Word similarity is a number between 0 to 1 which tells us how close two words are, semantically. This is done by finding similarity between word vectors in the vector space. spaCy, one of the fastest NLP libraries widely used today, provides a simple method for this task.
spaCy’s Model –
spaCy supports two methods to find word similarity: using context-sensitive tensors, and using word vectors. Below is the code to download these models.
# Downloading the small model containing tensors. python -m spacy download en_core_web_sm # Downloading over 1 million word vectors. python -m spacy download en_core_web_lg
Below is the code to find word similarity, which can be extended to sentences and documents.
import spacy nlp = spacy.load('en_core_web_md') print("Enter two space-separated words")words = input() tokens = nlp(words) for token in tokens: # Printing the following attributes of each token. # text: the word string, has_vector: if it contains # a vector representation in the model, # vector_norm: the algebraic norm of the vector, # is_oov: if the word is out of vocabulary. print(token.text, token.has_vector, token.vector_norm, token.is_oov) token1, token2 = tokens[0], tokens[1] print("Similarity:", token1.similarity(token2)) |
Output:
cat True 6.6808186 False dog True 7.0336733 False Similarity: 0.80168545
The ‘en_core_web_md’ model yields vectors of dimension 300*1 for ‘dog’ and ‘cat’. One may also use the larger model, ‘en_vectors_web_lg’ which yields vectors of larger dimension for the same two words.
Using Custom Language Models –
By simply switching the language model, we can find a similarity between Latin, French or German documents. spaCy supports a total of 49 languages at present. spaCy also allows one to fix word vectors for words as per user need. Below is an example.
import spacyimport numpy as npfrom spacy.vocab import Vocab nlp = spacy.load('en_core_web_md')new_word = 'bucrest' print('Before custom setting')print(vocab.get_vector('bucrest')) custom_vector = np.random.uniform(-1, 1, (300, )) vocab.set_vector(new_word, custom_vector) print('After custom setting')print(vocab.get_vector('bucrest')) |
Output:
Before custom setting array([0., 0., 0., 0., 0., 0., 0., 0., --- ]) After custom setting array([ 0.68106073, 0.6037007, 0.9526876, -0.25600302, -0.24049562, --- ])
