How to find similar words with FastText?

Question

I am playing around with FastText, https://pypi.python.org/pypi/fasttext,which is quite similar to Word2Vec. Since it seems to be a pretty new library with not to many built in functions yet, I was wondering how to extract morphological similar words.

For eg: model.similar_word("dog") -> dogs. But there is no function built-in.

If I type model["dog"]

I only get the vector, that might be used to compare cosine similarity. model.cosine_similarity(model["dog"], model["dogs"]]).

Do I have to make some sort of loop and do cosine_similarity on all possible pairs in a text? That would take time ...!!!

Aerophyte

When fasttext.skipgram('train.txt','model') is run, it creates a .bin & .vec file. Use these generated files and follow the process mentioned in the accepted answer. – Aerometry

@Prometheus Any ideas how to do something similar in Java? – Primero

Nope. Have never touched Java. However FYI, the .bin and .vec files are cross compatible. – Aerometry

Answer 1

Use Gensim, load fastText trained .vec file with load.word2vec models and use most_similiar() method to find similar words!

Delighted

Is their any API in fasttext that allows one to input two words and then returns their cosine similarity? Say something like (car,vehicle) and then returns something like 0.8? – Gloriagloriana

Answer 2

You can install pyfasttext library to extract the most similar or nearest words to a particualr word.

from pyfasttext import FastText
model = FastText('model.bin')
model.nearest_neighbors('dog', k=2000)

Or you can get the latest development version of fasttext, you can install from the github repository :

import fasttext
model = fasttext.load_model('model.bin')
model.get_nearest_neighbors('dog', k=100)

Bloodthirsty

Answer 3

You can install and import gensim library and then use gensim library to extract most similar words from the model that you downloaded from FastText.

Use this:

import gensim
model = gensim.models.KeyedVectors.load_word2vec_format('model.vec')
similar = model.most_similar(positive=['man'],topn=10)

And by topn parameter you get the top 10 most similar words.

Scottiescottish

Answer 4

You should use gensim to load the model.vec and then get similar words:

m = gensim.models.Word2Vec.load_word2vec_format('model.vec')
m.most_similar(...)

Wolgast

Answer 5

Use gensim,

from gensim.models import FastText

model = FastText.load(PATH_TO_MODEL)
model.wv.most_similar(positive=['dog'])

More info here

Joeyjoffre

Answer 6

Fasttext has a method called get_nearest_neighbors. nearest neighbor queries. One needs the model's .bin file to use this.

Ansley