chart.c

中心词驱动的短语结构句法分析器。该模型考虑了跟随介词短语的名词短语的中心词的作用。 有MIT大学Colling开发

	    k=add_edge(i,i);

	  }

      add_singles_stops(i,i);

    }
}
void add_singles(int s,int e,int si,int ei)
{
  int i,j,k;
  edge_type etemp;
  double prob,probl,probr;
  int nl,nr;
  int il,ir;
  int flaggap;
  int wh,th,p,ch;

  for(i=si;i<=ei;i++)
    if(edges[i].valid==1&&inbeam(i,s,e)&&edges[i].stop==1)
      for(j=0;j<unary_nums[edges[i].label];j++)
	if(unaries[edges[i].label][j]!=-1)
	{
	  etemp=edges[i];
	  etemp.stop=0;
	  etemp.type=1;
	  etemp.label=unaries[edges[i].label][j];
	  etemp.headch=i;

	  etemp.headlabel=edges[i].label;

	  if(etemp.label == NT_NPB)
	    etemp.hasverb = 0;

	  clear_args(&etemp.lc);
	  clear_args(&etemp.rc);
	  
	  etemp.lc.nvs=0;
	  etemp.lc.adj=1;
	  
	  etemp.rc.nvs=0;
	  etemp.rc.adj=1;

	  wh=current->wordnos[edges[i].head];
	  th=edges[i].headtag;
	  p=etemp.label;
	  ch=edges[i].label;

	  prob=get_unary_prob_witheffhash(ch,wh,th,p,&new_hash,&eff_hash)+edges[i].prob;

	  nl=lsubcats_counts[p][ch];
	  nr=rsubcats_counts[p][ch];

	  flaggap=(hasgap[etemp.label] && !hasgap[edges[i].label]);

	  for(il=0;il<nl;il++)
	    {
	      if(nl==1)
		probl=0;
	      else
		probl=get_subcat_prob_witheffhash(lsubcats[p][ch][il],ch,wh,th,p,0,&new_hash,&eff_hash);


	      find_args(&etemp.lc,lsubcats[p][ch][il]);
	      for(ir=0;ir<nr;ir++)
		{
		  if(nr==1)
		    probr=0;
		  else
		    probr=get_subcat_prob_witheffhash(rsubcats[p][ch][ir],ch,wh,th,p,1,&new_hash,&eff_hash);

		  find_args(&etemp.rc,rsubcats[p][ch][ir]);
		  etemp.prob=prob+probl+probr;

		  if(flaggap==0)
		    {
		      edges[numedges]=etemp;
		      if((k=add_edge(s,e))!=-1)
			{
			  edges[k].child1=numchilds;
			  edges[k].numchild=1;
			  childs[numchilds]=i;
			  numchilds++;
			}
		    }
		  else
		    {
		      edges[numedges].prob=prob+probl+probr;

		      /*case3 -- from child to left*/
		      edges[numedges]=etemp;
		      edges[numedges].lc.gap=1;
		      edges[numedges].prob+=get_gap_prob_witheffhash(3,ch,wh,th,p,&new_hash,&eff_hash);

		      if((k=add_edge(s,e))!=-1)
			{
			  edges[k].child1=numchilds;
			  edges[k].numchild=1;
			  childs[numchilds]=i;
			  numchilds++;
			}

		      /*case4 -- from child to right*/
		      edges[numedges]=etemp;
		      edges[numedges].rc.gap=1;
		      edges[numedges].prob+=get_gap_prob_witheffhash(4,ch,wh,th,p,&new_hash,&eff_hash);
/*		      printf("CHECK %g %g\n",edges[i].prob,edges[numedges].prob);
		      print_edge(i,0);*/
		      if((k=add_edge(s,e))!=-1)
			{
			  edges[k].child1=numchilds;
			  edges[k].numchild=1;
			  childs[numchilds]=i;
			  numchilds++;
			}
		    }
		}
	    }
	}
}


void print_edge(int e,int off)
{
  int i,j,newoff;
  int flag;

  flag=1;

  if(e==-1)
    {
      printf("TRACE T\n");
      return;
    }

  if( (edges[e].label == NT_NP || edges[e].label == NT_NPA)
     && edges[e].numchild ==1)
    {
      j= childs[edges[e].child1];
/*      if(edges[j].label == NT_NPB)
	flag=0;*/
    }
    

  if(flag)
    {
      printf("%s %g ",nts[edges[e].label],edges[e].prob);
      newoff=off+strlen(nts[edges[e].label])+1;
    }
  else
    newoff=off;

