From 4b8a716f3024a2e5c7c80518113f76c61b588bc0 Mon Sep 17 00:00:00 2001 From: Patrick Simianer Date: Fri, 26 Apr 2013 15:17:52 +0200 Subject: added (broken) feature function --- decoder/ff_source_syntax_p.cc | 245 ++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 245 insertions(+) create mode 100644 decoder/ff_source_syntax_p.cc (limited to 'decoder/ff_source_syntax_p.cc') diff --git a/decoder/ff_source_syntax_p.cc b/decoder/ff_source_syntax_p.cc new file mode 100644 index 00000000..cd081544 --- /dev/null +++ b/decoder/ff_source_syntax_p.cc @@ -0,0 +1,245 @@ +#include "ff_source_syntax_p.h" + +#include +#include +#include + +#include "sentence_metadata.h" +#include "array2d.h" +#include "filelib.h" + +using namespace std; + +// implements the source side syntax features described in Blunsom et al. (EMNLP 2008) +// source trees must be represented in Penn Treebank format, e.g. +// (S (NP John) (VP (V left))) + +// log transform to make long spans cluster together +// but preserve differences +inline int SpanSizeTransform(unsigned span_size) { + if (!span_size) return 0; + return static_cast(log(span_size+1) / log(1.39)) - 1; +} + +struct PSourceSyntaxFeaturesImpl { + PSourceSyntaxFeaturesImpl() {} + + PSourceSyntaxFeaturesImpl(const string& param) { + if (!(param.compare("") == 0)) { + string triggered_features_fn = param; + ReadFile triggered_features(triggered_features_fn); + string in; + while(getline(*triggered_features, in)) { + feature_filter.insert(FD::Convert(in)); + } + } + } + + void InitializeGrids(const string& tree, unsigned src_len) { + assert(tree.size() > 0); + //fids_cat.clear(); + fids_ef.clear(); + src_tree.clear(); + //fids_cat.resize(src_len, src_len + 1); + fids_ef.resize(src_len, src_len + 1); + src_tree.resize(src_len, src_len + 1, TD::Convert("XX")); + ParseTreeString(tree, src_len); + } + + void ParseTreeString(const string& tree, unsigned src_len) { + stack > stk; // first = i, second = category + pair cur_cat; cur_cat.first = -1; + unsigned i = 0; + unsigned p = 0; + while(p < tree.size()) { + const char cur = tree[p]; + if (cur == '(') { + stk.push(cur_cat); + ++p; + unsigned k = p + 1; + while (k < tree.size() && tree[k] != ' ') { ++k; } + cur_cat.first = i; + cur_cat.second = TD::Convert(tree.substr(p, k - p)); + // cerr << "NT: '" << tree.substr(p, k-p) << "' (i=" << i << ")\n"; + p = k + 1; + } else if (cur == ')') { + unsigned k = p; + while (k < tree.size() && tree[k] == ')') { ++k; } + const unsigned num_closes = k - p; + for (unsigned ci = 0; ci < num_closes; ++ci) { + // cur_cat.second spans from cur_cat.first to i + // cerr << TD::Convert(cur_cat.second) << " from " << cur_cat.first << " to " << i << endl; + // NOTE: unary rule chains end up being labeled with the top-most category + src_tree(cur_cat.first, i) = cur_cat.second; + cur_cat = stk.top(); + stk.pop(); + } + p = k; + while (p < tree.size() && (tree[p] == ' ' || tree[p] == '\t')) { ++p; } + } else if (cur == ' ' || cur == '\t') { + cerr << "Unexpected whitespace in: " << tree << endl; + abort(); + } else { // terminal symbol + unsigned k = p + 1; + do { + while (k < tree.size() && tree[k] != ')' && tree[k] != ' ') { ++k; } + // cerr << "TERM: '" << tree.substr(p, k-p) << "' (i=" << i << ")\n"; + ++i; + assert(i <= src_len); + while (k < tree.size() && tree[k] == ' ') { ++k; } + p = k; + } while (p < tree.size() && tree[p] != ')'); + } + } + // cerr << "i=" << i << " src_len=" << src_len << endl; + assert(i == src_len); // make sure tree specified in src_tree is + // the same length as the source sentence + } + + WordID FireFeatures(const TRule& rule, const int i, const int j, const WordID* ants, SparseVector* feats) { + //cerr << "fire features: " << rule.AsString() << " for " << i << "," << j << endl; + const WordID lhs = src_tree(i,j); + //int& fid_cat = fids_cat(i,j); + int& fid_ef = fids_ef(i,j)[&rule]; + if (fid_ef <= 0) { + ostringstream os; + //ostringstream os2; + os << "SYN:" << TD::Convert(lhs); + //os2 << "SYN:" << TD::Convert(lhs) << '_' << SpanSizeTransform(j - i); + //fid_cat = FD::Convert(os2.str()); + os << ':'; + unsigned ntc = 0; + for (unsigned k = 0; k < rule.f_.size(); ++k) { + if (k > 0) os << '_'; + int fj = rule.f_[k]; + if (fj <= 0) { + os << '[' << TD::Convert(ants[ntc++]) << ']'; + } else { + os << TD::Convert(fj); + } + } + os << ':'; + for (unsigned k = 0; k < rule.e_.size(); ++k) { + const int ei = rule.e_[k]; + if (k > 0) os << '_'; + if (ei <= 0) + os << '[' << (1-ei) << ']'; + else + os << TD::Convert(ei); + } + fid_ef = FD::Convert(os.str()); + } + //if (fid_cat > 0) + // feats->set_value(fid_cat, 1.0); + if (fid_ef > 0 && (feature_filter.find(fid_ef) != feature_filter.end())) + feats->set_value(fid_ef, 1.0); + return lhs; + } + + Array2D src_tree; // src_tree(i,j) NT = type + // mutable Array2D fids_cat; // this tends to overfit baddly + mutable Array2D > fids_ef; // fires for fully lexicalized + tr1::unordered_set feature_filter; +}; + +PSourceSyntaxFeatures::PSourceSyntaxFeatures(const string& param) : + FeatureFunction(sizeof(WordID)) { + impl = new PSourceSyntaxFeaturesImpl(param); +} + +PSourceSyntaxFeatures::~PSourceSyntaxFeatures() { + delete impl; + impl = NULL; +} + +void PSourceSyntaxFeatures::TraversalFeaturesImpl(const SentenceMetadata& smeta, + const Hypergraph::Edge& edge, + const vector& ant_contexts, + SparseVector* features, + SparseVector* estimated_features, + void* context) const { + WordID ants[8]; + for (unsigned i = 0; i < ant_contexts.size(); ++i) + ants[i] = *static_cast(ant_contexts[i]); + + *static_cast(context) = + impl->FireFeatures(*edge.rule_, edge.i_, edge.j_, ants, features); +} + +void PSourceSyntaxFeatures::PrepareForInput(const SentenceMetadata& smeta) { + impl->InitializeGrids(smeta.GetSGMLValue("src_tree"), smeta.GetSourceLength()); +} + +struct PSourceSpanSizeFeaturesImpl { + PSourceSpanSizeFeaturesImpl() {} + + void InitializeGrids(unsigned src_len) { + fids.clear(); + fids.resize(src_len, src_len + 1); + } + + int FireFeatures(const TRule& rule, const int i, const int j, const WordID* ants, SparseVector* feats) { + if (rule.Arity() > 0) { + int& fid = fids(i,j)[&rule]; + if (fid <= 0) { + ostringstream os; + os << "SSS:"; + unsigned ntc = 0; + for (unsigned k = 0; k < rule.f_.size(); ++k) { + if (k > 0) os << '_'; + int fj = rule.f_[k]; + if (fj <= 0) { + os << '[' << TD::Convert(-fj) << ants[ntc++] << ']'; + } else { + os << TD::Convert(fj); + } + } + os << ':'; + for (unsigned k = 0; k < rule.e_.size(); ++k) { + const int ei = rule.e_[k]; + if (k > 0) os << '_'; + if (ei <= 0) + os << '[' << (1-ei) << ']'; + else + os << TD::Convert(ei); + } + fid = FD::Convert(os.str()); + } + if (fid > 0) + feats->set_value(fid, 1.0); + } + return SpanSizeTransform(j - i); + } + + mutable Array2D > fids; +}; + +PSourceSpanSizeFeatures::PSourceSpanSizeFeatures(const string& param) : + FeatureFunction(sizeof(char)) { + impl = new PSourceSpanSizeFeaturesImpl; +} + +PSourceSpanSizeFeatures::~PSourceSpanSizeFeatures() { + delete impl; + impl = NULL; +} + +void PSourceSpanSizeFeatures::TraversalFeaturesImpl(const SentenceMetadata& smeta, + const Hypergraph::Edge& edge, + const vector& ant_contexts, + SparseVector* features, + SparseVector* estimated_features, + void* context) const { + int ants[8]; + for (unsigned i = 0; i < ant_contexts.size(); ++i) + ants[i] = *static_cast(ant_contexts[i]); + + *static_cast(context) = + impl->FireFeatures(*edge.rule_, edge.i_, edge.j_, ants, features); +} + +void PSourceSpanSizeFeatures::PrepareForInput(const SentenceMetadata& smeta) { + impl->InitializeGrids(smeta.GetSourceLength()); +} + + -- cgit v1.2.3