List<List<CoreLabel>> allWords = new ArrayList<List<CoreLabel>>();
List<Tree> allTrees = new ArrayList<Tree>();
List<List<Mention>> allGoldMentions = new ArrayList<List<Mention>>();
List<List<Mention>> allPredictedMentions;
List<CoreMap> allSentences = new ArrayList<CoreMap>();
Annotation docAnno = new Annotation("");
Pattern docPattern = Pattern.compile("<DOC>(.*?)</DOC>", Pattern.DOTALL+Pattern.CASE_INSENSITIVE);
Pattern sentencePattern = Pattern.compile("(<s>|<hl>|<dd>|<DATELINE>)(.*?)(</s>|</hl>|</dd>|</DATELINE>)", Pattern.DOTALL+Pattern.CASE_INSENSITIVE);
Matcher docMatcher = docPattern.matcher(fileContents);
if (! docMatcher.find(currentOffset)) return null;
currentOffset = docMatcher.end();
String doc = docMatcher.group(1);
Matcher sentenceMatcher = sentencePattern.matcher(doc);
String ner = null;
//Maintain current document ID.
Pattern docIDPattern = Pattern.compile("<DOCNO>(.*?)</DOCNO>", Pattern.DOTALL+Pattern.CASE_INSENSITIVE);
Matcher docIDMatcher = docIDPattern.matcher(doc);
if(docIDMatcher.find()) currentDocumentID = docIDMatcher.group(1);
else currentDocumentID = "documentAfter " + currentDocumentID;
while (sentenceMatcher.find()) {
String sentenceString = sentenceMatcher.group(2);
List<CoreLabel> words = tokenizerFactory.getTokenizer(new StringReader(sentenceString)).tokenize();
// FIXING TOKENIZATION PROBLEMS
for (int i = 0; i < words.size(); i++) {
CoreLabel w = words.get(i);
if (i > 0 && w.word().equals("$")) {
if(!words.get(i-1).word().endsWith("PRP") && !words.get(i-1).word().endsWith("WP"))
continue;
words.get(i-1).set(CoreAnnotations.TextAnnotation.class, words.get(i-1).word()+"$");
words.remove(i);
i--;
} else if (w.word().equals("\\/")) {
if(words.get(i-1).word().equals("</COREF>"))
continue;
w.set(CoreAnnotations.TextAnnotation.class, words.get(i-1).word()+"\\/"+words.get(i+1).word());
words.remove(i+1);
words.remove(i-1);
}
}
// END FIXING TOKENIZATION PROBLEMS
List<CoreLabel> sentence = new ArrayList<CoreLabel>();
// MUC accepts embedded coref mentions, so we need to keep a stack for the mentions currently open
Stack<Mention> stack = new Stack<Mention>();
List<Mention> mentions = new ArrayList<Mention>();
allWords.add(sentence);
allGoldMentions.add(mentions);
for (CoreLabel word : words) {
String w = word.get(CoreAnnotations.TextAnnotation.class);
// found regular token: WORD/POS
if (!w.startsWith("<") && w.contains("\\/") && w.lastIndexOf("\\/") != w.length()-2) {
int i = w.lastIndexOf("\\/");
String w1 = w.substring(0, i);
// we do NOT set POS info here. We take the POS tags from the parser!
word.set(CoreAnnotations.TextAnnotation.class, w1);
word.remove(CoreAnnotations.OriginalTextAnnotation.class);
if(Constants.USE_GOLD_NE) {
if (ner != null) {
word.set(CoreAnnotations.NamedEntityTagAnnotation.class, ner);
} else {
word.set(CoreAnnotations.NamedEntityTagAnnotation.class, "O");
}
}
sentence.add(word);
}
// found the start SGML tag for a NE, e.g., "<ORGANIZATION>"
else if (w.startsWith("<") && !w.startsWith("<COREF") && !w.startsWith("</")) {
Pattern nerPattern = Pattern.compile("<(.*?)>");
Matcher m = nerPattern.matcher(w);
m.find();
ner = m.group(1);
}
// found the end SGML tag for a NE, e.g., "</ORGANIZATION>"
else if (w.startsWith("</") && !w.startsWith("</COREF")) {
Pattern nerPattern = Pattern.compile("</(.*?)>");
Matcher m = nerPattern.matcher(w);
m.find();
String ner1 = m.group(1);
if (ner != null && !ner.equals(ner1)) throw new RuntimeException("Unmatched NE labels in MUC file: " + ner + " v. " + ner1);
ner = null;
}
// found the start SGML tag for a coref mention
else if (w.startsWith("<COREF")) {
Mention mention = new Mention();
// position of this mention in the sentence
mention.startIndex = sentence.size();
// extract GOLD info about this coref chain. needed for eval
Pattern idPattern = Pattern.compile("ID=\"(.*?)\"");
Pattern refPattern = Pattern.compile("REF=\"(.*?)\"");
Matcher m = idPattern.matcher(w);
m.find();
mention.mentionID = Integer.parseInt(m.group(1));
m = refPattern.matcher(w);
if (m.find()) {
mention.originalRef = Integer.parseInt(m.group(1));
}
// open mention. keep track of all open mentions using the stack
stack.push(mention);
}
// found the end SGML tag for a coref mention
else if (w.equals("</COREF>")) {
Mention mention = stack.pop();
mention.endIndex = sentence.size();
// this is a closed mention. add it to the final list of mentions
// System.err.printf("Found MENTION: ID=%d, REF=%d\n", mention.mentionID, mention.originalRef);
mentions.add(mention);
} else {
word.remove(CoreAnnotations.OriginalTextAnnotation.class);
if(Constants.USE_GOLD_NE){
if (ner != null) {
word.set(CoreAnnotations.NamedEntityTagAnnotation.class, ner);
} else {
word.set(CoreAnnotations.NamedEntityTagAnnotation.class, "O");
}
}
sentence.add(word);
}
}
StringBuilder textContent = new StringBuilder();
for (int i=0 ; i<sentence.size(); i++){
CoreLabel w = sentence.get(i);
w.set(CoreAnnotations.IndexAnnotation.class, i+1);
w.set(CoreAnnotations.UtteranceAnnotation.class, 0);
if(i>0) textContent.append(" ");
textContent.append(w.getString(CoreAnnotations.TextAnnotation.class));
}
CoreMap sentCoreMap = new Annotation(textContent.toString());
allSentences.add(sentCoreMap);
sentCoreMap.set(CoreAnnotations.TokensAnnotation.class, sentence);
}
// assign goldCorefClusterID
Map<Integer, Mention> idMention = Generics.newHashMap(); // temporary use
for (List<Mention> goldMentions : allGoldMentions) {