List<Util.Pair<Integer, Integer>> getCandidates(
Document document, int arg2Line, int connStart, int connEnd, int arg1Line) {
List<Util.Pair<Integer, Integer>> candidates = new ArrayList<Util.Pair<Integer, Integer>>();
int distance = 10;
Sentence arg2Sentence = document.getSentence(arg2Line);
// String conn = arg2Sentence.toString(connStart, connEnd).toLowerCase();
// String category = connAnalyzer.getCategory(conn);
int connHeadPos = connAnalyzer.getHeadWord(arg2Sentence.getParseTree(), connStart, connEnd);
SimpleDepGraph arg2DepGraph = document.getDepGraph(arg2Line);
List<Integer> reachable = arg2DepGraph.getReachableIndices(connHeadPos, false, distance);
for (Integer i : reachable) {
if (arg2Sentence.get(i).getTag("POS").matches("VB.*|NNS?|JJ.*|MD")) {
candidates.add(new Util.Pair<Integer, Integer>(arg2Line, i));
}
}
Tree mainHead = headAnalyzer.getCollinsHead(arg2Sentence.getParseTree().getChild(0));
if (mainHead != null) {
int mainHeadPos = treeAnalyzer.getLeafPosition(arg2Sentence.getParseTree(), mainHead);
List<String> pathAsList = arg2DepGraph.getPathAsList(connHeadPos, mainHeadPos, false);
if (pathAsList != null) {
distance = distance - (1 + pathAsList.size());
} else {
// System.out.println("No path from connHead to mainHead!");
distance--;
}
}
// if (arg1Line == arg2Line) return candidates;
for (int i = arg2Line - 1; i >= 0 && distance >= 0; i--) {
Sentence sentence = document.getSentence(i);
SimpleDepGraph depGraph = document.getDepGraph(i);
mainHead = headAnalyzer.getCollinsHead(sentence.getParseTree().getChild(0));
if (mainHead == null) {
distance--;
continue;
}
int mainHeadPos = treeAnalyzer.getLeafPosition(sentence.getParseTree(), mainHead);
reachable = depGraph.getReachableIndices(mainHeadPos, false, distance);
if (reachable == null) {
distance--;
continue;
}
for (Integer j : reachable) {
if (sentence.get(j).getTag("POS").matches("VB.*|NNS?|JJ.*|MD")) {
candidates.add(new Util.Pair<Integer, Integer>(i, j));
}
}
distance -= 2;
}
return candidates;
}
private PropertyList addDependencyFeatures(
PropertyList pl,
Document doc,
Pair<Integer, Integer> candidate,
int arg2Line,
int arg2HeadPos,
int connStart,
int connEnd) {
Sentence arg2Sentence = doc.getSentence(arg2Line);
String conn = arg2Sentence.toString(connStart, connEnd);
String category = connAnalyzer.getCategory(conn.toLowerCase());
int connHeadPos = connAnalyzer.getHeadWord(arg2Sentence.getParseTree(), connStart, connEnd);
int arg1Line = candidate.first();
Tree arg1Tree = doc.getTree(arg1Line);
int arg1HeadPos = candidate.second();
List<String> path = new ArrayList<String>();
if (arg1Line == arg2Line) {
SimpleDepGraph depGraph = doc.getDepGraph(arg1Line);
List<String> tmpPath = depGraph.getPathAsList(connHeadPos, arg1HeadPos, false);
if (tmpPath != null) {
path.addAll(tmpPath);
} else {
path.add("null");
}
} else {
Tree arg2Root = arg2Sentence.getParseTree();
Tree mainHead = headAnalyzer.getCollinsHead(arg2Root.getChild(0));
int mainHeadPos = treeAnalyzer.getLeafPosition(arg2Root, mainHead);
if (mainHeadPos != -1) {
SimpleDepGraph depGraph = doc.getDepGraph(arg2Line);
