public override float CalculateScore(IData feature)
{
_mixtureComponentSet.UpdateTopScores(feature);
float ascore = 0;
for (var i = 0; i < MixtureWeights.StreamsNum; i++)
{
var logTotal = LogMath.LogZero;
for (var j = 0; j < _mixtureComponentSet.TopGauNum; j++)
{
var topGauScore = _mixtureComponentSet.GetTopGauScore(i, j);
var topGauId = _mixtureComponentSet.GetTopGauId(i, j);
var mixtureWeightx = MixtureWeights.Get(_Id, i, topGauId);
logTotal = LogMath.AddAsLinear(logTotal, topGauScore + mixtureWeightx);
}
ascore += logTotal;
}
return(ascore);
}
public override float CalculateScore(IData feature)
{
if (feature is DoubleData)
this.LogInfo("DoubleData conversion required on mixture level!");
var featureVector = FloatData.ToFloatData(feature).Values;
var logTotal = LogMath.LogZero;
for (var i = 0; i < _mixtureComponents.Length; i++)
{
// In linear form, this would be:
//
// Total += Mixture[i].score * MixtureWeight[i]
logTotal = LogMath.AddAsLinear(logTotal,
_mixtureComponents[i].GetScore(featureVector) + MixtureWeights.Get(_Id, 0, i));
}
return logTotal;
}
/// <summary>
/// Apply the unigram weight to the set of unigrams
/// </summary>
private void ApplyUnigramWeight()
{
var logUnigramWeight = _logMath.LinearToLog(_unigramWeight);
var logNotUnigramWeight = _logMath.LinearToLog(1.0f - _unigramWeight);
var logUniform = _logMath.LinearToLog(1.0f / (_numberNGrams[0]));
var logWip = _logMath.LinearToLog(_wip);
var p2 = logUniform + logNotUnigramWeight;
for (var i = 0; i < _numberNGrams[0]; i++)
{
var unigram = Unigrams[i];
var p1 = unigram.LogProbability;
if (i != _startWordID)
{
p1 += logUnigramWeight;
p1 = _logMath.AddAsLinear(p1, p2);
}
if (_applyLanguageWeightAndWip)
{
p1 = p1 * _languageWeight + logWip;
unigram.SetLogBackoff(unigram.LogBackoff * _languageWeight);
}
if (unigram.WordID == 2520)
{
this.LogInfo("TEST");
}
unigram.SetLogProbability(p1);
}
}
/// <summary>
/// Loads the language model from the given location.
/// </summary>
/// <param name="location">The URL location of the model.</param>
/// <param name="unigramWeightValue">The unigram weight.</param>
/// <param name="dictionaryValue">The dictionary.</param>
private void Load(URL location, float unigramWeightValue, IDictionary dictionaryValue)
{
string line;
float logUnigramWeight = _logMath.LinearToLog(unigramWeightValue);
float inverseLogUnigramWeight = _logMath
.LinearToLog(1.0 - unigramWeightValue);
Open(location);
// look for beginning of data
ReadUntil("\\data\\");
// look for ngram statements
List <int> ngramList = new List <int>();
while ((line = ReadLine()) != null)
{
if (line.StartsWith("ngram"))
{
StringTokenizer st = new StringTokenizer(line, " \t\n\r\f=");
if (st.countTokens() != 3)
{
Corrupt("corrupt ngram field " + line + ' '
+ st.countTokens());
}
st.nextToken();
int index = int.Parse(st.nextToken(), CultureInfo.InvariantCulture.NumberFormat);
int count = int.Parse(st.nextToken(), CultureInfo.InvariantCulture.NumberFormat);
ngramList.Insert(index - 1, count);
MaxDepth = Math.Max(index, MaxDepth);
}
else if (line.Equals("\\1-grams:"))
{
break;
}
}
int numUnigrams = ngramList[0] - 1;
// -log(x) = log(1/x)
float logUniformProbability = -_logMath.LinearToLog(numUnigrams);
for (int index = 0; index < ngramList.Count; index++)
{
int ngram = index + 1;
int ngramCount = ngramList[index];
for (int i = 0; i < ngramCount; i++)
{
StringTokenizer tok = new StringTokenizer(ReadLine());
int tokenCount = tok.countTokens();
if (tokenCount != ngram + 1 && tokenCount != ngram + 2)
