spaCy/website/docs/usage/customizing-tokenizer.jade
2017-09-22 12:15:48 +08:00

250 lines
9.4 KiB
Plaintext

//- 💫 DOCS > USAGE > TOKENIZER
include ../../_includes/_mixins
p
| Tokenization is the task of splitting a text into meaningful segments,
| called #[em tokens]. The input to the tokenizer is a unicode text, and
| the output is a #[+api("doc") #[code Doc]] object. To construct a
| #[code Doc] object, you need a #[+api("vocab") #[code Vocab]] instance,
| a sequence of #[code word] strings, and optionally a sequence of
| #[code spaces] booleans, which allow you to maintain alignment of the
| tokens into the original string.
+aside("See Also")
| If you haven't read up on spaCy's #[+a("data-model") data model] yet,
| you should probably have a look. The main point to keep in mind is that
| spaCy's #[code Doc] doesn't copy or refer to the original string. The
| string is reconstructed from the tokens when required.
+h(2, "special-cases") Adding special case tokenization rules
p
| Most domains have at least some idiosyncrasies that require custom
| tokenization rules. Here's how to add a special case rule to an existing
| #[+api("tokenizer") #[code Tokenizer]] instance:
+code.
import spacy
from spacy.symbols import ORTH, LEMMA, POS
nlp = spacy.load('en')
assert [w.text for w in nlp(u'gimme that')] == [u'gimme', u'that']
nlp.tokenizer.add_special_case(u'gimme',
[
{
ORTH: u'gim',
LEMMA: u'give',
POS: u'VERB'},
{
ORTH: u'me'}])
assert [w.text for w in nlp(u'gimme that')] == [u'gim', u'me', u'that']
# Pronoun lemma is returned as -PRON-
# More details please see: https://spacy.io/docs/usage/troubleshooting#pron-lemma
assert [w.lemma_ for w in nlp(u'gimme that')] == [u'give', u'-PRON-', u'that']
p
| The special case doesn't have to match an entire whitespace-delimited
| substring. The tokenizer will incrementally split off punctuation, and
| keep looking up the remaining substring:
+code.
assert 'gimme' not in [w.text for w in nlp(u'gimme!')]
assert 'gimme' not in [w.text for w in nlp(u'("...gimme...?")')]
p
| The special case rules have precedence over the punctuation splitting:
+code.
nlp.tokenizer.add_special_case(u'...gimme...?',
[{
ORTH: u'...gimme...?', LEMMA: u'give', TAG: u'VB'}])
assert len(nlp(u'...gimme...?')) == 1
p
| Because the special-case rules allow you to set arbitrary token
| attributes, such as the part-of-speech, lemma, etc, they make a good
| mechanism for arbitrary fix-up rules. Having this logic live in the
| tokenizer isn't very satisfying from a design perspective, however, so
| the API may eventually be exposed on the
| #[+api("language") #[code Language]] class itself.
+h(2, "how-tokenizer-works") How spaCy's tokenizer works
p
| spaCy introduces a novel tokenization algorithm, that gives a better
| balance between performance, ease of definition, and ease of alignment
| into the original string.
p
| After consuming a prefix or infix, we consult the special cases again.
| We want the special cases to handle things like "don't" in English, and
| we want the same rule to work for "(don't)!". We do this by splitting
| off the open bracket, then the exclamation, then the close bracket, and
| finally matching the special-case. Here's an implementation of the
| algorithm in Python, optimized for readability rather than performance:
+code.
def tokenizer_pseudo_code(text, find_prefix, find_suffix,
find_infixes, special_cases):
tokens = []
for substring in text.split(' '):
suffixes = []
while substring:
if substring in special_cases:
tokens.extend(special_cases[substring])
substring = ''
elif find_prefix(substring) is not None:
split = find_prefix(substring)
tokens.append(substring[:split])
substring = substring[split:]
elif find_suffix(substring) is not None:
split = find_suffix(substring)
suffixes.append(substring[split:])
substring = substring[:split]
elif find_infixes(substring):
infixes = find_infixes(substring)
offset = 0
for match in infixes:
tokens.append(substring[i : match.start()])
tokens.append(substring[match.start() : match.end()])
offset = match.end()
substring = substring[offset:]
else:
tokens.append(substring)
substring = ''
tokens.extend(reversed(suffixes))
return tokens
p
| The algorithm can be summarized as follows:
+list("numbers")
+item Iterate over space-separated substrings
+item
| Check whether we have an explicitly defined rule for this substring.
| If we do, use it.
