lucene - 在lucene中获取两个文档之间的余弦相似度

Question

我在 Lucene 中建立了一个索引。我想不指定查询，只是为了获得索引中两个文档之间的分数（余弦相似度或其他距离？）。

例如，我从以前打开的 IndexReader 中获取 id 为 2 和 4 的文档。 Document d1 = ir.document(2); 文档 d2 = ir.document(4);

如何获得这两个文档之间的余弦相似度？

谢谢

Answer 1

正如 Julia 指出的那样， Sujit Pal 的示例非常有用，但是 Lucene 4 API 有很大的变化。这是为 Lucene 4 重写的版本。

import java.io.IOException;
import java.util.*;

import org.apache.commons.math3.linear.*;
import org.apache.lucene.analysis.Analyzer;
import org.apache.lucene.analysis.core.SimpleAnalyzer;
import org.apache.lucene.document.*;
import org.apache.lucene.document.Field.Store;
import org.apache.lucene.index.*;
import org.apache.lucene.store.*;
import org.apache.lucene.util.*;

public class CosineDocumentSimilarity {

    public static final String CONTENT = "Content";

    private final Set<String> terms = new HashSet<>();
    private final RealVector v1;
    private final RealVector v2;

    CosineDocumentSimilarity(String s1, String s2) throws IOException {
        Directory directory = createIndex(s1, s2);
        IndexReader reader = DirectoryReader.open(directory);
        Map<String, Integer> f1 = getTermFrequencies(reader, 0);
        Map<String, Integer> f2 = getTermFrequencies(reader, 1);
        reader.close();
        v1 = toRealVector(f1);
        v2 = toRealVector(f2);
    }

    Directory createIndex(String s1, String s2) throws IOException {
        Directory directory = new RAMDirectory();
        Analyzer analyzer = new SimpleAnalyzer(Version.LUCENE_CURRENT);
        IndexWriterConfig iwc = new IndexWriterConfig(Version.LUCENE_CURRENT,
                analyzer);
        IndexWriter writer = new IndexWriter(directory, iwc);
        addDocument(writer, s1);
        addDocument(writer, s2);
        writer.close();
        return directory;
    }

    /* Indexed, tokenized, stored. */
    public static final FieldType TYPE_STORED = new FieldType();

    static {
        TYPE_STORED.setIndexed(true);
        TYPE_STORED.setTokenized(true);
        TYPE_STORED.setStored(true);
        TYPE_STORED.setStoreTermVectors(true);
        TYPE_STORED.setStoreTermVectorPositions(true);
        TYPE_STORED.freeze();
    }

    void addDocument(IndexWriter writer, String content) throws IOException {
        Document doc = new Document();
        Field field = new Field(CONTENT, content, TYPE_STORED);
        doc.add(field);
        writer.addDocument(doc);
    }

    double getCosineSimilarity() {
        return (v1.dotProduct(v2)) / (v1.getNorm() * v2.getNorm());
    }

    public static double getCosineSimilarity(String s1, String s2)
            throws IOException {
        return new CosineDocumentSimilarity(s1, s2).getCosineSimilarity();
    }

    Map<String, Integer> getTermFrequencies(IndexReader reader, int docId)
            throws IOException {
        Terms vector = reader.getTermVector(docId, CONTENT);
        TermsEnum termsEnum = null;
        termsEnum = vector.iterator(termsEnum);
        Map<String, Integer> frequencies = new HashMap<>();
        BytesRef text = null;
        while ((text = termsEnum.next()) != null) {
            String term = text.utf8ToString();
            int freq = (int) termsEnum.totalTermFreq();
            frequencies.put(term, freq);
            terms.add(term);
        }
        return frequencies;
    }

    RealVector toRealVector(Map<String, Integer> map) {
        RealVector vector = new ArrayRealVector(terms.size());
        int i = 0;
        for (String term : terms) {
            int value = map.containsKey(term) ? map.get(term) : 0;
            vector.setEntry(i++, value);
        }
        return (RealVector) vector.mapDivide(vector.getL1Norm());
    }
}

Answer 2

索引时，可以选择存储词频向量。

在运行时，使用 IndexReader.getTermFreqVector() 查找两个文档的词频向量，并使用 IndexReader.docFreq() 查找每个词的文档频率数据。这将为您提供计算两个文档之间的余弦相似度所需的所有组件。

一种更简单的方法可能是将文档 A 作为查询提交（将所有单词作为 OR 词添加到查询中，按词频提升每个词）并在结果集中查找文档 B。

Answer 3

我知道问题已经得到解答，但是对于将来可能来到这里的人来说，可以在这里找到解决方案的好例子：

http://sujitpal.blogspot.ch/2011/10/computing-document-similarity-using.html

Answer 4

这是 Mark Butler 的一个非常好的解决方案，但是 tf/idf 权重的计算是错误的！

Term-Frequency (tf)：该术语在本文档中出现的次数（并非所有文档，如带有 termsEnum.totalTermFreq() 的代码中的所有文档）。

