API and code to convert text into indexable/searchable tokens. Covers {@link Lucene.Net.Analysis.Analyzer} and related classes.

Parsing? Tokenization? Analysis!

Lucene, indexing and search library, accepts only plain text input.

Parsing

Applications that build their search capabilities upon Lucene may support documents in various formats – HTML, XML, PDF, Word – just to name a few. Lucene does not care about the Parsing of these and other document formats, and it is the responsibility of the application using Lucene to use an appropriate Parser to convert the original format into plain text before passing that plain text to Lucene.

Tokenization

Plain text passed to Lucene for indexing goes through a process generally called tokenization. Tokenization is the process of breaking input text into small indexing elements – tokens. The way input text is broken into tokens heavily influences how people will then be able to search for that text. For instance, sentences beginnings and endings can be identified to provide for more accurate phrase and proximity searches (though sentence identification is not provided by Lucene).

In some cases simply breaking the input text into tokens is not enough – a deeper Analysis may be needed. There are many post tokenization steps that can be done, including (but not limited to):

Core Analysis

The analysis package provides the mechanism to convert Strings and Readers into tokens that can be indexed by Lucene. There are three main classes in the package from which all analysis processes are derived. These are:

Lucene 2.9 introduces a new TokenStream API. Please see the section "New TokenStream API" below for more details.

Hints, Tips and Traps

The synergy between {@link Lucene.Net.Analysis.Analyzer} and {@link Lucene.Net.Analysis.Tokenizer} is sometimes confusing. To ease on this confusion, some clarifications:

Lucene Java provides a number of analysis capabilities, the most commonly used one being the {@link Lucene.Net.Analysis.Standard.StandardAnalyzer}. Many applications will have a long and industrious life with nothing more than the StandardAnalyzer. However, there are a few other classes/packages that are worth mentioning:

  1. {@link Lucene.Net.Analysis.PerFieldAnalyzerWrapper} – Most Analyzers perform the same operation on all {@link Lucene.Net.Documents.Field}s. The PerFieldAnalyzerWrapper can be used to associate a different Analyzer with different {@link Lucene.Net.Documents.Field}s.
  2. The contrib/analyzers library located at the root of the Lucene distribution has a number of different Analyzer implementations to solve a variety of different problems related to searching. Many of the Analyzers are designed to analyze non-English languages.
  3. The contrib/snowball library located at the root of the Lucene distribution has Analyzer and TokenFilter implementations for a variety of Snowball stemmers. See http://snowball.tartarus.org for more information on Snowball stemmers.
  4. There are a variety of Tokenizer and TokenFilter implementations in this package. Take a look around, chances are someone has implemented what you need.

Analysis is one of the main causes of performance degradation during indexing. Simply put, the more you analyze the slower the indexing (in most cases). Perhaps your application would be just fine using the simple {@link Lucene.Net.Analysis.WhitespaceTokenizer} combined with a {@link Lucene.Net.Analysis.StopFilter}. The contrib/benchmark library can be useful for testing out the speed of the analysis process.

Invoking the Analyzer

Applications usually do not invoke analysis – Lucene does it for them:

However an application might invoke Analysis of any text for testing or for any other purpose, something like:
      Analyzer analyzer = new StandardAnalyzer(); // or any other analyzer
      TokenStream ts = analyzer.tokenStream("myfield",new StringReader("some text goes here"));
      while (ts.incrementToken()) {
        System.out.println("token: "+ts));
        t = ts.next();
      }
  

Indexing Analysis vs. Search Analysis

Selecting the "correct" analyzer is crucial for search quality, and can also affect indexing and search performance. The "correct" analyzer differs between applications. Lucene java's wiki page AnalysisParalysis provides some data on "analyzing your analyzer". Here are some rules of thumb:

  1. Test test test... (did we say test?)
  2. Beware of over analysis – might hurt indexing performance.
  3. Start with same analyzer for indexing and search, otherwise searches would not find what they are supposed to...
  4. In some cases a different analyzer is required for indexing and search, for instance: This might sometimes require a modified analyzer – see the next section on how to do that.

Implementing your own Analyzer

Creating your own Analyzer is straightforward. It usually involves either wrapping an existing Tokenizer and set of TokenFilters to create a new Analyzer or creating both the Analyzer and a Tokenizer or TokenFilter. Before pursuing this approach, you may find it worthwhile to explore the contrib/analyzers library and/or ask on the java-user@lucene.apache.org mailing list first to see if what you need already exists. If you are still committed to creating your own Analyzer or TokenStream derivation (Tokenizer or TokenFilter) have a look at the source code of any one of the many samples located in this package.