List<String> tmpPath = depGraph.getPathAsList(connHeadPos, mainHeadPos, false);
if (tmpPath != null) {
path.addAll(tmpPath);
} else {
path.add("null");
}
}
for (int i = 0; i < Math.abs(arg1Line - arg2Line); i++) {
path.add("SENT");
}
Tree arg1Root = arg1Tree;
mainHead = headAnalyzer.getCollinsHead(arg1Root.getChild(0));
mainHeadPos = treeAnalyzer.getLeafPosition(arg1Root, mainHead);
if (mainHeadPos != -1) {
SimpleDepGraph depGraph = doc.getDepGraph(arg1Line);
List<String> tmpPath = depGraph.getPathAsList(mainHeadPos, arg1HeadPos, false);
if (tmpPath != null) {
path.addAll(tmpPath);
} else {
path.add("null");
}
}
}
StringBuilder sbPath = new StringBuilder();
StringBuilder sbPathWithoutCC = new StringBuilder();
StringBuilder sbPathWithoutRep = new StringBuilder();
String prev = "";
for (String node : path) {
sbPath.append(node).append(":");
if (!node.matches("cc|-cc")) {
sbPathWithoutCC.append(node).append(":");
}
if (!node.equals(prev)) {
sbPathWithoutRep.append(node).append(":");
}
prev = node;
}
// M-dependency path
pl = PropertyList.add("M=" + sbPath.toString(), 1.0, pl);
// Q-M&C
pl = PropertyList.add("Q=" + "CONN-" + conn + '&' + "M-" + sbPath.toString(), 1.0, pl);
// T-M&R
pl = PropertyList.add("T=" + "CAT-" + category + '&' + "M-" + sbPath.toString(), 1.0, pl);
// O-collapsed path without cc
pl = PropertyList.add("O=" + sbPathWithoutCC.toString(), 1.0, pl);
// P-collapsed path without repetition
pl = PropertyList.add("P=" + sbPathWithoutRep.toString(), 1.0, pl);
return pl;
}
private PropertyList addConstituentFeatures(
PropertyList pl,
Document doc,
Pair<Integer, Integer> candidate,
int arg2Line,
int arg2HeadPos,
int connStart,
int connEnd) {
Sentence arg2Sentence = doc.getSentence(arg2Line);
String conn = arg2Sentence.toString(connStart, connEnd);
int connHeadPos = connAnalyzer.getHeadWord(arg2Sentence.getParseTree(), connStart, connEnd);
int arg1Line = candidate.first();
Tree arg1Tree = doc.getTree(arg1Line);
int arg1HeadPos = candidate.second();
List<String> path = new ArrayList<String>();
List<String> pathWithoutPOS = new ArrayList<String>();
if (arg1Line == arg2Line) {
Tree root = arg1Tree;
List<Tree> leaves = root.getLeaves();
List<Tree> treePath = root.pathNodeToNode(leaves.get(connHeadPos), leaves.get(arg1HeadPos));
if (treePath != null) {
for (Tree t : treePath) {
if (!t.isLeaf()) {
path.add(t.value());
if (!t.isPreTerminal()) {
pathWithoutPOS.add(t.value());
}
}
}
}
} else {
Tree arg2Root = arg2Sentence.getParseTree();
Tree mainHead = headAnalyzer.getCollinsHead(arg2Root.getChild(0));
List<Tree> leaves = arg2Root.getLeaves();
int mainHeadPos = treeAnalyzer.getLeafPosition(arg2Root, mainHead);
if (mainHeadPos != -1) {
List<Tree> treePath =
arg2Root.pathNodeToNode(leaves.get(connHeadPos), leaves.get(mainHeadPos));
if (treePath != null) {
for (Tree t : treePath) {
if (!t.isLeaf()) {
path.add(t.value());
if (!t.isPreTerminal()) {
pathWithoutPOS.add(t.value());
}
}
}
}
}
for (int i = 0; i < Math.abs(arg1Line - arg2Line); i++) {
path.add("SENT");
pathWithoutPOS.add("SENT");
}
Tree arg1Root = arg1Tree;
mainHead = headAnalyzer.getCollinsHead(arg1Root.getChild(0));
