levitate is based on the Python thefuzz (formerly fuzzywuzzy) package for fuzzy string matching. An R port of this already exists, but unlike fuzzywuzzyR, levitate is written entirely in R with no external dependencies on reticulate or Python. It also offers a couple of extra bells and whistles in the form of vectorised functions.

View the docs at https://www.lewinfox.com/levitate/.

Why “levitate”?

A common measure of string similarity is the Levenshtein distance, and the name was available on CRAN.

NOTE The default distance metric is Optimal String Alignment (OSA), not Levenshtein distance. This is the default method used by the stringdist package, which levitate uses for distance calculations. OSA allows transpositions whereas Levenshtein distance does not. To use Levenshtein distance pass method = "lv" to any lev_*() functions.

lev_distance("01", "10") # Transpositions allowed by the default `method = "osa"`
#> [1] 1

lev_distance("01", "10", method = "lv") # No transpositions
#> [1] 2

A full list of distance metrics is available in help("stringdist-metrics", package = stringdist).

Installation

Install the released version from CRAN:

install.packages("levitate")

Alternatively, you can install the development version from Github:

devtools::install_github("lewinfox/levitate")

Examples

lev_distance()

The edit distance is the number of additions, subtractions or substitutions needed to transform one string into another. Base R provides the adist() function to compute this. levitate provides lev_distance() which is powered by the stringdist package.

lev_distance("cat", "bat")
#> [1] 1

lev_distance("rat", "rats")
#> [1] 1

lev_distance("cat", "rats")
#> [1] 2

The function can accept vectorised input. Where the inputs have a length() greater than 1 the results are returned as a vector unless pairwise = FALSE, in which case a matrix is returned.

lev_distance(c("cat", "dog", "clog"), c("rat", "log", "frog"))
#> [1] 1 1 2

lev_distance(c("cat", "dog", "clog"), c("rat", "log", "frog"), pairwise = FALSE)
#>      rat log frog
#> cat    1   3    4
#> dog    3   1    2
#> clog   4   1    2

If at least one (or both) of the inputs is scalar (length 1) the result will be a vector. The elements of the vector are named based on the longer input (unless useNames = FALSE).

lev_distance(c("cat", "dog", "clog"), "rat")
#>  cat  dog clog 
#>    1    3    4

lev_distance("cat", c("rat", "log", "frog", "other"))
#>   rat   log  frog other 
#>     1     3     4     5

lev_distance("cat", c("rat", "log", "frog", "other"), useNames = FALSE)
#> [1] 1 3 4 5

lev_ratio()

More useful than the edit distance, lev_ratio() makes it easier to compare similarity across different strings. Identical strings will get a score of 1 and entirely dissimilar strings will get a score of 0.

This function behaves exactly like lev_distance():

lev_ratio("cat", "bat")
#> [1] 0.6666667

lev_ratio("rat", "rats")
#> [1] 0.75

lev_ratio("cat", "rats")
#> [1] 0.5

lev_ratio(c("cat", "dog", "clog"), c("rat", "log", "frog"))
#> [1] 0.6666667 0.6666667 0.5000000

lev_partial_ratio()

If a and b are different lengths, this function compares all the substrings of the longer string that are the same length as the shorter string and returns the highest lev_ratio() of all of them. E.g. when comparing "actor" and "tractor" we would compare "actor" with "tract", "racto" and "actor" and return the highest score (in this case 1).

lev_partial_ratio("actor", "tractor")
#> [1] 1

# What's actually happening is the max() of this result is being returned
lev_ratio("actor", c("tract", "racto", "actor"))
#> tract racto actor 
#>   0.2   0.6   1.0

lev_token_sort_ratio()

The inputs are tokenised and the tokens are sorted alphabetically, then the resulting strings are compared.

x <- "Episode IV - Star Wars: A New Hope"
y <- "Star Wars Episode IV - New Hope"

# Because the order of words is different the simple approach gives a low match ratio.
lev_ratio(x, y)
#> [1] 0.3529412

# The sorted token approach ignores word order.
lev_token_sort_ratio(x, y)
#> [1] 0.9354839

lev_token_set_ratio()

Similar to lev_token_sort_ratio() this function breaks the input down into tokens. It then identifies any common tokens between strings and creates three new strings:

x <- {common_tokens}
y <- {common_tokens}{remaining_unique_tokens_from_string_a}
z <- {common_tokens}{remaining_unique_tokens_from_string_b}

and performs three pairwise lev_ratio() calculations between them (x vs y, y vs z and x vs z). The highest of those three ratios is returned.

x <- "the quick brown fox jumps over the lazy dog"
y <- "my lazy dog was jumped over by a quick brown fox"

lev_ratio(x, y)
#> [1] 0.2916667

lev_token_sort_ratio(x, y)
#> [1] 0.6458333

lev_token_set_ratio(x, y)
#> [1] 0.7435897

lev_weighted_token_ratio()

The lev_weighted_*() family of functions work slightly differently from the others. They always tokenise their input, and they allow you to assign different weights to specific tokens. This allows you to exert some influence over parts of the input strings that are more or less interesting to you.

