private void balanceSentiment(
String file1,
String fieldDelim1,
int[] fields1,
String file2,
String fieldDelim2,
int[] fields2,
boolean append)
throws IOException {
Map<Integer, Integer> sentiFreq = TwitterCorpusStat.sentimentFreq(file1, fieldDelim1, fields1);
int[] freq = new int[3];
freq[0] = sentiFreq.get(OMTweet.POLARITY_POSITIVE);
freq[1] = sentiFreq.get(OMTweet.POLARITY_NEGATIVE);
freq[2] = sentiFreq.get(OMTweet.POLARITY_NEUTRAL);
int[][] indices = new int[3][];
for (int i = 0; i < 3; i++) {
indices[i] = null;
}
int sbjDiff;
if ((sbjDiff = Math.min(freq[0], freq[1]) * 2) > freq[2]) {
sbjDiff = (sbjDiff - freq[3]) / 2;
} else {
sbjDiff = 0;
}
if (freq[0] != freq[1] || sbjDiff > 0) {
int senti1 = 0;
int senti2 = 1;
if (freq[0] < freq[1]) {
senti1 = 1;
senti2 = 0;
}
indices[senti1] =
randomIndices(freq[senti1], freq[senti2] - sbjDiff, System.currentTimeMillis());
freq[senti1] = freq[senti2] - sbjDiff;
if (sbjDiff > 0) {
indices[senti2] =
randomIndices(freq[senti2], freq[senti2] - sbjDiff, System.currentTimeMillis());
freq[senti2] = freq[senti2] - sbjDiff;
}
}
int sbj = freq[0] + freq[1];
if (sbj < freq[2]) {
indices[2] = randomIndices(freq[2], sbj, System.currentTimeMillis());
freq[2] = sbj;
} else if (sbj > freq[2]) {
throw new IllegalStateException();
}
int[] idx = new int[3];
int[] cursor = new int[3];
for (int i = 0; i < 3; i++) {
idx[i] = 0;
cursor[i] = 0;
}
OMTwitterCorpusFileReader reader = new OMTwitterCorpusFileReader(file1, fieldDelim1, fields1);
OMTwitterCorpusFileWriter writer =
new OMTwitterCorpusFileWriter(file2, fieldDelim2, fields2, append);
int senti = 0;
while (reader.hasNext()) {
OMTweet tweet = reader.next();
switch (tweet.getPolarity()) {
case OMTweet.POLARITY_POSITIVE:
senti = 0;
break;
case OMTweet.POLARITY_NEGATIVE:
senti = 1;
break;
case OMTweet.POLARITY_NEUTRAL:
senti = 2;
break;
}
if (indices[senti] == null) {
writer.write(tweet);
} else if (cursor[senti] < indices[senti].length
&& indices[senti][cursor[senti]] == idx[senti]) {
writer.write(tweet);
cursor[senti]++;
}
idx[senti]++;
}
writer.close();
reader.close();
}
/* (non-Javadoc)
* @see org.apache.uima.collection.base_cpm.CasObjectProcessor#processCas(org.apache.uima.cas.CAS)
*/
public void processCas(CAS aCAS) throws ResourceProcessException {
JCas jcas;
try {
jcas = aCAS.getJCas();
} catch (CASException e) {
logger.log(Level.SEVERE, e.getMessage());
throw new ResourceProcessException(e);
}
TweetAnnotation tweetAnn =
(TweetAnnotation) jcas.getAnnotationIndex(TweetAnnotation.type).iterator().next();
OMTweet answerTweet = evalCorpusReader.next();
if (!answerTweet.getId().equals(tweetAnn.getId())) {
logger.log(
Level.SEVERE,
"target corpus and evaluation corpus don't match to each other - "
+ answerTweet.getId()
+ ", "
+ tweetAnn.getId());
throw new ResourceProcessException();
}
String[] entity = extractEntityTags(answerTweet.getText());
String classified = null;
String prevClassified = null;
StringBuffer sb = new StringBuffer();
try {
sb.append("\n[");
sb.append(answerTweet.getPolarityString());
sb.append("=>");
sb.append(tweetAnn.getPolarity());