The following sections discuss some aspects of implementing your own analyzer.

Field Section Boundaries

When {@link Lucene.Net.Documents.Document#add(Lucene.Net.Documents.Fieldable) document.add(field)} is called multiple times for the same field name, we could say that each such call creates a new section for that field in that document. In fact, a separate call to {@link Lucene.Net.Analysis.Analyzer#tokenStream(java.lang.String, java.io.Reader) tokenStream(field,reader)} would take place for each of these so called "sections". However, the default Analyzer behavior is to treat all these sections as one large section. This allows phrase search and proximity search to seamlessly cross boundaries between these "sections". In other words, if a certain field "f" is added like this:

      document.add(new Field("f","first ends",...);
      document.add(new Field("f","starts two",...);
      indexWriter.addDocument(document);
  
Then, a phrase search for "ends starts" would find that document. Where desired, this behavior can be modified by introducing a "position gap" between consecutive field "sections", simply by overriding {@link Lucene.Net.Analysis.Analyzer#getPositionIncrementGap(java.lang.String) Analyzer.getPositionIncrementGap(fieldName)}:
      Analyzer myAnalyzer = new StandardAnalyzer() {
         public int getPositionIncrementGap(String fieldName) {
           return 10;
         }
      };
  

Token Position Increments

By default, all tokens created by Analyzers and Tokenizers have a {@link Lucene.Net.Analysis.Tokenattributes.PositionIncrementAttribute#getPositionIncrement() position increment} of one. This means that the position stored for that token in the index would be one more than that of the previous token. Recall that phrase and proximity searches rely on position info.

If the selected analyzer filters the stop words "is" and "the", then for a document containing the string "blue is the sky", only the tokens "blue", "sky" are indexed, with position("sky") = 1 + position("blue"). Now, a phrase query "blue is the sky" would find that document, because the same analyzer filters the same stop words from that query. But also the phrase query "blue sky" would find that document.

If this behavior does not fit the application needs, a modified analyzer can be used, that would increment further the positions of tokens following a removed stop word, using {@link Lucene.Net.Analysis.Tokenattributes.PositionIncrementAttribute#setPositionIncrement(int)}. This can be done with something like:

      public TokenStream tokenStream(final String fieldName, Reader reader) {
        final TokenStream ts = someAnalyzer.tokenStream(fieldName, reader);
        TokenStream res = new TokenStream() {
          TermAttribute termAtt = (TermAttribute) addAttribute(TermAttribute.class);
          PositionIncrementAttribute posIncrAtt = (PositionIncrementAttribute) addAttribute(PositionIncrementAttribute.class);
        
          public boolean incrementToken() throws IOException {
            int extraIncrement = 0;
            while (true) {
              boolean hasNext = ts.incrementToken();
              if (hasNext) {
                if (stopWords.contains(termAtt.term())) {
                  extraIncrement++; // filter this word
                  continue;
                } 
                if (extraIncrement>0) {
                  posIncrAtt.setPositionIncrement(posIncrAtt.getPositionIncrement()+extraIncrement);
                }
              }
              return hasNext;
            }
          }
        };
        return res;
      }
   
Now, with this modified analyzer, the phrase query "blue sky" would find that document. But note that this is yet not a perfect solution, because any phrase query "blue w1 w2 sky" where both w1 and w2 are stop words would match that document.

Few more use cases for modifying position increments are:

  1. Inhibiting phrase and proximity matches in sentence boundaries – for this, a tokenizer that identifies a new sentence can add 1 to the position increment of the first token of the new sentence.
  2. Injecting synonyms – here, synonyms of a token should be added after that token, and their position increment should be set to 0. As result, all synonyms of a token would be considered to appear in exactly the same position as that token, and so would they be seen by phrase and proximity searches.

New TokenStream API

With Lucene 2.9 we introduce a new TokenStream API. The old API used to produce Tokens. A Token has getter and setter methods for different properties like positionIncrement and termText. While this approach was sufficient for the default indexing format, it is not versatile enough for Flexible Indexing, a term which summarizes the effort of making the Lucene indexer pluggable and extensible for custom index formats.

A fully customizable indexer means that users will be able to store custom data structures on disk. Therefore an API is necessary that can transport custom types of data from the documents to the indexer.

