@Override
public void debug() {
System.err.println("###################");
System.err.printf(
"hypothesis [class=%s,id=%d,pos=%d,expired=%s,pending=%d]: %s\n",
getClass(),
System.identityHashCode(this),
featurizable.targetPosition,
hasExpired,
pendingPhrases.size(),
this);
System.err.printf(
"parent hypothesis [class=%s,id=%d,pos=%d,expired=%s]: %s\n",
preceedingDerivation.getClass(),
System.identityHashCode(preceedingDerivation),
preceedingDerivation.featurizable.targetPosition,
preceedingDerivation.hasExpired(),
preceedingDerivation);
System.err.println("pendingPhrasesCost: " + pendingPhrasesCost);
DTUHypothesis<TK, FV> hyp = this;
if (hyp.isDone() != hyp.featurizable.done) {
System.err.println("Error in AbstractBeamInferer with: " + hyp);
System.err.println("isDone(): " + hyp.isDone());
System.err.println("pending phrases: " + hyp.pendingPhrases.size());
System.err.println("f.done: " + hyp.featurizable.done);
Derivation<TK, FV> curHyp = hyp;
while (curHyp != null) {
System.err.println(" " + curHyp.toString());
curHyp = curHyp.preceedingDerivation;
}
throw new RuntimeException();
}
}
private LispTree exampleToLispTree(ParserState state) {
LispTree tree = LispTree.proto.newList();
tree.addChild("example");
// Basic information
Example ex = state.ex;
if (ex.id != null) tree.addChild(LispTree.proto.newList("id", ex.id));
if (ex.utterance != null) tree.addChild(LispTree.proto.newList("utterance", ex.utterance));
if (ex.targetFormula != null)
tree.addChild(LispTree.proto.newList("targetFormula", ex.targetFormula.toLispTree()));
if (ex.targetValue != null)
tree.addChild(LispTree.proto.newList("targetValue", ex.targetValue.toLispTree()));
if (ex.context != null) tree.addChild(ex.context.toLispTree());
// Language info
if (ex.languageInfo != null) {
if (ex.languageInfo.tokens != null)
tree.addChild(
LispTree.proto.newList("tokens", LispTree.proto.newList(ex.languageInfo.tokens)));
if (ex.languageInfo.lemmaTokens != null)
tree.addChild(
LispTree.proto.newList(
"lemmaTokens", LispTree.proto.newList(ex.languageInfo.lemmaTokens)));
if (ex.languageInfo.posTags != null)
tree.addChild(
LispTree.proto.newList("posTags", LispTree.proto.newList(ex.languageInfo.posTags)));
if (ex.languageInfo.nerTags != null)
tree.addChild(
LispTree.proto.newList("nerTags", LispTree.proto.newList(ex.languageInfo.nerTags)));
if (ex.languageInfo.nerValues != null)
tree.addChild(
LispTree.proto.newList("nerValues", LispTree.proto.newList(ex.languageInfo.nerValues)));
}
// Derivations
LispTree derivations = LispTree.proto.newList();
derivations.addChild("derivations");
List<Derivation> preds = state.predDerivations;
for (int i = 0; i < preds.size(); i++) {
Derivation deriv = preds.get(i);
if (!isPruned(deriv)) {
derivations.addChild(deriv.toLispTree());
}
}
tree.addChild(derivations);
return tree;
}
/**
* Returns true if the two original variables are related in a way that makes subsequence or
* subset detection not informative.
*/
public static boolean isObviousSubSequenceDynamically(
Invariant inv, VarInfo subvar, VarInfo supervar) {
VarInfo[] vis = {subvar, supervar};
ProglangType rep1 = subvar.rep_type;
ProglangType rep2 = supervar.rep_type;
if (!(((rep1 == ProglangType.INT_ARRAY) && (rep2 == ProglangType.INT_ARRAY))
|| ((rep1 == ProglangType.DOUBLE_ARRAY) && (rep2 == ProglangType.DOUBLE_ARRAY))
|| ((rep1 == ProglangType.STRING_ARRAY) && (rep2 == ProglangType.STRING_ARRAY))))
return false;
if (debug.isLoggable(Level.FINE)) {
debug.fine(
"Checking isObviousSubSequenceDynamically " + subvar.name() + " in " + supervar.name());
}
Object[] di = isObviousSubSequence(subvar, supervar);
if (di[1] != null) {
inv.log("ObvSubSeq- true from isObviousSubSequence: " + di[1]);
return true;
}
debug.fine(" not isObviousSubSequence(statically)");
PptTopLevel ppt_parent = subvar.ppt;
// If the elements of supervar are always the same (EltOneOf),
// we aren't going to learn anything new from this invariant,
// since each sequence should have an EltOneOf over it.
