Evaluation in a Spacy NER model

Question

I am trying to evaluate a trained NER Model created using spacy lib. Normally for these kind of problems you can use f1 score (a ratio between precision and recall). I could not find in the documentation an accuracy function for a trained NER model.

I am not sure if it's correct but I am trying to do it with the following way(example) and using f1_score from sklearn:

from sklearn.metrics import f1_score
import spacy
from spacy.gold import GoldParse


nlp = spacy.load("en") #load NER model
test_text = "my name is John" # text to test accuracy
doc_to_test = nlp(test_text) # transform the text to spacy doc format

# we create a golden doc where we know the tagged entity for the text to be tested
doc_gold_text= nlp.make_doc(test_text)
entity_offsets_of_gold_text = [(11, 15,"PERSON")]
gold = GoldParse(doc_gold_text, entities=entity_offsets_of_gold_text)

# bring the data in a format acceptable for sklearn f1 function
y_true = ["PERSON" if "PERSON" in x else 'O' for x in gold.ner]
y_predicted = [x.ent_type_ if x.ent_type_ !='' else 'O' for x in doc_to_test]
f1_score(y_true, y_predicted, average='macro')`[1]
> 1.0

Any thoughts are or insights are useful.

Answer 1

You can find different metrics including F-score, recall and precision in spaCy/scorer.py.

This example shows how you can use it:

import spacy
from spacy.gold import GoldParse
from spacy.scorer import Scorer

def evaluate(ner_model, examples):
    scorer = Scorer()
    for input_, annot in examples:
        doc_gold_text = ner_model.make_doc(input_)
        gold = GoldParse(doc_gold_text, entities=annot)
        pred_value = ner_model(input_)
        scorer.score(pred_value, gold)
    return scorer.scores

# example run

examples = [
    ('Who is Shaka Khan?',
     [(7, 17, 'PERSON')]),
    ('I like London and Berlin.',
     [(7, 13, 'LOC'), (18, 24, 'LOC')])
]

ner_model = spacy.load(ner_model_path) # for spaCy's pretrained use 'en_core_web_sm'
results = evaluate(ner_model, examples)

The scorer.scores returns multiple scores. When running the example, the result looks like this: (Note the low scores occuring because the examples classify London and Berlin as 'LOC' while the model classifies them as 'GPE'. You can figure this out by looking at the ents_per_type.)

{'uas': 0.0, 'las': 0.0, 'las_per_type': {'attr': {'p': 0.0, 'r': 0.0, 'f': 0.0}, 'root': {'p': 0.0, 'r': 0.0, 'f': 0.0}, 'compound': {'p': 0.0, 'r': 0.0, 'f': 0.0}, 'nsubj': {'p': 0.0, 'r': 0.0, 'f': 0.0}, 'dobj': {'p': 0.0, 'r': 0.0, 'f': 0.0}, 'cc': {'p': 0.0, 'r': 0.0, 'f': 0.0}, 'conj': {'p': 0.0, 'r': 0.0, 'f': 0.0}}, 'ents_p': 33.33333333333333, 'ents_r': 33.33333333333333, 'ents_f': 33.33333333333333, 'ents_per_type': {'PERSON': {'p': 100.0, 'r': 100.0, 'f': 100.0}, 'LOC': {'p': 0.0, 'r': 0.0, 'f': 0.0}, 'GPE': {'p': 0.0, 'r': 0.0, 'f': 0.0}}, 'tags_acc': 0.0, 'token_acc': 100.0, 'textcat_score': 0.0, 'textcats_per_cat': {}}

The example is taken from a spaCy example on github (link does not work anymore). It was last tested with spacy 2.2.4.