public static void RunEndToEndTest(IList <Tree> binarizedTrees, Oracle oracle)
{
for (int index = 0; index < binarizedTrees.Count; ++index)
{
State state = ShiftReduceParser.InitialStateFromGoldTagTree(binarizedTrees[index]);
while (!state.IsFinished())
{
OracleTransition gold = oracle.GoldTransition(index, state);
NUnit.Framework.Assert.IsTrue(gold.transition != null);
state = gold.transition.Apply(state);
}
NUnit.Framework.Assert.AreEqual(binarizedTrees[index], state.stack.Peek());
}
}
/// <summary>
/// Returns an attempt at a "gold" transition given the current state
/// while parsing a known gold tree.
/// </summary>
/// <remarks>
/// Returns an attempt at a "gold" transition given the current state
/// while parsing a known gold tree.
/// Tree is passed in by index so the oracle can precompute various
/// statistics about the tree.
/// If we already finalized, then the correct transition is to idle.
/// If the stack is empty, shift is the only possible answer.
/// If the first item on the stack is a correct span, correctly
/// labeled, and it has unaries transitions above it, then if we are
/// not doing compound unaries, the next unary up is the correct
/// answer. If we are doing compound unaries, and the state does not
/// already have a transition, then the correct answer is a compound
/// unary transition to the top of the unary chain.
/// If the first item is the entire tree, with no remaining unary
/// transitions, then we need to finalize.
/// If the first item is a correct span, with or without a correct
/// label, and there are no unary transitions to be added, then we
/// must look at the next parent. If it has the same left side, then
/// we return a shift transition. If it has the same right side,
/// then we look at the next subtree on the stack (which must exist).
/// If it is also correct, then the transition is to combine the two
/// subtrees with the correct label and side.
/// TODO: suppose the correct label is not either child label and the
/// children are binarized states? We should see what the
/// debinarizer does in that case. Perhaps a post-processing step
/// If the previous stack item is too small, then any binary reduce
/// action is legal, with no gold transition. TODO: can this be improved?
/// If the previous stack item is too large, perhaps because of
/// incorrectly attached PP/SBAR, for example, we still need to
/// binary reduce. TODO: is that correct? TODO: we could look back
/// further in the stack to find hints at a label that would work
/// better, for example
/// If the current item is an incorrect span, then look at the
/// containing item. If it has the same left side, shift. If it has
/// the same right side, binary reduce (producing an exact span if
/// possible). If neither edge is correct, then any of shift or
/// binary reduce are acceptable, with no gold transition. TODO: can
/// this be improved?
/// </remarks>
internal virtual OracleTransition GoldTransition(int index, State state)
{
if (state.finished)
{
return(new OracleTransition(new IdleTransition(), false, false, false));
}
if (state.stack.Size() == 0)
{
return(new OracleTransition(new ShiftTransition(), false, false, false));
}
IDictionary <Tree, Tree> parents = parentMaps[index];
Tree gold = binarizedTrees[index];
IList <Tree> leaves = leafLists[index];
Tree S0 = state.stack.Peek();
Tree enclosingS0 = GetEnclosingTree(S0, parents, leaves);
OracleTransition result = GetUnaryTransition(S0, enclosingS0, parents, compoundUnaries);
if (result != null)
{
return(result);
}
// TODO: we could interject that all trees must end with ROOT, for example
if (state.tokenPosition >= state.sentence.Count && state.stack.Size() == 1)
{
return(new OracleTransition(new FinalizeTransition(rootStates), false, false, false));
}
if (state.stack.Size() == 1)
{
return(new OracleTransition(new ShiftTransition(), false, false, false));
}
if (SpansEqual(S0, enclosingS0))
{
Tree parent = parents[enclosingS0];
// cannot be root
while (SpansEqual(parent, enclosingS0))
{
// in case we had missed unary transitions
enclosingS0 = parent;
parent = parents[parent];
}
if (parent.Children()[0] == enclosingS0)
{
// S0 is the left child of the correct tree
return(new OracleTransition(new ShiftTransition(), false, false, false));
}
// was the second (right) child. there must be something else on the stack...
