我正在尝试使用Weka在JAVA中构建文本分类器。 我已经阅读了一些教程,并且我正在尝试构建自己的分类器。
我有以下类别:
computer,sport,unknown
以及以下已经过培训的数据
cs belongs to computer
java -> computer
soccer -> sport
snowboard -> sport
例如,如果用户想要对单词java进行分类,它应该返回类别计算机(毫无疑问,java只存在于该类别中!)。
它确实编译,但会产生奇怪的输出。
输出结果为:
====== RESULT ====== CLASSIFIED AS: [0.5769230769230769, 0.2884615384615385, 0.1346153846153846]
====== RESULT ====== CLASSIFIED AS: [0.42857142857142855, 0.42857142857142855, 0.14285714285714285]
但是第一个要分类的文本是java,它只出现在类别计算机中,因此它应该是
[1.0 0.0 0.0]
而对于另一个它根本找不到,所以应该归类为未知
[0.0 0.0 1.0].
以下是代码:
import java.io.FileNotFoundException;
import java.io.Serializable;
import java.util.Arrays;
import weka.classifiers.Classifier;
import weka.classifiers.bayes.NaiveBayesMultinomialUpdateable;
import weka.core.Attribute;
import weka.core.FastVector;
import weka.core.Instance;
import weka.core.Instances;
import weka.filters.Filter;
import weka.filters.unsupervised.attribute.StringToWordVector;
public class TextClassifier implements Serializable {
private static final long serialVersionUID = -1397598966481635120L;
public static void main(String[] args) {
try {
TextClassifier cl = new TextClassifier(new NaiveBayesMultinomialUpdateable());
cl.addCategory("computer");
cl.addCategory("sport");
cl.addCategory("unknown");
cl.setupAfterCategorysAdded();
//
cl.addData("cs", "computer");
cl.addData("java", "computer");
cl.addData("soccer", "sport");
cl.addData("snowboard", "sport");
double[] result = cl.classifyMessage("java");
System.out.println("====== RESULT ====== \tCLASSIFIED AS:\t" + Arrays.toString(result));
result = cl.classifyMessage("asdasdasd");
System.out.println("====== RESULT ======\tCLASSIFIED AS:\t" + Arrays.toString(result));
} catch (Exception e) {
e.printStackTrace();
}
}
private Instances trainingData;
private StringToWordVector filter;
private Classifier classifier;
private boolean upToDate;
private FastVector classValues;
private FastVector attributes;
private boolean setup;
private Instances filteredData;
public TextClassifier(Classifier classifier) throws FileNotFoundException {
this(classifier, 10);
}
public TextClassifier(Classifier classifier, int startSize) throws FileNotFoundException {
this.filter = new StringToWordVector();
this.classifier = classifier;
// Create vector of attributes.
this.attributes = new FastVector(2);
// Add attribute for holding texts.
this.attributes.addElement(new Attribute("text", (FastVector) null));
// Add class attribute.
this.classValues = new FastVector(startSize);
this.setup = false;
}
public void addCategory(String category) {
category = category.toLowerCase();
// if required, double the capacity.
int capacity = classValues.capacity();
if (classValues.size() > (capacity - 5)) {
classValues.setCapacity(capacity * 2);
}
classValues.addElement(category);
}
public void addData(String message, String classValue) throws IllegalStateException {
if (!setup) {
throw new IllegalStateException("Must use setup first");
}
message = message.toLowerCase();
classValue = classValue.toLowerCase();
// Make message into instance.
Instance instance = makeInstance(message, trainingData);
// Set class value for instance.
instance.setClassValue(classValue);
// Add instance to training data.
trainingData.add(instance);
upToDate = false;
}
/**
* Check whether classifier and filter are up to date. Build i necessary.
* @throws Exception
*/
private void buildIfNeeded() throws Exception {
if (!upToDate) {
// Initialize filter and tell it about the input format.
filter.setInputFormat(trainingData);
// Generate word counts from the training data.
filteredData = Filter.useFilter(trainingData, filter);
// Rebuild classifier.
classifier.buildClassifier(filteredData);
upToDate = true;
}
}
public double[] classifyMessage(String message) throws Exception {
message = message.toLowerCase();
if (!setup) {
throw new Exception("Must use setup first");
}
// Check whether classifier has been built.
if (trainingData.numInstances() == 0) {
throw new Exception("No classifier available.");
}
buildIfNeeded();
Instances testset = trainingData.stringFreeStructure();
Instance testInstance = makeInstance(message, testset);
// Filter instance.
filter.input(testInstance);
Instance filteredInstance = filter.output();
return classifier.distributionForInstance(filteredInstance);
}
private Instance makeInstance(String text, Instances data) {
// Create instance of length two.
Instance instance = new Instance(2);
// Set value for message attribute
Attribute messageAtt = data.attribute("text");
instance.setValue(messageAtt, messageAtt.addStringValue(text));
// Give instance access to attribute information from the dataset.
instance.setDataset(data);
return instance;
}
public void setupAfterCategorysAdded() {
attributes.addElement(new Attribute("class", classValues));
// Create dataset with initial capacity of 100, and set index of class.
trainingData = new Instances("MessageClassificationProblem", attributes, 100);
trainingData.setClassIndex(trainingData.numAttributes() - 1);
setup = true;
}
}
顺便说一句,发现了一个很好的页面:
答案 0 :(得分:4)
贝叶斯分类器为您提供单词属于某个类别的(加权)概率。这几乎永远不会完全是0或1.您可以设置硬截止(例如0.5)并根据此确定类的成员资格,或检查计算的概率并根据该决定(即最高的地图为1,最低为0)。
答案 1 :(得分:1)
我想我只是提出你可以通过下载和使用http://lightsidelabs.com中的LightSIDE来完成大部分此类文本分类工作而无需编码。这个开源Java包包含WEKA,可用于Windows和Mac上的发行版 - 可以非常灵活地处理大多数WEKA友好数据集,允许您遍历各种模型,设置和参数,并为快照提供良好支持。在任何时候保存您的数据和模型以及分类结果,直到您构建了您满意的模型。该产品去年在Kaggle.com的ASAP竞赛中证明了这一点,并且获得了很大的关注。当然,人们总是需要或需要“自己动手”,但也许甚至可以作为支票,如果您正在编程WEKA解决方案,了解和使用LightSIDE可能非常方便。
答案 2 :(得分:0)
如果您尝试获取确定的类而不是分布,请尝试切换
return classifier.distributionForInstance(filteredInstance);
到
return classifier.classifyInstance(filteredInstance);