handle translation from the null word

git-svn-id: https://ws10smt.googlecode.com/svn/trunk@689 ec762483-ff6d-05da-a07a-a48fb63a330f

2 files changed, 44 insertions, 55 deletions

diff --git a/word-aligner/aligner.pl b/word-aligner/aligner.pl
index 7821560f..e23c2beb 100755
--- a/word-aligner/aligner.pl
+++ b/word-aligner/aligner.pl
@@ -6,9 +6,10 @@ use Getopt::Long;
 my $training_dir = "$SCRIPT_DIR/../training";
 die "Can't find training dir: $training_dir" unless -d $training_dir;
 
-my $mkcls = '/Users/redpony/software/giza/giza-pp/mkcls-v2/mkcls';
+my $mkcls = '/Users/cdyer/software/giza-pp/mkcls-v2/mkcls';
 my $num_classes = 50;
 my $nodes = 40;
+my $TRAINING_ITERATIONS = 2000;
 my $pmem = "2500mb";
 my $DECODER = "cdec";
 GetOptions("cdec=s" => \$DECODER,
@@ -16,15 +17,11 @@ GetOptions("cdec=s" => \$DECODER,
            "pmem=s" => \$pmem,
            "mkcls=s" => \$mkcls,
           ) or usage();
-usage() unless (scalar @ARGV == 3);
+usage() unless (scalar @ARGV == 1);
 die "Cannot find mkcls (specify with --mkcls=/path/to/mkcls) at $mkcls\n" unless -f $mkcls;
 die "Cannot execute mkcls at $mkcls\n" unless -x $mkcls;
 
 my $in_file = shift @ARGV;
-my $m4 = shift @ARGV;
-my $im4 = shift @ARGV;
-die "Can't find model4: $m4" unless -f $m4;
-die "Can't find inverse model4: $im4" unless -f $im4;
 
 die "Expected format corpus.l1-l2 where l1 & l2 are two-letter abbreviations\nfor the source and target language respectively\n" unless ($in_file =~ /^.+\.([a-z][a-z])-([a-z][a-z])$/);
 my $f_lang = $1;
@@ -32,13 +29,11 @@ my $e_lang = $2;
 
 print STDERR "Source language: $f_lang\n";
 print STDERR "Target language: $e_lang\n";
-print STDERR "  Model 4 align: $m4\n";
-print STDERR "InModel 4 align: $im4\n";
 print STDERR " Using mkcls in: $mkcls\n\n";
 die "Don't have an orthographic normalizer for $f_lang\n" unless -f "$SCRIPT_DIR/ortho-norm/$f_lang.pl";
 die "Don't have an orthographic normalizer for $e_lang\n" unless -f "$SCRIPT_DIR/ortho-norm/$e_lang.pl";
 
-my @stages = qw(nopos relpos markov);
+my @stages = qw(markov);
 my @directions = qw(f-e e-f);
 
 my $corpus = 'c';
@@ -67,12 +62,8 @@ die unless $? == 0;
 my @targets = qw(grammars);
 
 for my $direction (@directions) {
-  my $prev_stage = undef;
-  for my $stage (@stages) {
-    push @targets, "$stage-$direction";
-    make_stage($stage, $direction, $prev_stage);
-    $prev_stage = $stage;
-  }
+  push @targets, "model-$direction";
+  make_stage($direction);
 }
 
 open TOPLEVEL, ">$align_dir/Makefile" or die "Can't write $align_dir/Makefile: $!";
@@ -84,8 +75,6 @@ SCRIPT_DIR = $SCRIPT_DIR
 TRAINING_DIR = $training_dir
 MKCLS = $mkcls
 NCLASSES = $num_classes
-GIZAALIGN = $m4
-INVGIZAALIGN = $im4
 
 TARGETS = @targets
 PTRAIN = \$(TRAINING_DIR)/cluster-ptrain.pl --restart_if_necessary
@@ -113,13 +102,12 @@ print STDERR "Created alignment task. chdir to talign/ then type make.\n\n";
 exit 0;
 
 sub make_stage {
-  my ($stage, $direction, $prev_stage) = @_;
+  my ($direction) = @_;
   my $stage_dir = "$align_dir/model-$direction";
   my $first = $direction;
   $first =~ s/^(.+)-.*$/$1/;
   mkdir $stage_dir;
-  my $RELPOS = "feature_function=RelativeSentencePosition $align_dir/grammars/corpus.class.$first\n";
-  open CDEC, ">$stage_dir/cdec.$stage.ini" or die;
+  open CDEC, ">$stage_dir/cdec.ini" or die "Can't write $stage_dir/cdec.ini: $!";
   print CDEC <<EOT;
 formalism=lextrans
 intersection_strategy=full
@@ -127,23 +115,22 @@ grammar=$align_dir/grammars/corpus.$direction.lex-grammar.gz
 feature_function=LexicalPairIdentity
 feature_function=InputIdentity
 feature_function=OutputIdentity
+feature_function=RelativeSentencePosition $align_dir/grammars/corpus.class.$first
+feature_function=MarkovJump +b
+feature_function=MarkovJumpFClass $align_dir/grammars/corpus.class.$first
+feature_function=SourceBigram
+feature_function=SourcePOSBigram $align_dir/grammars/corpus.class.$first
 EOT
-  if ($stage =~ /relpos/) {
-    print CDEC "$RELPOS\n";
-  } elsif ($stage =~ /markov/) {
-    print CDEC "$RELPOS\n";
-    print CDEC "feature_function=MarkovJump\n";
-    print CDEC "feature_function=MarkovJumpFClass $align_dir/grammars/corpus.class.$first\n";
-    print CDEC "feature_function=SourceBigram\n";
-    print CDEC "feature_function=SourcePOSBigram $align_dir/grammars/corpus.class.$first\n";
-  }
   close CDEC;
+  open AGENDA, ">$stage_dir/agenda.txt" or die "Can't write $stage_dir/agenda.txt: $!";
+  print AGENDA "cdec.ini $TRAINING_ITERATIONS\n";
+  close AGENDA;
 }
 
 sub usage {
   die <<EOT;
 
