diff options
author | Chris Dyer <redpony@gmail.com> | 2014-04-20 22:25:15 -0400 |
---|---|---|
committer | Chris Dyer <redpony@gmail.com> | 2014-04-20 22:25:15 -0400 |
commit | 1748e9a095bcc3a1db8ab47eb7ac6a1f9568772b (patch) | |
tree | af9c6d12c49726a690d3a29eb80146ee46158b0e /word-aligner | |
parent | a3434180474e832eeea7d775893ce85244f8145c (diff) |
binary derivations with maximal arity-2
Diffstat (limited to 'word-aligner')
-rw-r--r-- | word-aligner/Makefile.am | 5 | ||||
-rw-r--r-- | word-aligner/binderiv.cc | 202 |
2 files changed, 206 insertions, 1 deletions
diff --git a/word-aligner/Makefile.am b/word-aligner/Makefile.am index 1f7f78ae..075ad009 100644 --- a/word-aligner/Makefile.am +++ b/word-aligner/Makefile.am @@ -1,8 +1,11 @@ -bin_PROGRAMS = fast_align +bin_PROGRAMS = fast_align binderiv fast_align_SOURCES = fast_align.cc ttables.cc da.h ttables.h fast_align_LDADD = ../utils/libutils.a +binderiv_SOURCES = binderiv.cc +binderiv_LDADD = ../utils/libutils.a + EXTRA_DIST = aligner.pl ortho-norm support makefiles stemmers AM_CPPFLAGS = -W -Wall -I$(top_srcdir) -I$(top_srcdir)/utils -I$(top_srcdir)/training diff --git a/word-aligner/binderiv.cc b/word-aligner/binderiv.cc new file mode 100644 index 00000000..8ebc1105 --- /dev/null +++ b/word-aligner/binderiv.cc @@ -0,0 +1,202 @@ +#include <iostream> +#include <string> +#include <queue> +#include <sstream> + +#include "alignment_io.h" +#include "tdict.h" + +using namespace std; + +enum CombinationType { + kNONE = 0, + kAXIOM, + kMONO, kSWAP, kCONTAINS_L, kCONTAINS_R, kINTERLEAVE +}; + +string nm(CombinationType x) { + switch (x) { + case kNONE: return "NONE"; + case kAXIOM: return "AXIOM"; + case kMONO: return "MONO"; + case kSWAP: return "SWAP"; + case kCONTAINS_L: return "CONTAINS_L"; + case kCONTAINS_R: return "CONTAINS_R"; + case kINTERLEAVE: return "INTERLEAVE"; + } +} + +string Substring(const vector<WordID>& s, unsigned i, unsigned j) { + ostringstream os; + for (unsigned k = i; k < j; ++k) { + if (k > i) os << ' '; + os << TD::Convert(s[k]); + } + return os.str(); +} + +inline int min4(int a, int b, int c, int d) { + int l = a; + if (b < a) l = b; + int l2 = c; + if (d < c) l2 = c; + return min(l, l2); +} + +inline int max4(int a, int b, int c, int d) { + int l = a; + if (b > a) l = b; + int l2 = c; + if (d > c) l2 = d; + return max(l, l2); +} + +struct State { + int s,t,u,v; + State() : s(), t(), u(), v() {} + State(int a, int b, int c, int d) : s(a), t(b), u(c), v(d) { + assert(s <= t); + assert(u <= v); + } + bool IsGood() const { + return (s != 0 || t != 0 || u != 0 || v != 0); + } + CombinationType operator&(const State& r) const { + if (r.s != t) return kNONE; + if (v <= r.u) return kMONO; + if (r.v <= u) return kSWAP; + if (v >= r.v && u <= r.u) return kCONTAINS_R; + if (r.v >= v && r.u <= u) return kCONTAINS_L; + return kINTERLEAVE; + } + State& operator*=(const State& r) { + assert(r.s == t); + t = r.t; + const int tu = min4(u, v, r.u, r.v); + v = max4(u, v, r.u, r.v); + u = tu; + return *this; + } +}; + +double score(CombinationType x) { + switch (x) { + case kNONE: return 0.0; + case kAXIOM: return 1.0; + case kMONO: return 16.0; + case kSWAP: return 8.0; + case kCONTAINS_R: return 4.0; + case kCONTAINS_L: return 2.0; + case kINTERLEAVE: return 1.0; + } +} + +State operator*(const State& l, const State& r) { + State res = l; + res *= r; + return res; +} + +ostream& operator<<(ostream& os, const State& s) { + return os << '[' << s.s << ", " << s.t << ", " << s.u << ", " << s.v << ']'; +} + +string NT(const State& s) { + bool decorate=true; + if (decorate) { + ostringstream os; + os << "[X_" << s.s << '_' << s.t << '_' << s.u << '_' << s.v << "]"; + return os.str(); + } else { + return "[X]"; + } +} + +void CreateEdge(const vector<WordID>& f, const vector<WordID>& e, CombinationType ct, const State& cur, const State& left, const State& right) { + switch(ct) { + case kINTERLEAVE: + case kAXIOM: + cerr << NT(cur) << " ||| " << Substring(f, cur.s, cur.t) << " ||| " << Substring(e, cur.u, cur.v) << "\n"; + break; + case kMONO: + cerr << NT(cur) << " ||| " << NT(left) << ' ' << NT(right) << " ||| [1] [2]\n"; + break; + case kSWAP: + cerr << NT(cur) << " ||| " << NT(left) << ' ' << NT(right) << " ||| [2] [1]\n"; + break; + case kCONTAINS_L: + cerr << NT(cur) << " ||| " << Substring(f, right.s, left.s) << ' ' << NT(left) << ' ' << Substring(f, left.t, right.t) << " ||| " << Substring(e, right.u, left.u) << " [1] " << Substring(e, left.v, right.v) << endl; + break; + case kCONTAINS_R: + cerr << NT(cur) << " ||| " << Substring(f, left.s, right.s) << ' ' << NT(right) << ' ' << Substring(f, right.t, left.t) << " ||| " << Substring(e, left.u, right.u) << " [1] " << Substring(e, right.v, left.v) << endl; + break; + } +} + +void BuildArity2Forest(const vector<WordID>& f, const vector<WordID>& e, const vector<State>& axioms) { + const unsigned n = f.size(); + Array2D<State> chart(n, n+1); + Array2D<CombinationType> ctypes(n, n+1); + Array2D<double> cscore(n, n+1); + Array2D<int> cmids(n, n+1, -1); + for (const auto& axiom : axioms) { + chart(axiom.s, axiom.t) = axiom; + ctypes(axiom.s, axiom.t) = kAXIOM; + cscore(axiom.s, axiom.t) = 1.0; + CreateEdge(f, e, kAXIOM, axiom, axiom, axiom); + //cerr << "AXIOM " << axiom.s << ", " << axiom.t << " : " << chart(axiom.s, axiom.t) << " : " << 1 << endl; + } + for (unsigned l = 2; l <= n; ++l) { + const unsigned i_end = n + 1 - l; + for (unsigned i = 0; i < i_end; ++i) { + const unsigned j = i + l; + for (unsigned k = i + 1; k < j; ++k) { + const State& left = chart(i, k); + const State& right = chart(k, j); + if (!left.IsGood() || !right.IsGood()) continue; + CombinationType comb = left & right; + if (comb != kNONE) { + double ns = cscore(i,k) + cscore(k,j) + score(comb); + if (ns > cscore(i,j)) { + cscore(i,j) = ns; + chart(i,j) = left * right; + cmids(i,j) = k; + ctypes(i,j) = comb; + //cerr << "PROVED " << chart(i,j) << " : " << cscore(i,j) << " [" << nm(comb) << " " << left << " * " << right << "]\n"; + } else { + //cerr << "SUBOPTIMAL " << (left*right) << " : " << ns << " [" << nm(comb) << " " << left << " * " << right << "]\n"; + } + CreateEdge(f, e, comb, left * right, left, right); + } else { + //cerr << "CAN'T " << left << " * " << right << endl; + } + } + } + } +} + +int main(int argc, char** argv) { + State s; + vector<WordID> e,f; + TD::ConvertSentence("B C that A", &e); + TD::ConvertSentence("A de B C", &f); + State w0(0,1,3,4), w1(1,2,2,3), w2(2,3,0,1), w3(3,4,1,2); + vector<State> al = {w0, w1, w2, w3}; + // f cannot have any unaligned words, however, multiple overlapping axioms are possible + // so you can write code to align unaligned words in all ways to surrounding words + BuildArity2Forest(f, e, al); + + TD::ConvertSentence("A B C D", &e); + TD::ConvertSentence("A B , C D", &f); + vector<State> al2 = {State(0,1,0,1), State(1,2,1,2), State(1,3,1,2), State(2,4,2,3), State(4,5,3,4)}; + BuildArity2Forest(f, e, al2); + + TD::ConvertSentence("A B C D", &e); + TD::ConvertSentence("C A D B", &f); + vector<State> al3 = {State(0,1,2,3), State(1,2,0,1), State(2,3,3,4), State(3,4,1,2)}; + BuildArity2Forest(f, e, al3); + + // things to do: run EM, do posterior inference with a Dirichlet prior, etc. + return 0; +} + |