在 WEKA API 中为单个文档计算 TF-IDF 以预测分类

Question

出于某种原因，我正在使用 WEKA API...

我已经为一组文档生成了 tf-idf 分数，

StringToWordVector filter = new StringToWordVector();
filter.setInputFormat(data);
filter.setIDFTransform(true);
filter.setStopwordsHandler(new StopWordsHandlerEN());//just a simple handler for stop words I created
filter.setLowerCaseTokens(true);
filter.setStemmer(new MyStemmer());//a stemmer I created
filter.setWordsToKeep(words2keep);
Instances result = Filter.useFilter(data, filter);

然后将它们分成训练和测试子集，进行训练、测试等等...

一旦我有了一个训练有素的现成分类模型，我想创建一个简单的 API 来对任何传入的文档进行分类。但问题是新的 tf-idf 分数需要根据 tf-idf 向量和起始文档集的词来计算，对吧？换句话说，如果我没记错的话，我需要加载 scikit-learn 的 tfidfvectorizer.

的对应项

我在 WEKA 中找不到类似的东西...有吗？..

Answer 1

StringToWordVector filter uses the weka.core.DictionaryBuilder class TF/IDF 计算的幕后花絮。

只要您使用要转换的文本创建一个 weka.core.Instance 对象，就可以使用生成器的 vectorizeInstance(Instance) 方法来实现。

编辑 1：

下面是一个基于您的代码（但使用 Weka classes）的示例，它显示了如何使用过滤器或 DictionaryBuilder 进行 TF/IDF 转换。两者都被序列化、反序列化和 re-used 以及证明这些 classes 是可序列化的：

import weka.core.DictionaryBuilder;
import weka.core.Instance;
import weka.core.Instances;
import weka.core.SerializationHelper;
import weka.core.converters.ConverterUtils;
import weka.core.stemmers.LovinsStemmer;
import weka.core.stopwords.Rainbow;
import weka.filters.Filter;
import weka.filters.unsupervised.attribute.StringToWordVector;

public class TFIDF {

  // just exposes the internal DictionaryBuilder member
  public static class StringToWordVectorExposed
    extends StringToWordVector {

    public DictionaryBuilder getDictionary() {
      return m_dictionaryBuilder;
    }
  }

  public static void main(String[] args) throws Exception {
    // load data
    Instances train = ConverterUtils.DataSource.read("/some/where/train.arff");
    train.setClassIndex(train.numAttributes() - 1);
    Instances test = ConverterUtils.DataSource.read("/some/where/test.arff");
    test.setClassIndex(test.numAttributes() - 1);
    // init filter
    StringToWordVectorExposed filter = new StringToWordVectorExposed();
    int words2keep = 100;
    filter.setInputFormat(train);
    filter.setIDFTransform(true);
    filter.setStopwordsHandler(new Rainbow());
    filter.setLowerCaseTokens(true);
    filter.setStemmer(new LovinsStemmer());
    filter.setWordsToKeep(words2keep);
    filter.setInputFormat(train);
    Instances trainFiltered = Filter.useFilter(train, filter);
    DictionaryBuilder builder = filter.getDictionary();
    // apply filter/dictionary
    Instances testFiltered = Filter.useFilter(test, filter);
    System.out.println(testFiltered.instance(0));
    Instance tfidf = builder.vectorizeInstance(test.instance(0));
    System.out.println(tfidf);
    // serialize
    SerializationHelper.write("/some/where/filter.ser", filter);
    SerializationHelper.write("/some/where/dictionary.ser", filter.getDictionary());
    // deserialize
    StringToWordVectorExposed filter2 = (StringToWordVectorExposed) SerializationHelper.read("/some/where/filter.ser");
    DictionaryBuilder builder2 = (DictionaryBuilder) SerializationHelper.read("/some/where/dictionary.ser");
    // re-apply
    testFiltered = Filter.useFilter(test, filter2);
    System.out.println(testFiltered.instance(0));
    tfidf = builder2.vectorizeInstance(test.instance(0));
    System.out.println(tfidf);
  }
}

在 WEKA API 中为单个文档计算 TF-IDF 以预测分类

Calculate TF-IDF in WEKA API for single document to predict classification

java

weka

tf-idf

tfidfvectorizer