ud.py

from transformers import TokenClassificationPipeline

class UniversalDependenciesPipeline(TokenClassificationPipeline):
  def _forward(self,model_inputs):
    import torch
    v=model_inputs["input_ids"][0].tolist()
    with torch.no_grad():
      e=self.model(input_ids=torch.tensor([v[0:i]+[self.tokenizer.mask_token_id]+v[i+1:]+[j] for i,j in enumerate(v[1:-1],1)],device=self.device))
    return {"logits":e.logits[:,1:-2,:],**model_inputs}
  def postprocess(self,model_outputs,**kwargs):
    import numpy
    e=model_outputs["logits"].numpy()
    r=[1 if i==0 else -1 if j.endswith("|root") else 0 for i,j in sorted(self.model.config.id2label.items())]
    e+=numpy.where(numpy.add.outer(numpy.identity(e.shape[0]),r)==0,0,numpy.nan)
    g=self.model.config.label2id["X|_|goeswith"]
    r=numpy.tri(e.shape[0])
    for i in range(e.shape[0]):
      for j in range(i+2,e.shape[1]):
        r[i,j]=r[i,j-1] if numpy.nanargmax(e[i,j-1])==g else 1
    e[:,:,g]+=numpy.where(r==0,0,numpy.nan)
    m,p=numpy.nanmax(e,axis=2),numpy.nanargmax(e,axis=2)
    h=self.chu_liu_edmonds(m)
    z=[i for i,j in enumerate(h) if i==j]
    if len(z)>1:
      k,h=z[numpy.nanargmax(m[z,z])],numpy.nanmin(m)-numpy.nanmax(m)
      m[:,z]+=[[0 if j in z and (i!=j or i==k) else h for i in z] for j in range(m.shape[0])]
      h=self.chu_liu_edmonds(m)
    t=model_outputs["sentence"].replace("\n"," ")
    v=[(s,e,c if c!=self.tokenizer.unk_token else t[s:e]) for (s,e),c in zip(model_outputs["offset_mapping"][0].tolist(),self.tokenizer.convert_ids_to_tokens(model_outputs["input_ids"][0].tolist())) if s<e]
    q=[self.model.config.id2label[p[j,i]].split("|") for i,j in enumerate(h)]
    g="aggregation_strategy" in kwargs and kwargs["aggregation_strategy"]!="none"
    if g:
      for i,j in reversed(list(enumerate(q[1:],1))):
        if j[-1]=="goeswith" and set([k[-1] for k in q[h[i]+1:i+1]])=={"goeswith"}:
          h=[b if i>b else b-1 for a,b in enumerate(h) if i!=a]
          s,e,c=v.pop(i)
          v[i-1]=(v[i-1][0],e,v[i-1][2]+c)
          q.pop(i)
    u="\n"
    z={"a":"ア","i":"イ","u":"ウ","e":"エ","o":"オ","k":"ㇰ","s":"ㇱ","t":"ㇳ","n":"ㇴ","h":"ㇷ","m":"ㇺ","r":"ㇽ","p":"ㇷ゚"}
    f=-1
    for i,(s,e,c) in reversed(list(enumerate(v))):
      if t[s]=="\u309a":
        s-=1
      w,x=[j for j in t[s:e]],""
      if i>0 and s<v[i-1][1]:
        w[0]=z[c[0]] if c[0] in z else "ッ"
        f=max(f,i)
      elif f>0:
        x="{}-{}\t{}\t_\t_\t_\t_\t_\t_\t_\t{}\n".format(i+1,f+1,t[s:v[f][1]],"_" if f+1<len(v) and v[f][1]<v[f+1][0] else "SpaceAfter=No")
        f=-1
      if i+1<len(v) and e>v[i+1][0]:
        w[-1]=z[c[-1]] if c[-1] in z else "ッ"
      l=("".join(w).replace(" ","") if max(w)<"z" else c).replace("sh","s").replace("ch","c").replace("au","aw").replace("iu","iw").replace("eu","ew").replace("ou","ow").replace("ai","ay").replace("ui","uy").replace("ei","ey").replace("oi","oy") if g else "_"
      u=x+"\t".join([str(i+1),"".join(w),l,q[i][0],"|".join(q[i][1:-1]),"_",str(0 if h[i]==i else h[i]+1),q[i][-1],"_","_" if i+1<len(v) and e<v[i+1][0] else "SpaceAfter=No"])+"\n"+u
    return "# text = "+t+"\n"+u
  def chu_liu_edmonds(self,matrix):
    import numpy
    h=numpy.nanargmax(matrix,axis=0)
    x=[-1 if i==j else j for i,j in enumerate(h)]
    for b in [lambda x,i,j:-1 if i not in x else x[i],lambda x,i,j:-1 if j<0 else x[j]]:
      y=[]
      while x!=y:
        y=list(x)
        for i,j in enumerate(x):
          x[i]=b(x,i,j)
      if max(x)<0:
        return h
    y,x=[i for i,j in enumerate(x) if j==max(x)],[i for i,j in enumerate(x) if j<max(x)]
    z=matrix-numpy.nanmax(matrix,axis=0)
    m=numpy.block([[z[x,:][:,x],numpy.nanmax(z[x,:][:,y],axis=1).reshape(len(x),1)],[numpy.nanmax(z[y,:][:,x],axis=0),numpy.nanmax(z[y,y])]])
    k=[j if i==len(x) else x[j] if j<len(x) else y[numpy.nanargmax(z[y,x[i]])] for i,j in enumerate(self.chu_liu_edmonds(m))]
    h=[j if i in y else k[x.index(i)] for i,j in enumerate(h)]
    i=y[numpy.nanargmax(z[x[k[-1]],y] if k[-1]<len(x) else z[y,y])]
    h[i]=x[k[-1]] if k[-1]<len(x) else i
    return h