Using fuzzy string matching to find duplicate tags

I’m a big fan of keyword tagging as a way to organise my digital data. (For an explanation why, see my post about how I scan and store my paperwork.)

Tags work best if I use them consistently – same spelling, same wording, same everything. In practice, I don’t always get that right – I make typos, mistakes, or I forget that I already have a tag for a particular concept, and I create another tag. Here’s a few similar tags from my notes folder:

Amazon S3

books I want to read
books to read
list: books to read

books I've read
books i’ve read

Anybody can see that these tags mean the same thing, but if I search for any one of them I won’t find anything tagged with the alternative spellings.

I could consolidate each of these into a single tag – but I can only do that if I know I’ve created these similar tags. How do I find these inconsistencies? I often have hundreds of tags, and looking through the list by hand isn’t practical.

This isn’t a new problem, and I found a great article from Seatgeek (a ticketing platform) that solves something very similar. They were trying to find similar descriptions of the same event. The post explains various approaches to comparing text – string similarity, partial string similarity, token ratios, and so on – and then they packaged their work in a library called FuzzyWuzzy.

There’s a variant of FuzzyWuzzy called RapidFuzz, which does similar calculations but performs a lot faster – anecdotally, I see a 10× speedup when using RapidFuzz over FuzzyWuzzy.

To find all possible pairs of tags in a collection, I use itertools.combinations.

Putting itertools and RapidFuzz together, here’s the function I use:

import itertools

from rapidfuzz import fuzz

def find_similar_pairs(tags, *, required_similarity=80):
    Find pairs of similar-looking tags in the collection ``tags``.

    Increase ``required_similarity`` for stricter matching (=> less results).
    for t1, t2 in itertools.combinations(sorted(tags), 2):
        if fuzz.ratio(t1, t2) > required_similarity:
            yield (t1, t2)

The required similarity often needs a bit of tuning. Sometimes I do have tags that look very similar but are genuinely different (e.g. mental health and dental health), and I don’t want too many of those in the results – but I do want to find all the duplicates. I start at 80%, but I often adjust up or down to get less or more results.

Comparing every pair is O(N2) complexity in the number of tags. This could be slow with lots of tags, but my personal collections don’t have enough tags for that to be an issue. Comparing the ~1100 tags in my Pinboard account takes 1.5s on a fairly old laptop; it’s plenty fast enough.

When I call this function, I usually have a tally-like dictionary as input. For each similar pair, I print the number of times I’ve used the tag, which helps me see what the canonical spelling of a duplicate tag is. If I’ve used one spelling once, and one spelling 50 times, I know which way to correct.

if __name__ == "__main__":
    tags = get_tags()  # {"Amazon S3": 50, "amazon-s3": 1, …}

    for t1, t2 in find_similar_pairs(tags):
        print("%3d\t%s" % (tags[t1], t1))
        print("%3d\t%s" % (tags[t2], t2))

This sort of cleanup task makes my tagging system more powerful. Having a consistent set of tags makes it easier to find things. Having an automated script makes this sort of cleanup task possible.

If you’re dealing with a pile of messy, human-generated data, I can recommend using something like FuzzyWuzzy or RapidFuzz to cut through the noise.