leaves = arg1Root.getLeaves();
mainHeadPos = treeAnalyzer.getLeafPosition(arg1Root, mainHead);
if (mainHeadPos != -1) {
List<Tree> treePath =
arg1Root.pathNodeToNode(leaves.get(mainHeadPos), leaves.get(arg1HeadPos));
if (treePath != null) {
for (Tree t : treePath) {
if (!t.isLeaf()) {
path.add(t.value());
if (!t.isPreTerminal()) {
pathWithoutPOS.add(t.value());
}
}
}
}
}
}
// H-full path
// L-C&H
StringBuilder fullPath = new StringBuilder();
for (String node : path) {
fullPath.append(node).append(":");
}
pl = PropertyList.add("H=" + fullPath.toString(), 1.0, pl);
pl = PropertyList.add("L=CONN-" + conn + "&" + "H-" + fullPath.toString(), 1.0, pl);
// I-length of path
pl = PropertyList.add("I=" + path.size(), 1.0, pl);
// J-collapsed path without part of speech
// K-collapsed path without repititions
fullPath = new StringBuilder();
StringBuilder collapsedPath = new StringBuilder();
String prev = "";
for (String node : pathWithoutPOS) {
fullPath.append(node).append(":");
if (!node.equals(prev)) {
collapsedPath.append(node).append(":");
}
prev = node;
}
pl = PropertyList.add("J=" + fullPath.toString(), 1.0, pl);
pl = PropertyList.add("K=" + collapsedPath.toString(), 1.0, pl);
return pl;
}
private PropertyList addBaselineFeatures(
PropertyList pl,
Document doc,
Pair<Integer, Integer> candidate,
int arg2Line,
int arg2HeadPos,
int connStart,
int connEnd) {
Sentence arg2Sentence = doc.getSentence(arg2Line);
int arg1Line = candidate.first();
Sentence arg1Sentence = doc.getSentence(arg1Line);
int arg1HeadPos = candidate.second();
String conn = arg2Sentence.toString(connStart, connEnd);
// R-connective type
String category = connAnalyzer.getCategory(conn.toLowerCase());
pl = PropertyList.add("R=" + category, 1.0, pl);
// A-position of the connective
String position = "Medial";
if (connStart < 4) position = "Initial";
else if (connEnd >= (arg1Sentence.size() - 3)) position = "Terminal";
pl = PropertyList.add("A=" + position, 1.0, pl);
// S-A & R
pl = PropertyList.add("S=" + position + "&" + category, 1.0, pl);
// C-connective phrase
pl = PropertyList.add("C=" + conn, 1.0, pl);
// D-downcase conn phrase
pl = PropertyList.add("D=" + conn.toLowerCase(), 1.0, pl);
// E-argument head word
pl = PropertyList.add("E=" + arg1Sentence.get(arg1HeadPos).word(), 1.0, pl);
// B-same sentence or not
pl = PropertyList.add("B=" + (arg1Line == arg2Line), 1.0, pl);
// G-A&B
pl = PropertyList.add("A=" + position + "&" + "B=" + (arg1Line == arg2Line), 1.0, pl);
// F-arg1 head prior or after conn
if (arg1Line < arg2Line || arg1HeadPos < connStart) {
pl = PropertyList.add("F=<", 1.0, pl);
} else {
pl = PropertyList.add("F=>", 1.0, pl);
}
// if (1 < 2) return pl;
// Z1-relative position of arg1-conn-arg2
String z = null;
if (arg1Line < arg2Line) {
if (arg2HeadPos < connStart) z = "ARG1-ARG2-CONN";
else z = "ARG1-CONN-ARG2";
} else if (arg1HeadPos < connStart) {
if (arg2HeadPos < arg1HeadPos) z = "ARG2-ARG1-CONN";
else if (arg2HeadPos < connStart) z = "ARG1-ARG2-CONN";
else z = "ARG1-CONN-ARG2";
} else if (arg2HeadPos < connStart) {
z = "ARG2-CONN-ARG1";
} else if (arg2HeadPos < arg1HeadPos) {
