Examples with structures._ParentDoc4DCM structures._ParentDoc4DCM used on opensource projects

Example 6 with structures._ParentDoc4DCM

use of structures._ParentDoc4DCM in project IR_Base by Linda-sunshine.

the class weightedCorrespondenceModel method updateEta4Child.

public void updateEta4Child(_ParentDoc4DCM pDoc) {
    for (_ChildDoc cDoc : pDoc.m_childDocs) {
        _SparseFeature[] fvs = cDoc.getSparse();
        for (int n = 0; n < fvs.length; n++) {
            int wId = fvs[n].getIndex();
            double wVal = fvs[n].getValue();
            for (int k = 0; k < number_of_topics; k++) {
                cDoc.m_phi[n][k] = Utils.digamma(pDoc.m_sstat[k]) + Utils.digamma(cDoc.m_sstat[k]);
                cDoc.m_phi[n][k] += Utils.digamma(pDoc.m_lambda_stat[k][wId]) - Utils.digamma(pDoc.m_lambda_topicStat[k]);
            }
            double logSum = logSumOfExponentials(cDoc.m_phi[n]);
            if (Double.isInfinite(logSum)) {
                System.out.println("infinite");
                System.out.println("this doc\t" + cDoc.getName() + "\t" + "this word has a total biased probability assignment\t" + m_corpus.getFeature(wId));
            }
            if (Double.isNaN(logSum)) {
                System.out.println("nan");
                for (int k = 0; k < number_of_topics; k++) System.out.println("cDoc.m_phi\t" + cDoc.m_phi[n][k]);
            }
            double phiSum = 0;
            for (int k = 0; k < number_of_topics; k++) {
                if ((cDoc.m_phi[n][k] - logSum) < -200) {
                    cDoc.m_phi[n][k] = 1e-20;
                } else {
                    cDoc.m_phi[n][k] = Math.exp(cDoc.m_phi[n][k] - logSum);
                }
                phiSum += cDoc.m_phi[n][k];
            }
            if (Math.abs(phiSum - 1) > 1) {
                System.out.println("phiSum for comment\t" + phiSum);
                for (int k = 0; k < number_of_topics; k++) System.out.println("m_phi\t" + cDoc.m_phi[n][k]);
            }
            if (Double.isNaN(phiSum)) {
                for (int k = 0; k < number_of_topics; k++) {
                    System.out.println("pDoc.m_sstat[k]\t" + pDoc.m_sstat[k]);
                    System.out.println("cDoc.m_sstat[k]\t" + cDoc.m_sstat[k]);
                    System.out.println("pDoc.m_lambda_stat[k][wId]\t" + pDoc.m_lambda_stat[k][wId]);
                    System.out.println("pDoc.m_lambda_topicStat[k]\t" + pDoc.m_lambda_topicStat[k]);
                    System.out.println("cDoc.m_phi[n][k]\t" + cDoc.m_phi[n][k]);
                }
            }
        }
    }
}

Example 7 with structures._ParentDoc4DCM

use of structures._ParentDoc4DCM in project IR_Base by Linda-sunshine.

the class DCMCorrLDA method initialAlphaBeta.

protected void initialAlphaBeta() {
    double parentDocNum = 0;
    double childDocNum = 0;
    Arrays.fill(m_sstat, 0);
    Arrays.fill(m_alphaAuxilary, 0);
    for (int k = 0; k < number_of_topics; k++) {
        Arrays.fill(topic_term_probabilty[k], 0);
        Arrays.fill(word_topic_sstat[k], 0);
    }
    for (_Doc d : m_trainSet) {
        if (d instanceof _ParentDoc4DCM) {
            _ParentDoc4DCM pDoc = (_ParentDoc4DCM) d;
            for (int k = 0; k < number_of_topics; k++) {
                double tempProb = pDoc.m_sstat[k] / pDoc.getTotalDocLength();
                m_sstat[k] += tempProb;
                if (pDoc.m_sstat[k] == 0)
                    continue;
                for (int v = 0; v < vocabulary_size; v++) {
                    tempProb = pDoc.m_wordTopic_stat[k][v] / pDoc.m_topic_stat[k];
                    topic_term_probabilty[k][v] += tempProb;
                }
            }
            parentDocNum += 1;
            for (_ChildDoc cDoc : pDoc.m_childDocs) {
                for (int k = 0; k < number_of_topics; k++) {
                    double tempProb = cDoc.m_sstat[k] / cDoc.getTotalDocLength();
                    m_alphaAuxilary[k] += tempProb;
                }
                childDocNum += 1;
            }
        }
    }
    for (int k = 0; k < number_of_topics; k++) {
        m_sstat[k] /= parentDocNum;
        m_alphaAuxilary[k] /= childDocNum;
        for (int v = 0; v < vocabulary_size; v++) {
            topic_term_probabilty[k][v] /= (parentDocNum + childDocNum);
        }
    }
    for (int k = 0; k < number_of_topics; k++) {
        m_alpha[k] = m_sstat[k];
        m_alpha_c[k] = m_alphaAuxilary[k];
        for (int v = 0; v < vocabulary_size; v++) m_beta[k][v] = topic_term_probabilty[k][v] + d_beta;
    }
    m_totalAlpha = Utils.sumOfArray(m_alpha);
    m_totalAlpha_c = Utils.sumOfArray(m_alpha_c);
    for (int k = 0; k < number_of_topics; k++) {
        m_totalBeta[k] = Utils.sumOfArray(m_beta[k]);
    }
}

