use custom NER webservice to set_entities in doc

[aakashb95]

Hi,

I have a webservice for NER that gives me:
(entity, entity_label, (entity_char_idx))
for example:
(John Snow , PERSON, (0,8))

How can i use doc.set_ents to set the entity_label from my webservice to tokens in spacy document? I want to do this so that i can access them with token.ent_type_ for my downstream tasks.

[ljvmiranda921]

Hi @aakashb95 !

You can use char_span so that you can slice through character indices.
Also, the character offsets are incorrect: for start it should be the index of the first character, but for end, it must be the last character after the span:

import spacy

data = [
    ("John Snow", "PERSON", (0, 9)),  # last character after the span
]

nlp = spacy.blank("en")
for text, label, (start, end) in data:
    doc = nlp(text)
    span = doc.char_span(start, end, label=label)
    doc.ents = [span]  

Usually, you'd have multiple entities in a given document, that's why we pass a list to doc.ents. You can adapt the snippet above to fit your use-case 👍 Also, you can check the char_span docs for more info on the API.

[aakashb95]

Thanks for the snippet. I have a follow up question.

Consider:
example = "Mohamed Salah, the 'Egyptian King' scored a hat-trick at Old Trafford. Manchester United defenders had no answer to his movements around the penalty area. On the other hand, Cristiano Ronaldo, the Portuguese superstar, had a goal disallowed. He had very few chances in the game."

I am using AllenNLP's coreference resolution module to identify mentions that are coreferent to each other:

# pip install allennlp==2.1.0 allennlp-models==2.1.0
from allennlp.predictors.predictor import Predictor
import allennlp_models.tagging
import spacy

nlp = spacy.load("en_core_web_sm")

predictor = Predictor.from_path("https://storage.googleapis.com/allennlp-public-models/coref-spanbert-large-2021.03.10.tar.gz")
res = predictor.predict(example)

clusters = res['clusters']
doc = nlp(example)

all_clusters = []
for cluster_idx in clusters:
    temp_clusters = []
    for coref_idx in cluster_idx:
        start_word_idx, end_word_idx = coref_idx
        mention = doc[start_word_idx : end_word_idx + 1]
        temp_clusters.append(mention)
    all_clusters.append(clusters)

Using the above code I get:

all_clusters = [[Mohamed Salah, the 'Egyptian King', his],
 [Cristiano Ronaldo, the Portuguese superstar, He]]

the mentions in these clusters are all spacy Spans.
Now, using my NER webservice, I get the following output:

[('Old Trafford', 'FAC', (57, 69)),
 ('Portuguese', 'NORP', (197, 207)),
 ('Manchester United', 'ORG', (71, 88)),
 ('Mohamed Salah', 'PERSON', (0, 13)),
 ('Cristiano Ronaldo', 'PERSON', (174, 191))]

Now I want to use my NER output and using doc.set_ent, and set the entities to those mentions that are inside of coreference clusters. I want to preserve the token word index given by spacy whilst attaching the named entity from my custom webservice.

In your snippet

for text, label, (start, end) in data:
    doc = nlp(text)

we're creating doc for the text identified by my NER, whereas I want to set the entities in the original example text.
Instead of running the webservice again for my downstream tasks later on, I want to be able to see ('Cristiano Ronaldo', 'PERSON') in doc.ents because I manually set them.

How should i use char_span in this scenario? Please let me know if you need any clarifications.

P.S I don't want to use spacy's inbuilt NER and don't have the bandwidth to train spacy NER since we already have a webservice

[ljvmiranda921]

Hi @aakashb95 ,

we're creating doc for the text identified by my NER, whereas I want to set the entities in the original example text.
Instead of running the webservice again for my downstream tasks later on, I want to be able to see ('Cristiano Ronaldo', 'PERSON') in doc.ents because I manually set them.

It's possible to create the doc once, or reuse the one you already have. Loop through the entities from the NER service, then attach them into doc.ents afterwards:

import spacy

data = [
 ('Old Trafford', 'FAC', (57, 69)),
 ('Portuguese', 'NORP', (197, 207)),
 ('Manchester United', 'ORG', (71, 88)),
 ('Mohamed Salah', 'PERSON', (0, 13)),
 ('Cristiano Ronaldo', 'PERSON', (174, 191))
]

original_text = "...."
nlp = spacy.load("en_core_web_sm")
doc = nlp(original_text)

ents = []
for text, label, (start, end) in data:
    span = doc.char_span(start, end, label=label)
    ents.append(span)
doc.ents = ents

If you want to set it on the spans returned by your clusters, it may be a bit tricky since you have to realign your offsets. It may be possible to just do a heuristic based on the length of the word (i.e., start is always 0, and end is len(word)+1), but may not be foolproof.

[aakashb95]

This is great. Thanks. It worked on smaller sentences and articles but failed with E[1010] on a text which had NERs like so:

('G Kishan Reddy', 'PERSON', [1151, 1165]),
('Reddy', 'PERSON', [1160, 1165])

[E1010] Unable to set entity information for token 223 which is included in more than one span in entities, blocked, missing or outside.

Would you suggest i write a checking utility based on start and end character indexes to eliminate the smaller of the two spans? Is there any other way to bypass this?

[ljvmiranda921]

For this, I suggest looking into various alignment modes in char_span. Or most likely, check what's wrong in the annotations.

[aakashb95]

I just checked this discussion:
#4966 (comment)

How can i incorporate this for my example?

[ljvmiranda921]

Hi @aakashb95 ! This follow-up is starting to stray from the original question. For overlapping entities, you can check SpanGroup. The discussion you linked was from v2, and SpanGroups aren't implemented at that time.

[aakashb95]

Sorry, I thought it was a sub-problem of my original question as it dealt with overlapping entities.
You have been super helpful!
If i face issues, i'll raise a different discussion on this specific topic. Thanks!