if (false) {
PptSlice1 slice = ppt_parent.findSlice(supervar);
if (slice == null) {
System.out.println("No slice: parent =" + ppt_parent);
} else {
System.out.println("Slice var =" + slice.var_infos[0]);
for (Invariant superinv : slice.invs) {
System.out.println("Inv = " + superinv);
if (superinv instanceof EltOneOf) {
EltOneOf eltinv = (EltOneOf) superinv;
if (eltinv.num_elts() > 0) {
inv.log(" obvious because of " + eltinv.format());
return true;
}
}
}
}
}
// Obvious if subvar is always just []
if (true) {
PptSlice1 slice = ppt_parent.findSlice(subvar);
if (slice != null) {
for (Invariant subinv : slice.invs) {
if (subinv instanceof OneOfSequence) {
OneOfSequence seqinv = (OneOfSequence) subinv;
if (seqinv.num_elts() == 1) {
Object elt = seqinv.elt();
if (elt instanceof long[] && ((long[]) elt).length == 0) {
Debug.log(
debug, inv.getClass(), inv.ppt, vis, "ObvSubSeq- True from subvar being []");
return true;
}
if (elt instanceof double[] && ((double[]) elt).length == 0) {
inv.log("ObvSubSeq- True from subvar being []");
return true;
}
}
}
}
}
}
// Check for a[0..i] subseq a[0..j] but i < j.
VarInfo subvar_super = subvar.isDerivedSubSequenceOf();
VarInfo supervar_super = supervar.isDerivedSubSequenceOf();
if (subvar_super != null && subvar_super == supervar_super) {
// both sequences are derived from the same supersequence
if ((subvar.derived instanceof SequenceScalarSubsequence
|| subvar.derived instanceof SequenceScalarArbitrarySubsequence)
&& (supervar.derived instanceof SequenceScalarSubsequence
|| supervar.derived instanceof SequenceScalarArbitrarySubsequence)) {
VarInfo sub_left_var = null,
sub_right_var = null,
super_left_var = null,
super_right_var = null;
// I'm careful not to access foo_shift unless foo_var has been set
// to a non-null value, but Java is too stupid to recognize that.
int sub_left_shift = 42,
sub_right_shift = 69,
super_left_shift = 1492,
super_right_shift = 1776;
if (subvar.derived instanceof SequenceScalarSubsequence) {
SequenceScalarSubsequence sub = (SequenceScalarSubsequence) subvar.derived;
if (sub.from_start) {
sub_right_var = sub.sclvar();
sub_right_shift = sub.index_shift;
} else {
sub_left_var = sub.sclvar();
sub_left_shift = sub.index_shift;
}
} else if (subvar.derived instanceof SequenceScalarArbitrarySubsequence) {
SequenceScalarArbitrarySubsequence sub =
(SequenceScalarArbitrarySubsequence) subvar.derived;
sub_left_var = sub.startvar();
sub_left_shift = (sub.left_closed ? 0 : 1);
sub_right_var = sub.endvar();
sub_right_shift = (sub.right_closed ? 0 : -1);
} else {
Assert.assertTrue(false);
}
if (supervar.derived instanceof SequenceScalarSubsequence) {
SequenceScalarSubsequence super_ = (SequenceScalarSubsequence) supervar.derived;
if (super_.from_start) {
super_right_var = super_.sclvar();
super_right_shift = super_.index_shift;
} else {
super_left_var = super_.sclvar();
super_left_shift = super_.index_shift;
}
} else if (supervar.derived instanceof SequenceScalarArbitrarySubsequence) {
SequenceScalarArbitrarySubsequence super_ =
(SequenceScalarArbitrarySubsequence) supervar.derived;
super_left_var = super_.startvar();
super_left_shift = (super_.left_closed ? 0 : 1);
super_right_var = super_.endvar();
super_right_shift = (super_.right_closed ? 0 : -1);
} else {
Assert.assertTrue(false);
}
boolean left_included, right_included;
if (super_left_var == null) left_included = true;
else if (sub_left_var == null) // we know super_left_var != null here
left_included = false;
else
left_included =
VarInfo.compare_vars(
super_left_var, super_left_shift, sub_left_var, sub_left_shift, true /* <= */);
if (super_right_var == null) right_included = true;
else if (sub_right_var == null) // we know super_right_var != null here
right_included = false;
else
right_included =
VarInfo.compare_vars(
super_right_var,
super_right_shift,
sub_right_var,
sub_right_shift,
false /* >= */);
// System.out.println("Is " + subvar.name() + " contained in "
// + supervar.name()
// + "? left: " + left_included + ", right: "
// + right_included);
if (left_included && right_included) {
inv.log("ObvSubSeq- True a[0..i] subseq a[0..j] and i < j");
return true;
}
} else if ((subvar.derived instanceof SequenceStringSubsequence)
&& (supervar.derived instanceof SequenceStringSubsequence)) {
// Copied from just above
SequenceStringSubsequence sss1 = (SequenceStringSubsequence) subvar.derived;
SequenceStringSubsequence sss2 = (SequenceStringSubsequence) supervar.derived;
VarInfo index1 = sss1.sclvar();
int shift1 = sss1.index_shift;
boolean start1 = sss1.from_start;
VarInfo index2 = sss2.sclvar();
int shift2 = sss2.index_shift;
boolean start2 = sss2.from_start;
if (start1 == start2)
if (VarInfo.compare_vars(index1, shift1, index2, shift2, start1)) {
inv.log("True from comparing indices");
return true;
}
} else {
Assert.assertTrue(
false,
"how can this happen? "
+ subvar.name()
+ " "
+ subvar.derived.getClass()
+ " "
+ supervar.name()
+ " "
+ supervar.derived.getClass());
}
}
// Also need to check A[0..i] subseq A[0..j] via compare_vars.