Example 8 with structures._ParentDoc4DCM

use of structures._ParentDoc4DCM in project IR_Base by Linda-sunshine.

the class DCMCorrLDA method calculate_log_likelihood.

protected double calculate_log_likelihood(_ParentDoc4DCM d) {
    double docLogLikelihood = 0;
    int docID = d.getID();
    double parentDocLength = d.getTotalDocLength();
    for (int k = 0; k < number_of_topics; k++) {
        double term = Utils.lgamma(d.m_sstat[k] + m_alpha[k]);
        docLogLikelihood += term;
        term = Utils.lgamma(m_alpha[k]);
        docLogLikelihood -= term;
    }
    docLogLikelihood += Utils.lgamma(m_totalAlpha);
    docLogLikelihood -= Utils.lgamma(parentDocLength + m_totalAlpha);
    for (int k = 0; k < number_of_topics; k++) {
        for (int v = 0; v < vocabulary_size; v++) {
            double term = Utils.lgamma(d.m_wordTopic_stat[k][v] + m_beta[k][v]);
            docLogLikelihood += term;
            term = Utils.lgamma(m_beta[k][v]);
            docLogLikelihood -= term;
        }
        docLogLikelihood += Utils.lgamma(m_totalBeta[k]);
        docLogLikelihood -= Utils.lgamma(d.m_topic_stat[k] + m_totalBeta[k]);
    }
    for (_ChildDoc cDoc : d.m_childDocs) {
        double muDp = cDoc.getMu() / parentDocLength;
        docLogLikelihood += Utils.digamma(m_totalAlpha_c + cDoc.getMu());
        docLogLikelihood += Utils.digamma(m_totalAlpha_c + cDoc.getMu() + cDoc.getTotalDocLength());
        for (int k = 0; k < number_of_topics; k++) {
            double term = Utils.digamma(m_alpha_c[k] + muDp * d.m_sstat[k] + cDoc.m_sstat[k]);
            term -= Utils.digamma(m_alpha_c[k] + muDp * d.m_sstat[k]);
            docLogLikelihood += term;
        }
    }
    return docLogLikelihood;
}

Example 9 with structures._ParentDoc4DCM

use of structures._ParentDoc4DCM in project IR_Base by Linda-sunshine.

the class DCMCorrLDA method initialize_probability.

protected void initialize_probability(Collection<_Doc> collection) {
    m_alpha_c = new double[number_of_topics];
    m_alphaAuxilary = new double[number_of_topics];
    m_alpha = new double[number_of_topics];
    m_beta = new double[number_of_topics][vocabulary_size];
    m_totalAlpha = 0;
    m_totalAlpha_c = 0;
    m_totalBeta = new double[number_of_topics];
    m_topic_word_prob = new double[number_of_topics][vocabulary_size];
    for (_Doc d : collection) {
        if (d instanceof _ParentDoc4DCM) {
            _ParentDoc4DCM pDoc = (_ParentDoc4DCM) d;
            pDoc.setTopics4Gibbs(number_of_topics, 0, vocabulary_size);
            for (_Stn stnObj : d.getSentences()) {
                stnObj.setTopicsVct(number_of_topics);
            }
            for (_ChildDoc cDoc : pDoc.m_childDocs) {
                cDoc.setTopics4Gibbs_LDA(number_of_topics, 0);
                for (_Word w : cDoc.getWords()) {
                    int wid = w.getIndex();
                    int tid = w.getTopic();
                    pDoc.m_wordTopic_stat[tid][wid]++;
                    pDoc.m_topic_stat[tid]++;
                }
                computeMu4Doc(cDoc);
            }
        }
    }
    initialAlphaBeta();
    imposePrior();
}

Example 10 with structures._ParentDoc4DCM

use of structures._ParentDoc4DCM in project IR_Base by Linda-sunshine.

the class DCMCorrLDA method sampleInParentDoc.

protected void sampleInParentDoc(_Doc d) {
    _ParentDoc4DCM pDoc = (_ParentDoc4DCM) d;
    int wid, tid;
    double normalizedProb;
    for (_Word w : pDoc.getWords()) {
        tid = w.getTopic();
        wid = w.getIndex();
        pDoc.m_sstat[tid]--;
        pDoc.m_topic_stat[tid]--;
        pDoc.m_wordTopic_stat[tid][wid]--;
        normalizedProb = 0;
        for (tid = 0; tid < number_of_topics; tid++) {
            double pWordTopic = parentWordByTopicProb(tid, wid, pDoc);
            double pTopicPDoc = parentTopicInDocProb(tid, pDoc);
            double pTopicCDoc = parentChildInfluenceProb(tid, pDoc);
            m_topicProbCache[tid] = pWordTopic * pTopicPDoc * pTopicCDoc;
            normalizedProb += m_topicProbCache[tid];
        }
        normalizedProb *= m_rand.nextDouble();
        for (tid = 0; tid < number_of_topics; tid++) {
            normalizedProb -= m_topicProbCache[tid];
            if (normalizedProb <= 0)
                break;
        }
        if (tid == number_of_topics)
            tid--;
        w.setTopic(tid);
        pDoc.m_sstat[tid]++;
        pDoc.m_topic_stat[tid]++;
        pDoc.m_wordTopic_stat[tid][wid]++;
    }
}