public CrossValidationResults Evaluate()
{
Helpers.XorShiftRandom random = new Helpers.XorShiftRandom(RandomSeed);
List <Sentence> sentences = Randomize ? Sentences.OrderBy(s => random.NextUInt()).ToList() : Sentences;
List <Sentence>[] folds = new List <Sentence> [Folds];
for (int i = 0; i < Folds; i++)
{
folds[i] = new List <Sentence>();
}
for (int i = 0; i < sentences.Count; i++)
{
folds[i % Folds].Add(sentences[i]);
}
CrossValidationResults results = new CrossValidationResults();
DateTime start = DateTime.Now;
int iterationsCount = 0;
Parallel.For(0, Folds, (i) =>
{
CrossValidationResults.Result result = new CrossValidationResults.Result();
result.Fold = i + 1;
result.Test = folds[i].Select(sentence => sentence.Clone()).ToList();
result.Train = new List <Sentence>();
for (int j = 0; j < Folds; j++)
{
if (j != i)
{
result.Train.AddRange(folds[j].Select(sentence => sentence.Clone()));
}
}
ITrainableTagger tagger = Activator.CreateInstance <T>();
if (tagger is Tagging.Perceptron.PerceptronTagger)
{
Tagging.Perceptron.PerceptronTagger ptagger = tagger as Tagging.Perceptron.PerceptronTagger;
ptagger.IterationFinished += (it) =>
{
lock (folds) iterationsCount++;
if (Progress != null)
{
Progress(((double)iterationsCount / (ptagger.Iterations * Folds)));
}
};
}
tagger.Train(result.Train);
tagger.Tag(result.Test);
results.Add(result);
});
results.Duration = DateTime.Now - start;
return(results);
}
public void Train(IEnumerable <Sentence> sentences)
{
perceptronMsd = new Perceptron();
perceptronLemma = new PerceptronString();
IndexedSentence[] normalizedSentences =
sentences.Select(s => new IndexedSentence((Reverse ? (s as IEnumerable <Token>).Reverse() : s).Select(t => Normalize(t)))).ToArray();
HashSet <Tag> alltags = new HashSet <Tag>();
foreach (IndexedSentence sentence in normalizedSentences)
{
foreach (Token token in sentence)
{
alltags.Add(new Tag(token.CorrectTag.Msd, null));
foreach (Tag tag in token.PossibleTags)
{
alltags.Add(new Tag(tag.Msd, null));
}
}
}
for (int iteration = 0; iteration < Iterations; iteration++)
{
if (IterationStarted != null)
{
IterationStarted(iteration + 1);
}
foreach (IndexedSentence sentence in normalizedSentences.OrderBy(s => random.NextUInt()))
{
foreach (Token token in sentence)
{
if (token.PossibleTags != null && token.PossibleTags.Length == 1)
{
token.PredictedTag = token.CorrectTag; // todo: possibleTags[0]?? what if it's not correct tag?
continue;
}
IEnumerable <Tag> possibleTags = token.PossibleTags.OrderBy(t => t.Msd).ToArray();
if (possibleTags == null || !possibleTags.Any())
{
possibleTags = alltags;
}
double? bestMsdScore = null;
Tag bestMsd = null;
Features bestMsdFeatures = null;
Features localFeaturesTag = GetFeatures(featureTemplatesTag, token, sentence, true);
foreach (Tag tag in possibleTags)
{
Tag tagMsd = new Tag(tag.Msd);
token.PredictedTag = tagMsd;
Features featuresTag = GetFeatures(featureTemplatesTag, token, sentence, false);
featuresTag.AddRange(localFeaturesTag);
double score = perceptronMsd.Score(featuresTag, tagMsd);
if (bestMsdScore == null || score > bestMsdScore.Value)
{
bestMsdScore = score;
bestMsd = tagMsd;
bestMsdFeatures = featuresTag;
}
}
perceptronMsd.Update(bestMsdFeatures, new Tag(token.CorrectTag.Msd), bestMsd);
if (token.CorrectTag.Lemma != null)
{
double? bestLemmaScore = null;
string bestLemma = null;
Features bestLemmaFeatures = null;
token.PredictedTag = new Tag(token.CorrectTag.Msd);
Features localFeaturesLemma = GetFeatures(featureTemplatesLemma, token, sentence, true);
foreach (Tag tag in possibleTags)
{
if (tag.Msd != bestMsd.Msd || tag.Lemma == null)
{
continue;
}
token.PredictedTag = new Tag(token.CorrectTag.Msd, tag.Lemma);
Features featuresLemma = GetFeatures(featureTemplatesLemma, token, sentence, false);
featuresLemma.AddRange(localFeaturesLemma);
double scoreLemma = perceptronLemma.Score(featuresLemma, tag.Lemma);
if (bestLemmaScore == null || scoreLemma > bestLemmaScore.Value)
{
bestLemmaScore = scoreLemma;
bestLemma = tag.Lemma;
bestLemmaFeatures = featuresLemma;
}
}
if (bestLemmaFeatures != null)
{
perceptronLemma.Update(bestLemmaFeatures, token.CorrectTag.Lemma, bestLemma);
}
}
token.PredictedTag = token.CorrectTag;
}
}
if (IterationFinished != null)
{
IterationFinished(iteration + 1);
}
}
if (Average)
{
perceptronMsd.AverageWeights();
perceptronLemma.AverageWeights();
}
perceptronMsd.RemoveInsignificantWeights(WeightThreshold);
perceptronLemma.RemoveInsignificantWeights(WeightThreshold);
}
