Exclude math and modifier symbols as tokens

Former-commit-id: 8f3c9f576c
This commit is contained in:
Robyn Speer 2015-08-25 11:43:22 -04:00
parent 8637aaef9e
commit 070c89c00c

View File

@ -4,10 +4,11 @@ import unicodedata
# Here's what the following regular expression is looking for: # Here's what the following regular expression is looking for:
# #
# At the start, it looks for a character in the set [\S--\p{punct}]. \S # At the start, it looks for a character in the set \S -- the set of
# contains non-space characters, and then it subtracts the set of Unicode # non-punctuation -- with various characters subtracted out, including punctuation
# punctuation characters from that set. This is slightly different from \w, # and most of the 'symbol' categories. (We leave So, "Symbol - Other", because
# because it leaves symbols (such as emoji) as tokens. # it contains things like emoji that have interesting frequencies. This is why
# we don't just insist on the token starting with a "word" character, \w.)
# #
# After it has found one such character, the rest of the token is (?:\B\S)*, # After it has found one such character, the rest of the token is (?:\B\S)*,
# which continues to consume characters as long as the next character does not # which continues to consume characters as long as the next character does not
@ -26,7 +27,7 @@ import unicodedata
# correct behavior for word-wrapping, but it's an ugly failure mode for NLP # correct behavior for word-wrapping, but it's an ugly failure mode for NLP
# tokenization. # tokenization.
TOKEN_RE = regex.compile(r'[\S--\p{punct}](?:\B\S|[\p{IsIdeo}\p{Script=Hiragana}])*', regex.V1 | regex.WORD) TOKEN_RE = regex.compile(r'[\S--[\p{punct}\p{Sm}\p{Sc}\p{Sk}]](?:\B\S|[\p{IsIdeo}\p{Script=Hiragana}])*', regex.V1 | regex.WORD)
ARABIC_MARK_RE = regex.compile(r'[[\p{Mn}&&\p{Block=Arabic}]\N{ARABIC TATWEEL}]', regex.V1) ARABIC_MARK_RE = regex.compile(r'[[\p{Mn}&&\p{Block=Arabic}]\N{ARABIC TATWEEL}]', regex.V1)