z = "CONN-ARG2-ARG1";
} else {
z = "CONN-ARG1-ARG2";
}
pl = PropertyList.add("Z=" + z, 1.0, pl);
// Z2-Conn&Z1
pl = PropertyList.add("CONN=" + conn + "&" + "Z=" + z, 1.0, pl);
return pl;
}
private void addInstancesThroughPipe(
PDTBRelation relation,
Document document,
int arg1Line,
int arg2Line,
InstanceList instanceList) {
// System.out.println("Relation: " + relation.toString());
// System.out.println("arg1Line: " + arg1Line);
// System.out.println("arg2Line: " + arg2Line);
String connectiveGornAddress = relation.getConnectiveGornAddress();
Tree arg2Tree = document.getTree(arg2Line);
List<Tree> connHeadLeaves =
connAnalyzer.getConnHeadLeaves(arg2Tree, connectiveGornAddress, relation.getConnHead());
if (connHeadLeaves.isEmpty()) return;
int connStart = treeAnalyzer.getLeafPosition(arg2Tree, connHeadLeaves.get(0));
int connEnd =
treeAnalyzer.getLeafPosition(arg2Tree, connHeadLeaves.get(connHeadLeaves.size() - 1));
if ((connEnd - connStart) > 4) { // handle if..else, etc.
connEnd = connStart;
}
// consider only the first sentence in case of multi-line argument1
String arg1GornAddress = relation.getArg1GornAddress();
Tree arg1Tree = document.getTree(arg1Line);
List<Tree> arg1GornNodes = getArgGornNodes(arg1Tree, arg1Line, arg1GornAddress);
Tree syntacticHead = headAnalyzer.getSyntacticHead(arg1Tree, arg1GornNodes);
int arg1HeadPos = treeAnalyzer.getLeafPosition(arg1Tree, syntacticHead);
String arg2GornAddress = relation.getArg2GornAddress();
List<Tree> arg2GornNodes = getArgGornNodes(arg2Tree, arg2Line, arg2GornAddress);
Tree arg2SyntacticHead = headAnalyzer.getSyntacticHead(arg2Tree, arg2GornNodes);
int arg2HeadPos = treeAnalyzer.getLeafPosition(arg2Tree, arg2SyntacticHead);
if (arg2HeadPos == -1) {
System.out.println("arg2Head == -1");
return;
}
if (arg1HeadPos == -1) {
System.out.println("arg1Head == -1");
return;
}
int trueCandidate = -1;
List<Pair<Integer, Integer>> candidates =
getCandidates(document, arg2Line, connStart, connEnd, arg1Line);
for (int i = 0; i < candidates.size(); i++) {
Pair<Integer, Integer> candidate = candidates.get(i);
if (candidate.first() == arg1Line && candidate.second() == arg1HeadPos) {
trueCandidate = i;
break;
}
}
if (trueCandidate == -1) {
// trueCandidate = candidates.size();
// candidates.add(new Pair<Integer, Integer>(arg1Line, arg1HeadPos));
// System.out.println("Covered!");
System.out.println("true candidate == -1!!!");
System.out.println(syntacticHead.value());
} else {
int extractArg2 =
ARG2_EXTRACTOR.extractArg2(
document.getSentence(arg2Line),
document.getTree(arg2Line),
document.getDepGraph(arg2Line),
connStart,
connEnd);
if (extractArg2 == -1) {
extractArg2 = 0;
System.out.println("Arg2 == -1!!!!!!!!!!!!!!!!!");
}
// Arg1RankInstance instance = new Arg1RankInstance(document, candidates, arg2Line,
// extractArg2, connStart, connEnd, trueCandidate);
Arg1RankInstance instance =
new Arg1RankInstance(
document, candidates, arg2Line, arg2HeadPos, connStart, connEnd, trueCandidate);
instanceList.addThruPipe(instance);
}
}
/**
* Shows accuracy according to Ben Wellner's definition of accuracy
*
