SpaCy — intra-word hyphens. How to treat them one word?

问题

Following is the code provided as answer to the question;

import spacy
from spacy.tokenizer import Tokenizer
from spacy.util import compile_prefix_regex, compile_infix_regex, compile_suffix_regex
import re

nlp = spacy.load('en')

infixes = nlp.Defaults.prefixes + (r"[./]", r"[-]~", r"(.'.)")

infix_re = spacy.util.compile_infix_regex(infixes)

def custom_tokenizer(nlp):
    return Tokenizer(nlp.vocab, infix_finditer=infix_re.finditer)

nlp.tokenizer = custom_tokenizer(nlp)

s1 = "Marketing-Representative- won't die in car accident."
s2 = "Out-of-box implementation"

for s in s1,s2:
    doc = nlp("{}".format(s))
    print([token.text for token in doc])

Result

$python3 /tmp/nlp.py  
['Marketing-Representative-', 'wo', "n't", 'die', 'in', 'car', 'accident', '.']  
['Out-of-box', 'implementation']

What are the first (r"[./]") and the last (r"(.'.)") patterns used for in the following?

infixes = nlp.Defaults.prefixes + (r"[./]", r"[-]~", r"(.'.)")

Edit: I expect the splits to be as follows;

That

Yahya

laptop-cover

I want spacy to treat intra-hyphen word as one token without impacting negatively on other split rules.

"That is Yahya's laptop-cover. 3.14!"

["That", "is", "Yahya", "'s", "laptop-cover", ".", "3.14", "!"] (EXPECTED)

By default,

import spacy
nlp = spacy.load('en_core_web_md')
for token in nlp("That is Yahya's laptop-cover. 3.14!"):
    print (token.text)

SpaCy gives;

["That", "is", "Yahya", "'s", "laptop", "-", "cover", ".", "3.14", "!"]

However,

from spacy.util import compile_infix_regex
infixes = nlp.Defaults.prefixes + tuple([r"[-]~"])
infix_re = spacy.util.compile_infix_regex(infixes)
nlp.tokenizer = spacy.tokenizer.Tokenizer(nlp.vocab, infix_finditer=infix_re.finditer)
for token in nlp("That is Yahya's laptop-cover. 3.14!"):
    print (token.text)

gives;

["That", "is", "Yahya", "'", "s", "laptop-cover.", "3.14", "!"]

回答1:

NOTE: To see the custom tokenizer that keeps the hyphenated words see the botton of the answer.

Here, a custom tokenizer is defined, that tokenizes text into tokens using a set of built-in (nlp.Defaults.prefixes) and custom ([./], [-]~, (.'.)) patterns.

The nlp.Defaults.prefixes + (r"[./]", r"[-]~", r"(.'.)") is tuple concatenation operation, the result looks like

