spaCy/spacy/lang/ja/__init__.py
Adriane Boyd c5de9b463a
Update custom tokenizer APIs and pickling (#8972)
* Fix incorrect pickling of Japanese and Korean pipelines, which led to
the entire pipeline being reset if pickled

* Enable pickling of Vietnamese tokenizer

* Update tokenizer APIs for Chinese, Japanese, Korean, Thai, and
Vietnamese so that only the `Vocab` is required for initialization
2021-08-19 14:37:47 +02:00

300 lines
11 KiB
Python

from typing import Optional, Union, Dict, Any
from pathlib import Path
import srsly
from collections import namedtuple
from .stop_words import STOP_WORDS
from .syntax_iterators import SYNTAX_ITERATORS
from .tag_map import TAG_MAP
from .tag_orth_map import TAG_ORTH_MAP
from .tag_bigram_map import TAG_BIGRAM_MAP
from ...errors import Errors
from ...language import Language
from ...scorer import Scorer
from ...symbols import POS
from ...tokens import Doc
from ...training import validate_examples
from ...util import DummyTokenizer, registry, load_config_from_str
from ...vocab import Vocab
from ... import util
DEFAULT_CONFIG = """
[nlp]
[nlp.tokenizer]
@tokenizers = "spacy.ja.JapaneseTokenizer"
split_mode = null
"""
@registry.tokenizers("spacy.ja.JapaneseTokenizer")
def create_tokenizer(split_mode: Optional[str] = None):
def japanese_tokenizer_factory(nlp):
return JapaneseTokenizer(nlp.vocab, split_mode=split_mode)
return japanese_tokenizer_factory
class JapaneseTokenizer(DummyTokenizer):
def __init__(self, vocab: Vocab, split_mode: Optional[str] = None) -> None:
self.vocab = vocab
self.split_mode = split_mode
self.tokenizer = try_sudachi_import(self.split_mode)
def __reduce__(self):
return JapaneseTokenizer, (self.vocab, self.split_mode)
def __call__(self, text: str) -> Doc:
# convert sudachipy.morpheme.Morpheme to DetailedToken and merge continuous spaces
sudachipy_tokens = self.tokenizer.tokenize(text)
dtokens = self._get_dtokens(sudachipy_tokens)
dtokens, spaces = get_dtokens_and_spaces(dtokens, text)
# create Doc with tag bi-gram based part-of-speech identification rules
words, tags, inflections, lemmas, readings, sub_tokens_list = (
zip(*dtokens) if dtokens else [[]] * 6
)
sub_tokens_list = list(sub_tokens_list)
doc = Doc(self.vocab, words=words, spaces=spaces)
next_pos = None # for bi-gram rules
for idx, (token, dtoken) in enumerate(zip(doc, dtokens)):
token.tag_ = dtoken.tag
if next_pos: # already identified in previous iteration
token.pos = next_pos
next_pos = None
else:
token.pos, next_pos = resolve_pos(
token.orth_,
dtoken.tag,
tags[idx + 1] if idx + 1 < len(tags) else None,
)
# if there's no lemma info (it's an unk) just use the surface
token.lemma_ = dtoken.lemma if dtoken.lemma else dtoken.surface
doc.user_data["inflections"] = inflections
doc.user_data["reading_forms"] = readings
doc.user_data["sub_tokens"] = sub_tokens_list
return doc
def _get_dtokens(self, sudachipy_tokens, need_sub_tokens: bool = True):
sub_tokens_list = (
self._get_sub_tokens(sudachipy_tokens) if need_sub_tokens else None
)
dtokens = [
DetailedToken(
token.surface(), # orth
"-".join([xx for xx in token.part_of_speech()[:4] if xx != "*"]), # tag
",".join([xx for xx in token.part_of_speech()[4:] if xx != "*"]), # inf
token.dictionary_form(), # lemma
token.reading_form(), # user_data['reading_forms']
sub_tokens_list[idx]
if sub_tokens_list
else None, # user_data['sub_tokens']
)
for idx, token in enumerate(sudachipy_tokens)
if len(token.surface()) > 0
# remove empty tokens which can be produced with characters like … that
]
# Sudachi normalizes internally and outputs each space char as a token.
# This is the preparation for get_dtokens_and_spaces() to merge the continuous space tokens
return [
t
for idx, t in enumerate(dtokens)
if idx == 0
or not t.surface.isspace()
or t.tag != "空白"
or not dtokens[idx - 1].surface.isspace()
or dtokens[idx - 1].tag != "空白"
]
def _get_sub_tokens(self, sudachipy_tokens):
if (
self.split_mode is None or self.split_mode == "A"
): # do nothing for default split mode
return None
sub_tokens_list = [] # list of (list of list of DetailedToken | None)
for token in sudachipy_tokens:
sub_a = token.split(self.tokenizer.SplitMode.A)
if len(sub_a) == 1: # no sub tokens
sub_tokens_list.append(None)
elif self.split_mode == "B":
sub_tokens_list.append([self._get_dtokens(sub_a, False)])
else: # "C"
sub_b = token.split(self.tokenizer.SplitMode.B)
if len(sub_a) == len(sub_b):
dtokens = self._get_dtokens(sub_a, False)
sub_tokens_list.append([dtokens, dtokens])
else:
sub_tokens_list.append(
[
self._get_dtokens(sub_a, False),
self._get_dtokens(sub_b, False),
]
)
return sub_tokens_list
def score(self, examples):
validate_examples(examples, "JapaneseTokenizer.score")
return Scorer.score_tokenization(examples)
def _get_config(self) -> Dict[str, Any]:
return {"split_mode": self.split_mode}
