Poor NER performance?

[stajdini]

I am using Stanza to identify NER is short pieces of text like (business names/brand names). Here is one example:

 # BUILDING THE MODELS
#-----stanza
sen = stanza.Pipeline ("en")
smlp = stanza.MultilingualPipeline()

# TESTING THE MODELS
name = 'The Port of Peri Peri'

print('stanza sen')
doc = sen(name)
for sent in doc.sentences:
   for token in sent.tokens:
       for word in token.words:
print('-----------------')

print('stanza smlp')
doc = smlp(name)
for sent in doc.sentences:
   for token in sent.tokens:
       for word in token.words:
           print(word.text, word.xpos, word.upos, word.deprel, token.ner)#, word.feats)
print('-----------------')


stanza sen
The DT DET det B-PERSON
Port NNP PROPN root I-PERSON
of IN ADP case I-PERSON
Peri NNP PROPN nmod I-PERSON
Peri NNP PROPN nmod E-PERSON
-----------------
stanza smlp
The DT DET det B-PERSON
Port NNP PROPN root I-PERSON
of IN ADP case I-PERSON
Peri NNP PROPN nmod I-PERSON
Peri NNP PROPN nmod E-PERSON
-----------------

Obviously, 'The Port of Peri Peri' does not even look like a person's name. Is there any way I can improve Stanza's performance?

[AngledLuffa]

If you gather up a few such examples, we can retrain the model with those examples and get them right in the future. Please bear in mind that statistical models will have some sort of errors in them