{
Corrupt("Bad format");
}
float log10Prob = float.Parse(tok.nextToken(), CultureInfo.InvariantCulture.NumberFormat);
float log10Backoff = 0.0f;
// construct the WordSequence for this N-Gram
List <Word> wordList = new List <Word>(MaxDepth);
for (int j = 0; j < ngram; j++)
{
string word = tok.nextToken();
_vocabulary.Add(word);
Word wordObject = dictionaryValue.GetWord(word);
if (wordObject == null)
{
wordObject = Word.Unknown;
}
wordList.Add(wordObject);
}
WordSequence wordSequence = new WordSequence(wordList);
if (tok.hasMoreTokens())
{
log10Backoff = float.Parse(tok.nextToken(), CultureInfo.InvariantCulture.NumberFormat);
}
float logProb = _logMath.Log10ToLog(log10Prob);
float logBackoff = _logMath.Log10ToLog(log10Backoff);
// Apply unigram weights if this is a unigram probability
if (ngram == 1)
{
float p1 = logProb + logUnigramWeight;
float p2 = logUniformProbability + inverseLogUnigramWeight;
logProb = _logMath.AddAsLinear(p1, p2);
// System.out
// .println("p1 " + p1 + " p2 " + p2 + " luw "
// + logUnigramWeight + " iluw "
// + inverseLogUnigramWeight + " lup "
// + logUniformProbability + " logprog "
// + logProb);
}
Put(wordSequence, logProb, logBackoff);
}
if (index < ngramList.Count - 1)
{
string next = "\\" + (ngram + 1) + "-grams:";
ReadUntil(next);
}
}
ReadUntil("\\end\\");
Close();
}
public override void Allocate()
{
_vocabulary.Clear();
_logProbs.Clear();
_logBackoffs.Clear();
HashMap <WordSequence, Integer> unigrams = new HashMap <WordSequence, Integer>();
HashMap <WordSequence, Integer> bigrams = new HashMap <WordSequence, Integer>();
HashMap <WordSequence, Integer> trigrams = new HashMap <WordSequence, Integer>();
int wordCount = 0;
foreach (string sentence in _sentences)
{
string[] textWords = sentence.Split("\\s+");
var words = new List <Word>();
words.Add(_dictionary.GetSentenceStartWord());
foreach (String wordString in textWords)
{
if (wordString.Length == 0)
{
continue;
}
_vocabulary.Add(wordString);
Word word = _dictionary.GetWord(wordString);
if (word == null)
{
words.Add(Word.Unknown);
}
else
{
words.Add(word);
}
}
words.Add(_dictionary.GetSentenceEndWord());
if (words.Count > 0)
{
AddSequence(unigrams, new WordSequence(words[0]));
wordCount++;
}
if (words.Count > 1)
{
wordCount++;
AddSequence(unigrams, new WordSequence(words[1]));
AddSequence(bigrams, new WordSequence(words[0], words[1]));
}
for (int i = 2; i < words.Count; ++i)
{
wordCount++;
AddSequence(unigrams, new WordSequence(words[i]));
AddSequence(bigrams, new WordSequence(words[i - 1], words[i]));
AddSequence(trigrams, new WordSequence(words[i - 2], words[i - 1], words[i]));
}
}
float discount = .5f;
float deflate = 1 - discount;
var uniprobs = new HashMap <WordSequence, Float>();
foreach (var e in unigrams)
{
uniprobs.Put(e.Key, e.Value * deflate / wordCount);
}
LogMath lmath = LogMath.GetLogMath();
float logUnigramWeight = lmath.LinearToLog(_unigramWeight);
float invLogUnigramWeight = lmath.LinearToLog(1 - _unigramWeight);
float logUniformProb = -lmath.LinearToLog(uniprobs.Count);
var sorted1Grams = new SortedSet <WordSequence>(unigrams.Keys);
var iter = new SortedSet <WordSequence>(bigrams.KeySet()).GetEnumerator();
WordSequence ws = iter.MoveNext() ? iter.Current : null;
foreach (WordSequence unigram in sorted1Grams)
{
float p = lmath.LinearToLog(uniprobs.Get(unigram));
p += logUnigramWeight;
p = lmath.AddAsLinear(p, logUniformProb + invLogUnigramWeight);
_logProbs.Put(unigram, p);
float sum = 0f;
while (ws != null)
{
