模型:
KoichiYasuoka/bert-base-japanese-wikipedia-ud-head
这是一个在日本维基百科文本上进行依赖分析(检测长单位词的头)的BERT模型,作为问答系统进行预训练,来源于 bert-base-japanese-char-extended 和 UD_Japanese-GSDLUW 。在上下文中使用[MASK]以避免在提问时出现多次使用的单词的歧义。
from transformers import AutoTokenizer,AutoModelForQuestionAnswering,QuestionAnsweringPipeline tokenizer=AutoTokenizer.from_pretrained("KoichiYasuoka/bert-base-japanese-wikipedia-ud-head") model=AutoModelForQuestionAnswering.from_pretrained("KoichiYasuoka/bert-base-japanese-wikipedia-ud-head") qap=QuestionAnsweringPipeline(tokenizer=tokenizer,model=model,align_to_words=False) print(qap(question="国語",context="全学年にわたって小学校の国語の教科書に挿し絵が用いられている"))
或者(配合 ufal.chu-liu-edmonds 使用)
class TransformersUD(object): def __init__(self,bert): import os from transformers import (AutoTokenizer,AutoModelForQuestionAnswering, AutoModelForTokenClassification,AutoConfig,TokenClassificationPipeline) self.tokenizer=AutoTokenizer.from_pretrained(bert) self.model=AutoModelForQuestionAnswering.from_pretrained(bert) x=AutoModelForTokenClassification.from_pretrained if os.path.isdir(bert): d,t=x(os.path.join(bert,"deprel")),x(os.path.join(bert,"tagger")) else: from transformers.utils import cached_file c=AutoConfig.from_pretrained(cached_file(bert,"deprel/config.json")) d=x(cached_file(bert,"deprel/pytorch_model.bin"),config=c) s=AutoConfig.from_pretrained(cached_file(bert,"tagger/config.json")) t=x(cached_file(bert,"tagger/pytorch_model.bin"),config=s) self.deprel=TokenClassificationPipeline(model=d,tokenizer=self.tokenizer, aggregation_strategy="simple") self.tagger=TokenClassificationPipeline(model=t,tokenizer=self.tokenizer) def __call__(self,text): import numpy,torch,ufal.chu_liu_edmonds w=[(t["start"],t["end"],t["entity_group"]) for t in self.deprel(text)] z,n={t["start"]:t["entity"].split("|") for t in self.tagger(text)},len(w) r,m=[text[s:e] for s,e,p in w],numpy.full((n+1,n+1),numpy.nan) v,c=self.tokenizer(r,add_special_tokens=False)["input_ids"],[] for i,t in enumerate(v): q=[self.tokenizer.cls_token_id]+t+[self.tokenizer.sep_token_id] c.append([q]+v[0:i]+[[self.tokenizer.mask_token_id]]+v[i+1:]+[[q[-1]]]) b=[[len(sum(x[0:j+1],[])) for j in range(len(x))] for x in c] with torch.no_grad(): d=self.model(input_ids=torch.tensor([sum(x,[]) for x in c]), token_type_ids=torch.tensor([[0]*x[0]+[1]*(x[-1]-x[0]) for x in b])) s,e=d.start_logits.tolist(),d.end_logits.tolist() for i in range(n): for j in range(n): m[i+1,0 if i==j else j+1]=s[i][b[i][j]]+e[i][b[i][j+1]-1] h=ufal.chu_liu_edmonds.chu_liu_edmonds(m)[0] if [0 for i in h if i==0]!=[0]: i=([p for s,e,p in w]+["root"]).index("root") j=i+1 if i<n else numpy.nanargmax(m[:,0]) m[0:j,0]=m[j+1:,0]=numpy.nan h=ufal.chu_liu_edmonds.chu_liu_edmonds(m)[0] u="# text = "+text.replace("\n"," ")+"\n" for i,(s,e,p) in enumerate(w,1): p="root" if h[i]==0 else "dep" if p=="root" else p u+="\t".join([str(i),r[i-1],"_",z[s][0][2:],"_","|".join(z[s][1:]), str(h[i]),p,"_","_" if i<n and e<w[i][0] else "SpaceAfter=No"])+"\n" return u+"\n" nlp=TransformersUD("KoichiYasuoka/bert-base-japanese-wikipedia-ud-head") print(nlp("全学年にわたって小学校の国語の教科書に挿し絵が用いられている"))