| 1 |
;;; |
| 2 |
;;; util.levenshtein - Calculate edit distance between two sequences |
| 3 |
;;; |
| 4 |
;;; Copyright (c) 2016-2024 Shiro Kawai <[email protected]> |
| 5 |
;;; |
| 6 |
;;; Redistribution and use in source and binary forms, with or without |
| 7 |
;;; modification, are permitted provided that the following conditions |
| 8 |
;;; are met: |
| 9 |
;;; |
| 10 |
;;; 1. Redistributions of source code must retain the above copyright |
| 11 |
;;; notice, this list of conditions and the following disclaimer. |
| 12 |
;;; |
| 13 |
;;; 2. Redistributions in binary form must reproduce the above copyright |
| 14 |
;;; notice, this list of conditions and the following disclaimer in the |
| 15 |
;;; documentation and/or other materials provided with the distribution. |
| 16 |
;;; |
| 17 |
;;; 3. Neither the name of the authors nor the names of its contributors |
| 18 |
;;; may be used to endorse or promote products derived from this |
| 19 |
;;; software without specific prior written permission. |
| 20 |
;;; |
| 21 |
;;; THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS |
| 22 |
;;; "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT |
| 23 |
;;; LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR |
| 24 |
;;; A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT |
| 25 |
;;; OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, |
| 26 |
;;; SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED |
| 27 |
;;; TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR |
| 28 |
;;; PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF |
| 29 |
;;; LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING |
| 30 |
;;; NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS |
| 31 |
;;; SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. |
| 32 |
;;; |
| 33 |
|
| 34 |
(define-module util.levenshtein |
| 35 |
(use scheme.list) |
| 36 |
(use scheme.vector) |
| 37 |
(use gauche.array) |
| 38 |
(use gauche.sequence) |
| 39 |
(export l-distance l-distances |
| 40 |
re-distance re-distances |
| 41 |
dl-distance dl-distances)) |
| 42 |
(select-module util.levenshtein) |
| 43 |
|
| 44 |
;; Procedures to calculate Levenshtein, Damerau-Levenshtein, and |
| 45 |
;; Restricted edit distances of two sequences |
| 46 |
;; (not limited to strings). |
| 47 |
|
| 48 |
;; Note: |
| 49 |
;; Those widely known algorithms were extensively researched in 60s |
| 50 |
;; to 70s, and sometimes different researchers reported same or similar |
| 51 |
;; algorithms without attaching specific name. We took widely known |
| 52 |
;; names but the source we found didn't have that names. |
| 53 |
;; |
| 54 |
;; Levenshtein distance: Wagner & Fischer, The String-to-string correction |
| 55 |
;; problem, JACM 21(1), 1974, shows DP algorithm to calculate this, but |
| 56 |
;; there's no mention of "Levenshtein" in the paper. |
| 57 |
;; |
| 58 |
;; Damerau-Levenshtein-distance: Lowrance & Wagner, An extension of |
| 59 |
;; the string-to-string correction problem, JACM 22(2), 1975. Again, |
| 60 |
;; there's no mention of "Damerau" nor "Levenshtein". |
| 61 |
;; |
| 62 |
;; Restricted edit distance: Oommen & Loke, Pattern |
| 63 |
