MALLET从文档分类器中获取最具影响力的功能

时间:2014-11-18 00:51:13

标签: java nlp mallet

我已按照MALLET示例http://mallet.cs.umass.edu/classifier-devel.php

建立了文档分类分类器

我接下来要做的是为每个班级获得最具影响力的功能。我确信这很简单,但我还没能找到如何用Java做到这一点。

感谢任何帮助。

1 个答案:

答案 0 :(得分:0)

我遇到了同样的问题。这对我有用。 (不是完全独立的,例如,假设您已经有分类器,以及一些测试数据)

PrintWriter debugOut = new PrintWriter(new File(<filePath>));
InstanceList testInstances = new InstanceList(classifier.getInstancePipe());
CsvIterator reader = new CsvIterator(new FileReader(<path_to_testdata>), \\w+)\\s+(\\w+)\\s+(.*)", 3, 2, 1);  // (data, label, name) field indices         
testInstances.addThruPipe(reader);
PerLabelInfoGain plig = new PerLabelInfoGain (testInstances);
Alphabet alpha = classifier.getAlphabet();
LabelAlphabet la = classifier.getLabelAlphabet();
debugOut.println("debugging label numbers: " + la.size());
for (int q = 0 ; q < la.size(); q++){
    debugOut.println("Class: " + la.lookupLabel(q));
    for (int j = 0; j < 10; j++){
        int alphaId = plig.getInfoGain(q).getIndexAtRank(j);
        Object label = alpha.lookupObject(alphaId);
        debugOut.println(j + "\t" + plig.getInfoGain(q).getValueAtRank(i) + "\t" + label);
    }
    debugOut.println("===============");
}
debugOut.close();

导致:

debugging label numbers: 3
Class: sexism
0   0.1257616291393775  sexist
1   0.1257616291393775  rt
2   0.1257616291393775  female
3   0.1257616291393775  notsexist
4   0.1257616291393775  m
5   0.1257616291393775  women
6   0.1257616291393775  mt8_9
7   0.1257616291393775  sports
8   0.1257616291393775  islam
9   0.1257616291393775  men
===============
Class: none
0   0.09383300761779656 sexist
1   0.09383300761779656 mkr
2   0.09383300761779656 female
3   0.09383300761779656 muslims
4   0.09383300761779656 rt
5   0.09383300761779656 notsexist
6   0.09383300761779656 women
7   0.09383300761779656 islam
8   0.09383300761779656 mt8_9
9   0.09383300761779656 mohammed
===============
Class: racism
0   0.062072998255453926    islam
1   0.062072998255453926    muslims
2   0.062072998255453926    mkr
3   0.062072998255453926    mohammed
4   0.062072998255453926    muslim
5   0.062072998255453926    maxblumenthal
6   0.062072998255453926    quran
7   0.062072998255453926    years
8   0.062072998255453926    prophet
9   0.062072998255453926    1400
===============

编辑:plig.getInfoGain(q).getValueAtRank( i )显然应该是plig.getInfoGain(q).getValueAtRank( j

相关问题