Finding the Word Analogy from given words using Word2Vec embeddings

In many placement exam rounds, we often encounter a basic question to find word analogies. In the word analogy task, we complete the sentence “a is to b as c is to ___ “, which is often represented as a : b :: c : d and we have to find the word ‘d’. A  sample question can be like: ‘man is to woman as king is to ___‘. 

The human brain can recognize that the blank must be filled with the word ‘queen‘. But for a machine to understand this pattern and fill the blank with the most appropriate word requires a lot of training to be done. What if we can use a Machine Learning algorithm to automate this task of finding the word analogy. In this tutorial, we will be using Word2Vec model and a pre-trained model named ‘GoogleNews-vectors-negative300.bin‘ which is trained on over 50 Billion words by Google. Each word inside the pre-trained dataset is embedded in a 300-dimensional space and the words which are similar in context/meaning are placed closer to each other in the space.

Methodology to find out the analogous word:

In this problem, our goal is to find a word d, such that the associated word vectors va, vb, vc, vd are related to each other in the following relationship: ‘vb – va = vd – vc‘. We will measure the similarity between vb-va and vd-vc using cosine similarity.

Importing important libraries:

We need to install an additional gensim library, to use word2vec model, to install gensim use the command ‘pip install gensim‘ on your terminal/command prompt.

Python3

import numpy as np
import gensim
from gensim.models import word2vec,KeyedVectors
from sklearn.metrics.pairwise import cosine_similarity

Loading the word vectors using the pre-trained model:

Python3

vector_word_notations = KeyedVectors.load_word2vec_format('GoogleNews-vectors-negative300.bin',binary=True)

Defining a function to predict analogous word:

Python3

def analogous_word(word_1,word_2,word_3,vector_word_notations):
    ''' The function accepts a triad of words, word_1, word_2, word_3 and returns word_4 such that word_1:word_2::word_3:word_4 '''

    # converting each word to its lowercase
    word_1,word_2,word_3 = word_1.lower(),word_2.lower(),word_3.lower()

    # Similarity between |word_2-word_1| = |word_4-word_3| should be maximum
    maximum_similarity = -99999

    word_4 = None

    words = vector_word_notations.vocab.keys()

    va,vb,vc = vector_word_notations[word_1],\
    vector_word_notations[word_2],vector_word_notations[word_3]

    # to find word_4 such that similarity
    # (|word_2 - word_1|, |word_4 - word_3|) should be maximum

    for i in words:
        if i in [word_1,word_2,word_3]:
            continue

        wvec = vector_word_notations[i]
        similarity = cosine_similarity(,[wvec-vc])

        if similarity > maximum_similarity:
            maximum_similarity = similarity
            word_4 = i

    return word_4

Testing our model:

Python3

triad_1 = ("Man","Woman","King")
# *triad_1 is written to unpack the elements in the tuple
output = analogous_word(*triad_1,word_vectors)
print(output)

# The output will be shown as queen