文档频率（df）：该词出现的文档总数。

逆文档频率：idf = log(N/df)，其中 N 是文档总数。

Tf/idf 权重 = tf * idf，对于给定的术语和给定的文档。

我希望使用 Lucene 进行有效的计算！我无法找到正确的 if/idf 权重的有效计算。

编辑：我制作了这段代码来计算权重作为 tf/idf 权重，而不是纯词频。它工作得很好，但我想知道是否有更有效的方法。

import java.io.IOException;
import java.util.HashMap;
import java.util.HashSet;
import java.util.Map;
import java.util.Set;

import org.apache.commons.math3.linear.ArrayRealVector;
import org.apache.commons.math3.linear.RealVector;
import org.apache.lucene.analysis.Analyzer;
import org.apache.lucene.analysis.core.SimpleAnalyzer;
import org.apache.lucene.document.Document;
import org.apache.lucene.document.Field;
import org.apache.lucene.document.FieldType;
import org.apache.lucene.index.DirectoryReader;
import org.apache.lucene.index.DocsEnum;
import org.apache.lucene.index.IndexReader;
import org.apache.lucene.index.IndexWriter;
import org.apache.lucene.index.IndexWriterConfig;
import org.apache.lucene.index.Term;
import org.apache.lucene.index.Terms;
import org.apache.lucene.index.TermsEnum;
import org.apache.lucene.search.DocIdSetIterator;
import org.apache.lucene.store.Directory;
import org.apache.lucene.store.RAMDirectory;
import org.apache.lucene.util.BytesRef;
import org.apache.lucene.util.Version;

public class CosineSimeTest {

    public static void main(String[] args) {
        try {
            CosineSimeTest cosSim = new 
                    CosineSimeTest( "This is good", 
                            "This is good" );
            System.out.println( cosSim.getCosineSimilarity() );
        } catch (IOException e) {
            e.printStackTrace();
        }
    }

    public static final String CONTENT = "Content";
    public static final int N = 2;//Total number of documents

    private final Set<String> terms = new HashSet<>();
    private final RealVector v1;
    private final RealVector v2;

    CosineSimeTest(String s1, String s2) throws IOException {
        Directory directory = createIndex(s1, s2);
        IndexReader reader = DirectoryReader.open(directory);
        Map<String, Double> f1 = getWieghts(reader, 0);
        Map<String, Double> f2 = getWieghts(reader, 1);
        reader.close();
        v1 = toRealVector(f1);
        System.out.println( "V1: " +v1 );
        v2 = toRealVector(f2);
        System.out.println( "V2: " +v2 );
    }

    Directory createIndex(String s1, String s2) throws IOException {
        Directory directory = new RAMDirectory();
        Analyzer analyzer = new SimpleAnalyzer(Version.LUCENE_CURRENT);
        IndexWriterConfig iwc = new IndexWriterConfig(Version.LUCENE_CURRENT,
                analyzer);
        IndexWriter writer = new IndexWriter(directory, iwc);
        addDocument(writer, s1);
        addDocument(writer, s2);
        writer.close();
        return directory;
    }

    /* Indexed, tokenized, stored. */
    public static final FieldType TYPE_STORED = new FieldType();

    static {
        TYPE_STORED.setIndexed(true);
        TYPE_STORED.setTokenized(true);
        TYPE_STORED.setStored(true);
        TYPE_STORED.setStoreTermVectors(true);
        TYPE_STORED.setStoreTermVectorPositions(true);
        TYPE_STORED.freeze();
    }

    void addDocument(IndexWriter writer, String content) throws IOException {
        Document doc = new Document();
        Field field = new Field(CONTENT, content, TYPE_STORED);
        doc.add(field);
        writer.addDocument(doc);
    }

    double getCosineSimilarity() {
        double dotProduct = v1.dotProduct(v2);
        System.out.println( "Dot: " + dotProduct);
        System.out.println( "V1_norm: " + v1.getNorm() + ", V2_norm: " + v2.getNorm() );
        double normalization = (v1.getNorm() * v2.getNorm());
        System.out.println( "Norm: " + normalization);
        return dotProduct / normalization;
    }


