5

Consider these words:

reminder - remainder

They are very similar in sound and letters. Yet they are far from each other in meaning.

In computational linguistics, do we have algorithms to measure the amount of similarity between words?

4

The first type of similarity, phonetic and lexical similarity, can be measured using an edit distance metric, for instance the widely used Levenshtein distance. This basically uses an algorithm to determine the number of deletions and insertions of characters one has to make to get from the first word to the second word. The more similar the words, the smaller the edit distance. This is usually considered a quite simple and straight-forward thing to do.

Semantic similarity is a different beast. Now we are no longer interested in how a word looks on the surface, but what information a word brings to the table, and even more important, how to quantify this difference. This is a very active research area called Distributional Semantics that has been looked into in the computational linguistic community for at least two decades. Recently, this has also become popularized through the deep learning community, but now called word embeddings.

The basic idea is that you can represent a word as a vector in a vector space, and having words of similar meaning to be close to each other in this space. The words get their place in the vector space by having an algorithm reading large corpora and taking note of the context of each word. For instance, we can expect tea and coffee both having the word drink as a common neighbor and therefore this makes the distributional semantic model place them close to each other in the vector space.

This works well for some words using some algorithms some of the time, but there is still a lot of work to be done here.

2
  • 1
    Levenshtein distance measures orthographic similarity and doesn't measure phonetic similarity. For that, you would need to convert to a phonetic transcription and then do a symbol-to-feature mapping (to identify that there is a single difference between [θ] and [ð] but two differences between [θ] and [d]). – user6726 Dec 2 '15 at 20:41
  • @user6726 Right, thanks for clarifying! – Jimmy C Dec 2 '15 at 23:43

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.