/// <summary>The search algorithm, starting with a full sentence and iteratively shortening it to its entailed sentences.</summary>
/// <returns>A list of search results, corresponding to shortenings of the sentence.</returns>
private IList <ForwardEntailerSearchProblem.SearchResult> SearchImplementation()
{
// Pre-process the tree
SemanticGraph parseTree = new SemanticGraph(this.parseTree);
System.Diagnostics.Debug.Assert(Edu.Stanford.Nlp.Naturalli.Util.IsTree(parseTree));
// (remove common determiners)
IList <string> determinerRemovals = new List <string>();
parseTree.GetLeafVertices().Stream().Filter(null).ForEach(null);
// (cut conj_and nodes)
ICollection <SemanticGraphEdge> andsToAdd = new HashSet <SemanticGraphEdge>();
foreach (IndexedWord vertex in parseTree.VertexSet())
{
if (parseTree.InDegree(vertex) > 1)
{
SemanticGraphEdge conjAnd = null;
foreach (SemanticGraphEdge edge in parseTree.IncomingEdgeIterable(vertex))
{
if ("conj:and".Equals(edge.GetRelation().ToString()))
{
conjAnd = edge;
}
}
if (conjAnd != null)
{
parseTree.RemoveEdge(conjAnd);
System.Diagnostics.Debug.Assert(Edu.Stanford.Nlp.Naturalli.Util.IsTree(parseTree));
andsToAdd.Add(conjAnd);
}
}
}
// Clean the tree
Edu.Stanford.Nlp.Naturalli.Util.CleanTree(parseTree);
System.Diagnostics.Debug.Assert(Edu.Stanford.Nlp.Naturalli.Util.IsTree(parseTree));
// Find the subject / object split
// This takes max O(n^2) time, expected O(n*log(n)) time.
// Optimal is O(n), but I'm too lazy to implement it.
BitSet isSubject = new BitSet(256);
foreach (IndexedWord vertex_1 in parseTree.VertexSet())
{
// Search up the tree for a subj node; if found, mark that vertex as a subject.
IEnumerator <SemanticGraphEdge> incomingEdges = parseTree.IncomingEdgeIterator(vertex_1);
SemanticGraphEdge edge = null;
if (incomingEdges.MoveNext())
{
edge = incomingEdges.Current;
}
int numIters = 0;
while (edge != null)
{
if (edge.GetRelation().ToString().EndsWith("subj"))
{
System.Diagnostics.Debug.Assert(vertex_1.Index() > 0);
isSubject.Set(vertex_1.Index() - 1);
break;
}
incomingEdges = parseTree.IncomingEdgeIterator(edge.GetGovernor());
if (incomingEdges.MoveNext())
{
edge = incomingEdges.Current;
}
else
{
edge = null;
}
numIters += 1;
if (numIters > 100)
{
// log.error("tree has apparent depth > 100");
return(Java.Util.Collections.EmptyList);
}
}
}
// Outputs
IList <ForwardEntailerSearchProblem.SearchResult> results = new List <ForwardEntailerSearchProblem.SearchResult>();
if (!determinerRemovals.IsEmpty())
{
if (andsToAdd.IsEmpty())
{
double score = Math.Pow(weights.DeletionProbability("det"), (double)determinerRemovals.Count);
System.Diagnostics.Debug.Assert(!double.IsNaN(score));
System.Diagnostics.Debug.Assert(!double.IsInfinite(score));
results.Add(new ForwardEntailerSearchProblem.SearchResult(parseTree, determinerRemovals, score));
}
else
{
SemanticGraph treeWithAnds = new SemanticGraph(parseTree);
System.Diagnostics.Debug.Assert(Edu.Stanford.Nlp.Naturalli.Util.IsTree(treeWithAnds));
foreach (SemanticGraphEdge and in andsToAdd)
{
treeWithAnds.AddEdge(and.GetGovernor(), and.GetDependent(), and.GetRelation(), double.NegativeInfinity, false);
}
System.Diagnostics.Debug.Assert(Edu.Stanford.Nlp.Naturalli.Util.IsTree(treeWithAnds));
results.Add(new ForwardEntailerSearchProblem.SearchResult(treeWithAnds, determinerRemovals, Math.Pow(weights.DeletionProbability("det"), (double)determinerRemovals.Count)));
}
}
// Initialize the search
System.Diagnostics.Debug.Assert(Edu.Stanford.Nlp.Naturalli.Util.IsTree(parseTree));
IList <IndexedWord> topologicalVertices;
try
{
topologicalVertices = parseTree.TopologicalSort();
}
catch (InvalidOperationException)
{
// log.info("Could not topologically sort the vertices! Using left-to-right traversal.");
topologicalVertices = parseTree.VertexListSorted();
}
if (topologicalVertices.IsEmpty())
{
return(results);
}
Stack <ForwardEntailerSearchProblem.SearchState> fringe = new Stack <ForwardEntailerSearchProblem.SearchState>();
fringe.Push(new ForwardEntailerSearchProblem.SearchState(new BitSet(256), 0, parseTree, null, null, 1.0));
// Start the search
int numTicks = 0;
while (!fringe.IsEmpty())
{
// Overhead with popping a node.
