From 700b2abf48bf0a455064d6cf08754cbfd4e3a383 Mon Sep 17 00:00:00 2001 From: Chris Dyer Date: Mon, 12 Sep 2011 19:22:59 +0100 Subject: source syntax features ~ blunsom emnlp 2008 --- decoder/Makefile.am | 1 + decoder/cdec_ff.cc | 2 + decoder/ff_source_syntax.cc | 157 ++++++++++++++++++++++++++++++++++++++++++++ decoder/ff_source_syntax.h | 24 +++++++ utils/stringlib.cc | 7 +- 5 files changed, 190 insertions(+), 1 deletion(-) create mode 100644 decoder/ff_source_syntax.cc create mode 100644 decoder/ff_source_syntax.h diff --git a/decoder/Makefile.am b/decoder/Makefile.am index e5f7505f..ede1cff0 100644 --- a/decoder/Makefile.am +++ b/decoder/Makefile.am @@ -72,6 +72,7 @@ libcdec_a_SOURCES = \ ff_wordalign.cc \ ff_csplit.cc \ ff_tagger.cc \ + ff_source_syntax.cc \ ff_bleu.cc \ ff_factory.cc \ freqdict.cc \ diff --git a/decoder/cdec_ff.cc b/decoder/cdec_ff.cc index 588842f1..d562bc3a 100644 --- a/decoder/cdec_ff.cc +++ b/decoder/cdec_ff.cc @@ -14,6 +14,7 @@ #include "ff_bleu.h" #include "ff_lm_fsa.h" #include "ff_sample_fsa.h" +#include "ff_source_syntax.h" #include "ff_register.h" #include "ff_charset.h" #include "ff_wordset.h" @@ -55,6 +56,7 @@ void register_feature_functions() { ff_registry.Register("SpanFeatures", new FFFactory()); ff_registry.Register("NgramFeatures", new FFFactory()); ff_registry.Register("RuleIdentityFeatures", new FFFactory()); + ff_registry.Register("SourceSyntaxFeatures", new FFFactory); ff_registry.Register("RuleNgramFeatures", new FFFactory()); ff_registry.Register("CMR2008ReorderingFeatures", new FFFactory()); ff_registry.Register("KLanguageModel", new KLanguageModelFactory()); diff --git a/decoder/ff_source_syntax.cc b/decoder/ff_source_syntax.cc new file mode 100644 index 00000000..99acbd87 --- /dev/null +++ b/decoder/ff_source_syntax.cc @@ -0,0 +1,157 @@ +#include "ff_source_syntax.h" + +#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))) + +struct SourceSyntaxFeaturesImpl { + SourceSyntaxFeaturesImpl() {} + + void InitializeGrids(const string& tree, unsigned src_len) { + assert(tree.size() > 0); + fids_cat.clear(); + fids_fonly.clear(); + fids_ef.clear(); + src_tree.clear(); + fids_cat.resize(src_len, src_len + 1); + fids_fonly.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_fonly = fids_fonly(i,j)[&rule]; + int& fid_ef = fids_ef(i,j)[&rule]; + if (fid_ef <= 0) { + ostringstream os; + os << "SYN:" << TD::Convert(lhs); + fid_cat = FD::Convert(os.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); + } + } + fid_fonly = FD::Convert(os.str()); + 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_fonly > 0) + feats->set_value(fid_fonly, 1.0); + if (fid_ef > 0) + feats->set_value(fid_ef, 1.0); + return lhs; + } + + Array2D src_tree; // src_tree(i,j) NT = type + mutable Array2D fids_cat; // fires for an LHS match + mutable Array2D > fids_fonly; // fires for an f-string + mutable Array2D > fids_ef; // fires for fully lexicalized +}; + +SourceSyntaxFeatures::SourceSyntaxFeatures(const string& param) : + FeatureFunction(sizeof(WordID)) { + impl = new SourceSyntaxFeaturesImpl; +} + +SourceSyntaxFeatures::~SourceSyntaxFeatures() { + delete impl; + impl = NULL; +} + +void SourceSyntaxFeatures::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 SourceSyntaxFeatures::PrepareForInput(const SentenceMetadata& smeta) { + impl->InitializeGrids(smeta.GetSGMLValue("src_tree"), smeta.GetSourceLength()); +} + diff --git a/decoder/ff_source_syntax.h b/decoder/ff_source_syntax.h new file mode 100644 index 00000000..1e890736 --- /dev/null +++ b/decoder/ff_source_syntax.h @@ -0,0 +1,24 @@ +#ifndef _FF_SOURCE_TOOLS_H_ +#define _FF_SOURCE_TOOLS_H_ + +#include "ff.h" + +struct SourceSyntaxFeaturesImpl; + +class SourceSyntaxFeatures : public FeatureFunction { + public: + SourceSyntaxFeatures(const std::string& param); + ~SourceSyntaxFeatures(); + protected: + virtual void TraversalFeaturesImpl(const SentenceMetadata& smeta, + const Hypergraph::Edge& edge, + const std::vector& ant_contexts, + SparseVector* features, + SparseVector* estimated_features, + void* context) const; + virtual void PrepareForInput(const SentenceMetadata& smeta); + private: + SourceSyntaxFeaturesImpl* impl; +}; + +#endif diff --git a/utils/stringlib.cc b/utils/stringlib.cc index 7aaee9f0..ade02ca9 100644 --- a/utils/stringlib.cc +++ b/utils/stringlib.cc @@ -32,7 +32,12 @@ void ParseTranslatorInput(const string& line, string* input, string* ref) { void ProcessAndStripSGML(string* pline, map* out) { map& meta = *out; string& line = *pline; - string lline = LowercaseString(line); + string lline = *pline; + if (lline.find(" must be lowercase!\n"; + cerr << " " << *pline << endl; + abort(); + } if (lline.find(""); if (close == string::npos) return; // error -- cgit v1.2.3