def __init__(self,
category: str,
midasi: str,
sentence: BList,
mid_range: range,
mrph2dmid: Dict[Morpheme, int]):
self.category: str = category
self.midasi: str = midasi
self.sid: str = sentence.sid
self.mid_range: range = mid_range
dmid_start = mrph2dmid[sentence.mrph_list()[mid_range[0]]]
dmid_end = mrph2dmid[sentence.mrph_list()[mid_range[-1]]]
self.dmid_range: range = range(dmid_start, dmid_end + 1)
def split_kc(input_dir: Path, output_dir: Path, max_subword_length: int,
tokenizer: BertTokenizer):
"""
各文書を，tokenize したあとの長さが max_subword_length 以下になるように複数の文書に分割する．
1文に分割しても max_subword_length を超えるような長い文はそのまま出力する
"""
did2sids: Dict[str, List[str]] = defaultdict(list)
did2cumlens: Dict[str, List[int]] = {}
sid2knp: Dict[str, str] = {}
for knp_file in input_dir.glob('*.knp'):
with knp_file.open() as fin:
did = knp_file.stem
did2cumlens[did] = [0]
buff = ''
for line in fin:
buff += line
if line.strip() == 'EOS':
blist = BList(buff)
did2sids[did].append(blist.sid)
did2cumlens[did].append(did2cumlens[did][-1] + len(
tokenizer.tokenize(' '.join(
m.midasi for m in blist.mrph_list()))))
sid2knp[blist.sid] = buff
buff = ''
for did, sids in did2sids.items():
cum: List[int] = did2cumlens[did]
end = 1
# end を探索
while end < len(sids) and cum[end + 1] - cum[0] <= max_subword_length:
end += 1
idx = 0
while end < len(sids) + 1:
start = 0
# start を探索
while cum[end] - cum[start] > max_subword_length:
start += 1
if start == end - 1:
break
with output_dir.joinpath(f'{did}-{idx:02}.knp').open(
mode='w') as fout:
fout.write(''.join(
sid2knp[sid]
for sid in sids[start:end])) # start から end まで書き出し
idx += 1
end += 1
class Sentence:
""" KWDLC(または Kyoto Corpus)の1文書を扱うクラス
Attributes:
blist (BList): KNPのBListオブジェクト
doc_id (str): 文書ID
bps (List[BasePhrase]): 含まれる基本句のリスト
"""
def __init__(
self,
knp_string: str,
dtid_offset: int,
dmid_offset: int,
doc_id: str,
) -> None:
"""
Args:
knp_string(str): 1文についてのKNPのtab出力
dtid_offset (int): 文書中でこの文が始まるまでの文書レベル基本句ID
dmid_offset (int): 文書中でこの文が始まるまでの文書レベル形態素ID
doc_id(str): 文書ID
"""
self.blist = BList(knp_string)
self.doc_id: str = doc_id
self.bps: List[BasePhrase] = []
dtid = dtid_offset
dmid = dmid_offset
for tag in self.blist.tag_list():
base_phrase = BasePhrase(tag, dmid, dtid, self.blist.sid, doc_id)
self.bps.append(base_phrase)
dtid += 1
dmid += len(base_phrase)
self._mrph2dmid: Dict[Morpheme, int] = dict(
ChainMap(*(bp.mrph2dmid for bp in self.bps)))
for bp in self.bps:
if bp.tag.parent_id >= 0:
bp.parent = self.bps[bp.tag.parent_id]
for child in bp.tag.children:
bp.children.append(self.bps[child.tag_id])
@property
def sid(self) -> str:
"""文ID"""
return self.blist.sid
@property
def dtids(self) -> List[int]:
return [bp.dtid for bp in self.bps]
@property
def mrph2dmid(self) -> Dict[Morpheme, int]:
"""形態素とその文書レベルIDを紐付ける辞書"""
return self._mrph2dmid
@property
def surf(self) -> str:
"""表層表現"""
return ''.join(bp.surf for bp in self.bps)
def bnst_list(self):
return self.blist.bnst_list()
def tag_list(self):
return self.blist.tag_list()
def mrph_list(self):
return self.blist.mrph_list()
def __len__(self) -> int:
"""含まれる基本句の数"""
return len(self.bps)
def __getitem__(self, tid: int) -> Optional[BasePhrase]:
if 0 <= tid < len(self):
return self.bps[tid]
else:
logger.error(f'base phrase: {tid} out of range')
return None
def __iter__(self) -> Iterator[BasePhrase]:
return iter(self.bps)
def __eq__(self, other: 'Sentence') -> bool:
return self.sid == other.sid
def __str__(self) -> str:
return self.surf
def __repr__(self) -> str:
return f'Sentence(\'{self.surf}\', sid: {self.sid})'
class Sentence:
"""A class to represent a single sentence.
Attributes:
blist (BList): BList object of pyknp.
doc_id (str): The document ID of this sentence.
bps (List[BasePhrase]): Base phrases in this sentence.
