/** Mark twin mentions: All mention boundaries should be matched */
private void findTwinMentionsStrict() {
for (int sentNum = 0; sentNum < goldOrderedMentionsBySentence.size(); sentNum++) {
List<Mention> golds = goldOrderedMentionsBySentence.get(sentNum);
List<Mention> predicts = predictedOrderedMentionsBySentence.get(sentNum);
// For CoNLL training there are some documents with gold mentions with the same position
// offsets
// See
// /scr/nlp/data/conll-2011/v2/data/train/data/english/annotations/nw/wsj/09/wsj_0990.v2_auto_conll
// (Packwood - Roth)
CollectionValuedMap<IntPair, Mention> goldMentionPositions =
new CollectionValuedMap<IntPair, Mention>();
for (Mention g : golds) {
IntPair ip = new IntPair(g.startIndex, g.endIndex);
if (goldMentionPositions.containsKey(ip)) {
StringBuilder existingMentions = new StringBuilder();
for (Mention eg : goldMentionPositions.get(ip)) {
if (existingMentions.length() > 0) {
existingMentions.append(",");
}
existingMentions.append(eg.mentionID);
}
SieveCoreferenceSystem.logger.warning(
"WARNING: gold mentions with the same offsets: "
+ ip
+ " mentions="
+ g.mentionID
+ ","
+ existingMentions
+ ", "
+ g.spanToString());
}
// assert(!goldMentionPositions.containsKey(ip));
goldMentionPositions.add(new IntPair(g.startIndex, g.endIndex), g);
}
for (Mention p : predicts) {
IntPair pos = new IntPair(p.startIndex, p.endIndex);
if (goldMentionPositions.containsKey(pos)) {
Collection<Mention> cm = goldMentionPositions.get(pos);
Mention g = cm.iterator().next();
cm.remove(g);
p.mentionID = g.mentionID;
p.twinless = false;
g.twinless = false;
}
}
// temp: for making easy to recognize twinless mention
for (Mention p : predicts) {
if (p.twinless) p.mentionID += 10000;
}
}
}
/** Mark twin mentions: heads of the mentions are matched */
private void findTwinMentionsRelaxed() {
for (int sentNum = 0; sentNum < goldOrderedMentionsBySentence.size(); sentNum++) {
List<Mention> golds = goldOrderedMentionsBySentence.get(sentNum);
List<Mention> predicts = predictedOrderedMentionsBySentence.get(sentNum);
Map<IntPair, Mention> goldMentionPositions = Generics.newHashMap();
Map<Integer, LinkedList<Mention>> goldMentionHeadPositions = Generics.newHashMap();
for (Mention g : golds) {
goldMentionPositions.put(new IntPair(g.startIndex, g.endIndex), g);
if (!goldMentionHeadPositions.containsKey(g.headIndex)) {
goldMentionHeadPositions.put(g.headIndex, new LinkedList<Mention>());
}
goldMentionHeadPositions.get(g.headIndex).add(g);
}
List<Mention> remains = new ArrayList<Mention>();
for (Mention p : predicts) {
IntPair pos = new IntPair(p.startIndex, p.endIndex);
if (goldMentionPositions.containsKey(pos)) {
Mention g = goldMentionPositions.get(pos);
p.mentionID = g.mentionID;
p.twinless = false;
g.twinless = false;
goldMentionHeadPositions.get(g.headIndex).remove(g);
if (goldMentionHeadPositions.get(g.headIndex).isEmpty()) {
goldMentionHeadPositions.remove(g.headIndex);
}
} else remains.add(p);
}
for (Mention r : remains) {
if (goldMentionHeadPositions.containsKey(r.headIndex)) {
Mention g = goldMentionHeadPositions.get(r.headIndex).poll();
r.mentionID = g.mentionID;
r.twinless = false;
g.twinless = false;
if (goldMentionHeadPositions.get(g.headIndex).isEmpty()) {
goldMentionHeadPositions.remove(g.headIndex);
}
}
}
}
}