i have built an index in Lucene. I want without specifying a query, just to get a score (cosine similarity or another distance?) between two documents in the index.
For example i am getting from previously opened IndexReader ir the documents with ids 2 and 4. Document d1 = ir.document(2); Document d2 = ir.document(4);
How can i get the cosine similarity between these two documents?
Thank you
When indexing, there's an option to store term frequency vectors.
During runtime, look up the term frequency vectors for both documents using IndexReader.getTermFreqVector(), and look up document frequency data for each term using IndexReader.docFreq(). That will give you all the components necessary to calculate the cosine similarity between the two docs.
An easier way might be to submit doc A as a query (adding all words to the query as OR terms, boosting each by term frequency) and look for doc B in the result set.
As Julia points out Sujit Pal's example is very useful but the Lucene 4 API has substantial changes. Here is a version rewritten for Lucene 4.
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());
}
}
I know question has been answered, but for the people who might come here in future, nice example of the solution can be found here:
http://sujitpal.blogspot.ch/2011/10/computing-document-similarity-using.html
It's a very good solution by Mark Butler, however the calculations of the tf/idf weights are WRONG!
Term-Frequency (tf): how much this term appeared in this document (NOT all documents as in the code with termsEnum.totalTermFreq()).
Document Frequency (df): the total number of documents that this term appeared in.
Inverse Document Frequency: idf = log(N/df), where N is the total number of documents.
Tf/idf weight = tf * idf, for a given term and a given document.
I was hoping for an efficient calculation using Lucene! I'm unable to find a an efficient calculation for the correct if/idf weights.
EDIT: I made this code to calculate the weights as tf/idf weights and not as pure term-frequency. It works pretty well, but I wonder if there's a more efficient way.
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) );
}
}
}
you can find better solution @ http://darakpanand.wordpress.com/2013/06/01/document-comparison-by-cosine-methodology-using-lucene/#more-53 . following are the steps
Calculating Cosine Similarity in Lucene Version 4.x is different than those from 3.x. Following post has detailed explanation with all the necessary code for calculating cosine similarity in Lucene 4.10.2. ComputerGodzilla: Calculated Cosine Similarity in Lucene!
If you don't need to store documents to Lucene and just want to calculate similarity between two docs, here's the faster code (Scala, from my blog 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
}
So, to calculate similarity between text1 and text2 just call similarity(extractTerms(text1), extractTerms(text2))
