Lucene custom scoring for numeric fields

I would like to have, in addition to standard term search with tf-idf similarity over text content field, scoring based on "similarity" of numeric fields. This similarity will be depending on distance between the value in query and in document (e.g. gaussian with m= [user input], s= 0.5)

I.e. let's say documents represent people, and person document have two fields:

description (full text)
age (numeric).

I want to find documents like

description:(x y z) age:30

but age to be not the filter, but rather part of score (for person of age 30 multiplier will be 1.0, for 25-year-old person 0.8 etc.)

Can this be achieved in a sensible manner?

EDIT: Finally I found out this can be done by wrapping ValueSourceQuery and TermQuery with CustomScoreQuery. See my solution below.

EDIT 2: With fast-changing versions of Lucene, I just want to add that it was tested on Lucene 3.0 (Java).

回答1:

Okay, so here's (a bit verbose) proof-of-concept as a full JUnit test. Haven't tested its efficiency yet for large index, but from what I've read probably after a warm-up it should perform well, providing there's enough RAM available to cache numeric fields.

  package tests;

  import org.apache.lucene.analysis.Analyzer;
  import org.apache.lucene.analysis.WhitespaceAnalyzer;
  import org.apache.lucene.document.Document;
  import org.apache.lucene.document.Field;
  import org.apache.lucene.document.NumericField;
  import org.apache.lucene.index.IndexWriter;
  import org.apache.lucene.queryParser.QueryParser;
  import org.apache.lucene.search.IndexSearcher;
  import org.apache.lucene.search.Query;
  import org.apache.lucene.search.ScoreDoc;
  import org.apache.lucene.search.TopDocs;
  import org.apache.lucene.search.function.CustomScoreQuery;
  import org.apache.lucene.search.function.IntFieldSource;
  import org.apache.lucene.search.function.ValueSourceQuery;
  import org.apache.lucene.store.Directory;
  import org.apache.lucene.store.RAMDirectory;
  import org.apache.lucene.util.Version;

  import junit.framework.TestCase;

  public class AgeAndContentScoreQueryTest extends TestCase
  {
     public class AgeAndContentScoreQuery extends CustomScoreQuery
     {
        protected float peakX;
        protected float sigma;

        public AgeAndContentScoreQuery(Query subQuery, ValueSourceQuery valSrcQuery, float peakX, float sigma) {
           super(subQuery, valSrcQuery);
           this.setStrict(true); // do not normalize score values from ValueSourceQuery!
           this.peakX = peakX;   // age for which the age-relevance is best
           this.sigma = sigma;
        }

        @Override
        public float customScore(int doc, float subQueryScore, float valSrcScore){
           // subQueryScore is td-idf score from content query
           float contentScore = subQueryScore;

           // valSrcScore is a value of date-of-birth field, represented as a float
           // let's convert age value to gaussian-like age relevance score
           float x = (2011 - valSrcScore); // age
           float ageScore = (float) Math.exp(-Math.pow(x - peakX, 2) / 2*sigma*sigma);

           float finalScore = ageScore * contentScore;

           System.out.println("#contentScore: " + contentScore);
           System.out.println("#ageValue:     " + (int)valSrcScore);
           System.out.println("#ageScore:     " + ageScore);
           System.out.println("#finalScore:   " + finalScore);
           System.out.println("+++++++++++++++++");

           return finalScore;
        }
     }

     protected Directory directory;
     protected Analyzer analyzer = new WhitespaceAnalyzer();
     protected String fieldNameContent = "content";
     protected String fieldNameDOB = "dob";

     protected void setUp() throws Exception
     {
        directory = new RAMDirectory();
        analyzer = new WhitespaceAnalyzer();

        // indexed documents
        String[] contents = {"foo baz1", "foo baz2 baz3", "baz4"};
        int[] dobs = {1991, 1981, 1987}; // date of birth

        IndexWriter writer = new IndexWriter(directory, analyzer, IndexWriter.MaxFieldLength.UNLIMITED);
        for (int i = 0; i < contents.length; i++) 
        {
           Document doc = new Document();
           doc.add(new Field(fieldNameContent, contents[i], Field.Store.YES, Field.Index.ANALYZED)); // store & index
           doc.add(new NumericField(fieldNameDOB, Field.Store.YES, true).setIntValue(dobs[i]));      // store & index
           writer.addDocument(doc);
        }
        writer.close();
     }

     public void testSearch() throws Exception
     {
        String inputTextQuery = "foo bar";
        float peak = 27.0f;
        float sigma = 0.1f;

        QueryParser parser = new QueryParser(Version.LUCENE_30, fieldNameContent, analyzer);
        Query contentQuery = parser.parse(inputTextQuery);

        ValueSourceQuery dobQuery = new ValueSourceQuery( new IntFieldSource(fieldNameDOB) );
         // or: FieldScoreQuery dobQuery = new FieldScoreQuery(fieldNameDOB,Type.INT);

        CustomScoreQuery finalQuery = new AgeAndContentScoreQuery(contentQuery, dobQuery, peak, sigma);

        IndexSearcher searcher = new IndexSearcher(directory);
        TopDocs docs = searcher.search(finalQuery, 10);

        System.out.println("\nDocuments found:\n");
        for(ScoreDoc match : docs.scoreDocs)
        {
           Document d = searcher.doc(match.doc);
           System.out.println("CONTENT: " + d.get(fieldNameContent) );
           System.out.println("D.O.B.:  " + d.get(fieldNameDOB) );
           System.out.println("SCORE:   " + match.score );
           System.out.println("-----------------");
        }
     }
  }

回答2:

This can be achieved using Solr's FunctionQuery