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/.
levitate
”?A common measure of string similarity is the Levenshtein distance, and the name was available on CRAN.
Install the released version from CRAN:
install.packages("levitate")
Alternatively, you can install the development version from Github:
::install_github("lewinfox/levitate") devtools
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.
<- "Episode IV - Star Wars: A New Hope"
x <- "Star Wars Episode IV - New Hope"
y
# 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.
<- "the quick brown fox jumps over the lazy dog"
x <- "my lazy dog was jumped over by a quick brown fox"
y
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
lev_weighted_token_ratio()
has a key limitation: tokens
will only be weighted if:
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.
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
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.