adverb forms using Wordnet

You might have stumbled in your NLP application development upon situations when you needed to get the “closest” adjective to a noun, or maybe you needed to “nounify” a verb. After poking around Wordnet I found a simple and pretty effective way to do this. Keep in mind that it is not error proof, but for most of my needs, I found it to perform pretty well. We’ll be using NLTK Wordnet wrapper for this. Let’s have a look at the code:

from nltk.corpus import wordnet as wn

# Just to make it a bit more readable

WN_NOUN = 'n'

WN_VERB = 'v'

WN_ADJECTIVE = 'a'

WN_ADJECTIVE_SATELLITE = 's'

WN_ADVERB = 'r'

def convert(word, from_pos, to_pos):

""" Transform words given from/to POS tags """

synsets = wn.synsets(word, pos=from_pos)

# Word not found

if not synsets:

return []

# Get all lemmas of the word (consider 'a'and 's' equivalent)

lemmas = [l for s in synsets

for l in s.lemmas

if s.name.split('.')[1] == from_pos

or from_pos in (WN_ADJECTIVE, WN_ADJECTIVE_SATELLITE)

and s.name.split('.')[1] in (WN_ADJECTIVE, WN_ADJECTIVE_SATELLITE)]

# Get related forms

derivationally_related_forms = [(l, l.derivationally_related_forms()) for l in lemmas]

# filter only the desired pos (consider 'a' and 's' equivalent)

related_noun_lemmas = [l for drf in derivationally_related_forms

for l in drf[1]

if l.synset.name.split('.')[1] == to_pos

or to_pos in (WN_ADJECTIVE, WN_ADJECTIVE_SATELLITE)

and l.synset.name.split('.')[1] in (WN_ADJECTIVE, WN_ADJECTIVE_SATELLITE)]

# Extract the words from the lemmas

words = [l.name for l in related_noun_lemmas]

len_words = len(words)

# Build the result in the form of a list containing tuples (word, probability)

result = [(w, float(words.count(w))/len_words) for w in set(words)]

result.sort(key=lambda w: -w[1])

# return all the possibilities sorted by probability

return result

This is pretty much it. Because most of the time, there is not just one correct answer to the question, I’ve decided to return a list with possible answers with attached probabilities. The probabilities are computed in a simple manner: no_times_lemma_appears/no_possible_answers. Let’s take a look at some results: