bazarr/libs/ftfy/chardata.py

316 lines
12 KiB
Python
Raw Normal View History

2018-11-01 00:08:29 +08:00
"""
This gives other modules access to the gritty details about characters and the
encodings that use them.
"""
import html
import itertools
2018-11-01 00:08:29 +08:00
import re
import unicodedata
2018-11-01 00:08:29 +08:00
# These are the encodings we will try to fix in ftfy, in the
# order that they should be tried.
CHARMAP_ENCODINGS = [
"latin-1",
"sloppy-windows-1252",
"sloppy-windows-1251",
"sloppy-windows-1250",
"sloppy-windows-1253",
"sloppy-windows-1254",
"iso-8859-2",
"macroman",
"cp437",
2018-11-01 00:08:29 +08:00
]
SINGLE_QUOTE_RE = re.compile("[\u02bc\u2018-\u201b]")
DOUBLE_QUOTE_RE = re.compile("[\u201c-\u201f]")
2018-11-01 00:08:29 +08:00
def _build_regexes():
"""
ENCODING_REGEXES contain reasonably fast ways to detect if we
could represent a given string in a given encoding. The simplest one is
the 'ascii' detector, which of course just determines if all characters
2018-11-01 00:08:29 +08:00
are between U+0000 and U+007F.
"""
# Define a regex that matches ASCII text.
encoding_regexes = {"ascii": re.compile("^[\x00-\x7f]*$")}
2018-11-01 00:08:29 +08:00
for encoding in CHARMAP_ENCODINGS:
# Make a sequence of characters that bytes \x80 to \xFF decode to
# in each encoding, as well as byte \x1A, which is used to represent
# the replacement character <20> in the sloppy-* encodings.
byte_range = bytes(list(range(0x80, 0x100)) + [0x1A])
charlist = byte_range.decode(encoding)
2018-11-01 00:08:29 +08:00
# The rest of the ASCII bytes -- bytes \x00 to \x19 and \x1B
# to \x7F -- will decode as those ASCII characters in any encoding we
# support, so we can just include them as ranges. This also lets us
# not worry about escaping regex special characters, because all of
# them are in the \x1B to \x7F range.
regex = "^[\x00-\x19\x1b-\x7f{0}]*$".format(charlist)
2018-11-01 00:08:29 +08:00
encoding_regexes[encoding] = re.compile(regex)
return encoding_regexes
2018-11-01 00:08:29 +08:00
ENCODING_REGEXES = _build_regexes()
def _build_html_entities():
entities = {}
# Create a dictionary based on the built-in HTML5 entity dictionary.
# Add a limited set of HTML entities that we'll also decode if they've
# been case-folded to uppercase, such as decoding &NTILDE; as "Ñ".
for name, char in html.entities.html5.items():
if name.endswith(";"):
entities["&" + name] = char
# Restrict the set of characters we can attempt to decode if their
# name has been uppercased. If we tried to handle all entity names,
# the results would be ambiguous.
if name == name.lower():
name_upper = name.upper()
entity_upper = "&" + name_upper
if html.unescape(entity_upper) == entity_upper:
entities[entity_upper] = char.upper()
return entities
HTML_ENTITY_RE = re.compile(r"&#?[0-9A-Za-z]{1,24};")
HTML_ENTITIES = _build_html_entities()
def possible_encoding(text, encoding):
2018-11-01 00:08:29 +08:00
"""
Given text and a single-byte encoding, check whether that text could have
been decoded from that single-byte encoding.
2018-11-01 00:08:29 +08:00
In other words, check whether it can be encoded in that encoding, possibly
sloppily.
"""
return bool(ENCODING_REGEXES[encoding].match(text))
def _build_control_char_mapping():
"""
Build a translate mapping that strips likely-unintended control characters.
See :func:`ftfy.fixes.remove_control_chars` for a description of these
codepoint ranges and why they should be removed.
2018-11-01 00:08:29 +08:00
"""
control_chars = {}
for i in itertools.chain(
range(0x00, 0x09),
[0x0B],
range(0x0E, 0x20),
[0x7F],
range(0x206A, 0x2070),
[0xFEFF],
range(0xFFF9, 0xFFFD),
):
control_chars[i] = None
return control_chars
CONTROL_CHARS = _build_control_char_mapping()
