Discovering taxonomic relationships with Spacy - spacy

Question

How can we make some generic inference of taxonomic relation between entities from text? Looking for words near 'type of' in the word2vec of en_core_web_lg model, they all seem unrelated. The words near 'type' however are more similar to it. But how can I use common phrases in my text and apply some generic similarity for inferring taxonomy from SVO triples etc.? Can do a Sense2Vec type approach, but wondering if something existing can be used without new training.
Output of code below:
['eradicate', 'wade', 'equator', 'educated', 'lcd', 'byproducts', 'two', 'propensity', 'rhinos', 'procrastinate']
def get_related(word):
filtered_words = [w for w in word.vocab if w.is_lower == word.is_lower and w.prob >= -15]
similarity = sorted(filtered_words, key=lambda w: word.similarity(w), reverse=True)
return similarity[:10]
print ([w.lower_ for w in get_related(nlp.vocab[u'type_of'])])

Answer 1

All the similarities your code retrieves are 0.0, so sorting the list has no effect.
You are treating "type_of" as a word (more accurately, a lexeme), and assuming spaCy will understand it as the phrase "type of". Note that the first has an underscore, while the second one does not; however even without the underscore, it is not a lexeme in the model's vocabulary. Since the model does not have sufficient data on "type_of" for a similarity score, the score is 0.0 for every word you compare to it.
Instead, you can create a Span of the words "type of" and call similarity() on that. This requires only a small change to your code:
import spacy
def get_related(span): # this now expects a Span instead of a Lexeme
filtered_words = [w for w in span.vocab if
w.is_lower == span.text.islower()
and w.prob >= -15] # filter by probability and case
# (use the lowercase words if and only if the whole Span is in lowercase)
similarity = sorted(filtered_words,
key=lambda w: span.similarity(w),
reverse=True) # sort by the similarity of each word to the whole Span
return similarity[:10] # return the 10 most similar words
nlp = spacy.load('en_core_web_lg') # load the model
print([w.lower_ for w in get_related(nlp(u'type')[:])]) # print related words for "type"
print([w.lower_ for w in get_related(nlp(u'type of')[:])]) # print related words for "type of"
Output:
['type', 'types', 'kind', 'sort', 'specific', 'example', 'particular', 'similar', 'different', 'style']
['type', 'of', 'types', 'kind', 'particular', 'sort', 'different', 'such', 'same', 'associated']
As you can see, all the words are related to the input to some degree, and the output is similar but not identical for "type" and "type of".