"""Zero-shot Classification related modeling class"""
from typing import Dict, List, Optional
from pororo.tasks.utils.base import PororoBiencoderBase, PororoFactoryBase
[docs]class PororoZeroShotFactory(PororoFactoryBase):
"""
Zero-shot topic classification
See also:
https://joeddav.github.io/blog/2020/05/29/ZSL.html
Korean (`brainbert.base.ko.kornli`)
- dataset: KorNLI (Ham et al. 2020)
- metric: N/A
English (`roberta.base.en.nli`)
- dataset: MNLI (Adina Williams et al. 2017)
- metric: N/A
Japanese (`jaberta.base.ja.nli`)
- dataset: XNLI (Alexis Conneau et al. 2018)
- metric: N/A
Chinese (`zhberta.base.zh.nli`)
- dataset: XNLI (Alexis Conneau et al. 2018)
- metric: N/A
Examples:
>>> zsl = Pororo(task="zero-topic")
>>> zsl("Who are you voting for in 2020?", ["business", "art & culture", "politics"])
{'business': 33.23, 'art & culture': 8.33, 'politics': 96.12}
>>> zsl = Pororo(task="zero-topic", lang="ko")
>>> zsl('''라리가 사무국, 메시 아닌 바르사 지지..."바이 아웃 유효" [공식발표]''', ["스포츠", "사회", "정치", "경제", "생활/문화", "IT/과학"])
{'스포츠': 94.15, '사회': 37.11, '정치': 74.26, '경제': 39.18, '생활/문화': 71.15, 'IT/과학': 34.71}
>>> zsl('''장제원, 김종인 당무감사 추진에 “참 잔인들 하다”···정강정책 개정안은 “졸작”''', ["스포츠", "사회", "정치", "경제", "생활/문화", "IT/과학"])
{'스포츠': 2.18, '사회': 56.1, '정치': 88.24, '경제': 16.17, '생활/문화': 66.13, 'IT/과학': 11.2}
>>> zsl = Pororo(task="zero-topic", lang="ja")
>>> zsl("香川 真司は、兵庫県神戸市垂水区出身のプロサッカー選手。元日本代表。ポジションはMF、FW。ボルシア・ドルトムント時代の2010-11シーズンでリーグ前半期17試合で8得点を記録し9シーズンぶりのリーグ優勝に貢献。キッカー誌が選定したブンデスリーガの年間ベスト イレブンに名を連ねた。", ["スポーツ", "政治", "技術"])
{'スポーツ': 0.2, '政治': 99.71, '技術': 68.9}
>>> zsl = Pororo(task="zero-topic", lang="zh")
>>> zsl("商务部14日发布数据显示，今年前10个月，我国累计对外投资904.6亿美元，同比增长5.9%。", ["政治", "经济", "国际化"])
{'政治': 33.72, '经济': 3.9, '国际化': 13.67}
"""
def __init__(self, task: str, lang: str, model: Optional[str]):
super().__init__(task, lang, model)
[docs] @staticmethod
def get_available_langs():
return ["en", "ko", "ja", "zh"]
[docs] @staticmethod
def get_available_models():
return {
"ko": ["brainbert.base.ko.kornli"],
"ja": ["jaberta.base.ja.nli"],
"zh": ["zhberta.base.zh.nli"],
"en": ["roberta.base.en.nli"],
}
[docs] def load(self, device: str):
"""
Load user-selected task-specific model
Args:
device (str): device information
Returns:
object: User-selected task-specific model
"""
if "brainbert" in self.config.n_model:
from pororo.models.brainbert import BrainRobertaModel
model = BrainRobertaModel.load_model(
f"bert/{self.config.n_model}",
self.config.lang,
).eval().to(device)
return PororoBertZeroShot(model, self.config)
if "jaberta" in self.config.n_model:
from pororo.models.brainbert import JabertaModel
model = JabertaModel.load_model(
f"bert/{self.config.n_model}",
self.config.lang,
).eval().to(device)
return PororoBertZeroShot(model, self.config)
if "zhberta" in self.config.n_model:
from pororo.models.brainbert import ZhbertaModel
model = ZhbertaModel.load_model(
f"bert/{self.config.n_model}",
self.config.lang,
).eval().to(device)
return PororoBertZeroShot(model, self.config)
if "roberta" in self.config.n_model:
from pororo.models.brainbert import CustomRobertaModel
model = CustomRobertaModel.load_model(
f"bert/{self.config.n_model}",
self.config.lang,
).eval().to(device)
return PororoBertZeroShot(model, self.config)
[docs]class PororoBertZeroShot(PororoBiencoderBase):
def __init__(self, model, config):
super().__init__(config)
self._model = model
self._template = {
"ko": "이 문장은 {label}에 관한 것이다.",
"ja": "この文は、{label}に関するものである。",
"zh": "这句话是关于{label}的。",
"en": "This sentence is about {label}.",
}
[docs] def predict(
self,
sent: str,
labels: List[str],
**kwargs,
) -> Dict[str, float]:
"""
Conduct zero-shot classification
Args:
sent (str): sentence to be classified
labels (List[str]): candidate labels
Returns:
List[Tuple(str, float)]: confidence scores corresponding to each input label
"""
cands = [
self._template[self.config.lang].format(label=label)
for label in labels
]
result = dict()
for label, cand in zip(labels, cands):
if self.config.lang == "ko":
tokens = self._model.encode(
sent,
cand,
add_special_tokens=True,
no_separator=False,
)
else:
tokens = self._model.encode(
sent,
cand,
no_separator=False,
)
# throw away "neutral" (dim 1) and take the probability of "entail" (2) as the probability of the label being true
pred = self._model.predict(
"sentence_classification_head",
tokens,
return_logits=True,
)[:, [0, 2]]
prob = pred.softmax(dim=1)[:, 1].item() * 100
result[label] = round(prob, 2)
return result