2018-11-01 00:08:29 +08:00
# Recognize UTF-8 sequences that would be valid if it weren't for a b'\xa0'
# that some Windows-1252 program converted to a plain space.
#
# The smaller values are included on a case-by-case basis, because we don't want
# to decode likely input sequences to unlikely characters. These are the ones
# that *do* form likely characters before 0xa0:
#
# 0xc2 -> U+A0 NO-BREAK SPACE
# 0xc3 -> U+E0 LATIN SMALL LETTER A WITH GRAVE
# 0xc5 -> U+160 LATIN CAPITAL LETTER S WITH CARON
# 0xce -> U+3A0 GREEK CAPITAL LETTER PI
# 0xd0 -> U+420 CYRILLIC CAPITAL LETTER ER
# 0xd9 -> U+660 ARABIC-INDIC DIGIT ZERO
#
# In three-character sequences, we exclude some lead bytes in some cases.
#
# When the lead byte is immediately followed by 0xA0, we shouldn't accept
# a space there, because it leads to some less-likely character ranges:
#
# 0xe0 -> Samaritan script
# 0xe1 -> Mongolian script (corresponds to Latin-1 'á' which is too common)
#
# We accept 0xe2 and 0xe3, which cover many scripts. Bytes 0xe4 and
# higher point mostly to CJK characters, which we generally don't want to
# decode near Latin lowercase letters.
2018-11-01 00:08:29 +08:00
#
# In four-character sequences, the lead byte must be F0, because that accounts
# for almost all of the usage of high-numbered codepoints (tag characters whose
# UTF-8 starts with the byte F3 are only used in some rare new emoji sequences).
#
# This is meant to be applied to encodings of text that tests true for `is_bad`.
# Any of these could represent characters that legitimately appear surrounded by
# spaces, particularly U+C5 (Å), which is a word in multiple languages!
2018-11-01 00:08:29 +08:00
#
# We should consider checking for b'\x85' being converted to ... in the future.
# I've seen it once, but the text still wasn't recoverable.
ALTERED_UTF8_RE = re.compile(
b"[\xc2\xc3\xc5\xce\xd0\xd9][ ]"
b"|[\xe2\xe3][ ][\x80-\x84\x86-\x9f\xa1-\xbf]"
b"|[\xe0-\xe3][\x80-\x84\x86-\x9f\xa1-\xbf][ ]"
b"|[\xf0][ ][\x80-\xbf][\x80-\xbf]"
b"|[\xf0][\x80-\xbf][ ][\x80-\xbf]"
b"|[\xf0][\x80-\xbf][\x80-\xbf][ ]"
)
2018-11-01 00:08:29 +08:00
# This expression matches UTF-8 and CESU-8 sequences where some of the
# continuation bytes have been lost. The byte 0x1a (sometimes written as ^Z) is
# used within ftfy to represent a byte that produced the replacement character
# \ufffd. We don't know which byte it was, but we can at least decode the UTF-8
# sequence as \ufffd instead of failing to re-decode it at all.
#
# In some cases, we allow the ASCII '?' in place of \ufffd, but at most once per
# sequence.
2018-11-01 00:08:29 +08:00
LOSSY_UTF8_RE = re.compile(
b"[\xc2-\xdf][\x1a]"
b"|[\xc2-\xc3][?]"
b"|\xed[\xa0-\xaf][\x1a?]\xed[\xb0-\xbf][\x1a?\x80-\xbf]"
b"|\xed[\xa0-\xaf][\x1a?\x80-\xbf]\xed[\xb0-\xbf][\x1a?]"
b"|[\xe0-\xef][\x1a?][\x1a\x80-\xbf]"
b"|[\xe0-\xef][\x1a\x80-\xbf][\x1a?]"
b"|[\xf0-\xf4][\x1a?][\x1a\x80-\xbf][\x1a\x80-\xbf]"
b"|[\xf0-\xf4][\x1a\x80-\xbf][\x1a?][\x1a\x80-\xbf]"
b"|[\xf0-\xf4][\x1a\x80-\xbf][\x1a\x80-\xbf][\x1a?]"
b"|\x1a"
2018-11-01 00:08:29 +08:00
)
# This regex matches C1 control characters, which occupy some of the positions
# in the Latin-1 character map that Windows assigns to other characters instead.
C1_CONTROL_RE = re.compile(r"[\x80-\x9f]")
2018-11-01 00:08:29 +08:00
# A translate mapping that breaks ligatures made of Latin letters. While
# ligatures may be important to the representation of other languages, in Latin
# letters they tend to represent a copy/paste error. It omits ligatures such
# as æ that are frequently used intentionally.
2018-11-01 00:08:29 +08:00
#
# This list additionally includes some Latin digraphs that represent two
# characters for legacy encoding reasons, not for typographical reasons.
#
# Ligatures and digraphs may also be separated by NFKC normalization, but that
# is sometimes more normalization than you want.
2018-11-01 00:08:29 +08:00
LIGATURES = {
ord("IJ"): "IJ", # Dutch ligatures
ord("ij"): "ij",
ord("ʼn"): "ʼn", # Afrikaans digraph meant to avoid auto-curled quote
ord("DZ"): "DZ", # Serbian/Croatian digraphs for Cyrillic conversion
ord("Dz"): "Dz",
ord("dz"): "dz",
ord("DŽ"): "",
ord("Dž"): "",
ord("dž"): "",
ord("LJ"): "LJ",
ord("Lj"): "Lj",
ord("lj"): "lj",
ord("NJ"): "NJ",
ord("Nj"): "Nj",
ord("nj"): "nj",
ord(""): "ff", # Latin typographical ligatures
ord(""): "fi",
ord(""): "fl",
ord(""): "ffi",
ord(""): "ffl",
ord(""): "ſt",
ord(""): "st",
2018-11-01 00:08:29 +08:00
}
def _build_width_map():
"""
Build a translate mapping that replaces halfwidth and fullwidth forms
with their standard-width forms.
