void align( const AlignMatrix& w, const SentenceValues& huLength, const SentenceValues& enLength,
Trail& bestTrail, AlignMatrix& v )
{
const int huBookSize = w.size();
const int enBookSize = w.otherSize();
const int thickness = w.thickness();
massert(w.size()+1 == v.size());
massert(w.otherSize()+1 == v.otherSize());
TrelliMatrix trellis( huBookSize+1,enBookSize+1,thickness, Dead );
buildDynProgMatrix( w, huLength, enLength, v, trellis );
//x std::cout << std::endl;
//x dumpAlignMatrix(v);
//x std::cout << std::endl;
//x dumpTrelliMatrix(trellis);
//x exit(-1);
std::cerr << "Matrix built." << std::endl;
trelliToLadder( trellis, bestTrail );
std::cerr << "Trail found." << std::endl;
}
// Fills the complement of the radius of the trail with minus infties.
// The return value true means success. Failure means that during the fill,
// we intersected the outside of the quasidiagonal area.
// In this case, the operation is not finished.
bool borderDetailedAlignMatrix( AlignMatrix& alignMatrix, const Trail& trail, int radius )
{
int huBookSize = alignMatrix.size();
int enBookSize = alignMatrix.otherSize();
int huPos, enPos;
for ( huPos=0; huPos<huBookSize; ++huPos )
{
int rowStart = alignMatrix.rowStart(huPos);
int rowEnd = alignMatrix.rowEnd(huPos);
for ( enPos=rowStart; enPos<rowEnd; ++enPos )
{
alignMatrix.cell(huPos,enPos) = outsideOfRadiusValue;
}
}
// We seriously use the fact that many-to-zero segments are subdivided into one-to-zero segments.
// Inside setBox, an exception is thrown if we try to write outside the quasidiagonal.
// If we catch such an exception, it means that the quasidiagonal is not thick enough.
// In this case, we abandon the whole align, just to be sure.
try
{
for ( int i=0; i<trail.size(); ++i )
{
setBox( alignMatrix, trail[i].first, trail[i].second, radius, insideOfRadiusValue );
}
}
catch ( const char* errorType )
{
massert( std::string(errorType) == "out of quasidiagonal" )
return false;
}
bool verify = true;
if (verify)
{
int numberOfEvaluatedItems(0);
for ( huPos=0; huPos<huBookSize; ++huPos )
{
int rowStart = alignMatrix.rowStart(huPos);
int rowEnd = alignMatrix.rowEnd(huPos);
for ( enPos=rowStart; enPos<rowEnd; ++enPos )
{
if (alignMatrix[huPos][enPos]==insideOfRadiusValue)
{
++numberOfEvaluatedItems;
}
}
}
std::cerr << numberOfEvaluatedItems << " items inside the border." << std::endl;
}
return true;
}
void setBox( AlignMatrix& m, int huPos, int enPos, int radius, int insideOfRadiusValue )
{
for ( int x=huPos-radius; x<=huPos+radius; ++x )
{
for ( int y=enPos-radius; y<=enPos+radius; ++y )
{
if ( (x>=0) && (x<m.size()) && (y>=0) && (y<m.otherSize()) )
{
m.cell(x,y) = insideOfRadiusValue ;
}
}
}
}
void buildDynProgMatrix( const AlignMatrix& w, const SentenceValues& huLength, const SentenceValues& enLength,
QuasiDiagonal<double>& v, TrelliMatrix& trellis )
{
const int huBookSize = w.size();
const int enBookSize = w.otherSize();
int huPos,enPos;
// v[huPos][enPos] gives the similarity of the [0,huPos) and [0,enPos) intervals.
// The smaller value, the better similarity. (Unlike in the original similarity matrix w, where bigger is better.)
double infinity = 1e6;
for ( huPos=0; huPos<=huBookSize; ++huPos )
{
int rowStart = v.rowStart(huPos);
int rowEnd = v.rowEnd(huPos);
for ( enPos=rowStart; enPos<rowEnd; ++enPos )
{
double& val = v.cell(huPos,enPos);
unsigned char& trail = trellis.cell(huPos,enPos);
bool quasiglobal_knightsMoveAllowed = true;
if (quasiglobal_knightsMoveAllowed)
{
double lengthFitness(0);
bool quasiglobal_lengthFitnessApplied = true;
// The array is indexed by the step directions. The smaller value, the better.
double values[Dead];
int i;
for ( i=1; i<Dead; ++i )
values[i] = infinity;
if (huPos>0)
{
values[HuSkip] = v[huPos-1][enPos] - skipScore;
}
if (enPos>0)
{
values[EnSkip] = v[huPos][enPos-1] - skipScore;
}
if ((huPos>0) && (enPos>0))
{
if (quasiglobal_lengthFitnessApplied)
{
lengthFitness = closeness(huLength[huPos-1], enLength[enPos-1]);
}
else
{
lengthFitness = 0;
}
values[Diag] = v[huPos-1][enPos-1] - w[huPos-1][enPos-1] - lengthFitness ;
}
const double dotLength = 2.0 ;
if ((huPos>1) && (enPos>0))
{
if (quasiglobal_lengthFitnessApplied)
{
lengthFitness = closeness(huLength[huPos-2]+huLength[huPos-1]+dotLength, enLength[enPos-1]);
}
else
{
lengthFitness = 0;
}
const double& a = w[huPos-1][enPos-1] ;
const double& b = w[huPos-2][enPos-1] ;
double lengthSimilarity =
values[HuHuEnSkip] = v[huPos-2][enPos-1] - ( a<b ? a : b ) - skipScore - lengthFitness ; // The worse of the two crossed square.
}
if ((huPos>0) && (enPos>1))
{
if (quasiglobal_lengthFitnessApplied)
{
// Attention, the two-sentence length is the first argument. Usually the Hungarian is the first argument, but not here.
lengthFitness = closeness(enLength[enPos-2]+enLength[enPos-1]+dotLength, huLength[huPos-1]);
}
else
{
lengthFitness = 0;
}
const double& a = w[huPos-1][enPos-1] ;
const double& b = w[huPos-1][enPos-2] ;
values[HuEnEnSkip] = v[huPos-1][enPos-2] - ( a<b ? a : b ) - skipScore - lengthFitness ; // The worse of the two crossed square.
}
unsigned char direction = Dead;
double bestValue = infinity;
for ( i=1; i<Dead; ++i )
{
if (values[i]<bestValue)
{
bestValue = values[i];
direction = i;
}
}
trail = direction;
if (direction==Dead)
{
val = 0;
}
else
{
val = bestValue;
}
}
else // (!quasiglobal_knightsMoveAllowed)
{
int borderCase = ( (huPos==0) ? 0 : 2 ) + ( (enPos==0) ? 0 : 1 ) ;
switch (borderCase)
{
case 0:
{
val = 0;
trail = Dead;
break;
}
case 1: // huPos==0
{
val = v[0][enPos-1] - skipScore ;
trail = EnSkip;
break;
}
case 2: // enPos==0
{
val = v[huPos-1][0] - skipScore ;
trail = HuSkip;
break;
}
case 3:
{
double x = v[huPos-1][enPos] - skipScore ;
double y = v[huPos] [enPos-1] - skipScore ;
double xy = v[huPos-1][enPos-1] - w[huPos-1][enPos-1] ;
double best = xy;
trail = Diag;
if (x<best)
{
best = x;
trail = HuSkip;
}
if (y<best)
{
best = y;
trail = EnSkip;
}
val = best;
break;
}
}
}
}
}
}