/**
* <summary>Constructor for the {@link NeuralNetwork} class. Gets corpus and network parameters as input and sets the
* corresponding parameters first. After that, initializes the network with random weights between -0.5 and 0.5.
* Constructs vector update matrix and prepares the exp table.</summary>
* <param name="corpus">Corpus used to train word vectors using Word2Vec algorithm.</param>
* <param name="parameter">Parameters of the Word2Vec algorithm.</param>
*/
public NeuralNetwork(Corpus.Corpus corpus, WordToVecParameter parameter)
{
this._vocabulary = new Vocabulary(corpus);
this._parameter = parameter;
this._corpus = corpus;
_wordVectors = new Matrix(_vocabulary.Size(), parameter.GetLayerSize(), -0.5, 0.5, new Random());
_wordVectorUpdate = new Matrix(_vocabulary.Size(), parameter.GetLayerSize());
PrepareExpTable();
}
/**
* <summary>Main method for training the CBow version of Word2Vec algorithm.</summary>
*/
private void TrainCbow()
{
var iteration = new Iteration(_corpus, _parameter);
var currentSentence = _corpus.GetSentence(iteration.GetSentenceIndex());
var random = new Random();
var outputs = new Vector(_parameter.GetLayerSize(), 0);
var outputUpdate = new Vector(_parameter.GetLayerSize(), 0);
_corpus.ShuffleSentences(1);
while (iteration.GetIterationCount() < _parameter.GetNumberOfIterations())
{
iteration.AlphaUpdate();
var wordIndex = _vocabulary.GetPosition(currentSentence.GetWord(iteration.GetSentencePosition()));
var currentWord = _vocabulary.GetWord(wordIndex);
outputs.Clear();
outputUpdate.Clear();
var b = random.Next(_parameter.GetWindow());
var cw = 0;
int lastWordIndex;
for (var a = b; a < _parameter.GetWindow() * 2 + 1 - b; a++)
{
var c = iteration.GetSentencePosition() - _parameter.GetWindow() + a;
if (a != _parameter.GetWindow() && currentSentence.SafeIndex(c))
{
lastWordIndex = _vocabulary.GetPosition(currentSentence.GetWord(c));
outputs.Add(_wordVectors.GetRow(lastWordIndex));
cw++;
}
}
if (cw > 0)
{
outputs.Divide(cw);
int l2;
double f;
double g;
if (_parameter.IsHierarchicalSoftMax())
{
for (var d = 0; d < currentWord.GetCodeLength(); d++)
{
l2 = currentWord.GetPoint(d);
f = outputs.DotProduct(_wordVectorUpdate.GetRow(l2));
if (f <= -MAX_EXP || f >= MAX_EXP)
{
continue;
}
f = _expTable[(int)((f + MAX_EXP) * (EXP_TABLE_SIZE / MAX_EXP / 2))];
g = (1 - currentWord.GetCode(d) - f) * iteration.GetAlpha();
outputUpdate.Add(_wordVectorUpdate.GetRow(l2).Product(g));
_wordVectorUpdate.Add(l2, outputs.Product(g));
}
}
else
{
for (var d = 0; d < _parameter.GetNegativeSamplingSize() + 1; d++)
{
int target;
int label;
if (d == 0)
{
target = wordIndex;
label = 1;
}
else
{
target = _vocabulary.GetTableValue(random.Next(_vocabulary.GetTableSize()));
if (target == 0)
{
target = random.Next(_vocabulary.Size() - 1) + 1;
}
if (target == wordIndex)
{
continue;
}
label = 0;
}
l2 = target;
f = outputs.DotProduct(_wordVectorUpdate.GetRow(l2));
g = CalculateG(f, iteration.GetAlpha(), label);
outputUpdate.Add(_wordVectorUpdate.GetRow(l2).Product(g));
_wordVectorUpdate.Add(l2, outputs.Product(g));
}
}
for (var a = b; a < _parameter.GetWindow() * 2 + 1 - b; a++)
{
var c = iteration.GetSentencePosition() - _parameter.GetWindow() + a;
if (a != _parameter.GetWindow() && currentSentence.SafeIndex(c))
{
lastWordIndex = _vocabulary.GetPosition(currentSentence.GetWord(c));
_wordVectors.Add(lastWordIndex, outputUpdate);
}
}
}
currentSentence = iteration.SentenceUpdate(currentSentence);
}
}