summaryrefslogtreecommitdiff
path: root/decoder
diff options
context:
space:
mode:
authorChris Dyer <cdyer@cs.cmu.edu>2011-03-03 22:44:42 -0500
committerChris Dyer <cdyer@cs.cmu.edu>2011-03-03 22:44:42 -0500
commitc14605cf041a3c5887c74e9c1e7815b053e4b82a (patch)
treef0856095ad81c8fe14c26065dad62ba5a56245fb /decoder
parentab307856cf1fe713f236e122231771b4b79a8657 (diff)
new span features
Diffstat (limited to 'decoder')
-rw-r--r--decoder/ff_spans.cc45
-rw-r--r--decoder/ff_spans.h7
2 files changed, 38 insertions, 14 deletions
diff --git a/decoder/ff_spans.cc b/decoder/ff_spans.cc
index 1cf72be9..17e1d650 100644
--- a/decoder/ff_spans.cc
+++ b/decoder/ff_spans.cc
@@ -13,6 +13,13 @@
using namespace std;
+// log transform to make long spans cluster together
+// but preserve differences
+int SpanSizeTransform(unsigned span_size) {
+ if (!span_size) return 0;
+ return static_cast<int>(log(span_size+1) / log(1.39)) - 1;
+}
+
SpanFeatures::SpanFeatures(const string& param) :
kS(TD::Convert("S") * -1),
kX(TD::Convert("X") * -1),
@@ -41,8 +48,8 @@ SpanFeatures::SpanFeatures(const string& param) :
}
word2class_[v[0]] = v[1];
}
- word2class_[TD::Convert("<s>")] = TD::Convert("BOS");
- word2class_[TD::Convert("</s>")] = TD::Convert("EOS");
+ word2class_[TD::Convert("BOS")] = TD::Convert("BOS");
+ word2class_[TD::Convert("EOS")] = TD::Convert("EOS");
oov_ = TD::Convert("OOV");
}
@@ -88,10 +95,15 @@ void SpanFeatures::TraversalFeaturesImpl(const SentenceMetadata& smeta,
} else { // non-collapsed features:
features->set_value(end_span_ids_[edge.j_], 1);
features->set_value(beg_span_ids_[edge.i_], 1);
- if (edge.rule_->lhs_ == kS)
+ features->set_value(end_bigram_ids_[edge.j_], 1);
+ features->set_value(beg_bigram_ids_[edge.i_], 1);
+ if (edge.rule_->lhs_ == kS) {
features->set_value(span_feats_(edge.i_,edge.j_).second, 1);
- else
+ features->set_value(len_span_feats_(edge.i_,edge.j_).second, 1);
+ } else {
features->set_value(span_feats_(edge.i_,edge.j_).first, 1);
+ features->set_value(len_span_feats_(edge.i_,edge.j_).first, 1);
+ }
}
}
@@ -104,11 +116,14 @@ WordID SpanFeatures::MapIfNecessary(const WordID& w) const {
void SpanFeatures::PrepareForInput(const SentenceMetadata& smeta) {
const Lattice& lattice = smeta.GetSourceLattice();
- const WordID eos = TD::Convert("</s>");
- const WordID bos = TD::Convert("<s>");
+ const WordID eos = TD::Convert("EOS"); // right of the last source word
+ const WordID bos = TD::Convert("BOS"); // left of the first source word
beg_span_ids_.resize(lattice.size() + 1);
end_span_ids_.resize(lattice.size() + 1);
span_feats_.resize(lattice.size() + 1, lattice.size() + 1);
+ beg_bigram_ids_.resize(lattice.size() + 1);
+ end_bigram_ids_.resize(lattice.size() + 1);
+ len_span_feats_.resize(lattice.size() + 1, lattice.size() + 1);
if (use_collapsed_features_) {
beg_span_vals_.resize(lattice.size() + 1);
end_span_vals_.resize(lattice.size() + 1);
@@ -126,6 +141,12 @@ void SpanFeatures::PrepareForInput(const SentenceMetadata& smeta) {