;; recognition of strings with substitutions, insertions, deletions |
| 64 |
;; and generalized transpositions, Pattern Recognition 30(5), 1997. |
| 65 |
;; This one is mentioned as Optimal string alignment distance in |
| 66 |
;; Wikipedia. The paper doesn't give itself a name, but claims it's |
| 67 |
;; "less restrictive" than Lowrance & Wagner since it removes the cost |
| 68 |
;; restriction that Lowrance & Wagner requires to work. On the other |
| 69 |
;; hand, this one restricts how transposition is applied and is not |
| 70 |
;; fully compatible to Damerau-Levenshtein. |
| 71 |
|
| 72 |
;; It is often explained using (N+k)x(M+k) array for dynamic programming |
| 73 |
;; (k=1 or 2), but we only need to refer to look back at most k rows, |
| 74 |
;; so we can run the algorithm with k+1 rows and rotating them. |
| 75 |
;; |
| 76 |
;; s e q u e n c e - A |
| 77 |
;; . . . . . . . . . . . . |
| 78 |
;; . . . . . . . . . . . . |
| 79 |
;; s . . . . . . . . . . . . |
| 80 |
;; e . . . . . . . . . . . . |
| 81 |
;; q . . . . . . . . . . . . "active" rows |
| 82 |
;; u . . . . . . . . . . . . < row-2 |
| 83 |
;; e . . . . . . . . . . . . < row-1 |
| 84 |
;; n . . . . . . . . X . . . < row-0 X @ (i,j) |
| 85 |
;; c . . . . . . . . . . . . |
| 86 |
;; e . . . . . . . . . . . . |
| 87 |
;; | . . . . . . . . . . . . |
| 88 |
;; B . . . . . . . . . . . . |
| 89 |
;; |
| 90 |
;; |
| 91 |
;; Three algorithms only differ slightly in details and it's tempting to |
| 92 |
;; factor the common part out. However such arrangement introduces overhead |
| 93 |
;; and the resulting code isn't that clean. So we decided to implement |
| 94 |
;; each separately, allowing some redundancy among their code. |
| 95 |
;; |
| 96 |
;; NB: Usually, the algorithm is described as sequences with 1-base index, |
| 97 |
;; and array index start from 0 or -1 up to the length of the sequence. |
| 98 |
;; We use 0-base vector, so the row index is offset by 1 or 2. |
| 99 |
;; |
| 100 |
;; The base implementation of every algorithm has the following signature: |
| 101 |
;; |
| 102 |
;; (base A Bs elt= cutoff) |
| 103 |
;; |
| 104 |
;; This calculates distance between sequence A and each sequence in the |
| 105 |
;; list of sequence Bs, and returns the list of distances. |
| 106 |
;; elt= is used to compare elements in the sequence. cutoff can be #f |
| 107 |
;; or a positive integer, and if it's positive integer, we stop calculating |
| 108 |
;; the distance of a particular pair as soon as the minimum distance |
| 109 |
;; exceeds the value; the distance for that pair becomes #f. |
| 110 |
|
| 111 |
;; Basic Levenshtein |
| 112 |
(define (l-base A Bs elt= cutoff) |
| 113 |
(let* ([alen (size-of A)] |
| 114 |
[A (if (length>=? Bs 3) |
| 115 |
(coerce-to <list> A) |
| 116 |
A)] |
| 117 |
[rows (circular-list (make-vector (+ alen 1)) |
| 118 |
(make-vector (+ alen 1)))]) |
| 119 |
(define (run B check) |
| 120 |
(vector-unfold! identity (cadr rows) 0 (+ alen 1)) |
| 121 |
(for-each-with-index |
| 122 |
(^[j b] |
| 123 |
(let ([row-0 (car rows)] |
| 124 |
[row-1 (cadr rows)] |
| 125 |
[dmin (+ j 1)]) |
| 126 |
(vector-set! row-0 0 (+ j 1)) |
| 127 |
(for-each-with-index |
| 128 |
