Agfdhyk

I have a set of keywords I am already matching for. It is a medical context so I've made up a equivalent scenario at least for the parsing I'm trying to do:

I have a car with chrome 1000-inch rims.

Let's say I want to return as a phrase all children words/tokens of the keyword rims where rims is already marked as an entity by SpaCy as CARPART.

In python this is what I'm doing:

test_phrases = nlp("""I have a car with chrome 100-inch rims.""")

print(test_phrases.cats)

for t in test_phrases:

    print('Token: {} || POS: {} || DEP: {} CHILDREN: {} || ent_type: {}'.format(t,t.pos_,t.dep_,[c for c in t.children],t.ent_type_))



Token: I || POS: PRON || DEP: nsubj CHILDREN:  || ent_type: 

Token: have || POS: VERB || DEP: ROOT CHILDREN: [I, car, .] || 

ent_type: 

Token: a || POS: DET || DEP: det CHILDREN:  || ent_type: 

Token: car || POS: NOUN || DEP: dobj CHILDREN: [a, with] || ent_type: 

Token: with || POS: ADP || DEP: prep CHILDREN: [rims] || ent_type: 

Token: chrome || POS: ADJ || DEP: amod CHILDREN:  || ent_type: 

Token: 100-inch || POS: NOUN || DEP: compound CHILDREN:  || ent_type: 

Token: rims || POS: NOUN || DEP: pobj CHILDREN: [chrome, 100-inch] || 

ent_type: 

Token: . || POS: PUNCT || DEP: punct CHILDREN:  || ent_type: CARPART

So, what I want to do is use is something like:

test_matcher = Matcher(nlp.vocab)



test_phrase = ['']

patterns = [[{'ENT':'CARPART',????}] for kp in test_phrase]

test_matcher.add('CARPHRASE', None, *patterns)

call the test_matcher on test_doc have it return:

chrome 100-inch rims

I think I found a satisfactory solution that will work when creating a Spacy Class object. You can test this out to make sure it works with your solution then add to something like this in Spacy pipeline:

from spacy.matcher import Matcher



keyword_list = ['rims']

patterns = [[{'LOWER':kw}] for kw in keyword_list]



test_matcher.add('TESTPHRASE',None, *patterns)





 def add_children_matches(doc,keyword_matcher):

     '''Add children to match on original single-token keyword.'''

    matches = keyword_matcher(doc)

    for match_id, start, end in matches:

        tokens = doc[start:end]

        print('keyword:',tokens)

        # Since we are getting children for keyword, there should only be one token

        if len(tokens) != 1:

            print('Skipping {}. Too many tokens to match.'.format(tokens))

            continue

        keyword_token = tokens[0]

        sorted_children = sorted([c.i for c in keyword_token.children] + [keyword_token.i],reverse=False)

        print('keyphrase:',doc[min(sorted_children):max(sorted_children)+1])







doc = nlp("""I have a car with chrome 1000-inch rims.""")

add_children_matches(doc,test_matcher)

This gives:

keyword: rims

keyphrase: chrome 1000-inch rims

Edit: To fully answer my own question you'd have to use something like:

 def add_children_matches(doc,keyword_matcher):

     '''Add children to match on original single-token keyword.'''

    matches = keyword_matcher(doc)

    spans = 

    for match_id, start, end in matches:

        tokens = doc[start:end]

        print('keyword:',tokens)

        # Since we are getting children for keyword, there should only be one token

        if len(tokens) != 1:

            print('Skipping {}. Too many tokens to match.'.format(tokens))

            continue

        keyword_token = tokens[0]

        sorted_children = sorted([c.i for c in keyword_token.children] + [keyword_token.i],reverse=False)

        print('keyphrase:',doc[min(sorted_children):max(sorted_children)+1])





    start_char = doc[min(sorted_children):max(sorted_children)+1].start_char

    end_char = doc[min(sorted_children):max(sorted_children)+1].end_char



    span = doc.char_span(start_char, end_char,label='CARPHRASE')

    if span != None:

        spans.append(span)



    return doc