* @param classifier
* @param instanceList
*/
private void showAccuracy(Classifier classifier, InstanceList instanceList) throws IOException {
int total = instanceList.size();
int correct = 0;
HashMap<String, Integer> errorMap = new HashMap<String, Integer>();
FileWriter errorWriter = new FileWriter("arg1Error.log");
for (Instance instance : instanceList) {
Classification classification = classifier.classify(instance);
if (classification.bestLabelIsCorrect()) {
correct++;
} else {
Arg1RankInstance rankInstance = (Arg1RankInstance) instance;
Document doc = rankInstance.getDocument();
Sentence s = doc.getSentence(rankInstance.getArg2Line());
String conn =
s.toString(rankInstance.getConnStart(), rankInstance.getConnEnd()).toLowerCase();
// String category = connAnalyzer.getCategory(conn);
if (errorMap.containsKey(conn)) {
errorMap.put(conn, errorMap.get(conn) + 1);
} else {
errorMap.put(conn, 1);
}
int arg2Line = rankInstance.getArg2Line();
int arg1Line =
rankInstance.getCandidates().get(rankInstance.getTrueArg1Candidate()).first();
int arg1HeadPos =
rankInstance.getCandidates().get(rankInstance.getTrueArg1Candidate()).second();
int predictedCandidateIndex =
Integer.parseInt(classification.getLabeling().getBestLabel().toString());
if (arg1Line == arg2Line) {
errorWriter.write("FileName: " + doc.getFileName() + "\n");
errorWriter.write("Sentential\n");
errorWriter.write("Conn: " + conn + "\n");
errorWriter.write("Arg1Head: " + s.get(arg1HeadPos).word() + "\n");
errorWriter.write(s.toString() + "\n\n");
} else {
errorWriter.write("FileName: " + doc.getFileName() + "\n");
errorWriter.write("Inter-Sentential\n");
errorWriter.write("Arg1 in : " + arg1Line + "\n");
errorWriter.write("Arg2 in : " + arg2Line + "\n");
errorWriter.write("Conn: " + conn + "\n");
errorWriter.write(s.toString() + "\n");
Sentence s1 = doc.getSentence(arg1Line);
errorWriter.write("Arg1Head: " + s1.get(arg1HeadPos) + "\n");
errorWriter.write(s1.toString() + "\n\n");
}
int predictedArg1Line = rankInstance.getCandidates().get(predictedCandidateIndex).first();
int predictedArg1HeadPos =
rankInstance.getCandidates().get(predictedCandidateIndex).second();
Sentence pSentence = doc.getSentence(predictedArg1Line);
errorWriter.write(
"Predicted arg1 sentence: "
+ pSentence.toString()
+ " [Correct: "
+ (predictedArg1Line == arg1Line)
+ "]\n");
errorWriter.write("Predicted head: " + pSentence.get(predictedArg1HeadPos).word() + "\n\n");
}
}
errorWriter.close();
Set<Entry<String, Integer>> entrySet = errorMap.entrySet();
List<Entry<String, Integer>> list = new ArrayList<Entry<String, Integer>>(entrySet);
Collections.sort(
list,
new Comparator<Entry<String, Integer>>() {
@Override
public int compare(Entry<String, Integer> o1, Entry<String, Integer> o2) {
if (o1.getValue() > o2.getValue()) return -1;
else if (o1.getValue() < o2.getValue()) return 1;
return 0;
}
});
for (Entry<String, Integer> item : list) {
System.out.println(item.getKey() + "-" + item.getValue());
}
System.out.println("Total: " + total);
System.out.println("Correct: " + correct);
System.out.println("Accuracy: " + (1.0 * correct) / total);
}