama]

I have yet to try the larger model because of size difference and I'm assuming the runtime overhead wouldn't be worth it, but I will try it if you think I will get different results. My application is clinical so I'll post this example:

"Left adnexal cystic lesion measures 4.3 x 3.8 cm. It appears increased from the prior examination measurement of 2.0 x 2.0 cm. This is indeterminate but may represent a hemorrhagic cyst possibly endometrioma but this requires clinical correlation"

Again this might be a lost cause because I'm trying to use this on different text than the model would expect, but when run in Debug mode on the demo site if gives decent weight associations between lesion, this, and it.

[thomwolf]

Hi @kubicki14, yes the text looks very different from the training data so it's really hard to predict what the statistical model will guess. We will be releasing a version that let you more simply retrain the model on your data but that won't before a few months.
In the mean time I would recommend using the largest model en_coref_lg.

[devitoclic]

Same problem here. Online demo resolve corefs with custom words and expressions but using the latest neuralcoref package in local environment returns doc._.has_coref = False for the same example. How many examples do we need to adapt the model to custom data?