def _set_config(self, config: Dict[str, Any] = {}) -> None:
self.split_mode = config.get("split_mode", None)
def to_bytes(self, **kwargs) -> bytes:
serializers = {"cfg": lambda: srsly.json_dumps(self._get_config())}
return util.to_bytes(serializers, [])
def from_bytes(self, data: bytes, **kwargs) -> "JapaneseTokenizer":
deserializers = {"cfg": lambda b: self._set_config(srsly.json_loads(b))}
util.from_bytes(data, deserializers, [])
self.tokenizer = try_sudachi_import(self.split_mode)
return self
def to_disk(self, path: Union[str, Path], **kwargs) -> None:
path = util.ensure_path(path)
serializers = {"cfg": lambda p: srsly.write_json(p, self._get_config())}
return util.to_disk(path, serializers, [])
def from_disk(self, path: Union[str, Path], **kwargs) -> "JapaneseTokenizer":
path = util.ensure_path(path)
serializers = {"cfg": lambda p: self._set_config(srsly.read_json(p))}
util.from_disk(path, serializers, [])
self.tokenizer = try_sudachi_import(self.split_mode)
return self
class JapaneseDefaults(Language.Defaults):
config = load_config_from_str(DEFAULT_CONFIG)
stop_words = STOP_WORDS
syntax_iterators = SYNTAX_ITERATORS
writing_system = {"direction": "ltr", "has_case": False, "has_letters": False}
class Japanese(Language):
lang = "ja"
Defaults = JapaneseDefaults
# Hold the attributes we need with convenient names
DetailedToken = namedtuple(
"DetailedToken", ["surface", "tag", "inf", "lemma", "reading", "sub_tokens"]
)
def try_sudachi_import(split_mode="A"):
"""SudachiPy is required for Japanese support, so check for it.
It it's not available blow up and explain how to fix it.
split_mode should be one of these values: "A", "B", "C", None->"A"."""
try:
from sudachipy import dictionary, tokenizer
split_mode = {
None: tokenizer.Tokenizer.SplitMode.A,
"A": tokenizer.Tokenizer.SplitMode.A,
"B": tokenizer.Tokenizer.SplitMode.B,
"C": tokenizer.Tokenizer.SplitMode.C,
}[split_mode]
tok = dictionary.Dictionary().create(mode=split_mode)
return tok
except ImportError:
raise ImportError(
"Japanese support requires SudachiPy and SudachiDict-core "
"(https://github.com/WorksApplications/SudachiPy). "
"Install with `pip install sudachipy sudachidict_core` or "
"install spaCy with `pip install spacy[ja]`."
) from None
def resolve_pos(orth, tag, next_tag):
"""If necessary, add a field to the POS tag for UD mapping.
Under Universal Dependencies, sometimes the same Unidic POS tag can
be mapped differently depending on the literal token or its context
in the sentence. This function returns resolved POSs for both token
and next_token by tuple.
"""
# Some tokens have their UD tag decided based on the POS of the following
# token.
# apply orth based mapping
if tag in TAG_ORTH_MAP:
orth_map = TAG_ORTH_MAP[tag]
if orth in orth_map:
return orth_map[orth], None # current_pos, next_pos
# apply tag bi-gram mapping
if next_tag:
tag_bigram = tag, next_tag
if tag_bigram in TAG_BIGRAM_MAP:
current_pos, next_pos = TAG_BIGRAM_MAP[tag_bigram]
if current_pos is None: # apply tag uni-gram mapping for current_pos
return (
TAG_MAP[tag][POS],
next_pos,
) # only next_pos is identified by tag bi-gram mapping
else:
return current_pos, next_pos
# apply tag uni-gram mapping
return TAG_MAP[tag][POS], None
def get_dtokens_and_spaces(dtokens, text, gap_tag="空白"):
# Compare the content of tokens and text, first
words = [x.surface for x in dtokens]
if "".join("".join(words).split()) != "".join(text.split()):
raise ValueError(Errors.E194.format(text=text, words=words))
text_dtokens = []
text_spaces = []
text_pos = 0
# handle empty and whitespace-only texts
if len(words) == 0:
return text_dtokens, text_spaces
elif len([word for word in words if not word.isspace()]) == 0:
assert text.isspace()
text_dtokens = [DetailedToken(text, gap_tag, "", text, None, None)]
text_spaces = [False]
return text_dtokens, text_spaces
# align words and dtokens by referring text, and insert gap tokens for the space char spans
for i, (word, dtoken) in enumerate(zip(words, dtokens)):
# skip all space tokens
if word.isspace():
continue
try:
word_start = text[text_pos:].index(word)
except ValueError:
raise ValueError(Errors.E194.format(text=text, words=words)) from None
# space token
if word_start > 0:
w = text[text_pos : text_pos + word_start]
text_dtokens.append(DetailedToken(w, gap_tag, "", w, None, None))
text_spaces.append(False)
text_pos += word_start
# content word
text_dtokens.append(dtoken)
text_spaces.append(False)
text_pos += len(word)
# poll a space char after the word
if i + 1 < len(dtokens) and dtokens[i + 1].surface == " ":
text_spaces[-1] = True
text_pos += 1
# trailing space token
if text_pos < len(text):
w = text[text_pos:]
text_dtokens.append(DetailedToken(w, gap_tag, "", w, None, None))
text_spaces.append(False)
return text_dtokens, text_spaces
__all__ = ["Japanese"]