  /*POS tag case*/
  if(edges[e].type==0)
    {
      printf("%s\n",current->words[current->wordpos[edges[e].head]]);
      return;
    }

  print_edge(childs[edges[e].child1],newoff);

  for(i=edges[e].child1+1;i<edges[e].child1+edges[e].numchild;i++)
    {    
      for(j=0;j<newoff;j++)
	printf(" ");
      print_edge(childs[i],newoff);
    }

}

int find_childno(int e,int c)
{
  int i;

  for(i=0;i<edges[e].numchild;i++)
    {
      if(childs[edges[e].child1+i] == c)
	return i+1;
    }

  return -1;
}

void print_wholent(int e)
{

  printf("%s",nts[edges[e].label]);

  if(edges[e].type==4)
    {
      printf("_NA~%d",edges[e].numchild);
      return;
    }
  
  if(1)
    {
      printf("~%s",current->words[current->wordpos[edges[e].head]]);
      printf("~%d",edges[e].numchild);
      printf("~%d",find_childno(e,edges[e].headch));
    }

}

void print_edges_flat(int e)
{
  int i,j,w,next;
  int flag;

  flag=1;

  if(e==-1)
    {
      printf("T/TRACE ");
      return;
    }

if(TREEBANKOUTPUTFLAG &&
   (edges[e].label == NT_NP || edges[e].label == NT_NPA)
     && edges[e].numchild ==1)
    {
      j= childs[edges[e].child1];
      if(edges[j].label == NT_NPB)
	flag=0;
    }


  /*POS tag case*/

  /*POS tag - print preceding punctuation*/
  if(edges[e].type==0)
    {
      if(edges[e].head==0)
	for(i=0;i<current->wordpos[edges[e].head];i++)
	  printf("%s/PUNC%s ",current->words[i],current->tags[i]);

      printf("%s/%s ",current->words[current->wordpos[edges[e].head]]
	     ,nts[edges[e].headtag]);
      w=current->wordpos[edges[e].head];
      if(edges[e].head==current->nws_np-1)
	next=current->nws;
      else
	next=current->wordpos[edges[e].head+1];
      for(i=w+1;i<next;i++)
	printf("%s/PUNC%s ",current->words[i],current->tags[i]);
      return;
    }

  if(flag)
    {
      printf("(");
      print_wholent(e);
      printf(" ");
    }

  print_edges_flat(childs[edges[e].child1]);

  for(i=edges[e].child1+1;i<edges[e].child1+edges[e].numchild;i++)
    {    
      print_edges_flat(childs[i]);
    }

  if(flag)
    printf(") ");
}

void print_chart()
{
  int i;
  int dist,subcat;

  for(i=0;i<numedges;i++)
    {
      printf("EDGE %d %g %g %d %d %d ",i,edges[i].prob,edges[i].prob2,edges[i].stop,edges[i].type,edges[i].label);

      printf(" L ");
      onectxt_to_dist_subcat(&edges[i].lc,&dist,&subcat,1);
      printf("%d %d HV%d",dist,subcat,edges[i].hasverb);

      printf(" R ");
      onectxt_to_dist_subcat(&edges[i].rc,&dist,&subcat,0);
      printf("%d %d\n",dist,subcat);

      print_edge(i,0);
      printf("\n\n");
    }

}

void join_2_edges_follow(int e1,int s,int m,int e,int *e2s,int ne2s)
{
  int h2_is_verb,e2,i2;

  int e1hlabel;

  /*e3,e4,d3,d4 are variables for new edges/distances created when
    e1/e2 are respectively the heads

    e3flag,e4flag = 0/1 if e3/e4 haven't/have been created
  */

  edge_type e3;
  int e3flag=0;
  int e3flag2=0;

  int dist,subcat,cc,punc;
  int wcc,tcc,wpunc,tpunc;
  int wh,ch,th;
  int wm,cm,tm;
  int p;
  int i,k;
  double prob,combineprob;
  int e1c;

/*ccprob is used to prefer wide scope premodification of CCed phrases,
    and narrow scope post modification*/

  double ccprob_post;

  /* d3 is e1 as head, e2 as modifier
     d4 is e2 as head, e1 as modifier
     
     calulate the distance variables and heads - the parent depends on 
     later calls in this routine

