private static void Analyze(string corpusPath, LanguageModel lm)
{
Console.WriteLine("Corpus path: {0}", corpusPath);
Console.WriteLine();
Console.WriteLine("Splitting corpus.");
List<List<string>> splitCorpus = SplitCorpus(corpusPath,80, 10, 10);
Console.WriteLine("Splitted corpus as follow:");
Console.WriteLine("Training: {0}", splitCorpus[0].Count);
Console.WriteLine("Validate: {0}", splitCorpus[1].Count);
Console.WriteLine("Test: {0}", splitCorpus[2].Count);
Console.WriteLine("Training model.");
lm.Train(splitCorpus[0]);
Console.WriteLine("Calculate Perplextiy with validate set.");
PerplexityCalculator perplexityCalculator = new PerplexityCalculator(lm);
int unkWords = 0;
double perplexity = perplexityCalculator.GetPerplexity(splitCorpus[1], out unkWords);
Console.WriteLine("Found {0} unknown words:", unkWords);
Console.WriteLine("Perplexity of validation is {0}", perplexity);
perplexity = perplexityCalculator.GetPerplexity(splitCorpus[2], out unkWords);
Console.WriteLine("Found {0} unknown words:", unkWords);
Console.WriteLine("Perplexity of testing is {0}", perplexity);
}
public static List<DoubleCombination> GetOptimumCombination(int rangeOfTrial, LanguageModel languageModel, IEnumerable<string> corpus)
{
if (languageModel == null) throw new ArgumentNullException("backoffModel");
List<double> optimumBetas = new List<double>(languageModel.Settings.NGramOrder);
Dictionary<double, List<DoubleCombination>> perplexityResults = new Dictionary<double, List<DoubleCombination>>();
Dictionary<int, double> orderValues = languageModel is ExampleBackOffModelWithDiscounting
? languageModel.Settings.BackOffBetaPerOrder
: languageModel.Settings.LinearInterpolationLambdaPerOrder;
for (int i = 1; i < rangeOfTrial; i++)
{
for (int j = 1; j < languageModel.Settings.NGramOrder + 1; j++)
{
// Initialize round
for (int k = 1; k < languageModel.Settings.NGramOrder + 1; k++)
{
orderValues[k] = i / (double)rangeOfTrial;
}
// Set each of the values to try
for (int k = 1; k < rangeOfTrial; k++)
{
orderValues[j] = k / (double)rangeOfTrial;
double sumValues = 0;
for (int l = 1; l < languageModel.Settings.NGramOrder + 1; l++)
{
sumValues += orderValues[l];
}
if (sumValues != 1)
continue;
languageModel.ClearCacheForDifferentSettings();
PerplexityCalculator calculator = new PerplexityCalculator(languageModel);
int unkWords = 0;
double perplexity = calculator.GetPerplexity(corpus, out unkWords);
if (!perplexityResults.ContainsKey(perplexity))
{
perplexityResults[perplexity] = new List<DoubleCombination>();
}
DoubleCombination configuration = new DoubleCombination(languageModel.Settings.NGramOrder);
for (int l = 0; l < configuration.NOrder; l++)
{
configuration[l] = orderValues[l + 1];
}
perplexityResults[perplexity].Add(configuration);
}
}
}
return perplexityResults[perplexityResults.Keys.Min()];
}
public PerplexityCalculator(LanguageModel languageModel)
{
_languageModel = languageModel;
}