public void discoverSpecificComments() {
System.out.println("topic similarity");
String fileName = "./data/results/0108_9/topicSimilarity.txt";
try {
PrintWriter pw = new PrintWriter(new File(fileName));
for (_Doc doc : m_trainSet) {
if (doc instanceof _ParentDoc2) {
pw.print(doc.getName() + "\t");
double stnTopicSimilarity = 0.0;
double docTopicSimilarity = 0.0;
for (_ChildDoc2 cDoc : ((_ParentDoc2) doc).m_childDocs) {
pw.print(cDoc.getName() + ":");
docTopicSimilarity = computeSimilarity(((_ParentDoc2) doc).m_topics, cDoc.m_topics);
pw.print(docTopicSimilarity);
for (int i = 0; i < ((_ParentDoc2) doc).m_sentenceMap.size(); i++) {
_Stn stnObj = ((_ParentDoc2) doc).m_sentenceMap.get(i);
if (stnObj.m_stnLength == 0) {
// // some sentences are normalized into zero
// length sentences
pw.print(":0");
continue;
}
double[] stnTopics = stnObj.m_topics;
stnTopicSimilarity = computeSimilarity(stnTopics, cDoc.m_topics);
pw.print(":" + stnTopicSimilarity);
}
pw.print("\t");
}
pw.println();
} else {
continue;
}
}
pw.flush();
pw.close();
} catch (FileNotFoundException e) {
// TODO Auto-generated catch block
e.printStackTrace();
}
}
public double calChildLogLikelihood2(_ChildDoc2 cDoc) {
double likelihood = 0.0;
int tid = 0;
int wid = 0;
double term1 = 0.0;
double term2 = 0.0;
for (int n = 0; n < cDoc.getTotalDocLength(); n++) {
for (int k = 0; k < number_of_topics; k++) {
wid = cDoc.m_words[n];
// tid = cDoc.m_topicAssignment[n];
if (cDoc.getTotalDocLength() == 0) return 0;
term1 =
(m_parentWordTopicSstat[k][wid] + m_childWordTopicSstat[k][wid])
/ (double) (m_parentSstat[k] + m_childSstat[k]);
term2 =
(cDoc.m_xTopicSstat[0][k] + cDoc.m_xTopicSstat[1][k])
/ (double) cDoc.getTotalDocLength();
if ((term1 == 0)) {
double eps = 1e-9;
term1 += eps;
}
if ((term2 == 0)) {
double eps = 1e-9;
term2 += eps;
}
likelihood += Math.log(term1) + Math.log(term2);
}
}
return likelihood;
}
public void printChildTopicAssignment(_ChildDoc2 d) {
String topicAssignmentfile =
"./data/results/0108_9/childTopicAssignment/topicAssignment_" + d.getName() + "_.txt";
try {
PrintWriter pw = new PrintWriter(new File(topicAssignmentfile));
for (int i = 0; i < d.m_index.length; i++) {
int index = d.m_index[i];
String featureName = m_corpus.getFeature(index);
int positionInDoc = d.m_positionInDoc[i];
int topic = d.m_topicAssignment[positionInDoc];
if (index == d.m_words[positionInDoc]) pw.print(featureName + ":" + topic + "\t");
}
pw.flush();
pw.close();
} catch (FileNotFoundException e) {
// TODO Auto-generated catch block
e.printStackTrace();
}
}
void sampleChildDocTopic(_ChildDoc2 d) {
int wid, tid, xid;
double[][] xTopicProb = new double[2][number_of_topics];
double prob;
double normalizedProb = 0;
for (int i = 0; i < d.m_words.length; i++) {
int samplingX = 0;
int samplingTopic = 0;
prob = 0;
normalizedProb = 0;
wid = d.m_words[i];
tid = d.m_topicAssignment[i];
xid = d.m_xIndicator[i];
d.m_xTopicSstat[xid][tid]--;
d.m_xSstat[xid]--;
if (m_collectCorpusStats) {
m_childWordTopicSstat[tid][wid]--;
m_childSstat[tid]--;
}
// p(z=tid,x=1) from specific
for (tid = 0; tid < number_of_topics; tid++) {
double term1 =
(d_beta + m_parentWordTopicSstat[tid][wid] + m_childWordTopicSstat[tid][wid])
/ (d_beta * vocabulary_size + m_parentSstat[tid] + m_childSstat[tid]);
double term2 =
(d.m_xTopicSstat[1][tid] + d_alpha) / (number_of_topics * d_alpha + d.m_xSstat[1]);
// double term3 =
// (m_gamma[1]+d.m_xSstat[0])/(m_gamma[1]+m_gamma[2]+d.m_xSstat[0]+d.m_xSstat[1]);
double term3 = (m_gamma[1] + d.m_xSstat[1]);
xTopicProb[1][tid] = term1 * term2 * term3;
normalizedProb += xTopicProb[1][tid];
}
if (d.m_parentDoc2 == null) {
System.out.println("null parent in child doc" + d.getName());
}
double parentDocLen = d.m_parentDoc2.getTotalDocLength();
// double parentDocLen = d.m_parentDoc2.getTotalDocLength();
// p(z=tid, x=0) from background
for (tid = 0; tid < number_of_topics; tid++) {
double term1 =
(d_beta + m_parentWordTopicSstat[tid][wid] + m_childWordTopicSstat[tid][wid])
/ (d_beta * vocabulary_size + m_parentSstat[tid] + m_childSstat[tid]);
double term2 =
(d_alpha + m_mu * d.m_parentDoc2.m_sstat[tid] / parentDocLen + d.m_xTopicSstat[0][tid])
/ (number_of_topics * d_alpha + m_mu + d.m_xSstat[0]);
double term3 = (m_gamma[0] + d.m_xSstat[0]);
xTopicProb[0][tid] = term1 * term2 * term3;
normalizedProb += xTopicProb[0][tid];
}
boolean finishLoop = false;
prob = normalizedProb * m_rand.nextDouble();
for (xid = 0; xid < m_gamma.length; xid++) {
for (tid = 0; tid < number_of_topics; tid++) {
prob -= xTopicProb[xid][tid];
if (prob <= 0) {
finishLoop = true;
break;
}
}
if (finishLoop) {
break;
}
}
if (xid == 2) xid--;
if (tid == number_of_topics) tid--;
samplingX = xid;
samplingTopic = tid;
d.m_topicAssignment[i] = samplingTopic;
d.m_xIndicator[i] = samplingX;
d.m_xTopicSstat[samplingX][samplingTopic]++;
d.m_xSstat[samplingX]++;
if (m_collectCorpusStats) {
m_childWordTopicSstat[samplingTopic][wid]++;
m_childSstat[samplingTopic]++;
}
}
}
