/* EXPORT GetLMProb: return probability of word wd_id following pr_id[] */
float GetLMProb(LModel *lm, LabId prid[NSIZE], LabId wdid)
{
LabId cpid[NSIZE];
NEntry *ne;
SEntry *se;
lmId p, q, word, ndx[NSIZE];
LogFloat bowt,prob;
int i, s;
switch (lm->type) {
case boNGram:
word = (int)wdid->aux;
if (word==0 || word>lm->data.ngram->vocSize)
return(LZERO);
for (s=-1,i=0;i<NSIZE;i++)
if (prid[i]!=NULL)
ndx[i]=(int)prid[i]->aux, cpid[i]=prid[i], s=i;
else
ndx[i]=0, cpid[i]=NULL;
/* If no answer back-off to unigram */
if (s<0) {
if (word!=0)
return(lm->data.ngram->unigrams[word]);
else
return(log(1.0/lm->data.ngram->vocSize));
}
cpid[s]=0;
ne = GetNEntry(lm->data.ngram,ndx,FALSE);
if (ne) {
/* Replace with bsearch equivalent */
for (i=0, se=ne->se; i<ne->nse; i++,se++)
if (se->word==word)
return(se->prob); /* Ngram found */
bowt=ne->bowt;
}
else {
bowt=0.0;
}
if (s==0)
return(lm->data.ngram->unigrams[word]+bowt); /* Backoff to unigram */
else
return(bowt+GetLMProb(lm,cpid,wdid)); /* else recurse */
break;
case matBigram:
p=(int) prid[0]->aux;
q=(int) wdid->aux;
return(lm->data.matbi->bigMat[p][q]);
default:
prob=LZERO;
}
return(prob);
}
/* EXPORT->CreateBoNGram: Allocate and create basic NGram structures */
NGramLM *CreateBoNGram(LModel *lm,int vocSize, int counts[NSIZE])
{
lmId ndx[NSIZE];
int i,k;
NGramLM *nglm;
nglm = (NGramLM *) New(lm->heap, sizeof(NGramLM));
lm->data.ngram = nglm;
nglm->heap = lm->heap;
for (i=0;i<=NSIZE;i++) nglm->counts[i]=0;
for (i=1;i<=NSIZE;i++)
if (counts[i]==0) break;
else nglm->counts[i]=counts[i];
nglm->nsize=i-1;
/* Don't count final layer */
for (k=0,i=1;i<nglm->nsize;i++)
k+=nglm->counts[i];
/* Then use total to guess NEntry hash size */
if (k<25000)
nglm->hashsize=NGHSIZE1;
else if (k<250000)
nglm->hashsize=NGHSIZE2;
else
nglm->hashsize=NGHSIZE3;
nglm->hashtab=(NEntry **) New(lm->heap,sizeof(NEntry*)*nglm->hashsize);
for (i=0; i<nglm->hashsize; i++)
nglm->hashtab[i]=NULL;
nglm->vocSize = vocSize;
nglm->unigrams = CreateVector(lm->heap,nglm->vocSize);
nglm->wdlist = (LabId *) New(lm->heap,nglm->vocSize*sizeof(LabId)); nglm->wdlist--;
for (i=1;i<=nglm->vocSize;i++) nglm->wdlist[i]=NULL;
for (i=0;i<NSIZE;i++) ndx[i]=0;
GetNEntry(nglm,ndx,TRUE);
return(nglm);
}
/* OutputBoBigram: output ARPA/MIL-LL style back off bigram */
void OutputBoBigram(void)
{
LModel lm;
NGramLM *nglm;
NEntry *ne;
SEntry *se;
AEntry **aelists;
lmId ndx[NSIZE];
int i,tot,counts[NSIZE+1];
double uent,ent,bent;
lm.heap=&statHeap;
lm.type=boNGram;
counts[1]=lSize;counts[2]=nae;
for(i=3;i<NSIZE+1;i++)
counts[i]=0;
nglm=CreateBoNGram(&lm,lSize,counts); /* Give max size at creation */
for (i=1;i<=lSize;i++)
nglm->wdlist[i]=lTab[i].name;
aelists=(AEntry**)New(&tmpHeap,sizeof(AEntry*)*(lSize+1));
for (i=1;i<=lSize;i++) aelists[i]=NULL;
RebuildAETab(aelists); /* Un-hash hashtable */
for (i=1,tot=0.0;i<=lSize;i++) { /* Calculate unigrams first */
if (i==(int)enterId->aux)
nglm->unigrams[i]=0.0;