Attribute and AttributeSource

Lucene 2.9 therefore introduces a new pair of classes called {@link Lucene.Net.Util.Attribute} and {@link Lucene.Net.Util.AttributeSource}. An Attribute serves as a particular piece of information about a text token. For example, {@link Lucene.Net.Analysis.Tokenattributes.TermAttribute} contains the term text of a token, and {@link Lucene.Net.Analysis.Tokenattributes.OffsetAttribute} contains the start and end character offsets of a token. An AttributeSource is a collection of Attributes with a restriction: there may be only one instance of each attribute type. TokenStream now extends AttributeSource, which means that one can add Attributes to a TokenStream. Since TokenFilter extends TokenStream, all filters are also AttributeSources.

Lucene now provides six Attributes out of the box, which replace the variables the Token class has:

Using the new TokenStream API

There are a few important things to know in order to use the new API efficiently which are summarized here. You may want to walk through the example below first and come back to this section afterwards.
  1. Please keep in mind that an AttributeSource can only have one instance of a particular Attribute. Furthermore, if a chain of a TokenStream and multiple TokenFilters is used, then all TokenFilters in that chain share the Attributes with the TokenStream.

  2. Attribute instances are reused for all tokens of a document. Thus, a TokenStream/-Filter needs to update the appropriate Attribute(s) in incrementToken(). The consumer, commonly the Lucene indexer, consumes the data in the Attributes and then calls incrementToken() again until it retuns false, which indicates that the end of the stream was reached. This means that in each call of incrementToken() a TokenStream/-Filter can safely overwrite the data in the Attribute instances.

  3. For performance reasons a TokenStream/-Filter should add/get Attributes during instantiation; i.e., create an attribute in the constructor and store references to it in an instance variable. Using an instance variable instead of calling addAttribute()/getAttribute() in incrementToken() will avoid expensive casting and attribute lookups for every token in the document.

  4. All methods in AttributeSource are idempotent, which means calling them multiple times always yields the same result. This is especially important to know for addAttribute(). The method takes the type (Class) of an Attribute as an argument and returns an instance. If an Attribute of the same type was previously added, then the already existing instance is returned, otherwise a new instance is created and returned. Therefore TokenStreams/-Filters can safely call addAttribute() with the same Attribute type multiple times. Even consumers of TokenStreams should normally call addAttribute() instead of getAttribute(), because it would not fail if the TokenStream does not have this Attribute (getAttribute() would throw an IllegalArgumentException, if the Attribute is missing). More advanced code could simply check with hasAttribute(), if a TokenStream has it, and may conditionally leave out processing for extra performance.

Example

In this example we will create a WhiteSpaceTokenizer and use a LengthFilter to suppress all words that only have two or less characters. The LengthFilter is part of the Lucene core and its implementation will be explained here to illustrate the usage of the new TokenStream API.
Then we will develop a custom Attribute, a PartOfSpeechAttribute, and add another filter to the chain which utilizes the new custom attribute, and call it PartOfSpeechTaggingFilter.

Whitespace tokenization

public class MyAnalyzer extends Analyzer {

  public TokenStream tokenStream(String fieldName, Reader reader) {
    TokenStream stream = new WhitespaceTokenizer(reader);
    return stream;
  }
  
  public static void main(String[] args) throws IOException {
    // text to tokenize
    final String text = "This is a demo of the new TokenStream API";
    
    MyAnalyzer analyzer = new MyAnalyzer();
    TokenStream stream = analyzer.tokenStream("field", new StringReader(text));
    
    // get the TermAttribute from the TokenStream
    TermAttribute termAtt = (TermAttribute) stream.addAttribute(TermAttribute.class);

    stream.reset();
    
    // print all tokens until stream is exhausted
    while (stream.incrementToken()) {
      System.out.println(termAtt.term());
    }
    
    stream.end()
    stream.close();
  }
}
In this easy example a simple white space tokenization is performed. In main() a loop consumes the stream and prints the term text of the tokens by accessing the TermAttribute that the WhitespaceTokenizer provides. Here is the output:
This
is
a
demo
of
the
new
TokenStream
API

Adding a LengthFilter

We want to suppress all tokens that have 2 or less characters. We can do that easily by adding a LengthFilter to the chain. Only the tokenStream() method in our analyzer needs to be changed:
  public TokenStream tokenStream(String fieldName, Reader reader) {
    TokenStream stream = new WhitespaceTokenizer(reader);
    stream = new LengthFilter(stream, 3, Integer.MAX_VALUE);
    return stream;
  }
Note how now only words with 3 or more characters are contained in the output:
This
demo
the
new
TokenStream
API
Now let's take a look how the LengthFilter is implemented (it is part of Lucene's core):
public final class LengthFilter extends TokenFilter {

  final int min;
  final int max;
  
  private TermAttribute termAtt;

