public State(ClustererDataLoader.ClustererDoc doc)
{
currentDocId = doc.id;
this.doc = doc;
this.hashedScores = new Dictionary <Clusterer.MergeKey, bool>();
this.hashedCosts = new Dictionary <long, double>();
this.clusters = new List <Clusterer.Cluster>();
this.hash = 0;
mentionToCluster = new Dictionary <int, Clusterer.Cluster>();
foreach (int m in doc.mentions)
{
Clusterer.Cluster c = new Clusterer.Cluster(m);
clusters.Add(c);
mentionToCluster[m] = c;
hash ^= c.hash * 7;
}
IList <Pair <int, int> > allPairs = new List <Pair <int, int> >(doc.classificationScores.KeySet());
ICounter <Pair <int, int> > scores = UseRanking ? doc.rankingScores : doc.classificationScores;
allPairs.Sort(null);
int i = 0;
for (i = 0; i < allPairs.Count; i++)
{
double score = scores.GetCount(allPairs[i]);
if (score < MinPairwiseScore && i > MinPairs)
{
break;
}
if (i >= EarlyStopThreshold && i / score > EarlyStopVal)
{
break;
}
}
mentionPairs = allPairs.SubList(0, i);
ICounter <int> seenAnaphors = new ClassicCounter <int>();
ICounter <int> seenAntecedents = new ClassicCounter <int>();
globalFeatures = new List <Clusterer.GlobalFeatures>();
for (int j = 0; j < allPairs.Count; j++)
{
Pair <int, int> mentionPair = allPairs[j];
Clusterer.GlobalFeatures gf = new Clusterer.GlobalFeatures();
gf.currentIndex = j;
gf.anaphorSeen = seenAnaphors.ContainsKey(mentionPair.second);
gf.size = mentionPairs.Count;
gf.docSize = doc.mentions.Count / 300.0;
globalFeatures.Add(gf);
seenAnaphors.IncrementCount(mentionPair.second);
seenAntecedents.IncrementCount(mentionPair.first);
}
currentIndex = 0;
SetClusters();
}
private static ICounter <string> GetFeatures(ClustererDataLoader.ClustererDoc doc, Clusterer.Cluster c1, Clusterer.Cluster c2, Clusterer.GlobalFeatures gf)
{
Clusterer.MergeKey key = new Clusterer.MergeKey(c1, c2, gf.currentIndex);
CompressedFeatureVector cfv = featuresCache[key];
ICounter <string> features = cfv == null ? null : compressor.Uncompress(cfv);
if (features != null)
{
featuresCacheHits += isTraining;
return(features);
}
featuresCacheMisses += isTraining;
features = new ClassicCounter <string>();
if (gf.anaphorSeen)
{
features.IncrementCount("anaphorSeen");
}
features.IncrementCount("docSize", gf.docSize);
features.IncrementCount("percentComplete", gf.currentIndex / (double)gf.size);
features.IncrementCount("bias", 1.0);
int earliest1 = EarliestMention(c1, doc);
int earliest2 = EarliestMention(c2, doc);
if (doc.mentionIndices[earliest1] > doc.mentionIndices[earliest2])
{
int tmp = earliest1;
earliest1 = earliest2;
earliest2 = tmp;
}
features.IncrementCount("anaphoricity", doc.anaphoricityScores.GetCount(earliest2));
if (c1.mentions.Count == 1 && c2.mentions.Count == 1)
{
Pair <int, int> mentionPair = new Pair <int, int>(c1.mentions[0], c2.mentions[0]);
features.AddAll(AddSuffix(GetFeatures(doc, mentionPair, doc.classificationScores), "-classification"));
features.AddAll(AddSuffix(GetFeatures(doc, mentionPair, doc.rankingScores), "-ranking"));
features = AddSuffix(features, "-single");
}
else
{
IList <Pair <int, int> > between = new List <Pair <int, int> >();
foreach (int m1 in c1.mentions)
{
foreach (int m2 in c2.mentions)
{
between.Add(new Pair <int, int>(m1, m2));
}
}
features.AddAll(AddSuffix(GetFeatures(doc, between, doc.classificationScores), "-classification"));
features.AddAll(AddSuffix(GetFeatures(doc, between, doc.rankingScores), "-ranking"));
}
featuresCache[key] = compressor.Compress(features);
return(features);
}