('§', '%', '=', '—', '–', '\\+(?![0-9])', '…', '……', ',', ':', ';', '\\!', '\\?', '¿', '؟', '¡', '\\(', '\\)', '\\[', '\\]', '\\{', '\\}', '<', '>', '_', '#', '\\*', '&', '。', '？', '！', '，', '、', '；', '：', '～', '·', '।', '،', '؛', '٪', '\\.\\.+', '…', "\\'", '"', '”', '“', '`', '‘', '´', '’', '‚', ',', '„', '»', '«', '「', '」', '『', '』', '（', '）', '〔', '〕', '【', '】', '《', '》', '〈', '〉', '\\$', '£', '€', '¥', '฿', 'US\\$', 'C\\$', 'A\\$', '₽', '﷼', '₴', '[\\u00A6\\u00A9\\u00AE\\u00B0\\u0482\\u058D\\u058E\\u060E\\u060F\\u06DE\\u06E9\\u06FD\\u06FE\\u07F6\\u09FA\\u0B70\\u0BF3-\\u0BF8\\u0BFA\\u0C7F\\u0D4F\\u0D79\\u0F01-\\u0F03\\u0F13\\u0F15-\\u0F17\\u0F1A-\\u0F1F\\u0F34\\u0F36\\u0F38\\u0FBE-\\u0FC5\\u0FC7-\\u0FCC\\u0FCE\\u0FCF\\u0FD5-\\u0FD8\\u109E\\u109F\\u1390-\\u1399\\u1940\\u19DE-\\u19FF\\u1B61-\\u1B6A\\u1B74-\\u1B7C\\u2100\\u2101\\u2103-\\u2106\\u2108\\u2109\\u2114\\u2116\\u2117\\u211E-\\u2123\\u2125\\u2127\\u2129\\u212E\\u213A\\u213B\\u214A\\u214C\\u214D\\u214F\\u218A\\u218B\\u2195-\\u2199\\u219C-\\u219F\\u21A1\\u21A2\\u21A4\\u21A5\\u21A7-\\u21AD\\u21AF-\\u21CD\\u21D0\\u21D1\\u21D3\\u21D5-\\u21F3\\u2300-\\u2307\\u230C-\\u231F\\u2322-\\u2328\\u232B-\\u237B\\u237D-\\u239A\\u23B4-\\u23DB\\u23E2-\\u2426\\u2440-\\u244A\\u249C-\\u24E9\\u2500-\\u25B6\\u25B8-\\u25C0\\u25C2-\\u25F7\\u2600-\\u266E\\u2670-\\u2767\\u2794-\\u27BF\\u2800-\\u28FF\\u2B00-\\u2B2F\\u2B45\\u2B46\\u2B4D-\\u2B73\\u2B76-\\u2B95\\u2B98-\\u2BC8\\u2BCA-\\u2BFE\\u2CE5-\\u2CEA\\u2E80-\\u2E99\\u2E9B-\\u2EF3\\u2F00-\\u2FD5\\u2FF0-\\u2FFB\\u3004\\u3012\\u3013\\u3020\\u3036\\u3037\\u303E\\u303F\\u3190\\u3191\\u3196-\\u319F\\u31C0-\\u31E3\\u3200-\\u321E\\u322A-\\u3247\\u3250\\u3260-\\u327F\\u328A-\\u32B0\\u32C0-\\u32FE\\u3300-\\u33FF\\u4DC0-\\u4DFF\\uA490-\\uA4C6\\uA828-\\uA82B\\uA836\\uA837\\uA839\\uAA77-\\uAA79\\uFDFD\\uFFE4\\uFFE8\\uFFED\\uFFEE\\uFFFC\\uFFFD\\U00010137-\\U0001013F\\U00010179-\\U00010189\\U0001018C-\\U0001018E\\U00010190-\\U0001019B\\U000101A0\\U000101D0-\\U000101FC\\U00010877\\U00010878\\U00010AC8\\U0001173F\\U00016B3C-\\U00016B3F\\U00016B45\\U0001BC9C\\U0001D000-\\U0001D0F5\\U0001D100-\\U0001D126\\U0001D129-\\U0001D164\\U0001D16A-\\U0001D16C\\U0001D183\\U0001D184\\U0001D18C-\\U0001D1A9\\U0001D1AE-\\U0001D1E8\\U0001D200-\\U0001D241\\U0001D245\\U0001D300-\\U0001D356\\U0001D800-\\U0001D9FF\\U0001DA37-\\U0001DA3A\\U0001DA6D-\\U0001DA74\\U0001DA76-\\U0001DA83\\U0001DA85\\U0001DA86\\U0001ECAC\\U0001F000-\\U0001F02B\\U0001F030-\\U0001F093\\U0001F0A0-\\U0001F0AE\\U0001F0B1-\\U0001F0BF\\U0001F0C1-\\U0001F0CF\\U0001F0D1-\\U0001F0F5\\U0001F110-\\U0001F16B\\U0001F170-\\U0001F1AC\\U0001F1E6-\\U0001F202\\U0001F210-\\U0001F23B\\U0001F240-\\U0001F248\\U0001F250\\U0001F251\\U0001F260-\\U0001F265\\U0001F300-\\U0001F3FA\\U0001F400-\\U0001F6D4\\U0001F6E0-\\U0001F6EC\\U0001F6F0-\\U0001F6F9\\U0001F700-\\U0001F773\\U0001F780-\\U0001F7D8\\U0001F800-\\U0001F80B\\U0001F810-\\U0001F847\\U0001F850-\\U0001F859\\U0001F860-\\U0001F887\\U0001F890-\\U0001F8AD\\U0001F900-\\U0001F90B\\U0001F910-\\U0001F93E\\U0001F940-\\U0001F970\\U0001F973-\\U0001F976\\U0001F97A\\U0001F97C-\\U0001F9A2\\U0001F9B0-\\U0001F9B9\\U0001F9C0-\\U0001F9C2\\U0001F9D0-\\U0001F9FF\\U0001FA60-\\U0001FA6D]', '[/.]', '-~', "(.'.)")