sb.append("] ");
sb.append(tweetAnn.getCoveredText());
sb.append('\n');
FSIterator<Annotation> tokenAnnIter =
jcas.getAnnotationIndex(TokenAnnotation.type).iterator();
TokenAnnotation tokenAnn = null;
int i = 0;
int prevClassifiedIdx = labelNoneIdx;
int prevAnswerIdx = labelNoneIdx;
String classifiedEntityStr = "";
String answerEntityStr = "";
while (tokenAnnIter.hasNext()) {
tokenAnn = (TokenAnnotation) tokenAnnIter.next();
classified = tokenAnn.getEntityLabel();
String answer = entity[i];
boolean correct = false;
if (classified.equals(answer)) {
correct = true;
}
int classifiedIdx = 0;
int answerIdx = 0;
try {
answerIdx = map.get(answer);
} catch (Exception e) {
logger.log(
Level.SEVERE,
"wrong annotation on the evaluation corpus - tweet id: "
+ answerTweet.getId()
+ ", answerTag="
+ answer);
logger.log(Level.SEVERE, e.getMessage());
answerIdx = map.get(labelNone);
}
try {
classifiedIdx = map.get(classified);
} catch (Exception e) {
logger.log(
Level.SEVERE,
"wrong annotation from the NER - tweet id: "
+ answerTweet.getId()
+ ", classifiedTag="
+ classified);
logger.log(Level.SEVERE, e.getMessage());
classifiedIdx = map.get(labelNone);
}
stat[classifiedIdx][0]++;
stat[answerIdx][1]++;
if (correct) {
stat[classifiedIdx][2]++;
}
if (classifiedIdx != labelNoneIdx) {
if (classifiedIdx / 3 != prevClassifiedIdx / 3) {
classifiedEntityCnt[classifiedIdx / 3]++;
if (prevClassifiedIdx != labelNoneIdx) {
sb.append('\t');
sb.append(classifiedEntityStr);
sb.append(" -> ");
sb.append(prevClassified.substring(0, prevClassified.lastIndexOf('_')));
sb.append('\n');
}
classifiedEntityStr = tokenAnn.getCoveredText();
} else {
classifiedEntityStr += " " + tokenAnn.getCoveredText();
}
} else if (prevClassifiedIdx != labelNoneIdx) {
sb.append('\t');
sb.append(classifiedEntityStr);
sb.append(" -> ");
sb.append(prevClassified.substring(0, prevClassified.lastIndexOf('_')));
sb.append('\n');
classifiedEntityStr = "";
}
prevClassifiedIdx = classifiedIdx;
if (answerIdx != labelNoneIdx) {
if (answerIdx / 3 != prevAnswerIdx / 3) {
answerEntityCnt[answerIdx / 3]++;
answerEntityStr = tokenAnn.getCoveredText();
} else {
answerEntityStr += " " + tokenAnn.getCoveredText();
}
} else if (prevAnswerIdx != labelNoneIdx) {
answerEntityStr = "";
}
prevAnswerIdx = answerIdx;
prevClassified = classified;
i++;
}
if (prevClassifiedIdx != labelNoneIdx) {
sb.append('\t');
sb.append(classifiedEntityStr);
sb.append(" -> ");
sb.append(prevClassified.substring(0, prevClassified.lastIndexOf('_')));
sb.append('\n');
}
// senti
String answerSenti = answerTweet.getPolarityString();
boolean correct = false;
String classifiedSenti = tweetAnn.getPolarity();
if (classifiedSenti.equals(senti)) {
correct = true;
}
int classifiedIdx = sentiIdx(classifiedSenti);
int answerIdx = sentiIdx(answerSenti);
senti[classifiedIdx][0]++;
senti[answerIdx][1]++;
if (classifiedIdx == answerIdx) {
correct = true;
}
if (correct) {
senti[classifiedIdx][2]++;
}
cnt++;
logger.log(Level.INFO, sb.toString());
} catch (CASRuntimeException e) {
throw new ResourceProcessException(e);
}
}