int BilingualModel::trainSentence(const string& src_sent, const string& trg_sent) {
auto src_nodes = src_model.getNodes(src_sent); // same size as src_sent, OOV words are replaced by <UNK>
auto trg_nodes = trg_model.getNodes(trg_sent);
// counts the number of words that are in the vocabulary
int words = 0;
words += src_nodes.size() - count(src_nodes.begin(), src_nodes.end(), HuffmanNode::UNK);
words += trg_nodes.size() - count(trg_nodes.begin(), trg_nodes.end(), HuffmanNode::UNK);
if (config->subsampling > 0) {
src_model.subsample(src_nodes); // puts <UNK> tokens in place of the discarded tokens
trg_model.subsample(trg_nodes);
}
if (src_nodes.empty() || trg_nodes.empty()) {
return words;
}
// The <UNK> tokens are necessary to perform the alignment (the nodes vector should have the same size
// as the original sentence)
auto alignment = uniformAlignment(src_nodes, trg_nodes);
// remove <UNK> tokens
src_nodes.erase(
std::remove(src_nodes.begin(), src_nodes.end(), HuffmanNode::UNK),
src_nodes.end());
trg_nodes.erase(
std::remove(trg_nodes.begin(), trg_nodes.end(), HuffmanNode::UNK),
trg_nodes.end());
// Monolingual training
for (int src_pos = 0; src_pos < src_nodes.size(); ++src_pos) {
trainWord(src_model, src_model, src_nodes, src_nodes, src_pos, src_pos, alpha);
}
for (int trg_pos = 0; trg_pos < trg_nodes.size(); ++trg_pos) {
trainWord(trg_model, trg_model, trg_nodes, trg_nodes, trg_pos, trg_pos, alpha);
}
if (config->beta == 0)
return words;
// Bilingual training
for (int src_pos = 0; src_pos < src_nodes.size(); ++src_pos) {
// 1-1 mapping between src_nodes and trg_nodes
int trg_pos = alignment[src_pos];
if (trg_pos != -1) { // target word isn't OOV
trainWord(src_model, trg_model, src_nodes, trg_nodes, src_pos, trg_pos, alpha * config->beta);
trainWord(trg_model, src_model, trg_nodes, src_nodes, trg_pos, src_pos, alpha * config->beta);
}
}
return words; // returns the number of words processed (for progress estimation)
}
int MonolingualModel::trainSentence(const string& sent, int sent_id) {
auto nodes = getNodes(sent); // same size as sent, OOV words are replaced by <UNK>
// counts the number of words that are in the vocabulary
int words = nodes.size() - count(nodes.begin(), nodes.end(), HuffmanNode::UNK);
if (config.subsampling > 0) {
subsample(nodes); // puts <UNK> tokens in place of the discarded tokens
}
if (nodes.empty()) {
return words;
}
// remove <UNK> tokens
nodes.erase(
remove(nodes.begin(), nodes.end(), HuffmanNode::UNK),
nodes.end());
// Monolingual training
for (int pos = 0; pos < nodes.size(); ++pos) {
trainWord(nodes, pos, sent_id);
}
return words; // returns the number of words processed, for progress estimation
}