For example, maybe you’re comparing company names from different sources, trying to match them up.

lev_ratio("united widgets, ltd", "utd widgets, ltd") # Note the typos
#> [1] 0.8421053

These strings score quite highly already, but the "ltd" in each name isn’t very helpful. We can use lev_weighted_token_ratio() to reduce the impact of "ltd".

NOTE Because the tokenisation affects the score, we can’t compare the output of the lev_weighted_*() functions with the non-weighted versions. To get a baseline, call the weighted function without supplying a weights argument.

lev_weighted_token_ratio("united widgets, ltd", "utd widgets, ltd")
#> [1] 0.8125

lev_weighted_token_ratio("united widgets, ltd", "utd widgets, ltd", weights = list(ltd = 0.1))
#> [1] 0.7744361

De-weighting "ltd" has reduced the similarity score of the strings, which gives a more accurate impression of their similarity.

We can remove the effect of "ltd" altogether by setting its weight to zero.

lev_weighted_token_ratio("united widgets, ltd", "utd widgets, ltd", weights = list(ltd = 0))
#> [1] 0.7692308

lev_weighted_token_ratio("united widgets", "utd widgets")
#> [1] 0.7692308

De-weighting also works the other way - if the token to be weighted appears in one string but not the other, then de-weighting it increases the similarity score:

lev_weighted_token_ratio("utd widgets", "united widgets, ltd")
#> [1] 0.625

lev_weighted_token_ratio("utd widgets", "united widgets, ltd", weights = list(ltd = 0.1))
#> [1] 0.7518797

Limitations of token weighting

lev_weighted_token_ratio() has a key limitation: tokens will only be weighted if:

  • The token appears in the same position in both strings (i.e. it’s the first/second/third, etc. token in both)
  • OR the strings contain different numbers of tokens, and the corresponding token position in the other string is empty.

This is probably easiest to see by example.

lev_weighted_token_ratio("utd widgets limited", "united widgets, ltd")
#> [1] 0.65
lev_weighted_token_ratio("utd widgets limited", "united widgets, ltd", weights = list(ltd = 0.1, limited = 0.1))
#> [1] 0.65

In this case the weighting has had no effect. Why not? Internally, the function has tokenised the strings as follows:

token_1 token_2 token_3
“utd” “widgets” “limited”
“united” “widgets” “ltd”

Because the token "ltd" doesn’t appear in the same position in both strings, the function doesn’t apply any weights.

This is a deliberate decision; while in the example above it’s easy to say “well, clearly ltd and limited are the same thing so we ought to weight them”, how should we handle a less clear example?

lev_weighted_token_ratio("green eggs and ham", "spam spam spam spam")
#> [1] 0.1176471
lev_weighted_token_ratio("green eggs and ham", "spam spam spam spam", weights = list(spam = 0.1, eggs = 0.5))
#> [1] 0.1176471

In this case it’s hard to say what the “correct” approach would be. There isn’t a meaningful way of applying weights to dissimilar tokens. In situations like “ltd”/“limited”, a pre-cleaning or standardisation process might be helpful, but that is outside the scope of what levitate offers.

I recommend exploring lev_weighted_token_sort_ratio() and lev_weighted_token_set_ratio() as they may give more useful results for some problems. Remember, weighting is going to be most useful when compared to the unweighted output of the same function.

Ranking functions

A common problem in this area is “given a string x and a set of strings y, which string in y is most / least similar to x?”. levitate provides two functions to help with this: lev_score_multiple() and lev_best_match().

lev_score_multiple() returns a ranked list of candidates. By default the highest-scoring is first.

lev_score_multiple("bilbo", c("gandalf", "frodo", "legolas"))
#> $frodo
#> [1] 0.2
#> 
#> $legolas
#> [1] 0.1428571
#> 
#> $gandalf
#> [1] 0

lev_best_match() returns the best matched string without any score information.

lev_best_match("bilbo", c("gandalf", "frodo", "legolas"))
#> [1] "frodo"

Both functions take a .fn argument which allows you to select a different ranking function. The default is lev_ratio() but you can pick another or write your own. See ?lev_score_multiple for details.

You can also reverse the direction of sorting by using decreasing = FALSE. This reverses the sort direction so lower scoring items are preferred. This may be helpful if you’re using a distance rather than a similarity measure, or if you want to return least similar strings.

lev_score_multiple("bilbo", c("gandalf", "frodo", "legolas"), decreasing = FALSE)
#> $gandalf
#> [1] 0
#> 
#> $legolas
#> [1] 0.1428571
#> 
#> $frodo
#> [1] 0.2

Porting code from thefuzz or fuzzywuzzyR

Results differ between levitate and thefuzz, not least because stringdist offers several possible similarity measures. Be careful if you are porting code that relies on hard-coded or learned cutoffs for similarity measures.