Tree S1 = state.stack.Pop().Peek();
Tree enclosingS1 = GetEnclosingTree(S1, parents, leaves);
if (SpansEqual(S1, enclosingS1))
{
// the two subtrees should be combined
return(new OracleTransition(new BinaryTransition(parent.Value(), ShiftReduceUtils.GetBinarySide(parent)), false, false, false));
}
return(new OracleTransition(null, false, true, false));
}
if (ShiftReduceUtils.LeftIndex(S0) == ShiftReduceUtils.LeftIndex(enclosingS0))
{
return(new OracleTransition(new ShiftTransition(), false, false, false));
}
if (ShiftReduceUtils.RightIndex(S0) == ShiftReduceUtils.RightIndex(enclosingS0))
{
Tree S1 = state.stack.Pop().Peek();
Tree enclosingS1 = GetEnclosingTree(S1, parents, leaves);
if (enclosingS0 == enclosingS1)
{
// BinaryTransition with enclosingS0's label, either side, but preferring LEFT
return(new OracleTransition(new BinaryTransition(enclosingS0.Value(), BinaryTransition.Side.Left), false, false, true));
}
// S1 is smaller than the next tree S0 is supposed to be part of,
// so we must have a BinaryTransition
if (ShiftReduceUtils.LeftIndex(S1) > ShiftReduceUtils.LeftIndex(enclosingS0))
{
return(new OracleTransition(null, false, true, true));
}
// S1 is larger than the next tree. This is the worst case
return(new OracleTransition(null, true, true, true));
}
// S0 doesn't match either endpoint of the enclosing tree
return(new OracleTransition(null, true, true, true));
}
private Pair <int, int> TrainTree(int index, IList <Tree> binarizedTrees, IList <IList <ITransition> > transitionLists, IList <PerceptronModel.Update> updates, Oracle oracle)
{
int numCorrect = 0;
int numWrong = 0;
Tree tree = binarizedTrees[index];
ReorderingOracle reorderer = null;
if (op.TrainOptions().trainingMethod == ShiftReduceTrainOptions.TrainingMethod.ReorderOracle || op.TrainOptions().trainingMethod == ShiftReduceTrainOptions.TrainingMethod.ReorderBeam)
{
reorderer = new ReorderingOracle(op);
}
// TODO. This training method seems to be working in that it
// trains models just like the gold and early termination methods do.
// However, it causes the feature space to go crazy. Presumably
// leaving out features with low weights or low frequencies would
// significantly help with that. Otherwise, not sure how to keep
// it under control.
if (op.TrainOptions().trainingMethod == ShiftReduceTrainOptions.TrainingMethod.Oracle)
{
State state = ShiftReduceParser.InitialStateFromGoldTagTree(tree);
while (!state.IsFinished())
{
IList <string> features = featureFactory.Featurize(state);
ScoredObject <int> prediction = FindHighestScoringTransition(state, features, true);
if (prediction == null)
{
throw new AssertionError("Did not find a legal transition");
}
int predictedNum = prediction.Object();
ITransition predicted = transitionIndex.Get(predictedNum);
OracleTransition gold = oracle.GoldTransition(index, state);
if (gold.IsCorrect(predicted))
{
numCorrect++;
if (gold.transition != null && !gold.transition.Equals(predicted))
{
int transitionNum = transitionIndex.IndexOf(gold.transition);
if (transitionNum < 0)
{
// TODO: do we want to add unary transitions which are
// only possible when the parser has gone off the rails?
continue;
}
updates.Add(new PerceptronModel.Update(features, transitionNum, -1, learningRate));
}
}
else
{
numWrong++;
int transitionNum = -1;
if (gold.transition != null)
{
transitionNum = transitionIndex.IndexOf(gold.transition);
}
// TODO: this can theoretically result in a -1 gold
// transition if the transition exists, but is a
// CompoundUnaryTransition which only exists because the
// parser is wrong. Do we want to add those transitions?
updates.Add(new PerceptronModel.Update(features, transitionNum, predictedNum, learningRate));
}
state = predicted.Apply(state);
}
}
else
{
if (op.TrainOptions().trainingMethod == ShiftReduceTrainOptions.TrainingMethod.Beam || op.TrainOptions().trainingMethod == ShiftReduceTrainOptions.TrainingMethod.ReorderBeam)
{
if (op.TrainOptions().beamSize <= 0)
{
throw new ArgumentException("Illegal beam size " + op.TrainOptions().beamSize);
}
IList <ITransition> transitions = Generics.NewLinkedList(transitionLists[index]);
PriorityQueue <State> agenda = new PriorityQueue <State>(op.TrainOptions().beamSize + 1, ScoredComparator.AscendingComparator);
State goldState = ShiftReduceParser.InitialStateFromGoldTagTree(tree);
agenda.Add(goldState);
// int transitionCount = 0;
while (transitions.Count > 0)
{
ITransition goldTransition = transitions[0];
ITransition highestScoringTransitionFromGoldState = null;
double highestScoreFromGoldState = 0.0;
PriorityQueue <State> newAgenda = new PriorityQueue <State>(op.TrainOptions().beamSize + 1, ScoredComparator.AscendingComparator);
State highestScoringState = null;
State highestCurrentState = null;
foreach (State currentState in agenda)
{
bool isGoldState = (op.TrainOptions().trainingMethod == ShiftReduceTrainOptions.TrainingMethod.ReorderBeam && goldState.AreTransitionsEqual(currentState));
IList <string> features = featureFactory.Featurize(currentState);
ICollection <ScoredObject <int> > stateTransitions = FindHighestScoringTransitions(currentState, features, true, op.TrainOptions().beamSize, null);
foreach (ScoredObject <int> transition in stateTransitions)
{
State newState = transitionIndex.Get(transition.Object()).Apply(currentState, transition.Score());
newAgenda.Add(newState);
if (newAgenda.Count > op.TrainOptions().beamSize)
{
newAgenda.Poll();
}
if (highestScoringState == null || highestScoringState.Score() < newState.Score())
{
highestScoringState = newState;
highestCurrentState = currentState;
}
if (isGoldState && (highestScoringTransitionFromGoldState == null || transition.Score() > highestScoreFromGoldState))
{
highestScoringTransitionFromGoldState = transitionIndex.Get(transition.Object());
highestScoreFromGoldState = transition.Score();
}
}
}
// This can happen if the REORDER_BEAM method backs itself
// into a corner, such as transitioning to something that
// can't have a FinalizeTransition applied. This doesn't
// happen for the BEAM method because in that case the correct
// state (eg one with ROOT) isn't on the agenda so it stops.
