fuzzy search with lucene

tadumtada picture tadumtada · Jul 19, 2013 · Viewed 14.7k times · Source

I implemented a fuzzy search with lucene 4.3.1 but i'm not satisfied with the result. I would like to specify a number of results it should return. So for example if I want 10 results, it should return the 10 best matches, no matter how bad they are. Most of the time it returns nothing if the word I search for is very different from anything in the index. How can I achieve more/fuzzier results?

Here the code I have:

    public String[] luceneQuery(String query, int numberOfHits, String path)
        throws ParseException, IOException {

    File dir = new File(path);
    Directory index = FSDirectory.open(dir);

    query = query + "~";
    Query q = new QueryParser(Version.LUCENE_43, "label", analyzer)
            .parse(query);

    IndexReader reader = DirectoryReader.open(index);
    IndexSearcher searcher = new IndexSearcher(reader);

    Query fuzzyQuery = new FuzzyQuery(new Term("label", query), 2);

    ScoreDoc[] fuzzyHits = searcher.search(fuzzyQuery, numberOfHits).scoreDocs;
    String[] fuzzyResults = new String[fuzzyHits.length];

    for (int i = 0; i < fuzzyHits.length; ++i) {
        int docId = fuzzyHits[i].doc;
        Document d = searcher.doc(docId);
        fuzzyResults[i] = d.get("label");
    }

    reader.close();
    return fuzzyResults;
}

Answer

femtoRgon picture femtoRgon · Jul 19, 2013

large edit distances are no longer supported by FuzzyQuery in Lucene 4.x. The current implementation of FuzzyQuery is a huge improvement on performance from the Lucene 3.x implementation, but only supports two edits. Distances greater than 2 Damerau–Levenshtein edits are considered to rarely be really useful.

According to the FuzzyQuery documentation, if you really must have higher edit distances:

If you really want this, consider using an n-gram indexing technique (such as the SpellChecker in the suggest module) instead.

The strong implication is that you should rethink what your trying to accomplish, and find a more useful approach.