if (numTicks >= maxTicks)
{
return(results);
}
numTicks += 1;
if (results.Count >= maxResults)
{
return(results);
}
ForwardEntailerSearchProblem.SearchState state = fringe.Pop();
System.Diagnostics.Debug.Assert(state.score > 0.0);
IndexedWord currentWord = topologicalVertices[state.currentIndex];
// Push the case where we don't delete
int nextIndex = state.currentIndex + 1;
int numIters = 0;
while (nextIndex < topologicalVertices.Count)
{
IndexedWord nextWord = topologicalVertices[nextIndex];
System.Diagnostics.Debug.Assert(nextWord.Index() > 0);
if (!state.deletionMask.Get(nextWord.Index() - 1))
{
fringe.Push(new ForwardEntailerSearchProblem.SearchState(state.deletionMask, nextIndex, state.tree, null, state, state.score));
break;
}
else
{
nextIndex += 1;
}
numIters += 1;
if (numIters > 10000)
{
// log.error("logic error (apparent infinite loop); returning");
return(results);
}
}
// Check if we can delete this subtree
bool canDelete = !state.tree.GetFirstRoot().Equals(currentWord);
foreach (SemanticGraphEdge edge in state.tree.IncomingEdgeIterable(currentWord))
{
if ("CD".Equals(edge.GetGovernor().Tag()))
{
canDelete = false;
}
else
{
// Get token information
CoreLabel token = edge.GetDependent().BackingLabel();
OperatorSpec @operator;
NaturalLogicRelation lexicalRelation;
Polarity tokenPolarity = token.Get(typeof(NaturalLogicAnnotations.PolarityAnnotation));
if (tokenPolarity == null)
{
tokenPolarity = Polarity.Default;
}
// Get the relation for this deletion
if ((@operator = token.Get(typeof(NaturalLogicAnnotations.OperatorAnnotation))) != null)
{
lexicalRelation = @operator.instance.deleteRelation;
}
else
{
System.Diagnostics.Debug.Assert(edge.GetDependent().Index() > 0);
lexicalRelation = NaturalLogicRelation.ForDependencyDeletion(edge.GetRelation().ToString(), isSubject.Get(edge.GetDependent().Index() - 1));
}
NaturalLogicRelation projectedRelation = tokenPolarity.ProjectLexicalRelation(lexicalRelation);
// Make sure this is a valid entailment
if (!projectedRelation.ApplyToTruthValue(truthOfPremise).IsTrue())
{
canDelete = false;
}
}
}
if (canDelete)
{
// Register the deletion
Lazy <Pair <SemanticGraph, BitSet> > treeWithDeletionsAndNewMask = Lazy.Of(null);
// Compute the score of the sentence
double newScore = state.score;
foreach (SemanticGraphEdge edge_1 in state.tree.IncomingEdgeIterable(currentWord))
{
double multiplier = weights.DeletionProbability(edge_1, state.tree.OutgoingEdgeIterable(edge_1.GetGovernor()));
System.Diagnostics.Debug.Assert(!double.IsNaN(multiplier));
System.Diagnostics.Debug.Assert(!double.IsInfinite(multiplier));
newScore *= multiplier;
}
// Register the result
if (newScore > 0.0)
{
SemanticGraph resultTree = new SemanticGraph(treeWithDeletionsAndNewMask.Get().first);
andsToAdd.Stream().Filter(null).ForEach(null);
results.Add(new ForwardEntailerSearchProblem.SearchResult(resultTree, AggregateDeletedEdges(state, state.tree.IncomingEdgeIterable(currentWord), determinerRemovals), newScore));
// Push the state with this subtree deleted
nextIndex = state.currentIndex + 1;
numIters = 0;
while (nextIndex < topologicalVertices.Count)
{
IndexedWord nextWord = topologicalVertices[nextIndex];
BitSet newMask = treeWithDeletionsAndNewMask.Get().second;
SemanticGraph treeWithDeletions = treeWithDeletionsAndNewMask.Get().first;
if (!newMask.Get(nextWord.Index() - 1))
{
System.Diagnostics.Debug.Assert(treeWithDeletions.ContainsVertex(topologicalVertices[nextIndex]));
fringe.Push(new ForwardEntailerSearchProblem.SearchState(newMask, nextIndex, treeWithDeletions, null, state, newScore));
break;
}
else
{
nextIndex += 1;
}
numIters += 1;
if (numIters > 10000)
{
// log.error("logic error (apparent infinite loop); returning");
return(results);
}
}
}
}
}
// Return
return(results);
}
/// <summary>Annotate every token for its polarity, based on the operators found.</summary>
/// <remarks>
/// Annotate every token for its polarity, based on the operators found. This function will set the
/// <see cref="PolarityAnnotation"/>
/// for every token.