if (op.TrainOptions().trainingMethod == ShiftReduceTrainOptions.TrainingMethod.ReorderBeam && highestScoringTransitionFromGoldState == null)
{
break;
}
State newGoldState = goldTransition.Apply(goldState, 0.0);
// if highest scoring state used the correct transition, no training
// otherwise, down the last transition, up the correct
if (!newGoldState.AreTransitionsEqual(highestScoringState))
{
++numWrong;
IList <string> goldFeatures = featureFactory.Featurize(goldState);
int lastTransition = transitionIndex.IndexOf(highestScoringState.transitions.Peek());
updates.Add(new PerceptronModel.Update(featureFactory.Featurize(highestCurrentState), -1, lastTransition, learningRate));
updates.Add(new PerceptronModel.Update(goldFeatures, transitionIndex.IndexOf(goldTransition), -1, learningRate));
if (op.TrainOptions().trainingMethod == ShiftReduceTrainOptions.TrainingMethod.Beam)
{
// If the correct state has fallen off the agenda, break
if (!ShiftReduceUtils.FindStateOnAgenda(newAgenda, newGoldState))
{
break;
}
else
{
transitions.Remove(0);
}
}
else
{
if (op.TrainOptions().trainingMethod == ShiftReduceTrainOptions.TrainingMethod.ReorderBeam)
{
if (!ShiftReduceUtils.FindStateOnAgenda(newAgenda, newGoldState))
{
if (!reorderer.Reorder(goldState, highestScoringTransitionFromGoldState, transitions))
{
break;
}
newGoldState = highestScoringTransitionFromGoldState.Apply(goldState);
if (!ShiftReduceUtils.FindStateOnAgenda(newAgenda, newGoldState))
{
break;
}
}
else
{
transitions.Remove(0);
}
}
}
}
else
{
++numCorrect;
transitions.Remove(0);
}
goldState = newGoldState;
agenda = newAgenda;
}
}
else
{
if (op.TrainOptions().trainingMethod == ShiftReduceTrainOptions.TrainingMethod.ReorderOracle || op.TrainOptions().trainingMethod == ShiftReduceTrainOptions.TrainingMethod.EarlyTermination || op.TrainOptions().trainingMethod == ShiftReduceTrainOptions.TrainingMethod
.Gold)
{
State state = ShiftReduceParser.InitialStateFromGoldTagTree(tree);
IList <ITransition> transitions = transitionLists[index];
transitions = Generics.NewLinkedList(transitions);
bool keepGoing = true;
while (transitions.Count > 0 && keepGoing)
{
ITransition transition = transitions[0];
int transitionNum = transitionIndex.IndexOf(transition);
IList <string> features = featureFactory.Featurize(state);
int predictedNum = FindHighestScoringTransition(state, features, false).Object();
ITransition predicted = transitionIndex.Get(predictedNum);
if (transitionNum == predictedNum)
{
transitions.Remove(0);
state = transition.Apply(state);
numCorrect++;
}
else
{
numWrong++;
// TODO: allow weighted features, weighted training, etc
updates.Add(new PerceptronModel.Update(features, transitionNum, predictedNum, learningRate));
switch (op.TrainOptions().trainingMethod)
{
case ShiftReduceTrainOptions.TrainingMethod.EarlyTermination:
{
keepGoing = false;
break;
}
case ShiftReduceTrainOptions.TrainingMethod.Gold:
{
transitions.Remove(0);
state = transition.Apply(state);
break;
}
case ShiftReduceTrainOptions.TrainingMethod.ReorderOracle:
{
keepGoing = reorderer.Reorder(state, predicted, transitions);
if (keepGoing)
{
state = predicted.Apply(state);
}
break;
}
default:
{
throw new ArgumentException("Unexpected method " + op.TrainOptions().trainingMethod);
}
}
}
}
}
}
}
return(Pair.MakePair(numCorrect, numWrong));
}