int cmp = ws.GetOldest().CompareTo(unigram);
if (cmp > 0)
{
break;
}
if (cmp == 0)
{
sum += uniprobs.Get(ws.GetNewest());
}
ws = iter.MoveNext() ? iter.Current : null;
}
_logBackoffs.Put(unigram, lmath.LinearToLog(discount / (1 - sum)));
}
var biprobs = new HashMap <WordSequence, Float>();
foreach (var entry in bigrams)
{
int unigramCount = unigrams.Get(entry.Key.GetOldest());
biprobs.Put(entry.Key, entry.Value * deflate / unigramCount);
}
var sorted2Grams = new SortedSet <WordSequence>(bigrams.KeySet());
iter = new SortedSet <WordSequence>(trigrams.KeySet()).GetEnumerator();
ws = iter.MoveNext() ? iter.Current : null;
foreach (WordSequence biword in sorted2Grams)
{
_logProbs.Put(biword, lmath.LinearToLog(biprobs.Get(biword)));
float sum = 0f;
while (ws != null)
{
int cmp = ws.GetOldest().CompareTo(biword);
if (cmp > 0)
{
break;
}
if (cmp == 0)
{
sum += biprobs.Get(ws.GetNewest());
}
ws = iter.MoveNext() ? iter.Current : null;
}
_logBackoffs.Put(biword, lmath.LinearToLog(discount / (1 - sum)));
}
foreach (var e in trigrams)
{
float p = e.Value * deflate;
p /= bigrams.Get(e.Key.GetOldest());
_logProbs.Put(e.Key, lmath.LinearToLog(p));
}
}
/**
* /// Compute the utterance-level posterior for every node in the lattice, i.e. the probability that this node occurs
* /// on any path through the lattice. Uses a forward-backward algorithm specific to the nature of non-looping
* /// left-to-right lattice structures.
* /// <p/>
* /// Node posteriors can be retrieved by calling getPosterior() on Node objects.
*
* /// @param languageModelWeightAdjustment the weight multiplier that will be applied to language score already scaled by language weight
* /// @param useAcousticScoresOnly use only the acoustic scores to compute the posteriors, ignore the language weight
* /// and scores
*/
public void ComputeNodePosteriors(float languageModelWeightAdjustment, bool useAcousticScoresOnly)
{
if (InitialNode == null)
{
return;
}
//forward
InitialNode.ForwardScore = LogMath.LogOne;
InitialNode.ViterbiScore = LogMath.LogOne;
var sortedNodes = SortNodes();
Debug.Assert(sortedNodes[0] == InitialNode);
foreach (var currentNode in sortedNodes)
{
foreach (var edge in currentNode.LeavingEdges)
{
var forwardProb = edge.FromNode.ForwardScore;
var edgeScore = ComputeEdgeScore
(edge, languageModelWeightAdjustment, useAcousticScoresOnly);
forwardProb += edgeScore;
edge.ToNode.ForwardScore = LogMath.AddAsLinear
((float)forwardProb,
(float)edge.ToNode.ForwardScore);
var vs = edge.FromNode.ViterbiScore +
edgeScore;
if (edge.ToNode.BestPredecessor == null ||
vs > edge.ToNode.ViterbiScore)
{
edge.ToNode.BestPredecessor = currentNode;
edge.ToNode.ViterbiScore = vs;
}
}
}
//backward
TerminalNode.BackwardScore = LogMath.LogOne;
Debug.Assert(sortedNodes[sortedNodes.Count - 1] == TerminalNode);
//var n = sortedNodes.GetEnumerator().ListIterator(sortedNodes.Count-1);
var n = sortedNodes.Count;
while (n > 0)
{
var currentNode = sortedNodes[--n]; //TODO: Check behavior
var currentEdges = currentNode.LeavingEdges;
foreach (var edge in currentEdges)
{
var backwardProb = edge.ToNode.BackwardScore;
backwardProb += ComputeEdgeScore
(edge, languageModelWeightAdjustment, useAcousticScoresOnly);
edge.FromNode.BackwardScore = LogMath.AddAsLinear((float)backwardProb,
(float)edge.FromNode.BackwardScore);
}
}
//inner
var normalizationFactor = TerminalNode.ForwardScore;
foreach (var node in Nodes.Values)
{
node.Posterior = (node.ForwardScore +
node.BackwardScore) - normalizationFactor;
}
}