"""
def __init__(
self,
knp_string: str,
dtid_offset: int,
dmid_offset: int,
doc_id: str,
) -> None:
"""
Args:
knp_string(str): KNP format string of this sentence.
dtid_offset (int): The document-wide tag ID of the previous base phrase.
dmid_offset (int): The document-wide morpheme ID of the previous morpheme.
doc_id(str): The document ID of this sentence.
"""
self.blist = BList(knp_string)
self.doc_id: str = doc_id
self.bps: List[BasePhrase] = []
dtid = dtid_offset
dmid = dmid_offset
for tag in self.blist.tag_list():
base_phrase = BasePhrase(tag, dmid, dtid, self.blist.sid, doc_id)
self.bps.append(base_phrase)
dtid += 1
dmid += len(base_phrase)
self._mrph2dmid: Dict[Morpheme, int] = dict(
ChainMap(*(bp.mrph2dmid for bp in self.bps)))
for bp in self.bps:
if bp.tag.parent_id >= 0:
bp.parent = self.bps[bp.tag.parent_id]
for child in bp.tag.children:
bp.children.append(self.bps[child.tag_id])
@property
def sid(self) -> str:
"""A sentence ID."""
return self.blist.sid
@property
def dtids(self) -> List[int]:
"""A document-wide tag ID."""
return [bp.dtid for bp in self.bps]
@property
def mrph2dmid(self) -> Dict[Morpheme, int]:
"""A mapping from morpheme to its document-wide ID."""
return self._mrph2dmid
@property
def surf(self) -> str:
"""A surface expression"""
return ''.join(bp.surf for bp in self.bps)
def bnst_list(self):
"""Return list of Bunsetsu object in pyknp."""
return self.blist.bnst_list()
def tag_list(self):
"""Return list of Tag object in pyknp."""
return self.blist.tag_list()
def mrph_list(self):
"""Return list of Morpheme object in pyknp."""
return self.blist.mrph_list()
def __len__(self) -> int:
"""Number of base phrases in this sentence"""
return len(self.bps)
def __getitem__(self, tid: int) -> Optional[BasePhrase]:
if 0 <= tid < len(self):
return self.bps[tid]
else:
logger.error(f'base phrase: {tid} out of range')
return None
def __iter__(self) -> Iterator[BasePhrase]:
return iter(self.bps)
def __eq__(self, other: 'Sentence') -> bool:
return self.sid == other.sid
def __str__(self) -> str:
return self.surf
def __repr__(self) -> str:
return f'Sentence(\'{self.surf}\', sid: {self.sid})'
def split_kc(input_dir: Path, output_dir: Path, max_subword_length: int, tokenizer: TokenizeHandlerMeta):
"""
各文書を，tokenize したあとの長さが max_subword_length 以下になるように複数の文書に分割する．
1文に分割しても max_subword_length を超えるような長い文はそのまま出力する
"""
did2sids: Dict[str, List[str]] = defaultdict(list)
did2cumlens: Dict[str, List[int]] = {}
sid2knp: Dict[str, str] = {}
max_all_tokens_len = 0
for knp_file in input_dir.glob('*.knp'):
with knp_file.open() as fin:
did = knp_file.stem
did2cumlens[did] = [0]
buff = ''
for line in fin:
buff += line
if line.strip() == 'EOS':
blist = BList(buff)
did2sids[did].append(blist.sid)
all_tokens, *_ = tokenizer.get_tokenized_tokens(list(m.midasi for m in blist.mrph_list()))
max_all_tokens_len = max(max_all_tokens_len, len(all_tokens))
did2cumlens[did].append(
did2cumlens[did][-1] + len(all_tokens)
# did2cumlens[did][-1] + len(tokenizer.tokenize(' '.join(m.midasi for m in blist.mrph_list())))
)
sid2knp[blist.sid] = buff
buff = ''
print(f"max_tokens_length per sentence -> {max_all_tokens_len}")
# assert max_all_tokens_len <= max_subword_length
# if max_all_tokens_len > max_subword_length:
# raise ValueError(f"max_tokens_length exceeded max_subword_length\n{max_all_tokens_len}>{max_subword_length}")
document_divide_unit_list = []
for did, sids in did2sids.items():
cum: List[int] = did2cumlens[did]
end = 1
# end を探索
while end < len(sids) and cum[end+1] - cum[0] <= max_subword_length:
end += 1
idx = 0
while end < len(sids) + 1:
start = 0
# start を探索
while cum[end] - cum[start] > max_subword_length:
start += 1
if start == end - 1:
break
document_divide_unit_list.append(
DocumentDivideUnit(did, idx, start, end)
)
# with output_dir.joinpath(f'{did}-{idx:02}.knp').open('wt') as fout:
# fout.write(''.join(sid2knp[sid] for sid in sids[start:end])) # start から end まで書き出し
idx += 1
end += 1
_write_partial_document = partial(
write_partial_document,
did2sids=did2sids,
sid2knp=sid2knp,
output_dir=output_dir
)
with Pool() as pool:
list(pool.imap(_write_partial_document, document_divide_unit_list))