  */

  e1c = childs[edges[e1].child1+edges[e1].numchild-1];

  if(edges[e1].type==0)
    e1hlabel=edges[e1].headtag;
  else
    e1hlabel=edges[edges[e1].headch].label;

  if(edges[e1].type==1&&edges[e1].label==NT_NPB)
    e1hlabel=edges[e1c].label;

  if(edges[e1].stop==0 &&
     (edges[e1].type==3||edges[e1].type==1))
/*     &&tablef[edges[e1].label][e1hlabel][edges[e2].label])*/
    {

      for(i2=0;i2<ne2s;i2++)
	{
	  e2 = e2s[i2];

	  combineprob=edges[e1].prob+edges[e2].prob;
  
	  if(combineprob>=pthresh &&
	     (sindex[s][e]==-1 || combineprob>=(bestprobs[s][e]-BEAMPROB)) &&
	     edges[e2].stop ==1 &&
	     tablef[edges[e1].label][e1hlabel][edges[e2].label])
	    {
	      h2_is_verb=edges[e2].hasverb;
	      if(e3flag2==0)
		{
		  if(edges[e1].label!=NT_NPB)
		    {
		      onectxt_to_dist_subcat(&edges[e1].rc,&dist,&subcat,0);
		      punc = current->commaats[m];
		      if(punc)
			{
			  wpunc=current->commawords[m];
			  tpunc=current->commatags[m];
			}
		      else
			wpunc=tpunc=0;
		      wh=current->wordnos[edges[e1].head];
		      th=edges[e1].headtag;
		      ch=edges[e1].label;
		      ch=e1hlabel;
		    }
		  else
		    {
		      wh=current->wordnos[edges[e1c].head];
		      th=edges[e1c].headtag;
		      ch=edges[e1c].label;
		      
		      wm=current->wordnos[edges[e2].head];
		      tm=edges[e2].headtag;
		      cm=edges[e2].label;
		      dist = 10;
		      subcat = 0;
		      punc=current->commaats[m];
		      if(punc)
			{
			  wpunc=current->commawords[m];
			  tpunc=current->commatags[m];
			}
		      else
			wpunc=tpunc=0;
		    }
		  

		  wcc=tcc=cc=0;
		  
		  p=edges[e1].label;
		  e3flag2 = 1;
		}	  

	      wm=current->wordnos[edges[e2].head];
	      tm=edges[e2].headtag;
	      cm=edges[e2].label;

	      if(edges[e1].type!=0 && argmap[edges[e1].label]==argmap[edges[edges[e1].headch].label])
		ccprob_post=-CCPROBSMALL;
	      else
		ccprob_post=0;
	      
	      prob=get_dependency_prob_witheffhash(wm,tm,cm,wh,th,p,ch,dist,subcat,cc,wcc,tcc,punc,wpunc,tpunc,&new_hash,&eff_hash)+combineprob+ccprob_post;



	      if(prob>pthresh)
		{
		  calc_contexts(&edges[e1].lc,&edges[e1].rc,&edges[e2].lc,&edges[e2].rc,&e3.lc,&e3.rc,1,h2_is_verb,edges[e2].label);
		  e3.hasverb = edges[e1].hasverb || edges[e2].hasverb;

		  if(edges[e1].label==NT_NPB)
		    {
		      e3.lc.nvs=0;
		      e3.lc.adj=1;
		      e3.rc.nvs=0;		      
		      e3.rc.adj=1;
		      e3.hasverb=0;
		    }
		  
		  if(e3flag==0)
		    {
		      e3.head=edges[e1].head;
		      e3.headtag=edges[e1].headtag;
		      e3.type=3;
		      e3flag=1;
		      e3.headch=edges[e1].headch;
		      e3.label=edges[e1].label;
		      e3.headlabel=edges[e1].headlabel;
		      e3.valid=1;
		      e3.stop=0;
		      e3flag = 1;
		    }

		  e3.prob=prob;
		  
		  edges[numedges]=e3;
		  
		  if((k=add_edge(s,e))!=-1)
		    {
		      edges[k].numchild=edges[e1].numchild+1;
		      edges[k].child1=numchilds;
		      for(i=0;i<edges[e1].numchild;i++)
			childs[numchilds+i]=childs[edges[e1].child1+i];
		      childs[numchilds+edges[k].numchild-1]=e2;
		      /*	      printf("AAAA\n");
				      print_edge(k,0);*/
		      numchilds+=edges[k].numchild;
		    }
		}
	    }
	}
    }
}

void join_2_edges_precede(int e2,int s,int m,int e,int *e1s,int ne1)
{
  int h1_is_verb,e1;

  int e2hlabel;

  /*e3,e4,d3,d4 are variables for new edges/distances created when
    e1/e2 are respectively the heads

    e3flag,e4flag = 0/1 if e3/e4 haven't/have been created
  */

  edge_type e4;
  int e4flag=0;
  int e4flag2=0;

  int dist,subcat,cc,punc;
  int wh,ch,th;
  int wm,cm,tm;
  int p;
  int i,i2,k;

  int wcc,tcc,wpunc,tpunc;

  double prob,combineprob;
  int e2c;

