private void InitArcs()
{
for (int arcIndex = 0; arcIndex < _wordGraph.Arcs.Count; arcIndex++)
{
WordGraphArc arc = _wordGraph.Arcs[arcIndex];
// init ecm score info for each word of arc
EcmScoreInfo prevEsi = _stateEcmScoreInfos[arc.PrevState];
var esis = new List <EcmScoreInfo>();
foreach (string word in arc.Words)
{
var esi = new EcmScoreInfo();
_ecm.SetupEsi(esi, prevEsi, word);
esis.Add(esi);
prevEsi = esi;
}
_arcEcmScoreInfos.Add(esis);
// init best scores for the arc's successive state
UpdateStateBestScores(arcIndex, 0);
_statesInvolvedInArcs.Add(arc.PrevState);
_statesInvolvedInArcs.Add(arc.NextState);
}
}
private void UpdateStateBestScores(int arcIndex, int prefixDiffSize)
{
WordGraphArc arc = _wordGraph.Arcs[arcIndex];
List <EcmScoreInfo> arcEsis = _arcEcmScoreInfos[arcIndex];
EcmScoreInfo prevEsi = arcEsis.Count == 0 ? _stateEcmScoreInfos[arc.PrevState] : arcEsis[arcEsis.Count - 1];
double wordGraphScore = _stateWordGraphScores[arc.PrevState] + arc.Score;
List <double> nextStateBestScores = _stateBestScores[arc.NextState];
List <int> nextStateBestPrevArcs = _stateBestPrevArcs[arc.NextState];
var positions = new List <int>();
int startPos = prefixDiffSize == 0 ? 0 : prevEsi.Scores.Count - prefixDiffSize;
for (int i = startPos; i < prevEsi.Scores.Count; i++)
{
double newScore = (EcmWeight * -prevEsi.Scores[i]) + (WordGraphWeight * wordGraphScore);
if (i == nextStateBestScores.Count || nextStateBestScores[i] < newScore)
{
AddOrReplace(nextStateBestScores, i, newScore);
positions.Add(i);
AddOrReplace(nextStateBestPrevArcs, i, arcIndex);
}
}
_stateEcmScoreInfos[arc.NextState].UpdatePositions(prevEsi, positions);
if (wordGraphScore > _stateWordGraphScores[arc.NextState])
{
_stateWordGraphScores[arc.NextState] = wordGraphScore;
}
}
public void SetupInitialEsi(EcmScoreInfo initialEsi)
{
double score = _segmentEditDistance.Compute(new string[0], new string[0]);
initialEsi.Scores.Clear();
initialEsi.Scores.Add(score);
initialEsi.Operations.Clear();
}
public void SetupEsi(EcmScoreInfo esi, EcmScoreInfo prevEsi, string word)
{
double score = _segmentEditDistance.Compute(new string[] { word }, new string[0]);
esi.Scores.Clear();
esi.Scores.Add(prevEsi.Scores[0] + score);
esi.Operations.Clear();
esi.Operations.Add(EditOperation.None);
}
public void ExtendEsi(EcmScoreInfo esi, EcmScoreInfo prevEsi, string word, string[] prefixDiff,
bool isLastWordComplete)
{
IEnumerable <EditOperation> ops = _segmentEditDistance.IncrComputePrefix(esi.Scores, prevEsi.Scores, word,
prefixDiff, isLastWordComplete);
foreach (EditOperation op in ops)
{
esi.Operations.Add(op);
}
}
private void UpdateInitialStateBestScores()
{
EcmScoreInfo esi = _stateEcmScoreInfos[WordGraph.InitialState];
_stateWordGraphScores[WordGraph.InitialState] = _wordGraph.InitialStateScore;
List <double> bestScores = _stateBestScores[WordGraph.InitialState];
List <int> bestPrevArcs = _stateBestPrevArcs[WordGraph.InitialState];
bestScores.Clear();
bestPrevArcs.Clear();
foreach (double score in esi.Scores)
{
bestScores.Add((EcmWeight * -score) + (WordGraphWeight * _wordGraph.InitialStateScore));
bestPrevArcs.Add(int.MaxValue);
}
}
private void GetSubStateHypotheses(PriorityQueue <Hypothesis> queue)
{
for (int arcIndex = 0; arcIndex < _wordGraph.Arcs.Count; arcIndex++)
{
WordGraphArc arc = _wordGraph.Arcs[arcIndex];
if (arc.Words.Count > 1 && !IsArcPruned(arc))
{
double wordGraphScore = _stateWordGraphScores[arc.PrevState] + arc.Score;
for (int i = 0; i < arc.Words.Count - 1; i++)
{
EcmScoreInfo esi = _arcEcmScoreInfos[arcIndex][i];
double score = (WordGraphWeight * wordGraphScore)
+ (EcmWeight * -esi.Scores[esi.Scores.Count - 1])
+ (WordGraphWeight * _restScores[arc.NextState]);
queue.Enqueue(new Hypothesis(score, arc.NextState, arcIndex, i));
}
}
}
}
public void UpdatePositions(EcmScoreInfo prevEsi, List <int> positions)
{
while (Scores.Count < prevEsi.Scores.Count)
{
Scores.Add(0);
}
while (Operations.Count < prevEsi.Operations.Count)
{
Operations.Add(EditOperation.None);
}
for (int i = 0; i < positions.Count; i++)
{
Scores[positions[i]] = prevEsi.Scores[positions[i]];
if (prevEsi.Operations.Count > i)
{
Operations[positions[i]] = prevEsi.Operations[positions[i]];
}
}
}
private void ProcessWordGraphForPrefixDiff(string[] prefixDiff, bool isLastWordComplete)
{
if (prefixDiff.Length == 0)
{
return;
}
if (!_wordGraph.IsEmpty)
{
EcmScoreInfo prevInitialEsi = _stateEcmScoreInfos[WordGraph.InitialState];
_ecm.ExtendInitialEsi(_stateEcmScoreInfos[WordGraph.InitialState], prevInitialEsi, prefixDiff);
UpdateInitialStateBestScores();
}
for (int arcIndex = 0; arcIndex < _wordGraph.Arcs.Count; arcIndex++)
{
WordGraphArc arc = _wordGraph.Arcs[arcIndex];
// update ecm score info for each word of arc
EcmScoreInfo prevEsi = _stateEcmScoreInfos[arc.PrevState];
List <EcmScoreInfo> esis = _arcEcmScoreInfos[arcIndex];
while (esis.Count < arc.Words.Count)
{
esis.Add(new EcmScoreInfo());
}
for (int i = 0; i < arc.Words.Count; i++)
{
EcmScoreInfo esi = esis[i];
_ecm.ExtendEsi(esi, prevEsi, arc.IsUnknown ? string.Empty : arc.Words[i], prefixDiff,
isLastWordComplete);
prevEsi = esi;
}
// update best scores for the arc's successive state
UpdateStateBestScores(arcIndex, prefixDiff.Length);
}
}
public void ExtendInitialEsi(EcmScoreInfo initialEsi, EcmScoreInfo prevInitialEsi, string[] prefixDiff)
{
_segmentEditDistance.IncrComputePrefixFirstRow(initialEsi.Scores, prevInitialEsi.Scores, prefixDiff);
}