|
|
|
from __future__ import division
|
|
|
|
|
|
|
|
import unicodedata
|
|
|
|
from typing import Tuple
|
|
|
|
|
|
|
|
from uniseg.graphemecluster import grapheme_clusters
|
|
|
|
|
|
|
|
from qurator.dinglehopper.edit_distance import distance
|
|
|
|
|
|
|
|
|
|
|
|
def character_error_rate_n(reference, compared) -> Tuple[float, int]:
|
|
|
|
"""
|
|
|
|
Compute character error rate.
|
|
|
|
|
|
|
|
:return: character error rate and length of the reference
|
|
|
|
"""
|
|
|
|
d = distance(reference, compared)
|
|
|
|
# XXX
|
|
|
|
from .cli import ExtractedText
|
|
|
|
if isinstance(reference, ExtractedText):
|
|
|
|
reference = reference.text
|
|
|
|
n = len(list(grapheme_clusters(unicodedata.normalize('NFC', reference))))
|
|
|
|
|
|
|
|
if d == 0:
|
|
|
|
return 0, n
|
|
|
|
if n == 0:
|
|
|
|
return float('inf'), n
|
|
|
|
return d/n, n
|
|
|
|
|
|
|
|
# XXX Should we really count newlines here?
|
|
|
|
|
|
|
|
|
|
|
|
def character_error_rate(reference, compared) -> float:
|
|
|
|
"""
|
|
|
|
Compute character error rate.
|
|
|
|
|
|
|
|
:return: character error rate
|
|
|
|
"""
|
|
|
|
cer, _ = character_error_rate_n(reference, compared)
|
|
|
|
return cer
|