Let's say I have multiple tags that I need to compare semantically. For example:

tags = ['Python', 'Football', 'Programming', 'Handball', 'Chess', 'Cheese', 'board game']

I would like compare these tags (and many more) semantically to find a similarity value between 0 and 1. For example, I want to get values like these:

f('Chess','Cheese') = 0.0  # tags look similar, but means very different things
f('Chess', 'board game') = 0.9 # because chess is a board game
f('Football', 'Handball') = 0.3 # because both are sports with a ball
f('Python', 'Programming') = 0.9 # because Python is a programming language

So what is the state of the art approach to get a function f like this? I know that machine learning might be doing this, but this area is huge and overwhelming for me (on the first glance it looks like NLP focuses on other problems). So what would be the best approach for this specific problem?


It's a complex problem, the standard method is to train a model which represents this semantic similarity function using a large text corpus (i.e. not only the tags themselves), by considering that the meaning of a word is provided by its context words.

A more direct approach is to use a pre-trained word embeddings vectors that can be compared directly to find their similarity.

A similar idea is to use WordNet, which is a database of words with their semantic similarity to each other.

| improve this answer | |
  • $\begingroup$ Hi Erwan, thank you for your answer! I took a at WordNet, but I could not find out how to compare two words and get a similarity value out of it. Any idea? $\endgroup$ – DarkMath Jun 11 at 12:43
  • $\begingroup$ @DarkMath see for instance this example: nlpforhackers.io/wordnet-sentence-similarity $\endgroup$ – Erwan Jun 11 at 12:51
  • $\begingroup$ Very helpful, thank you! :) $\endgroup$ – DarkMath Jun 11 at 14:31

You can use word embeddings take a look at word2vec and glove. There are many good tutorials on word2vec, but very briefly word2vec represents a word as a vector, so that you can do a similarity comparison, and also allows addition and subtraction of vectors.

The embeddings are already available, for example here: https://nlp.stanford.edu/projects/glove/

Here is an example, using glove embeddings:

import numpy as np
from sklearn.metrics.pairwise import cosine_similarity

word_to_embeddings = dict()

with open("/path/to/glove.twitter.27B.25d.txt") as f:

    for line in f:
        word = line.split()[0]
        embeddings = np.asarray(line.split()[1:], dtype='float32')
        word_to_embeddings[word] = embeddings

w = word_to_embeddings

sim_1 = cosine_similarity([w["chess"]], [w["cheese"]])[0][0]
print(f"chess and cheese sim: {sim_1}")

sim_2 = cosine_similarity([w["chess"]], [w["board"] + w["game"]])[0][0]
print(f"chess and board + game sim: {sim_2}")

sim_3 = cosine_similarity([w["football"]], [w["handball"]])[0][0]
print(f"football and handball sim: {sim_3}")

sim_4 = cosine_similarity([w["python"]], [w["programming"]])[0][0]
print(f"python and programming sim: {sim_4}")

sim_5 = cosine_similarity([w["c++"]], [w["programming"]])[0][0]
print(f"c++ and programming sim: {sim_5}")


chess and cheese sim: 0.4501191973686218
chess and board + game sim: 0.7286249995231628
football and handball sim: 0.7467300891876221
python and programming sim: 0.7076062560081482
c++ and programming sim: 0.8262864351272583

The results are not going to be exactly what you expect, for example Python could also refer to a snake. You might want to play around with different embeddings, because they can give different results.

| improve this answer | |
  • $\begingroup$ Thank you very much! I'll look at it in detail! $\endgroup$ – DarkMath Jun 11 at 11:16

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.