Files
test-search-engine/app/typo.py
T

134 lines
4.3 KiB
Python

from __future__ import annotations
from functools import lru_cache
@lru_cache(maxsize=16_384)
def damerau_levenshtein_distance(left: str, right: str) -> int:
"""Return unrestricted Damerau-Levenshtein distance.
In addition to insertions, deletions and substitutions, adjacent character
transpositions cost one operation. The implementation is the unrestricted
variant, not the more limited optimal-string-alignment shortcut.
"""
if left == right:
return 0
if not left:
return len(right)
if not right:
return len(left)
left_length = len(left)
right_length = len(right)
maximum_distance = left_length + right_length
matrix = [
[0 for _ in range(right_length + 2)] for _ in range(left_length + 2)
]
matrix[0][0] = maximum_distance
for left_index in range(left_length + 1):
matrix[left_index + 1][0] = maximum_distance
matrix[left_index + 1][1] = left_index
for right_index in range(right_length + 1):
matrix[0][right_index + 1] = maximum_distance
matrix[1][right_index + 1] = right_index
last_row_by_character: dict[str, int] = {}
for left_index in range(1, left_length + 1):
last_matching_column = 0
for right_index in range(1, right_length + 1):
matching_row = last_row_by_character.get(right[right_index - 1], 0)
matching_column = last_matching_column
substitution_cost = 1
if left[left_index - 1] == right[right_index - 1]:
substitution_cost = 0
last_matching_column = right_index
matrix[left_index + 1][right_index + 1] = min(
matrix[left_index][right_index] + substitution_cost,
matrix[left_index + 1][right_index] + 1,
matrix[left_index][right_index + 1] + 1,
matrix[matching_row][matching_column]
+ (left_index - matching_row - 1)
+ 1
+ (right_index - matching_column - 1),
)
last_row_by_character[left[left_index - 1]] = left_index
return matrix[left_length + 1][right_length + 1]
def damerau_similarity(left: str, right: str) -> float:
if left == right:
return 1.0
maximum_length = max(len(left), len(right))
if maximum_length == 0:
return 1.0
return 1.0 - damerau_levenshtein_distance(left, right) / maximum_length
@lru_cache(maxsize=32_768)
def character_trigrams(value: str) -> frozenset[str]:
padded = f"^{value}$"
if len(padded) <= 3:
return frozenset({padded})
return frozenset(padded[index : index + 3] for index in range(len(padded) - 2))
def trigram_dice(left: str, right: str) -> float:
left_trigrams = character_trigrams(left)
right_trigrams = character_trigrams(right)
denominator = len(left_trigrams) + len(right_trigrams)
if denominator == 0:
return 1.0
return 2.0 * len(left_trigrams & right_trigrams) / denominator
@lru_cache(maxsize=16_384)
def lcs_similarity(left: str, right: str) -> float:
maximum_length = max(len(left), len(right))
if maximum_length == 0:
return 1.0
previous = [0] * (len(right) + 1)
for left_character in left:
current = [0]
for right_index, right_character in enumerate(right, start=1):
if left_character == right_character:
current.append(previous[right_index - 1] + 1)
else:
current.append(max(previous[right_index], current[-1]))
previous = current
return previous[-1] / maximum_length
def word_similarity(left: str, right: str) -> float:
if left == right:
return 1.0
if not left or not right:
return 0.0
shorter_length = min(len(left), len(right))
edit = damerau_similarity(left, right)
if shorter_length <= 4:
# Trigrams are unstable for very short words; edit distance is the main
# signal and exact/prefix matching is handled before this function.
return edit
trigram = trigram_dice(left, right)
lcs = lcs_similarity(left, right)
return edit * 0.60 + trigram * 0.30 + lcs * 0.10
def fuzzy_threshold(token_length: int) -> float:
if token_length <= 4:
return 0.84
if token_length <= 7:
return 0.72
if token_length <= 10:
return 0.61
return 0.53