Problems in detecting Spanish Date with es-core-news-lg v3.8.0

[schatterjee-rc]

Our org has a requirement for detecting spanish dates that are not in perfect format like dd/mm/yyyy (Predefined DateRecognizer is more then enough for that). But stuff like - "El acontecimiento histórico tuvo lugar el 23 de abril de 2021.". The english translation of it ("The historical event took place on April 23, 2021.") works quite well with en-core-web-lg, spacy identifies it well.

For spanish I am using es-core-news-lg version 3.8.0, but no luck with any kind of dates with months like the one above. Is the model not trained on such date datasets?

The following is my setup -

import spacy

nlp = spacy.load("es_core_news_lg")

doc = nlp("El acontecimiento histórico tuvo lugar el 23 de abril de 2021.")


print(doc.ents)
for ent in doc.ents:
    print(ent.text, ent.label_)

[apoorva-01]

The reason isn't the date format, it's that es_core_news_lg doesn't have a DATE label at all. The Spanish pipelines are trained on WikiNER + AnCora, which only annotate PER, LOC, ORG, MISC. The English models catch dates because they're trained on OntoNotes, which does have DATE/TIME/MONEY. So no matter how you phrase "23 de abril de 2021", the statistical NER will never emit a date entity. You can see it directly:

print(nlp.get_pipe("ner").labels)
# ('LOC', 'MISC', 'ORG', 'PER')

For natural-language Spanish dates, skip the statistical NER and go rule-based. dateparser handles exactly your case out of the box:

from dateparser.search import search_dates

search_dates(
    "El acontecimiento histórico tuvo lugar el 23 de abril de 2021.",
    languages=["es"],
)
# [('23 de abril de 2021', datetime.datetime(2021, 4, 23, 0, 0))]

If you want them back inside the doc as spans, wrap that in a custom component and push the matches onto doc.ents, or build an EntityRuler with patterns for the "DD de MONTH de YYYY" shape. That plus your existing DateRecognizer for the dd/mm/yyyy cases covers both.