(^[i a] |
| 129 |
(let1 d (min (+ (vector-ref row-0 i) 1) |
| 130 |
(+ (vector-ref row-1 (+ i 1)) 1) |
| 131 |
(+ (vector-ref row-1 i) (if (elt= a b) 0 1))) |
| 132 |
(vector-set! row-0 (+ i 1) d) |
| 133 |
(when (< d dmin) (set! dmin d)))) |
| 134 |
A) |
| 135 |
(when check (check dmin))) |
| 136 |
(pop! rows)) |
| 137 |
B) |
| 138 |
(rlet1 r (vector-ref (cadr rows) alen) |
| 139 |
(when check (check r)))) |
| 140 |
|
| 141 |
(define f |
| 142 |
(if cutoff |
| 143 |
(^[B] (let/cc break |
| 144 |
(run B (^d (if (< cutoff d) (break #f)))))) |
| 145 |
(cut run <> #f))) |
| 146 |
|
| 147 |
(map f Bs))) |
| 148 |
|
| 149 |
(define (l-distance A B :key (elt= eqv?) (cutoff #f)) |
| 150 |
(car (l-base A (list B) elt= cutoff))) |
| 151 |
|
| 152 |
(define (l-distances A Bs :key (elt= eqv?) (cutoff #f)) |
| 153 |
(l-base A Bs elt= cutoff)) |
| 154 |
|
| 155 |
;; Restricted Edit distance |
| 156 |
;; |
| 157 |
;; This one treats transposition as one operation, but does not allow |
| 158 |
;; editing a transposed pair afterwards. |
| 159 |
;; |
| 160 |
;; We need two columns/rows for the boundary, so when we're looking |
| 161 |
;; at A[i] and B[j], the matrix cell is (i+2, j+2) and we look at |
| 162 |
;; (i+1, j+2) - deletion |
| 163 |
;; (i+2, j+1) - insertion |
| 164 |
;; (i+1, j+1) - same or substitution |
| 165 |
;; (i, j) - transposition |
| 166 |
;; and j+2 = row-0, j+1 = row-1, j = row-2. |
| 167 |
|
| 168 |
(define (re-base A Bs elt= cutoff) |
| 169 |
(let* ([alen (size-of A)] |
| 170 |
[A (if (length>=? Bs 3) |
| 171 |
(coerce-to <vector> A) |
| 172 |
A)] |
| 173 |
[aref (cute (referencer A) A <>)] |
| 174 |
[rows (circular-list (make-vector (+ alen 2)) |
| 175 |
(make-vector (+ alen 2)) |
| 176 |
(make-vector (+ alen 2)))]) |
| 177 |
(define (run B check) |
| 178 |
(define blen (size-of B)) |
| 179 |
(define bref (cute (referencer B) B <>)) |
| 180 |
(define dmin-1 0) |
| 181 |
(vector-fill! (caddr rows) (+ alen blen)) |
| 182 |
(vector-unfold! (^i (if (zero? i) (+ alen blen) (- i 1))) |
| 183 |
(cadr rows) 0 (+ alen 2)) |
| 184 |
(for-each-with-index |
| 185 |
(^[j b] |
| 186 |
(let ([row-0 (car rows)] |
| 187 |
[row-1 (cadr rows)] |
| 188 |
[dmin-0 (+ j 1)]) |
| 189 |
(vector-set! row-0 0 (+ alen blen)) |
| 190 |
(vector-set! row-0 1 (+ j 1)) |
| 191 |
(for-each-with-index |
| 192 |
(^[i a] |
| 193 |
(let* ([d (min (+ (vector-ref row-0 (+ i 1)) 1) |
| 194 |
(+ (vector-ref row-1 (+ i 2)) 1) |
| 195 |
(+ (vector-ref row-1 (+ i 1)) (if (elt= a b) 0 1)))] |
| 196 |
[d (if (and (> i 1) (> j 1) |
| 197 |
(elt= (aref i) (bref (- j 1))) |
| 198 |
(elt= (aref (- i 1)) (bref j))) |
| 199 |
(min d (+ (vector-ref (caddr rows) i) |
| 200 |
(if (elt= a b) 0 1))) |
| 201 |
d)]) |
| 202 |
(when (< d dmin-0) (set! dmin-0 d)) |
| 203 |
(vector-set! row-0 (+ i 2) d))) |
| 204 |
A) |
| 205 |
(when check (check (min dmin-0 dmin-1))) |
| 206 |
(set! dmin-1 dmin-0)) |
| 207 |
(pop! rows) |
| 208 |
(pop! rows)) |
| 209 |
B) |
| 210 |
(rlet1 r (vector-ref (cadr rows) (+ alen 1)) |
| 211 |
(when check (check r)))) |
| 212 |
|
| 213 |
(define f |
| 214 |
(if cutoff |
| 215 |
(^[B] (let/cc break |
| 216 |
(run B (^d (if (< cutoff d) (break #f)))))) |
