Similar string search in Levenshtein-distance

Install

from source

git clone https://github.com/lovit/levenshtein_finder
cd levenshtein_finder
python setup.py install

from PyPI

pip install levenshtein_finder

Usage

Define Tokenizer

levenshtein_finder provides two basic tokenizers.

In [1]

from levenshtein_finder import CharacterTokenizer, BigramTokenizer

char_tokenizer = CharacterTokenizer()
bigram_tokenizer = BigramTokenizer()

char_tokenizer.tokenize("abc de")  # ['a', 'b', 'c', ' ', 'd', 'e']
bigram_tokenizer.tokenize("abc de")  # ['ab', 'bc', 'c ', ' d', 'de']

char_tokenizer.train("path/to/text")  # or
char_tokenizer.train(["list", "of", "words"])

When you use word piece tokenizer in Huggingface's tokenizers, wrap the class instance with WordpieceTokenizersWrapper

In [2]

from tokenizers import BertWordPieceTokenizer
from levenshtein_finder import WordpieceTokenizersWrapper

tokenizer = WordpieceTokenizersWrapper(BertWordPieceTokenizer())
tokenizer.train("path/to/text")
tokenizer.tokenize("abc")  # ["ab", "#c"]
tokenizer.encode("abc"  # [53, 2]
tokenizer.decode(tokenizer.tokenize("abc"))  # "abc"

Set tokenizer which Levenshtein Finder uses. Default is CharacterTokenizer

In [3]

from levenshtein_finder import LevenshteinFinder()

finder = LevenshteinFinder()
finder = LevenshteinFinder(BigramTokenizer())
finder = LevenshteinFinder(WordpieceTokenizersWrapper(BertWordPieceTokenizer()))

Indexing

In [4]

finder = LevenshteinFinder()
finder.indexing("path/to/text")

# when you use CharacterTokenizer or BigramTokenizer, you can index with list of words
finder.indexing(["list", "of", "words"])

Searching

Basic search. The form of similars is dict{term: distance}.

In [5]

finder.search("query")

If you want to check details in searching, set verbose=True

In [6]

finder.search("분식회계", verbose=True)

Out [6]

query               : 분식회계
tokens              : ['분', '식', '회', '계']
num data            : 132864
num 1st candidates  : 10137
num final candidates: 7
num similars        : 4
elapsed time        : 0.00542903 sec
[{'idx': 84243, 'data': '분식회계', 'distance': 0},
 {'idx': 36211, 'data': '분식회', 'distance': 1},
 {'idx': 113942, 'data': '분식회계설', 'distance': 1},
 {'idx': 114122, 'data': '분석회계', 'distance': 1}]

For decomposing Korean to Jamo (NFKD)

To decompose Korean character to jamo components (감 -> [ㄱ, ㅏ, ㅁ]), use Normalizers.

In [7]

normalizer = Normalizers.create_normalizer(unicodedata=True)
tokenizer = CharacterTokenizer(normalizer=normalizer)
finder = LevenshteinFinder(tokenizer=tokenizer)
finder.indexing("text.txt")
finder.search("분식회계", verbose=True)

Out [7]

query               : 분식회계
tokens              : ['ᄇ', 'ᅮ', 'ᆫ', 'ᄉ', 'ᅵ', 'ᆨ', 'ᄒ', 'ᅬ', 'ᄀ', 'ᅨ']
num data            : 132864
num 1st candidates  : 125925
num final candidates: 4
num similars        : 2
elapsed time        : 0.0863643 sec
[{'idx': 84243, 'data': '분식회계', 'distance': 0},
 {'idx': 114122, 'data': '분석회계', 'distance': 1}]

Demo

Searching similar words

LevenshteinFinder needs seconds.

In [8]

import json
with open("data/nouns_from_financial_news.json") as f:
    words = list(json.load(f).keys())

finder = LevenshteinFinder()
finder.indexing(words)
finder.search("분식회계", verbose=True)

Out [8]

elapsed time        : 0.00542903 sec

Brute-force distance computation needs 1.495 seconds

In [9]

import time
from levenshtein_finder import levenshtein

query = '분식회계'

begin_time = time.time()
distance = {word:levenshtein(word, query) for word in words}
search_time = time.time() - begin_time
print('search time = {} sec'.format('%.3f'%search_time))

similars = sorted(filter(lambda x:x[1] <= 1, distance.items()), key=lambda x:x[1])
elapsed_time = time.time() - begin_time
print('total elapsed time = {} sec'.format('%.3f'%elapsed_time))
print(similars)

Out [9]

search time = 1.480 sec
total elapsed time = 1.495 sec
[('분식회계', 0), ('분식회', 1), ('분식회계설', 1), ('분석회계', 1)]

Searching similar strings

Find similar sentence from NSMC corpus (20M).

In [10]

finder = LevenshteinFinder()
finder.indexing("~/Korpora/nsmc/corpus.txt")
finder.search("음악이 주가 된 최고의 음악영화", max_distance=1, verbose=True)

Out [10]

query               : 음악이 주가 된 최고의 음악영화
tokens              : ['음', '악', '이', ' ', '주', '가', ' ', '된', ' ', '최', '고', '의', ' ', '음', '악', '영', '화']
num data            : 200002
num 1st candidates  : 189663
num final candidates: 1
num similars        : 1
elapsed time        : 0.123322 sec
[{'idx': 6, 'data': '음악이 주가 된, 최고의 음악영화', 'distance': 1}]