/// </remarks>
/// <param name="sentence">
/// As in
/// <see cref="DoOneSentence(Edu.Stanford.Nlp.Pipeline.Annotation, Edu.Stanford.Nlp.Util.ICoreMap)"/>
/// </param>
private static void AnnotatePolarity(ICoreMap sentence)
{
// Collect all the operators in this sentence
IList <OperatorSpec> operators = new List <OperatorSpec>();
IList <CoreLabel> tokens = sentence.Get(typeof(CoreAnnotations.TokensAnnotation));
foreach (CoreLabel token in tokens)
{
OperatorSpec specOrNull = token.Get(typeof(NaturalLogicAnnotations.OperatorAnnotation));
if (specOrNull != null)
{
operators.Add(specOrNull);
}
}
// Make sure every node of the dependency tree has a polarity.
// This is separate from the code below in case the tokens in the dependency
// tree don't correspond to the tokens in the sentence. This happens at least
// when the constituency parser craps out on a long sentence, and the
// dependency tree is put together haphazardly.
if (sentence.ContainsKey(typeof(SemanticGraphCoreAnnotations.BasicDependenciesAnnotation)))
{
foreach (IndexedWord token_1 in sentence.Get(typeof(SemanticGraphCoreAnnotations.BasicDependenciesAnnotation)).VertexSet())
{
token_1.Set(typeof(NaturalLogicAnnotations.PolarityAnnotation), Polarity.Default);
}
}
if (sentence.ContainsKey(typeof(SemanticGraphCoreAnnotations.EnhancedDependenciesAnnotation)))
{
foreach (IndexedWord token_1 in sentence.Get(typeof(SemanticGraphCoreAnnotations.EnhancedDependenciesAnnotation)).VertexSet())
{
token_1.Set(typeof(NaturalLogicAnnotations.PolarityAnnotation), Polarity.Default);
}
}
if (sentence.ContainsKey(typeof(SemanticGraphCoreAnnotations.EnhancedPlusPlusDependenciesAnnotation)))
{
foreach (IndexedWord token_1 in sentence.Get(typeof(SemanticGraphCoreAnnotations.EnhancedPlusPlusDependenciesAnnotation)).VertexSet())
{
token_1.Set(typeof(NaturalLogicAnnotations.PolarityAnnotation), Polarity.Default);
}
}
// Set polarity for each token
for (int i = 0; i < tokens.Count; ++i)
{
CoreLabel token_1 = tokens[i];
// Get operators in scope
IList <Triple <int, Monotonicity, MonotonicityType> > inScope = new List <Triple <int, Monotonicity, MonotonicityType> >(4);
foreach (OperatorSpec @operator in operators)
{
if (i >= @operator.subjectBegin && i < @operator.subjectEnd)
{
inScope.Add(Triple.MakeTriple(@operator.subjectEnd - @operator.subjectBegin, @operator.instance.subjMono, @operator.instance.subjType));
}
else
{
if (i >= @operator.objectBegin && i < @operator.objectEnd)
{
inScope.Add(Triple.MakeTriple(@operator.objectEnd - @operator.objectBegin, @operator.instance.objMono, @operator.instance.objType));
}
}
}
// Sort the operators by their scope (approximated by the size of their argument span)
inScope.Sort(null);
// Create polarity
IList <Pair <Monotonicity, MonotonicityType> > info = new List <Pair <Monotonicity, MonotonicityType> >(inScope.Count);
foreach (Triple <int, Monotonicity, MonotonicityType> term in inScope)
{
info.Add(Pair.MakePair(term.second, term.third));
}
Polarity polarity = new Polarity(info);
// Set polarity
token_1.Set(typeof(NaturalLogicAnnotations.PolarityAnnotation), polarity);
}
// Set the PolarityDirectionAnnotation
foreach (CoreLabel token_2 in tokens)
{
Polarity polarity = token_2.Get(typeof(NaturalLogicAnnotations.PolarityAnnotation));
if (polarity != null)
{
if (polarity.IsUpwards())
{
token_2.Set(typeof(NaturalLogicAnnotations.PolarityDirectionAnnotation), "up");
}
else
{
if (polarity.IsDownwards())
{
token_2.Set(typeof(NaturalLogicAnnotations.PolarityDirectionAnnotation), "down");
}
else
{
token_2.Set(typeof(NaturalLogicAnnotations.PolarityDirectionAnnotation), "flat");
}
}
}
}
}