As you see, these are all regular expressions and are used to process in-word punctuation, infixes. See the Spacy tokenizer algorithm:

The algorithm can be summarized as follows:

Iterate over space-separated substrings

Check whether we have an explicitly defined rule for this substring. If we do, use it.

Otherwise, try to consume a prefix.

If we consumed a prefix, go back to the beginning of the loop, so that special-cases always get priority.

If we didn’t consume a prefix, try to consume a suffix.

If we can’t consume a prefix or suffix, look for “infixes” — stuff like hyphens etc.

Once we can’t consume any more of the string, handle it as a single token.

Now, when we are at infix handling step, these regular expressions are used to split text into tokens based also on these patterns.

E.g. [/.] is important because if you do not add it, abc.def/ghi will be a single token, but with the pattern added, it will be split into 'abc', '.', 'def', '/', 'ghi'.

The [-]~ (that is the same as -~) matches a - and wants to match ~ right after, but since it is not there, the - is skipped and no split occurs, you get the whole 'Marketing-Representative-' token. Note, however, if you have 'Marketing-~Representative-' in the sentence, and you use -~ regex you will get ['Marketing', '-~', 'Representative-'] as a result as there will be a match.

The .'. regex matches any char + ' + any char. A dot matches any char in regex. So, the rule just tokenizes (splits out) these tokens out of the sentence (e.g. n't, r'd, etc.)

Answer to edit

You should be very careful when adding new rules and check if they do not overlap with already added ones.

E.g. when you add r"\b's\b" to split out Genetive case apostrophe-s, you should "override" the "\\'" rule from nlp.Defaults.prefixes. Either remove it if you do not plan to match ' as infix, or give priority to your custom rules by appendng the nlp.Defaults.prefixes to those rules, not vice versa.

See a sample code:

import re
import spacy
from spacy.tokenizer import Tokenizer

nlp = spacy.load("en_core_web_md")
infixes = tuple([r"'s\b", r"(?<!\d)\.(?!\d)"]) +  nlp.Defaults.prefixes
infix_re = spacy.util.compile_infix_regex(infixes)

def custom_tokenizer(nlp):
    return Tokenizer(nlp.vocab, infix_finditer=infix_re.finditer)

nlp.tokenizer = custom_tokenizer(nlp)
doc = nlp(u"That is Yahya's laptop-cover. 3.14!")
print([t.text for t in doc])

Output: ['That', 'is', 'Yahya', "'s", 'laptop-cover', '.', '3.14', '!']

Details

r"'s\b" - matches 's that are followed with a word boundary
r"(?<!\d)\.(?!\d) - matches a . that is not preceded or followed with a digit.

And if you want to use a custom tokenizer that keeps hyphenated letter words as single tokens you will have to re-define the infixes: the r"(?<=[{a}])(?:{h})(?=[{a}])".format(a=ALPHA, h=HYPHENS), line accounts for that and you need to get rid of it. Since it is the only item that contains a -|–|—|--|---|——|~ string it will be easier to drop this item from the infixes and re-compile the infix pattern:

import spacy
from spacy.tokenizer import Tokenizer
from spacy.util import compile_infix_regex

nlp = spacy.load("en_core_web_sm")

inf = list(nlp.Defaults.infixes)
inf = [x for x in inf if '-|–|—|--|---|——|~' not in x] # remove the hyphen-between-letters pattern from infix patterns
infix_re = compile_infix_regex(tuple(inf))

def custom_tokenizer(nlp):
    return Tokenizer(nlp.vocab, prefix_search=nlp.tokenizer.prefix_search,
                                suffix_search=nlp.tokenizer.suffix_search,
                                infix_finditer=infix_re.finditer,
                                token_match=nlp.tokenizer.token_match,
                                rules=nlp.Defaults.tokenizer_exceptions)

nlp.tokenizer = custom_tokenizer(nlp)
doc = nlp("That is Yahya's laptop-cover. 3.14!")
print([t.text for t in doc])
# => ['That', 'is', 'Yahya', "'s", 'laptop-cover', '.', '3.14', '!']

来源：https://stackoverflow.com/questions/55241927/spacy-intra-word-hyphens-how-to-treat-them-one-word

标签

nlp

tokenize

spacy