为什么 spacy ner 的结果是高度不可预测的？

Question

I tried spacy for ner but the results are highly unpredictable.Sometimes spacy is not recognizing a particular country.Can anyone please explain why is it happening? I tried on some random sentences.

CASE 1:

nlp = spacy.load("en_core_web_sm")
print(nlp)
sent = "hello china hello japan"
doc = nlp(sent)
for i in doc.ents:
  print(i.text," ",i.label_)

OUTPUT:no output in this case.

CASE 2:

nlp = spacy.load("en_core_web_sm")
print(nlp)
sent = "china is a populous nation in East Asia whose vast landscape encompasses grassland, desert, mountains, lakes, rivers and more than 14,000km of coastline."
doc = nlp(sent)
for i in doc.ents:
  print(i.text," ",i.label_)

OUTPUT:

<spacy.lang.en.English object at 0x7f2213bde080>
china   GPE
East Asia   LOC
more than 14,000km   QUANTITY

Answer 1

Natural Language models, like spaCy NER, learn from the contextual structure of the sentence (surrounding words). Why is that? Let's take the word Anwarvic as an example which is a new word that you haven't seen before and probably the spaCy model hasn't seen it before either. Let's see how the NER model is going to act when the provided sentence change:

• "I love Anwarvic"

>>> nlp = spacy.load("en_core_web_sm")
>>> sent = "I love Anwarvic"
>>> doc = nlp(sent)
>>> for i in doc.ents:
...     print(i.text," ",i.label_)
Anwarvic   PERSON

• "Anwarvic is gigantic"

>>> nlp = spacy.load("en_core_web_sm")
>>> sent = "Anwarvic is gigantic"
>>> doc = nlp(sent)
>>> for i in doc.ents:
...     print(i.text," ",i.label_)
Anwarvic   ORG

• "Anwarvic is awesome"

>>> nlp = spacy.load("en_core_web_sm")
>>> sent = "Anwarvic is awesome"
>>> doc = nlp(sent)
>>> for i in doc.ents:
...     print(i.text," ",i.label_)

As we can see, the extracted entites vary when the contextual structure of Anwarvic varies. So, in the first sentece the verb love is very common with people. That's why spaCy model predicted it as a PERSON . And the same happens with the second sentence where we use gigantic to describe organizations like ORG . In the third sentece, awesome is a pretty generic adjective that can be used to describe basically anything. That's why the spaCy NER model was confused.

Sidenote

Actually, when I ran the first provided code on my machine, it extracts both china and japan like so:

china   GPE
japan   GPE

Answer 2

NER works normally like this: You let a POS-Tagger tag your sentence with Part-of-Speech labels such as verbs, adjectives and proper nouns. The NER then looks at the nouns more directly. A POS-Tagger gets better the more information there is to classify POS-tags correctly. That is longer sentences, grammatically correct sentences, and correct spelling.

Your first example sent = "hello china hello japan" is short, without verbs etc. which makes it difficult for the tagger to classify POS-Tags. And another information is missing: countries are normally written Upper-case: Try sent = "hello China hello Japan" and it will work.

In your second example, the model detects china correctly even though it is lower case, because there is much more information in the whole sentence.

I recomment you to read more about POS-Tagging, it's quite fun!

Answer 3

The first answer to your question is context and that has already been relayed. This would be in common with other NLP libraries as well.

The second answer which is spaCy specific is non-determinism. spaCy uses various internal components which are dependent upon a random seed. As stated on this form post , it may be necessary to set the numpy and cupy seeds at minimum to get predictable results. It can well happen that one machine will give you a different output than another machine with the same code.