+item Otherwise, try to consume a prefix.
+item
| If we consumed a prefix, go back to the beginning of the loop, so
| that special-cases always get priority.
+item If we didn't consume a prefix, try to consume a suffix.
+item
| If we can't consume a prefix or suffix, look for "infixes" — stuff
| like hyphens etc.
+item Once we can't consume any more of the string, handle it as a single token.
+h(2, "native-tokenizers") Customizing spaCy's Tokenizer class
p
| Let's imagine you wanted to create a tokenizer for a new language. There
| are four things you would need to define:
+list("numbers")
+item
| A dictionary of #[strong special cases]. This handles things like
| contractions, units of measurement, emoticons, certain
| abbreviations, etc.
+item
| A function #[code prefix_search], to handle
| #[strong preceding punctuation], such as open quotes, open brackets,
| etc
+item
| A function #[code suffix_search], to handle
| #[strong succeeding punctuation], such as commas, periods, close
| quotes, etc.
+item
| A function #[code infixes_finditer], to handle non-whitespace
| separators, such as hyphens etc.
p
| You shouldn't usually need to create a #[code Tokenizer] subclass.
| Standard usage is to use #[code re.compile()] to build a regular
| expression object, and pass its #[code .search()] and
| #[code .finditer()] methods:
+code.
import re
from spacy.tokenizer import Tokenizer
prefix_re = re.compile(r'''[\[\("']''')
suffix_re = re.compile(r'''[\]\)"']''')
infix_re = re.compile(r'''[-~]''')
def create_tokenizer(nlp):
return Tokenizer(nlp.vocab, rules={},
prefix_search=prefix_re.search,
suffix_search=suffix_re.search,
infix_finditer=infix_re.finditer)
nlp = spacy.load('en', create_make_doc=create_tokenizer)
p
| If you need to subclass the tokenizer instead, the relevant methods to
| specialize are #[code find_prefix], #[code find_suffix] and
| #[code find_infix].
+h(2, "custom-tokenizer") Hooking an arbitrary tokenizer into the pipeline
p
| You can pass a custom tokenizer using the #[code make_doc] keyword, when
| you're creating the pipeline:
+code.
import spacy
nlp = spacy.load('en', make_doc=my_tokenizer)
p
| However, this approach often leaves us with a chicken-and-egg problem.
| To construct the tokenizer, we usually want attributes of the #[code nlp]
| pipeline. Specifically, we want the tokenizer to hold a reference to the
| pipeline's vocabulary object. Let's say we have the following class as
| our tokenizer:
+code.
import spacy
from spacy.tokens import Doc
class WhitespaceTokenizer(object):
def __init__(self, nlp):
self.vocab = nlp.vocab
def __call__(self, text):
words = text.split(' ')
# All tokens 'own' a subsequent space character in this tokenizer
spaces = [True] * len(words)
return Doc(self.vocab, words=words, spaces=spaces)
p
| As you can see, we need a #[code vocab] instance to construct this — but
| we won't get the #[code vocab] instance until we get back the #[code nlp]
| object from #[code spacy.load()]. The simplest solution is to build the
| object in two steps:
+code.
nlp = spacy.load('en')
nlp.make_doc = WhitespaceTokenizer(nlp)
p
| You can instead pass the class to the #[code create_make_doc] keyword,
| which is invoked as callback once the #[code nlp] object is ready:
+code.
nlp = spacy.load('en', create_make_doc=WhitespaceTokenizer)
p
| Finally, you can of course create your own subclasses, and create a bound
| #[code make_doc] method. The disadvantage of this approach is that spaCy
| uses inheritance to give each language-specific pipeline its own class.
| If you're working with multiple languages, a naive solution will
| therefore require one custom class per language you're working with.
| This might be at least annoying. You may be able to do something more
| generic by doing some clever magic with metaclasses or mixins, if that's
| the sort of thing you're into.