else if (lTab[i].count<uniFloor)
nglm->unigrams[i]=uniFloor;
else
nglm->unigrams[i]=lTab[i].count;
tot+=nglm->unigrams[i];
}
for (i=1,uent=0.0;i<=lSize;i++,se++) {
nglm->unigrams[i]=nglm->unigrams[i]/tot;
uent-=ent2(nglm->unigrams[i]);
}
nglm->counts[1]=lSize; /* Calculate real sizes during build */
nglm->counts[2]=0;
for (i=0; i<NSIZE; i++) ndx[i]=0;
if (trace&T_BIG) {
printf("\n UNIGRAM NEntry - %4d foll, ent %.3f [= %.3f]\n\n",
lSize,uent,pow(2.0,uent));
printf(" BIGRAMS NEntries\n");
fflush(stdout);
}
for (i=1,bent=0.0;i<=lSize;i++) {
ndx[0]=i;
ne=GetNEntry(nglm,ndx,TRUE);
ne->user=aelists[i];
ent = BuildNEntry(ne,nglm->unigrams,uent);
nglm->counts[2]+=ne->nse;
if (trace&T_BIG)
if (i!=(int)exitId->aux){
if (i==(int)enterId->aux)
bent+=nglm->unigrams[(int)exitId->aux]*ent;
else
bent+=nglm->unigrams[i]*ent;
printf(" %-20s - %4d foll, ent %6.3f [= %6.2f]\n",
lTab[i].name->name,ne->nse,ent,pow(2.0,ent));
fflush(stdout);
}
}
Dispose(&tmpHeap,aelists);
if (trace&T_BIG) {
printf("\n BIGRAM: training data entropy %.3f (perplexity %.2f)\n",
bent,pow(2.0,bent));
fflush(stdout);
}
ndx[0]=0; /* Set up unigram nentry separately */
ne=GetNEntry(nglm,ndx,TRUE);
ne->nse=lSize;
se=ne->se=(SEntry*)New(nglm->heap,sizeof(SEntry)*lSize);
for (i=1;i<=lSize;i++,se++) {
se->word=i;
if (nglm->unigrams[i]>0)
se->prob=nglm->unigrams[i]=log(nglm->unigrams[i]);
else
se->prob=nglm->unigrams[i]=LZERO;
}
lm.name=CopyString(lm.heap,bigFile); /* Name and write to disk */
WriteLModel(&lm,bigFile,0);
}
HError(3019,"HBuild: Can only specifiy one of -m, -n, -w, -x");
bType = wordPair;
if (NextArg()!=STRINGARG)
HError(3019,"HBuild: Word pair grammar file name expected");
ipFn = GetStrArg();
break;
case 'x':
if (bType != unknown)
HError(3019,"HBuild: Can only specifiy one of -m, -n, -w, -x");
bType = multiLat;
if (NextArg()!=STRINGARG)
HError(3019,"HBuild: Multi-level lattice file name expected");
ipFn = GetStrArg();
break;
case 'z':
zapUnknown = TRUE; break;
case 'T':
trace = GetChkedInt(0,511,s); break;
default:
HError(3019,"HBuild: Unknown switch %s",s);
}
}
if (NextArg()!=STRINGARG)
HError(3019,"HBuild: Word List file name expected");
wordListFn = GetStrArg();
if (NextArg()!=STRINGARG)
HError(3019,"HBuild: output lattice file name expected");
latFn = GetStrArg();
if (bType == unknown) bType = wordLoop;
if (saveLatBin) format |= HLAT_LBIN;
/* Read the word-list into a Vocab data structure */
InitVocab(&voc);
if(ReadDict(wordListFn, &voc)<SUCCESS)
HError(3013,"HBuild: ReadDict failed");
switch (bType) {
case matBiGram:
if (trace & T_TOP)
printf("Reading bigram from file %s\n",ipFn);
bigramLm = ReadLModel(&gstack, ipFn);
if (bigramLm->type != matBigram)
HError(3030,"HBuild: File specified is not a matrix bigram");
lat = ProcessBiGram(&gstack,&voc,bigramLm);
SaveLattice(lat,latFn,format);
break;
case boBiGram:
if (trace & T_TOP)
/* LMTransProb_ngram
return logprob of transition from src labelled word. Also return dest state.