"""
# Though it's not listed as a fullwidth character, we'll want to convert
# U+3000 IDEOGRAPHIC SPACE to U+20 SPACE on the same principle, so start
# with that in the dictionary.
width_map = {0x3000: " "}
for i in range(0xFF01, 0xFFF0):
char = chr(i)
alternate = unicodedata.normalize("NFKC", char)
2018-11-01 00:08:29 +08:00
if alternate != char:
width_map[i] = alternate
return width_map
2018-11-01 00:08:29 +08:00
WIDTH_MAP = _build_width_map()
# Character classes that help us pinpoint embedded mojibake. These can
# include common characters, because we'll also check them for 'badness'.
UTF8_CLUES = {
# Letters that decode to 0xC2 - 0xDF in a Latin-1-like encoding
"utf8_first_of_2": (
"ÂÃÄÅÆÇÈÉÊËÌÍÎÏÐÑÒÓÔÕÖ×ØÙÚÛÜÝÞßĂĆČĎĐĘĚĞİĹŃŇŐŘŞŢŮŰ"
"ΒΓΔΕΖΗΘΙΚΛΜΝΞΟΠΡΣΤΥΦΧΨΩΪΫάέήίВГДЕЖЗИЙКЛМНОПРСТУФХЦЧШЩЪЫЬЭЮЯ"
),
# Letters that decode to 0xE0 - 0xEF in a Latin-1-like encoding
"utf8_first_of_3": ("àáâãäåæçèéêëìíîïăćčďęěĺŕΰαβγδεζηθικλμνξοабвгдежзийклмноп"),
# Letters that decode to 0xF0 or 0xF3 in a Latin-1-like encoding.
# (Other leading bytes correspond only to unassigned codepoints)
"utf8_first_of_4": ("ðóđğπσру"),
# Letters that decode to 0x80 - 0xBF in a Latin-1-like encoding,
# including a space standing in for 0xA0
"utf8_continuation": (
"\x80-\xbf"
"ĄąĽľŁłŒœŚśŞşŠšŤťŸŹźŻżŽžƒˆˇ˘˛˜˝΄΅"
"ΆΈΉΊΌΎΏЁЂЃЄЅІЇЈЉЊЋЌЎЏёђѓєѕіїјљњћќўџҐґ"
"–—―‘’‚“”„†‡•…‰‹›€№™"
" "
),
# Letters that decode to 0x80 - 0xBF in a Latin-1-like encoding,
# and don't usually stand for themselves when adjacent to mojibake.
# This excludes spaces, dashes, quotation marks, and ellipses.
"utf8_continuation_strict": (
"\x80-\xbf"
"ĄąĽľŁłŒœŚśŞşŠšŤťŸŹźŻżŽžƒˆˇ˘˛˜˝΄΅"
"ΆΈΉΊΌΎΏЁЂЃЄЅІЇЈЉЊЋЌЎЏёђѓєѕіїјљњћќўџҐґ"
"†‡•‰‹›€№™"
),
}
# This regex uses UTF8_CLUES to find sequences of likely mojibake.
# It matches them with + so that several adjacent UTF-8-looking sequences
# get coalesced into one, allowing them to be fixed more efficiently
# and not requiring every individual subsequence to be detected as 'badness'.
#
# We accept spaces in place of "utf8_continuation", because spaces might have
# been intended to be U+A0 NO-BREAK SPACE.
#
# We do a lookbehind to make sure the previous character isn't a
# "utf8_continuation_strict" character, so that we don't fix just a few
# characters in a huge garble and make the situation worse.
#
# Unfortunately, the matches to this regular expression won't show their
# surrounding context, and including context would make the expression much
# less efficient. The 'badness' rules that require context, such as a preceding
# lowercase letter, will prevent some cases of inconsistent UTF-8 from being
# fixed when they don't see it.
UTF8_DETECTOR_RE = re.compile(
"""
(?<! [{utf8_continuation_strict}])
(
[{utf8_first_of_2}] [{utf8_continuation}]
|
[{utf8_first_of_3}] [{utf8_continuation}]{{2}}
|
[{utf8_first_of_4}] [{utf8_continuation}]{{3}}
)+
""".format(
**UTF8_CLUES
),
re.VERBOSE,
)