This Julia package computes various distances between AbstractStrings

Installation

The package is registered in the General registry and so can be installed at the REPL with ] add StringDistances.

Syntax

The function compare returns a similarity score between two strings. The function always returns a score between 0 and 1, with a value of 0 being completely different and a value of 1 being completely similar.

using StringDistances
compare("martha", "martha", Hamming())
#> 1.0
compare("martha", "marhta", Hamming())
#> 0.6666666666666667

Distances

Edit Distances

• Hamming Distance Hamming()
• Jaro Distance Jaro()
• Levenshtein Distance Levenshtein()
• Damerau-Levenshtein Distance DamerauLevenshtein()
• RatcliffObershelp Distance RatcliffObershelp()

Q-Grams Distances

Q-gram distances compare the set of all substrings of length q in each string.

• QGram Distance Qgram(q)
• Cosine Distance Cosine(q)
• Jaccard Distance Jaccard(q)
• Overlap Distance Overlap(q)
• Sorensen-Dice Distance SorensenDice(q)

Distance Modifiers

The package includes distance "modifiers", that can be applied to any distance.

• Winkler boosts the similary score of strings with common prefixes. The Winkler adjustment was originally defined for the Jaro similarity score but this package defines it for any string distance.

  compare("martha", "marhta", Jaro())
  #> 0.9444444444444445
  compare("martha", "marhta", Winkler(Jaro()))
  #> 0.9611111111111111
  
  compare("william", "williams", QGram(2))
  #> 0.9230769230769231
  compare("william", "williams", Winkler(QGram(2)))
  #> 0.9538461538461539
  

• Modifiers from the Python library fuzzywuzzy, that can be applied to any distance.

  • Partial returns the maximal similarity score between the shorter string and substrings of the longer string.

    compare("New York Yankees", "Yankees", Levenshtein())
    #> 0.4375
    compare("New York Yankees", "Yankees", Partial(Levenshtein()))
    #> 1.0
    

  • TokenSort adjusts for differences in word orders by reording words alphabetically.

    compare("mariners vs angels", "angels vs mariners", RatcliffObershelp())
    #> 0.44444
    compare("mariners vs angels", "angels vs mariners", TokenSort(RatcliffObershelp())
    #> 1.0
    

  • TokenSet adjusts for differences in word orders and word numbers by comparing the intersection of two strings with each string.

    compare("mariners vs angels", "los angeles angels at seattle mariners", Jaro())
    #> 0.559904
    compare("mariners vs angels", "los angeles angels at seattle mariners", TokenSet(Jaro()))
    #> 0.944444
    

  • TokenMax combines scores using the base distance, the Partial, TokenSort and TokenSet modifiers, with penalty terms depending on string lengths.

    compare("mariners vs angels", "los angeles angels at seattle mariners", TokenMax(RatcliffObershelp()))
    #> 0.855
    

Find

find_best returns the element of an iterator with the highest similarity score

find_best("New York", ["NewYork", "Newark", "San Francisco"], Levenshtein())
#> "NewYork"

find_all returns all the elements of an iterator with a similarity score higher than a minimum value (default to 0.8)

find_all("New York", ["NewYork", "Newark", "San Francisco"], Levenshtein(); min_score = 0.8)
#> 1-element Array{String,1}:
#> "NewYork"

While these functions are defined for any distance, they are particularly optimized for Levenshtein and DamerauLevenshtein distances (as well as their modifications via Partial, TokenSort, TokenSet, or TokenMax)

Compare vs Evaluate

The function compare returns a similarity score: a value of 0 means completely different and a value of 1 means completely similar. In contrast, the function evaluate returns the litteral distance between two strings, with a value of 0 being completely similar. Some distances are between 0 and 1, while others are unbouded.

compare("New York", "New York", Levenshtein())
#> 1.0
evaluate(Levenshtein(), "New York", "New York")
#> 0

Which distance should I use?

As a rule of thumb,

• Standardize strings before comparing them (cases, whitespaces, accents, abbreviations...)
• The distance Tokenmax(Levenshtein()) is a good choice to link sequence of words (adresses, names) across datasets.

References

• The stringdist Package for Approximate String Matching Mark P.J. van der Loo
• fuzzywuzzy