ostringstream sfid;
sfid << "ES:" << TD::Convert(word);
end_span_ids_[i] = FD::Convert(sfid.str());
+ ostringstream esbiid;
+ esbiid << "EBI:" << TD::Convert(bword) << "_" << TD::Convert(word);
+ end_bigram_ids_[i] = FD::Convert(esbiid.str());
+ ostringstream bsbiid;
+ bsbiid << "BBI:" << TD::Convert(bword) << "_" << TD::Convert(word);
+ beg_bigram_ids_[i] = FD::Convert(bsbiid.str());
ostringstream bfid;
bfid << "BS:" << TD::Convert(bword);
beg_span_ids_[i] = FD::Convert(bfid.str());
@@ -145,9 +166,14 @@ void SpanFeatures::PrepareForInput(const SentenceMetadata& smeta) {
word = lattice[j][0].label;
word = MapIfNecessary(word);
ostringstream pf;
- pf << "SS:" << TD::Convert(bword) << "_" << TD::Convert(word);
+ pf << "S:" << TD::Convert(bword) << "_" << TD::Convert(word);
span_feats_(i,j).first = FD::Convert(pf.str());
span_feats_(i,j).second = FD::Convert("S_" + pf.str());
+ ostringstream lf;
+ const unsigned span_size = (i < j ? j - i : i - j);
+ lf << "LS:" << SpanSizeTransform(span_size) << "_" << TD::Convert(bword) << "_" << TD::Convert(word);
+ len_span_feats_(i,j).first = FD::Convert(lf.str());
+ len_span_feats_(i,j).second = FD::Convert("S_" + lf.str());
if (use_collapsed_features_) {
span_vals_(i,j).first = feat2val_[pf.str()];
span_vals_(i,j).second = feat2val_["S_" + pf.str()];
@@ -187,11 +213,6 @@ CMR2008ReorderingFeatures::CMR2008ReorderingFeatures(const std::string& param) :
}
}
-int CMR2008ReorderingFeatures::SpanSizeTransform(unsigned span_size) {
- if (!span_size) return 0;
- return static_cast<int>(log(span_size+1) / log(1.39)) - 1;
-}
-
void CMR2008ReorderingFeatures::TraversalFeaturesImpl(const SentenceMetadata& smeta,
const Hypergraph::Edge& edge,
const vector<const void*>& ant_contexts,
diff --git a/decoder/ff_spans.h b/decoder/ff_spans.h
index 9928d70f..67269d26 100644
--- a/decoder/ff_spans.h
+++ b/decoder/ff_spans.h
@@ -23,8 +23,11 @@ class SpanFeatures : public FeatureFunction {
const int kS;
const int kX;
Array2D<std::pair<int,int> > span_feats_; // first for X, second for S
+ Array2D<std::pair<int,int> > len_span_feats_; // first for X, second for S, including length
std::vector<int> end_span_ids_;
+ std::vector<int> end_bigram_ids_;
std::vector<int> beg_span_ids_;
+ std::vector<int> beg_bigram_ids_;
std::map<WordID, WordID> word2class_; // optional projection to coarser class
// collapsed feature values
@@ -35,7 +38,9 @@ class SpanFeatures : public FeatureFunction {
int fid_span_;
std::map<std::string, double> feat2val_;
std::vector<double> end_span_vals_;
+ std::vector<double> end_bigram_vals_;
std::vector<double> beg_span_vals_;
+ std::vector<double> beg_bigram_vals_;
Array2D<std::pair<double,double> > span_vals_;
WordID oov_;
@@ -52,8 +57,6 @@ class CMR2008ReorderingFeatures : public FeatureFunction {
SparseVector<double>* estimated_features,
void* context) const;
private:
- static int SpanSizeTransform(unsigned span_size);
-
const int kS;
std::pair<int, int> unconditioned_fids_; // first = monotone
// second = inverse