Create Example object for Classification with SPacy 3.x

[nicolalandro]

How to create correctly the Example? I try with this code:

 doc = nlp.make_doc(text)
examples = [Example.from_dict(doc, { "tags": [annotation] } )]

But it get the following error:

ValueError: [E971] Found incompatible lengths in `Doc.from_array`: 1 for the array and 7 for the Doc itself.

How I should create it?

[DuyguA]

How to create correctly the Example? I try with this code:

 doc = nlp.make_doc(text)
examples = [Example.from_dict(doc, { "tags": [annotation] } )]

But it get the following error:

ValueError: [E971] Found incompatible lengths in `Doc.from_array`: 1 for the array and 7 for the Doc itself.

How I should create it?

Thanks for your question!

There should be one tag per token, hence Doc length and annotation length should match. Here's an example, there are 2 tokens in my Doc:

>>> text = "you go"
>>> doc = nlp.make_doc(text)
>>> annotation = ["PRON", "VERB"]
>>> example = Example.from_dict(doc, {"tags": annotation})

[DuyguA]

Keep in mind that tokens are assigned by the Tokenizer (so don't count whitespaces while calculating the Doc length 🙂 ). For an example contracted forms are tokens on their own:

>>> doc = nlp.make_doc("It's a new day")
>>> [token for token in doc]
[It, 's, a, new, day]
>>> len(doc)
5

[nicolalandro]

But I want to do Text Classification and not Named entity recognition, so I have One class for whole string. do I make something different?

[DuyguA]

Ah OK, I saw a tags in your question and I thought you wanna make POS training. Then you can do the following

for binary textcat:

label = {"TOXIC": True}
example = Example.from_dict(nlp.make_doc(text), {"cats": label})

Above you want to make a binary classification and your label is called TOXIC .

for multi_textcat:

label = {"FIVE_STAR": True, "FOUR_STAR": False, "THREE_STAR": False}
example = Example.from_dict(nlp.make_doc(text), {"cats": label})

Here is similar, you want to make multilabel classification and your labels are FIVE_STAR , FOUR_STAR and THREE_STAR.

[nicolalandro]

Ok thank you perfect! But now I have another problem during update:

KeyError: "Parameter 'E' for model 'hashembed' has not been allocated yet."

I do the following:

import spacy
import spacy.cli
from spacy.pipeline.textcat import single_label_cnn_config, Config
from spacy.training.example import Example

nlp=spacy.load('it_core_news_lg')

from spacy.util import minibatch, compounding

# Disabling other components
other_pipes = [pipe for pipe in nlp.pipe_names if pipe != 'textcat']
with nlp.disable_pipes(*other_pipes): # only train textcat
    optimizer = nlp.resume_training()

# Performing training
for epoch in range(5):
    losses = {}
    batches = minibatch(train, size=compounding(4., 32., 1.001))
    for batch in batches:
        examples = [Example.from_dict(nlp.make_doc(text), {"cats": {annotation:True}}) for text, annotation in batch]
        nlp.update(examples, losses=losses, drop=0.3)

[DuyguA]

I think the issue is excluding the tok2vec pipeline component, hashembed is part of that component. Can we change your code a bit :

include_pipes = ["tok2vec", "textcat"]
other_pipes = [pipe for pipe in nlp.pipe_names if pipe not in include_pipes]