| 217 |
(cut run <> #f))) |
| 218 |
|
| 219 |
(map f Bs))) |
| 220 |
|
| 221 |
(define (re-distance A B :key (elt= eqv?) (cutoff #f)) |
| 222 |
(car (re-base A (list B) elt= cutoff))) |
| 223 |
|
| 224 |
(define (re-distances A Bs :key (elt= eqv?) (cutoff #f)) |
| 225 |
(re-base A Bs elt= cutoff)) |
| 226 |
|
| 227 |
;; Damerau-Levenshtein distance |
| 228 |
;; We need a way to look up the last character position seen in A. |
| 229 |
;; Typical DL algorithm uses an array indexed by character code, but |
| 230 |
;; that does not work with large character set or general objects. |
| 231 |
;; For the time being, we use an assoc list. It needs O(length(A)) |
| 232 |
;; to look up, but in general length(A) is small and the constant |
| 233 |
;; factor is very small for alist (we could go with tree map, for |
| 234 |
;; example, if an element comparator is provided, but the constant factor |
| 235 |
;; of it could surpass the log order benefit if length(A) is small.) |
| 236 |
|
| 237 |
(define (dl-base A Bs elt= cutoff) |
| 238 |
(let* ([alen (size-of A)] |
| 239 |
[A (if (length>=? Bs 3) |
| 240 |
(coerce-to <list> A) |
| 241 |
A)] |
| 242 |
[rows (let* ([maxb (apply max (map size-of Bs))] |
| 243 |
[rows (vector-tabulate (+ maxb 2) |
| 244 |
(^_ (make-vector (+ alen 2))))]) |
| 245 |
;; The second row is (almost) constant |
| 246 |
(vector-unfold! (^i (- i 1)) (vector-ref rows 1) 0 (+ alen 2)) |
| 247 |
rows)]) |
| 248 |
|
| 249 |
(define (run B) |
| 250 |
(define DA '()) ; alist to map element -> index in A (0-base) |
| 251 |
(define blen (size-of B)) |
| 252 |
(vector-unfold! (^_ (+ alen blen)) (vector-ref rows 0) 0 (+ alen 2)) |
| 253 |
(vector-for-each-with-index (^[j v] |
| 254 |
(unless (= j 0) |
| 255 |
(vector-set! v 0 (+ alen blen)) |
| 256 |
(vector-set! v 1 (- j 1)))) |
| 257 |
rows) |
| 258 |
(let-syntax |
| 259 |
([D (syntax-rules () |
| 260 |
[(_ i j) (vector-ref (vector-ref rows j) i)] |
| 261 |
[(_ i j v) (vector-set! (vector-ref rows j) i v)])]) |
| 262 |
(for-each-with-index |
| 263 |
(^[i a] |
| 264 |
(define DB -1) ; the last j such that A[i] == B[j] (0-base) |
| 265 |
(for-each-with-index |
| 266 |
(^[j b] |
| 267 |
(let* ([k (assoc-ref DA b -1 elt=)] |
| 268 |
[l DB] |
| 269 |
[cost (if (elt= a b) |
| 270 |
(begin (set! DB j) 0) |
| 271 |
1)] |
| 272 |
[d (min (+ (D (+ i 2) (+ j 1)) 1) |
| 273 |
(+ (D (+ i 1) (+ j 2)) 1) |
| 274 |
(+ (D (+ i 1) (+ j 1)) cost) |
| 275 |
(+ (D (+ k 1) (+ l 1)) |
| 276 |
(- i k) |
| 277 |
(- j l 1)))]) |
| 278 |
(D (+ i 2) (+ j 2) d))) |
| 279 |
B) |
| 280 |
(push! DA (cons a i))) |
| 281 |
A) |
| 282 |
(D (+ alen 1) (+ blen 1)))) |
| 283 |
|
| 284 |
;; Because of transposition propagation, early cutoff isn't trivial. |
| 285 |
;; for now, we just run the entire algorithm then check the result. |
| 286 |
(define f |
| 287 |
(if cutoff |
| 288 |
(^[B] (let1 r (run B) |
| 289 |
(and (<= r cutoff) r))) |
| 290 |
run)) |
| 291 |
|
| 292 |
(map f Bs))) |
| 293 |
|
| 294 |
(define (dl-distance A B :key (elt= eqv?) (cutoff #f)) |
| 295 |
(car (dl-base A (list B) elt= cutoff))) |
| 296 |
|
| 297 |
(define (dl-distances A Bs :key (elt= eqv?) (cutoff #f)) |
| 298 |
(dl-base A Bs elt= cutoff)) |