// A subseq B[0..n] => A subseq B
List<Derivation> derivees = supervar.derivees();
// For each variable derived from supervar ("B")
for (Derivation der : derivees) {
// System.out.println(" ... der = " + der.getVarInfo().name() + " " + der);
if (der instanceof SequenceScalarSubsequence) {
// If that variable is "B[0..n]"
VarInfo supervar_part = der.getVarInfo();
// Get the canonical version; being equal to it is good enough.
if (supervar_part.get_equalitySet_leader() == subvar) {
Debug.log(debug, inv.getClass(), inv.ppt, vis, "ObvSubSeq- True from canonical leader");
return true;
}
if (supervar_part.isCanonical()) {
if (subvar == supervar_part) {
System.err.println(
"Error: variables "
+ subvar.name()
+ " and "
+ supervar_part.name()
+ " are identical. Canonical");
System.err.println(subvar.isCanonical());
System.err.println(supervar_part.isCanonical());
throw new Error();
}
// Check to see if there is a subsequence over the supervar
if (ppt_parent.is_subsequence(subvar, supervar_part)) {
if (Debug.logOn())
inv.log(
"ObvSubSeq- true from A subseq B[0..n] "
+ subvar.name()
+ "/"
+ supervar_part.name());
return (true);
}
}
}
}
return false;
}
/** Constructor used for 1st segment of a discontinuous phrase. */
public DTUHypothesis(
int sourceInputId,
ConcreteRule<TK, FV> translationOpt,
int insertionPosition,
Derivation<TK, FV> baseHyp,
CombinedFeaturizer<TK, FV> featurizer,
Scorer<FV> scorer,
SearchHeuristic<TK, FV> heuristic) {
super(
sourceInputId,
translationOpt,
translationOpt.abstractRule,
insertionPosition,
baseHyp,
featurizer,
scorer,
heuristic, /*
* targetPhrase=
*/
getSegment(translationOpt.abstractRule, 0),
/* hasPendingPhrases= */ hasPendingPhrases(translationOpt, baseHyp, true, false),
/* segmentIdx= */ 0);
// Copy old pending phrases from parent hypothesis:
this.pendingPhrases = new TreeSet<PendingPhrase<TK, FV>>();
if (baseHyp instanceof DTUHypothesis) {
Set<PendingPhrase<TK, FV>> oldPhrases = ((DTUHypothesis<TK, FV>) baseHyp).pendingPhrases;
for (PendingPhrase<TK, FV> oldPhrase : oldPhrases) {
this.pendingPhrases.add(new PendingPhrase<TK, FV>(oldPhrase));
int lastPosition = oldPhrase.lastPosition;
if (lastPosition < this.length) this.hasExpired = true;
}
}
// First segment of a discontinuous phrase has both source and target:
this.segmentIdx = 0;
// If parent hypothesis has expired, so does the current:
if (baseHyp.hasExpired()) this.hasExpired = true;
// Add new pending phrases:
// assert (MAX_TARGET_PHRASE_SPAN >= 0);
if (translationOpt.abstractRule instanceof DTURule) {
PendingPhrase<TK, FV> newPhrase =
new PendingPhrase<TK, FV>(
translationOpt,
sourceInputId,
this,
featurizer,
scorer,
0,
this.length + MIN_GAP_SIZE,
this.length + MAX_TARGET_PHRASE_SPAN);
pendingPhrases.add(newPhrase);
}
// Too many pending phrases?:
if (pendingPhrases.size() > MAX_PENDING_PHRASES) this.hasExpired = true;
// Estimate future cost for pending phrases:
pendingPhrasesCost = costPendingPhrases();
checkExpiration();
}