    Map<String, Double> getWieghts(IndexReader reader, int docId)
            throws IOException {
        Terms vector = reader.getTermVector(docId, CONTENT);
        Map<String, Integer> docFrequencies = new HashMap<>();
        Map<String, Integer> termFrequencies = new HashMap<>();
        Map<String, Double> tf_Idf_Weights = new HashMap<>();
        TermsEnum termsEnum = null;
        DocsEnum docsEnum = null;


        termsEnum = vector.iterator(termsEnum);
        BytesRef text = null;
        while ((text = termsEnum.next()) != null) {
            String term = text.utf8ToString();
            int docFreq = termsEnum.docFreq();
            docFrequencies.put(term, reader.docFreq( new Term( CONTENT, term ) ));

            docsEnum = termsEnum.docs(null, null);
            while (docsEnum.nextDoc() != DocIdSetIterator.NO_MORE_DOCS) {
                termFrequencies.put(term, docsEnum.freq());
            }

            terms.add(term);
        }

        for ( String term : docFrequencies.keySet() ) {
            int tf = termFrequencies.get(term);
            int df = docFrequencies.get(term);
            double idf = ( 1 + Math.log(N) - Math.log(df) );
            double w = tf * idf;
            tf_Idf_Weights.put(term, w);
            //System.out.printf("Term: %s - tf: %d, df: %d, idf: %f, w: %f\n", term, tf, df, idf, w);
        }

        System.out.println( "Printing docFrequencies:" );
        printMap(docFrequencies);

        System.out.println( "Printing termFrequencies:" );
        printMap(termFrequencies);

        System.out.println( "Printing if/idf weights:" );
        printMapDouble(tf_Idf_Weights);
        return tf_Idf_Weights;
    }

    RealVector toRealVector(Map<String, Double> map) {
        RealVector vector = new ArrayRealVector(terms.size());
        int i = 0;
        double value = 0;
        for (String term : terms) {

            if ( map.containsKey(term) ) {
                value = map.get(term);
            }
            else {
                value = 0;
            }
            vector.setEntry(i++, value);
        }
        return vector;
    }

    public static void printMap(Map<String, Integer> map) {
        for ( String key : map.keySet() ) {
            System.out.println( "Term: " + key + ", value: " + map.get(key) );
        }
    }

    public static void printMapDouble(Map<String, Double> map) {
        for ( String key : map.keySet() ) {
            System.out.println( "Term: " + key + ", value: " + map.get(key) );
        }
    }

}

Answer 5

在 Lucene 版本 4.x 中计算余弦相似度与 3.x 中的不同。以下帖子详细解释了在 Lucene 4.10.2 中计算余弦相似度的所有必要代码。ComputerGodzilla：在 Lucene 中计算余弦相似度！

Answer 6

你可以找到更好的解决方案@http ://darakpanand.wordpress.com/2013/06/01/document-comparison-by-cosine-methodology-using-lucene/#more-53 。以下是步骤

• 在 Lucene 的帮助下从内容构建术语向量的 java 代码（检查： http: //lucene.apache.org/core/）。
• 通过使用 commons-math.jar 库完成两个文档之间的余弦计算。

Answer 7

如果您不需要将文档存储到 Lucene 并且只想计算两个文档之间的相似性，这里是更快的代码（Scala，来自我的博客http://chepurnoy.org/blog/2014/03/faster-cosine-similarity -between-two-dicuments-with-scala-and-lucene/ )

def extractTerms(content: String): Map[String, Int] = {    
     val analyzer = new StopAnalyzer(Version.LUCENE_46)
     val ts = new EnglishMinimalStemFilter(analyzer.tokenStream("c", content))
     val charTermAttribute = ts.addAttribute(classOf[CharTermAttribute])

     val m = scala.collection.mutable.Map[String, Int]()

     ts.reset()
     while (ts.incrementToken()) {
         val term = charTermAttribute.toString
         val newCount = m.get(term).map(_ + 1).getOrElse(1)
         m += term -> newCount       
     }

     m.toMap
 }

def similarity(t1: Map[String, Int], t2: Map[String, Int]): Double = {
     //word, t1 freq, t2 freq
     val m = scala.collection.mutable.HashMap[String, (Int, Int)]()

     val sum1 = t1.foldLeft(0d) {case (sum, (word, freq)) =>
         m += word ->(freq, 0)
         sum + freq
     }

     val sum2 = t2.foldLeft(0d) {case (sum, (word, freq)) =>
         m.get(word) match {
             case Some((freq1, _)) => m += word ->(freq1, freq)
             case None => m += word ->(0, freq)
         }
         sum + freq
     }

     val (p1, p2, p3) = m.foldLeft((0d, 0d, 0d)) {case ((s1, s2, s3), e) =>
         val fs = e._2
         val f1 = fs._1 / sum1
         val f2 = fs._2 / sum2
         (s1 + f1 * f2, s2 + f1 * f1, s3 + f2 * f2)
     }

     val cos = p1 / (Math.sqrt(p2) * Math.sqrt(p3))
     cos
 }  

因此，要计算 text1 和 text2 之间的相似度，只需调用similarity(extractTerms(text1), extractTerms(text2))