-Usage: $0 [OPTIONS] training_corpus.fr-en giza.en-fr.A3 giza.fr-en.A3
+Usage: $0 [OPTIONS] training_corpus.fr-en
 
 EOT
 }
diff --git a/word-aligner/support/make_lex_grammar.pl b/word-aligner/support/make_lex_grammar.pl
index 3926fd8d..fb9d0214 100755
--- a/word-aligner/support/make_lex_grammar.pl
+++ b/word-aligner/support/make_lex_grammar.pl
@@ -5,7 +5,8 @@ use strict;
 my $LIMIT_SIZE=30;
 
 my ($effile, $model1, $imodel1, $orthof, $orthoe, $class_e, $class_f, $gizaf2e, $gizae2f) = @ARGV;
-die "Usage: $0 corpus.fr-en corpus.f-e.model1 corpus.e-f.model1 corpus.orthonorm-dict.f corpus.orthnorm-dict.e class.e class.f\n" unless $effile && -f $effile && $model1 && -f $model1 && $imodel1 && -f $imodel1 && $orthof && -f $orthof && $orthoe && -f $orthoe && -f $class_e && -f $class_f && -f $gizaf2e && -f $gizae2f;
+die "Usage: $0 corpus.fr-en corpus.f-e.model1 corpus.e-f.model1 corpus.orthonorm-dict.f corpus.orthnorm-dict.e class.e class.f\n" unless $effile && -f $effile && $model1 && -f $model1 && $imodel1 && -f $imodel1 && $orthof && -f $orthof && $orthoe && -f $orthoe && -f $class_e && -f $class_f;
+
 
 my %eclass = ();
 my %fclass = ();
@@ -20,12 +21,12 @@ our %cache;
 open EF, "<$effile" or die;
 open M1, "<$model1" or die;
 open IM1, "<$imodel1" or die;
-open M4, "<$gizaf2e" or die;
-open IM4, "<$gizae2f" or die;
+#open M4, "<$gizaf2e" or die;
+#open IM4, "<$gizae2f" or die;
+#binmode(M4,":utf8");
+#binmode(IM4,":utf8");
 binmode(EF,":utf8");
 binmode(M1,":utf8");
-binmode(M4,":utf8");
-binmode(IM4,":utf8");
 binmode(IM1,":utf8");
 binmode(STDOUT,":utf8");
 my %model1;
@@ -105,7 +106,7 @@ my $ADD_DICE = 1;
 my $ADD_111 = 1;
 my $ADD_ID = 1;
 my $ADD_PUNC = 1;
-my $ADD_NULL = 0;
+my $ADD_NULL = 1;
 my $ADD_STEM_ID = 0;
 my $ADD_SYM = 0;
 my $BEAM_RATIO = 50;
@@ -115,6 +116,8 @@ my $BIN_IDENT = 1;
 my $BIN_DICE = 1;
 my $ADD_FIDENT = 0;
 
+if ($ADD_NULL) { $fclass{'<eps>'}='NUL'; $eclass{'<eps>'} ='NUL'; }
+
 my %fdict;
 my %fcounts;
 my %ecounts;
@@ -146,24 +149,24 @@ while(<EF>) {
 
 print STDERR "Loading Giza output...\n";
 my %model4;
-while(<M4>) {
-  my $en = <M4>; chomp $en;
-  my $zh = <M4>; chomp $zh;
-  die unless $zh =~ /^NULL \({/;
-  my @ewords = split /\s+/, $en;
-  my @chunks = split /\}\) ?/, $zh;
-
-  for my $c (@chunks) {
-    my ($zh, $taps) = split / \(\{ /, $c;
-    if ($zh eq 'NULL') { $zh = '<eps>'; }
-    my @aps = map { $ewords[$_ - 1]; } (split / /, $taps);
-    #print "$zh -> @aps\n";
-    for my $ap (@aps) {
-      $model4{$zh}->{$ap} += 1;
-    }
-  }
-}
-close M4;
+#while(<M4>) {
+#  my $en = <M4>; chomp $en;
+#  my $zh = <M4>; chomp $zh;
+#  die unless $zh =~ /^NULL \({/;
+#  my @ewords = split /\s+/, $en;
+#  my @chunks = split /\}\) ?/, $zh;
+#
+#  for my $c (@chunks) {
+#    my ($zh, $taps) = split / \(\{ /, $c;
+#    if ($zh eq 'NULL') { $zh = '<eps>'; }
+#    my @aps = map { $ewords[$_ - 1]; } (split / /, $taps);
+#    #print "$zh -> @aps\n";
+#    for my $ap (@aps) {
+#      $model4{$zh}->{$ap} += 1;
+#    }
+#  }
+#}
+#close M4;
 
 my $specials = 0;
 my $fc = 1000000;
@@ -207,7 +210,6 @@ for my $f (sort keys %fdict) {
     }
     my $is_null = undef;
     if ($ADD_NULL && $f eq '<eps>') {
-      push @feats, "IsNull=1";
       $is_null = 1;
     }
     if ($ADD_LEN) {