ngram case
*/
LogFloat LMTrans (LModel *lm, LMState src, LabId wdid, LMState *dest)
{
NGramLM *nglm;
LogFloat lmprob;
lmId hist[NSIZE] = {0}; /* initialise whole array to zero! */
int i, l;
NEntry *ne;
SEntry *se;
lmId word;
assert (lm->type == boNGram);
nglm = lm->data.ngram;
word = (int) wdid->aux;
if (word==0 || word>lm->data.ngram->vocSize) {
HError (-9999, "word %d not in LM wordlist", word);
*dest = NULL;
return (LZERO);
}
ne = src;
if (!src) { /* unigram case */
lmprob = nglm->unigrams[word];
}
else {
/* lookup prob p(word | src) */
/* try to find pronid in SEntry array */
se = FindSEntry (ne->se, word, 0, ne->nse - 1);
assert (!se || (se->word == word));
if (se) /* found */
lmprob = se->prob;
else { /* not found */
lmprob = 0.0;
l = 0;
hist[NSIZE-1] = 0;
for (i = 0; i < NSIZE-1; ++i) {
hist[i] = ne->word[i];
if (hist[i] != 0)
l = i;
} /* l is now the index of the last (oldest) non zero element */
for ( ; l > 0; --l) {
if (ne)
lmprob += ne->bowt;
hist[l] = 0; /* back-off: discard oldest word */
ne = GetNEntry (nglm, hist, FALSE);
if (ne) { /* skip over non existing hists. fix for weird LMs */
/* try to find pronid in SEntry array */
se = FindSEntry (ne->se, word, 0, ne->nse - 1);
assert (!se || (se->word == word));
if (se) { /* found it */
lmprob += se->prob;
l = -1;
break;
}
}
}
if (l == 0) { /* backed-off all the way to unigram */
assert (!se);
lmprob += ne->bowt;
lmprob += nglm->unigrams[word];
}
}
}
/* now determine dest state */
if (src) {
ne = (NEntry *) src;
l = 0;
hist[NSIZE-1] = 0;
for (i = 1; i < NSIZE-1; ++i) {
hist[i] = ne->word[i-1];
if (hist[i] != 0)
l = i;
} /* l is now the index of the last (oldest) non zero element */
}
else {
for (i = 1; i < NSIZE-1; ++i)
hist[i] = 0;
l = 1;
}
hist[0] = word;
ne = (LMState) GetNEntry (nglm, hist, FALSE);
for ( ; !ne && (l > 0); --l) {
hist[l] = 0; /* back off */
ne = (LMState) GetNEntry (nglm, hist, FALSE);
}
/* if we left the loop because l=0, then ne is still NULL, which is what we want */
*dest = ne;
#if 0
printf ("lmprob = %f dest %p\n", lmprob, *dest);
#endif
return (lmprob);
}
/* ReadNGrams: read n grams list from file */
static int ReadNGrams(NGramLM *nglm,int n,int count, Boolean bin)
{
float prob;
LabId wdid;
SEntry *cse;
char wd[255];
lmId ndx[NSIZE+1];
NEntry *ne,*le=NULL;
int i, g, idx, total;
unsigned char size, flags=0;
cse = (SEntry *) New(nglm->heap,count*sizeof(SEntry));
for (i=1;i<=NSIZE;i++) ndx[i]=0;
if (trace&T_TIO)
printf("\nn%1d ",n),fflush(stdout);
total=0;
for (g=1; g<=count; g++){
PROGRESS(g);
if (bin) {
size = GetCh (&source);
flags = GetCh (&source);
}
prob = GetFloat(bin)*LN10;
if (n==1) { /* unigram treated as special */
ReadLMWord(wd);
wdid = GetLabId(wd, TRUE);
if (wdid->aux != NULL)
HError(8150,"ReadNGrams: Duplicate word (%s) in 1-gram list",
wdid->name);
wdid->aux = (Ptr)g;
nglm->wdlist[g] = wdid;
nglm->unigrams[g] = prob;
ndx[0]=g;
} else { /* bigram, trigram, etc. */
for (i=0;i<n;i++) {
if (bin) {
if (flags & BIN_ARPA_INT_LMID) {
unsigned int ui;
if (!ReadInt (&source, (int *) &ui, 1, bin))
