README.md

---
language:
- "ja"
tags:
- "japanese"
- "question-answering"
- "dependency-parsing"
datasets:
- "universal_dependencies"
license: "cc-by-sa-4.0"
pipeline_tag: "question-answering"
inference:
  parameters:
    align_to_words: false
widget:
- text: "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"
- text: "教科書"
  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"
- text: "の"
  context: "全学年にわたって小学校の国語[MASK]教科書に挿し絵が用いられている"
---

# deberta-large-japanese-unidic-ud-head

## Model Description

This is a DeBERTa(V2) model pretrained on 青空文庫 for dependency-parsing (head-detection on long-unit-words) as question-answering, derived from [deberta-large-japanese-unidic](https://huggingface.co/KoichiYasuoka/deberta-large-japanese-unidic) and [UD_Japanese-GSDLUW](https://github.com/UniversalDependencies/UD_Japanese-GSDLUW). Use [MASK] inside `context` to avoid ambiguity when specifying a multiple-used word as `question`.

## How to Use

```py
import torch
from transformers import AutoTokenizer,AutoModelForQuestionAnswering
tokenizer=AutoTokenizer.from_pretrained("KoichiYasuoka/deberta-large-japanese-unidic-ud-head")
model=AutoModelForQuestionAnswering.from_pretrained("KoichiYasuoka/deberta-large-japanese-unidic-ud-head")
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"
inputs=tokenizer(question,context,return_tensors="pt")
outputs=model(**inputs)
start,end=torch.argmax(outputs.start_logits),torch.argmax(outputs.end_logits)
print(tokenizer.convert_ids_to_tokens(inputs["input_ids"][0,start:end+1]))
```

or

```py
from transformers import (AutoTokenizer,AutoModelForQuestionAnswering,
  AutoModelForTokenClassification,AutoConfig,TokenClassificationPipeline)
class TaggerPipeline(TokenClassificationPipeline):
  def __call__(self,text):
    d=super().__call__(text)
    if len(d)>0 and ("start" not in d[0] or d[0]["start"]==None):
      import spacy_alignments as tokenizations
      v=[x["word"].replace(" ","") for x in d]
      a2b,b2a=tokenizations.get_alignments(v,text)
      for i,t in enumerate(a2b):
        s,e=(0,0) if t==[] else (t[0],t[-1]+1)
        if v[i].startswith(self.tokenizer.unk_token):
          s=([[-1]]+[x for x in a2b[0:i] if x>[]])[-1][-1]+1
        if v[i].endswith(self.tokenizer.unk_token):
          e=([x for x in a2b[i+1:] if x>[]]+[[len(text)]])[0][0]
        d[i]["start"],d[i]["end"]=s,e
    return d
class TransformersSlowUD(object):
  def __init__(self,bert):
    import os
    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=TaggerPipeline(model=d,tokenizer=self.tokenizer,
      aggregation_strategy="simple")
    self.tagger=TaggerPipeline(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=TransformersSlowUD("KoichiYasuoka/deberta-large-japanese-unidic-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"))
```

[fugashi](https://pypi.org/project/fugashi) [unidic-lite](https://pypi.org/project/unidic-lite) [spacy-alignments](https://pypi.org/project/spacy-alignments) and [ufal.chu-liu-edmonds](https://pypi.org/project/ufal.chu-liu-edmonds/) required.