public BiGram Next()
{
var biGram = new BiGram()
{
First = split[pointer], Second = split[pointer + 1], Suffix = split[pointer + 2]
};
pointer++;
return(biGram);
}
/// <summary>
/// Train the model on a given corpus
/// </summary>
/// <param name="path">Path to the corpus database</param>
public void Train(string path)
{
EmbeddingCorpus corpus = new EmbeddingCorpus(path);
// Pre-add all words
foreach (List <string> sentence in corpus.Sentences)
{
foreach (string word in sentence)
{
Add(word);
}
}
// Learning
foreach (List <string> sentence in corpus.Sentences)
{
var _sentence = new List <String>(sentence.Where(w => !corpus.SubSampling(w)));
var _grams = BiGram.FromSentence(_sentence);
foreach (var gram in _grams)
{
Output[gram.Right] = Vectors.Random(EmbeddingSize, UniformRandom.Distribution).Map(p => p * 0.01);
// Positive sample
UpdateSample(gram.Left, gram.Right, true);
// Update negative samples
foreach (string word in Vocabulary.Keys)
{
if (word == gram.Left || !corpus.NegativeSampling(word))
{
continue;
}
UpdateSample(gram.Left, word, false);
}
// Update hidden layer
for (int i = 0; i < EmbeddingSize; i++)
{
Vocabulary[gram.Left][i] += Error[i];
}
}
}
}
public void Add(BiGram biGram)
{
if (biGram.First.Trim().Length > 0 && biGram.Second.Trim().Length > 0)
{
if (this.map.ContainsKey(biGram.Hash()))
{
var updateGram = this.map[biGram.Hash()];
if (updateGram.ContainsKey(biGram.Suffix))
{
updateGram[biGram.Suffix]++;
}
else
{
updateGram.Add(biGram.Suffix, 1);
}
}
else
{
this.map.Add(biGram.Hash(), new Dictionary <string, int>());
this.map[biGram.Hash()].Add(biGram.Suffix, 1);
}
}
}
