public virtual IList <Pair <int, int> > GetClusterMerges(ClustererDataLoader.ClustererDoc doc)
{
IList <Pair <int, int> > merges = new List <Pair <int, int> >();
Clusterer.State currentState = new Clusterer.State(doc);
while (!currentState.IsComplete())
{
Pair <int, int> currentPair = currentState.mentionPairs[currentState.currentIndex];
if (currentState.DoBestAction(classifier))
{
merges.Add(currentPair);
}
}
return(merges);
}
public virtual Pair <Clusterer.CandidateAction, Clusterer.CandidateAction> GetActions(Clusterer.ClustererClassifier classifier)
{
ICounter <string> mergeFeatures = GetFeatures(doc, c1, c2, globalFeatures[currentIndex]);
double mergeScore = Math.Exp(classifier.WeightFeatureProduct(mergeFeatures));
hashedScores[new Clusterer.MergeKey(c1, c2, currentIndex)] = mergeScore > 0.5;
Clusterer.State merge = new Clusterer.State(this);
merge.DoAction(true);
double mergeB3 = merge.GetFinalCost(classifier);
Clusterer.State noMerge = new Clusterer.State(this);
noMerge.DoAction(false);
double noMergeB3 = noMerge.GetFinalCost(classifier);
double weight = doc.mentions.Count / 100.0;
double maxB3 = Math.Max(mergeB3, noMergeB3);
return(new Pair <Clusterer.CandidateAction, Clusterer.CandidateAction>(new Clusterer.CandidateAction(mergeFeatures, weight * (maxB3 - mergeB3)), new Clusterer.CandidateAction(new ClassicCounter <string>(), weight * (maxB3 - noMergeB3))));
}
private double EvaluatePolicy(IList <ClustererDataLoader.ClustererDoc> docs, bool training)
{
isTraining = 0;
EvalUtils.B3Evaluator evaluator = new EvalUtils.B3Evaluator();
foreach (ClustererDataLoader.ClustererDoc doc in docs)
{
Clusterer.State currentState = new Clusterer.State(doc);
while (!currentState.IsComplete())
{
currentState.DoBestAction(classifier);
}
currentState.UpdateEvaluator(evaluator);
}
isTraining = 1;
double score = evaluator.GetF1();
Redwood.Log("scoref.train", string.Format("B3 F1 score on %s: %.4f", training ? "train" : "validate", score));
return(score);
}
private IList <Pair <Clusterer.CandidateAction, Clusterer.CandidateAction> > RunPolicy(ClustererDataLoader.ClustererDoc doc, double beta)
{
IList <Pair <Clusterer.CandidateAction, Clusterer.CandidateAction> > examples = new List <Pair <Clusterer.CandidateAction, Clusterer.CandidateAction> >();
Clusterer.State currentState = new Clusterer.State(doc);
while (!currentState.IsComplete())
{
Pair <Clusterer.CandidateAction, Clusterer.CandidateAction> actions = currentState.GetActions(classifier);
if (actions == null)
{
continue;
}
examples.Add(actions);
bool useExpert = random.NextDouble() < beta;
double action1Score = useExpert ? -actions.first.cost : classifier.WeightFeatureProduct(actions.first.features);
double action2Score = useExpert ? -actions.second.cost : classifier.WeightFeatureProduct(actions.second.features);
currentState.DoAction(action1Score >= action2Score);
}
return(examples);
}
public State(Clusterer.State state)
{
this.hashedScores = state.hashedScores;
this.hashedCosts = state.hashedCosts;
this.doc = state.doc;
this.hash = state.hash;
this.mentionPairs = state.mentionPairs;
this.currentIndex = state.currentIndex;
this.globalFeatures = state.globalFeatures;
this.clusters = new List <Clusterer.Cluster>();
this.mentionToCluster = new Dictionary <int, Clusterer.Cluster>();
foreach (Clusterer.Cluster c in state.clusters)
{
Clusterer.Cluster copy = new Clusterer.Cluster(c);
clusters.Add(copy);
foreach (int m in copy.mentions)
{
mentionToCluster[m] = copy;
}
}
SetClusters();
}