HError (9999, "ReadNGrams: failed reading int lm word id");
idx = ui;
}
else {
unsigned short us;
if (!ReadShort (&source, (short *) &us, 1, bin))
HError (9999, "ReadNGrams: failed reading short lm word id at");
idx = us;
}
}
else {
ReadLMWord(wd);
wdid = GetLabId(wd, FALSE);
idx = (wdid==NULL?0:(int)wdid->aux);
}
if (idx<1 || idx>nglm->vocSize)
HError(8150,"ReadNGrams: Unseen word (%s) in %dGram",wd,n);
ndx[n-1-i]=idx;
}
}
total++;
ne = GetNEntry(nglm,ndx+1,FALSE);
if (ne == NULL)
HError(8150,"ReadNGrams: Backoff weight not seen for %dth %dGram",g,n);
if (ne!=le) {
if (le != NULL && ne->se != NULL)
HError(8150,"ReadNGrams: %dth %dGrams out of order",g,n);
if (le != NULL) {
if (le->nse==0) {
le->se=NULL;
} else {
qsort(le->se,le->nse,sizeof(SEntry),se_cmp);
}
}
ne->se = cse;
ne->nse = 0;
le = ne;
}
cse->prob = prob;
cse->word = ndx[0];
ne->nse++; cse++;
/* read back-off weight */
if (bin) {
if (flags & BIN_ARPA_HAS_BOWT) {
ne = GetNEntry(nglm,ndx,TRUE);
ne->bowt = GetFloat (TRUE)*LN10;
}
}
else {
SkipWhiteSpace(&source);
if (!source.wasNewline) {
ne=GetNEntry(nglm,ndx,TRUE);
ne->bowt = GetFloat(FALSE)*LN10;
}
}
}
/* deal with the last accumulated set */
if (le != NULL) {
if (le->nse==0) {
le->se=NULL;
} else {
qsort(le->se,le->nse,sizeof(SEntry),se_cmp);
}
}
if (trace&T_TIO)
printf("\n"),fflush(stdout);
return(total);
}
/* WriteNGram: Write n grams to file */
static int WriteNGrams(FILE *file,NGramLM *nglm,int n,float scale)
{
NEntry *ne,*be,*ce,**neTab;
SEntry *se;
LogFloat prob;
lmId ndx[NSIZE+1];
int c,i,j,k,N,g=1,hash,neCnt,total;
if (trace&T_TIO)
printf("\nn%1d ",n),fflush(stdout);
fprintf(file,"\n\\%d-grams:\n",n);
N=VectorSize(nglm->unigrams);
neTab=(NEntry **) New(&gstack,sizeof(NEntry*)*nglm->counts[0]);
for (hash=neCnt=0;hash<nglm->hashsize;hash++)
for (ne=nglm->hashtab[hash]; ne!=NULL; ne=ne->link) {
for (i=1,ce=ne;i<n;i++)
if (ne->word[i-1]==0) {
ce=NULL;
break;
}
if (ce!=NULL)
for (i=n;i<NSIZE;i++)
if (ne->word[i-1]!=0) {
ce=NULL;
break;
}
if (ce!=NULL && ce->nse>0)
neTab[neCnt++]=ce;
}
qsort(neTab,neCnt,sizeof(NEntry*),nep_cmp);
total=0;
for (c=n;c<=NSIZE;c++) ndx[c]=0;
for (j=0;j<neCnt;j++) {
ne=neTab[j];
for (c=1;c<n;c++) ndx[c]=ne->word[c-1];
if (ne!=NULL && ne->nse>0) {
for (i=0,se=ne->se;i<ne->nse;i++,se++) {
if (trace&T_TIO) {
if ((g%25000)==0)
printf(". "),fflush(stdout);
if ((g%800000)==0)
printf("\n "),fflush(stdout);
g++;
}
ndx[0]=se->word;
if (n<nglm->nsize) be=GetNEntry(nglm,ndx,FALSE);
else be=NULL;
if (be==NULL || be->nse==0) be=NULL;
total++;
if (n==1) prob=nglm->unigrams[se->word];
else prob=se->prob;
if (prob*scale<-99.999)
fprintf(file,"%+6.3f",-99.999);
else
fprintf(file,"%+6.4f",prob*scale);
c='\t';
for (k=n-1;k>=0;k--)
if (rawMITFormat)
fprintf(file,"%c%s",c,nglm->wdlist[ndx[k]]->name),c=' ';
else
fprintf(file,"%c%s",c,
ReWriteString(nglm->wdlist[ndx[k]]->name,
NULL,ESCAPE_CHAR)),c=' ';
if (be!=NULL)
fprintf(file,"\t%+6.4f\n",be->bowt*scale);
else
fprintf(file,"\n");
}
}
}
Dispose(&gstack,neTab);
if (trace&T_TIO)
printf("\n"),fflush(stdout);
return(total);
}