/*ccprob is used to prefer wide scope premodification of CCed phrases,
    and narrow scope post modification*/

  double ccprob_pre;


/*  if(edges[e1].label==5) return;
  if(edges[e2].label==5) return;*/

  /* d3 is e1 as head, e2 as modifier
     d4 is e2 as head, e1 as modifier
     
     calulate the distance variables and heads - the parent depends on 
     later calls in this routine

  */

  /*now for the extend case*/

  e2c = childs[edges[e2].child1];
  
  if(edges[e2].type==0)
    e2hlabel=edges[e2].headtag;
  else
    e2hlabel=edges[edges[e2].headch].label;

  if(edges[e2].label == NT_NPB)
    e2hlabel=edges[e2c].label;

  /*now the extend case again*/

  if(edges[e2].stop==0 && 
     (edges[e2].type==3||edges[e2].type==1))
/*      &&tablep[edges[e2].label][e2hlabel][edges[e1].label])*/
    {
      for(i2=0;i2<ne1;i2++)
	{
	  e1 = e1s[i2];
	  
	  combineprob=edges[e1].prob+edges[e2].prob;

/*	  printf("BBB %g\n",combineprob);
	  
	  print_edge(e1,0);
	  print_edge(e2,0);*/

	  
	  if(tablep[edges[e2].label][e2hlabel][edges[e1].label] &&
	     combineprob>=pthresh &&
	     edges[e1].stop ==1 &&
	     (sindex[s][e]==-1 || combineprob>=(bestprobs[s][e]-BEAMPROB)))
	    {

	      h1_is_verb=edges[e1].hasverb;

	      if(e4flag2==0)
		{
		  if(edges[e2].type!=0 &&
		     argmap[edges[e2].label]==argmap[edges[edges[e2].headch].label])
		    ccprob_pre=0;
		  else
		    ccprob_pre=-CCPROBSMALL;

		  if(edges[e2].label!=NT_NPB)
		    {
		      onectxt_to_dist_subcat(&edges[e2].lc,&dist,&subcat,1);
		      punc = current->commaats[m];
		      if(punc)
			{
			  wpunc = current->commawords[m];
			  tpunc = current->commatags[m];
			}
		      else
			wpunc=tpunc=0;
		      
		      wh=current->wordnos[edges[e2].head];
		      th=edges[e2].headtag;
		      ch=edges[e2].label;
		      ch=e2hlabel;
		    }
		  else
		    {
		      wh=current->wordnos[edges[e2c].head];
		      th=edges[e2c].headtag;
		      ch=th;
		      cc=0;
		      
		      dist = 110;
		      subcat = 0;
		      punc=current->commaats[m];
		      if(punc)
			{
			  wpunc = current->commawords[m];
			  tpunc = current->commatags[m];
			}
		      else
			wpunc=tpunc=0;
		    }
		  wcc=tcc=cc=0;
		  
		  p=edges[e2].label;
		  		  
		  e4flag2 = 1;
		}


	      wm=current->wordnos[edges[e1].head];
	      tm=edges[e1].headtag;
	      cm=edges[e1].label;
	      	      
	      prob=get_dependency_prob_witheffhash(wm,tm,cm,wh,th,p,ch,dist,subcat,cc,wcc,tcc,punc,wpunc,tpunc,&new_hash,&eff_hash)+combineprob+ccprob_pre;

	      if(prob>pthresh)
		{
		  calc_contexts(&edges[e1].lc,&edges[e1].rc,&edges[e2].lc,&edges[e2].rc,&e4.lc,&e4.rc,2,h1_is_verb,edges[e1].label);
		  e4.hasverb=edges[e1].hasverb || edges[e2].hasverb;

		  if(edges[e2].label==NT_NPB)
		    {
		      e4.lc.nvs=0;
		      e4.lc.adj=1;
		      e4.rc.nvs=0;
		      e4.rc.adj=1;
		      e4.hasverb=0;
		    }

		  if(e4flag==0)
		    {
		      e4.head=edges[e2].head;
		      e4.headtag=edges[e2].headtag;
		      e4.type=3;
		      e4.headch=edges[e2].headch;
		      e4.label=edges[e2].label;
		      e4.headlabel=edges[e2].headlabel;
		      e4.valid=1;
		      e4.stop=0;
		      
		      e4flag=1;
		    }

		  e4.prob=prob;

		  edges[numedges]=e4;
		  
		  if((k=add_edge(s,e))!=-1)
		    {
		      edges[k].numchild=edges[e2].numchild+1;
		      edges[k].child1=numchilds;
		      for(i=0;i<edges[e2].numchild;i++)
			childs[numchilds+i+1]=childs[edges[e2].child1+i];
		      childs[numchilds]=e1;