You cannot select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
141 lines
4.8 KiB
Plaintext
141 lines
4.8 KiB
Plaintext
.. Copyright (C) 2001-2020 NLTK Project
|
|
.. For license information, see LICENSE.TXT
|
|
|
|
=======================================
|
|
Demonstrate word embedding using Gensim
|
|
=======================================
|
|
|
|
We demonstrate three functions:
|
|
- Train the word embeddings using brown corpus;
|
|
- Load the pre-trained model and perform simple tasks; and
|
|
- Pruning the pre-trained binary model.
|
|
|
|
>>> import gensim
|
|
|
|
---------------
|
|
Train the model
|
|
---------------
|
|
|
|
Here we train a word embedding using the Brown Corpus:
|
|
|
|
>>> from nltk.corpus import brown
|
|
>>> model = gensim.models.Word2Vec(brown.sents())
|
|
|
|
It might take some time to train the model. So, after it is trained, it can be saved as follows:
|
|
|
|
>>> model.save('brown.embedding')
|
|
>>> new_model = gensim.models.Word2Vec.load('brown.embedding')
|
|
|
|
The model will be the list of words with their embedding. We can easily get the vector representation of a word.
|
|
>>> len(new_model['university'])
|
|
100
|
|
|
|
There are some supporting functions already implemented in Gensim to manipulate with word embeddings.
|
|
For example, to compute the cosine similarity between 2 words:
|
|
|
|
>>> new_model.similarity('university','school') > 0.3
|
|
True
|
|
|
|
---------------------------
|
|
Using the pre-trained model
|
|
---------------------------
|
|
|
|
NLTK includes a pre-trained model which is part of a model that is trained on 100 billion words from the Google News Dataset.
|
|
The full model is from https://code.google.com/p/word2vec/ (about 3 GB).
|
|
|
|
>>> from nltk.data import find
|
|
>>> word2vec_sample = str(find('models/word2vec_sample/pruned.word2vec.txt'))
|
|
>>> model = gensim.models.KeyedVectors.load_word2vec_format(word2vec_sample, binary=False)
|
|
|
|
We pruned the model to only include the most common words (~44k words).
|
|
|
|
>>> len(model.vocab)
|
|
43981
|
|
|
|
Each word is represented in the space of 300 dimensions:
|
|
|
|
>>> len(model['university'])
|
|
300
|
|
|
|
Finding the top n words that are similar to a target word is simple. The result is the list of n words with the score.
|
|
|
|
>>> model.most_similar(positive=['university'], topn = 3)
|
|
[('universities', 0.70039...), ('faculty', 0.67809...), ('undergraduate', 0.65870...)]
|
|
|
|
Finding a word that is not in a list is also supported, although, implementing this by yourself is simple.
|
|
|
|
>>> model.doesnt_match('breakfast cereal dinner lunch'.split())
|
|
'cereal'
|
|
|
|
Mikolov et al. (2013) figured out that word embedding captures much of syntactic and semantic regularities. For example,
|
|
the vector 'King - Man + Woman' is close to 'Queen' and 'Germany - Berlin + Paris' is close to 'France'.
|
|
|
|
>>> model.most_similar(positive=['woman','king'], negative=['man'], topn = 1)
|
|
[('queen', 0.71181...)]
|
|
|
|
>>> model.most_similar(positive=['Paris','Germany'], negative=['Berlin'], topn = 1)
|
|
[('France', 0.78840...)]
|
|
|
|
We can visualize the word embeddings using t-SNE (http://lvdmaaten.github.io/tsne/). For this demonstration, we visualize the first 1000 words.
|
|
|
|
| import numpy as np
|
|
| labels = []
|
|
| count = 0
|
|
| max_count = 1000
|
|
| X = np.zeros(shape=(max_count,len(model['university'])))
|
|
|
|
|
| for term in model.vocab:
|
|
| X[count] = model[term]
|
|
| labels.append(term)
|
|
| count+= 1
|
|
| if count >= max_count: break
|
|
|
|
|
| # It is recommended to use PCA first to reduce to ~50 dimensions
|
|
| from sklearn.decomposition import PCA
|
|
| pca = PCA(n_components=50)
|
|
| X_50 = pca.fit_transform(X)
|
|
|
|
|
| # Using TSNE to further reduce to 2 dimensions
|
|
| from sklearn.manifold import TSNE
|
|
| model_tsne = TSNE(n_components=2, random_state=0)
|
|
| Y = model_tsne.fit_transform(X_50)
|
|
|
|
|
| # Show the scatter plot
|
|
| import matplotlib.pyplot as plt
|
|
| plt.scatter(Y[:,0], Y[:,1], 20)
|
|
|
|
|
| # Add labels
|
|
| for label, x, y in zip(labels, Y[:, 0], Y[:, 1]):
|
|
| plt.annotate(label, xy = (x,y), xytext = (0, 0), textcoords = 'offset points', size = 10)
|
|
|
|
|
| plt.show()
|
|
|
|
------------------------------
|
|
Prune the trained binary model
|
|
------------------------------
|
|
|
|
Here is the supporting code to extract part of the binary model (GoogleNews-vectors-negative300.bin.gz) from https://code.google.com/p/word2vec/
|
|
We use this code to get the `word2vec_sample` model.
|
|
|
|
| import gensim
|
|
| from gensim.models.word2vec import Word2Vec
|
|
| # Load the binary model
|
|
| model = Word2Vec.load_word2vec_format('GoogleNews-vectors-negative300.bin.gz', binary = True);
|
|
|
|
|
| # Only output word that appear in the Brown corpus
|
|
| from nltk.corpus import brown
|
|
| words = set(brown.words())
|
|
| print (len(words))
|
|
|
|
|
| # Output presented word to a temporary file
|
|
| out_file = 'pruned.word2vec.txt'
|
|
| f = open(out_file,'wb')
|
|
|
|
|
| word_presented = words.intersection(model.vocab.keys())
|
|
| f.write('{} {}\n'.format(len(word_presented),len(model['word'])))
|
|
|
|
|
| for word in word_presented:
|
|
| f.write('{} {}\n'.format(word, ' '.join(str(value) for value in model[word])))
|
|
|
|
|
| f.close()
|