  /**
   * Build a filter that removes words that are too long or too
   * short from the text.
   */
  public LengthFilter(TokenStream in, int min, int max)
  {
    super(in);
    this.min = min;
    this.max = max;
    termAtt = (TermAttribute) addAttribute(TermAttribute.class);
  }
  
  /**
   * Returns the next input Token whose term() is the right len
   */
  public final boolean incrementToken() throws IOException
  {
    assert termAtt != null;
    // return the first non-stop word found
    while (input.incrementToken()) {
      int len = termAtt.termLength();
      if (len >= min && len <= max) {
          return true;
      }
      // note: else we ignore it but should we index each part of it?
    }
    // reached EOS -- return null
    return false;
  }
}
The TermAttribute is added in the constructor and stored in the instance variable termAtt. Remember that there can only be a single instance of TermAttribute in the chain, so in our example the addAttribute() call in LengthFilter returns the TermAttribute that the WhitespaceTokenizer already added. The tokens are retrieved from the input stream in the incrementToken() method. By looking at the term text in the TermAttribute the length of the term can be determined and too short or too long tokens are skipped. Note how incrementToken() can efficiently access the instance variable; no attribute lookup or downcasting is neccessary. The same is true for the consumer, which can simply use local references to the Attributes.

Adding a custom Attribute

Now we're going to implement our own custom Attribute for part-of-speech tagging and call it consequently PartOfSpeechAttribute. First we need to define the interface of the new Attribute:
  public interface PartOfSpeechAttribute extends Attribute {
    public static enum PartOfSpeech {
      Noun, Verb, Adjective, Adverb, Pronoun, Preposition, Conjunction, Article, Unknown
    }
  
    public void setPartOfSpeech(PartOfSpeech pos);
  
    public PartOfSpeech getPartOfSpeech();
  }
Now we also need to write the implementing class. The name of that class is important here: By default, Lucene checks if there is a class with the name of the Attribute with the postfix 'Impl'. In this example, we would consequently call the implementing class PartOfSpeechAttributeImpl.
This should be the usual behavior. However, there is also an expert-API that allows changing these naming conventions: {@link Lucene.Net.Util.AttributeSource.AttributeFactory}. The factory accepts an Attribute interface as argument and returns an actual instance. You can implement your own factory if you need to change the default behavior.

Now here is the actual class that implements our new Attribute. Notice that the class has to extend {@link Lucene.Net.Util.AttributeImpl}:
public final class PartOfSpeechAttributeImpl extends AttributeImpl 
                            implements PartOfSpeechAttribute{
  
  private PartOfSpeech pos = PartOfSpeech.Unknown;
  
  public void setPartOfSpeech(PartOfSpeech pos) {
    this.pos = pos;
  }
  
  public PartOfSpeech getPartOfSpeech() {
    return pos;
  }

  public void clear() {
    pos = PartOfSpeech.Unknown;
  }

  public void copyTo(AttributeImpl target) {
    ((PartOfSpeechAttributeImpl) target).pos = pos;
  }

  public boolean equals(Object other) {
    if (other == this) {
      return true;
    }
    
    if (other instanceof PartOfSpeechAttributeImpl) {
      return pos == ((PartOfSpeechAttributeImpl) other).pos;
    }
 
    return false;
  }

  public int hashCode() {
    return pos.ordinal();
  }
}
This is a simple Attribute implementation has only a single variable that stores the part-of-speech of a token. It extends the new AttributeImpl class and therefore implements its abstract methods clear(), copyTo(), equals(), hashCode(). Now we need a TokenFilter that can set this new PartOfSpeechAttribute for each token. In this example we show a very naive filter that tags every word with a leading upper-case letter as a 'Noun' and all other words as 'Unknown'.
  public static class PartOfSpeechTaggingFilter extends TokenFilter {
    PartOfSpeechAttribute posAtt;
    TermAttribute termAtt;
    
    protected PartOfSpeechTaggingFilter(TokenStream input) {
      super(input);
      posAtt = (PartOfSpeechAttribute) addAttribute(PartOfSpeechAttribute.class);
      termAtt = (TermAttribute) addAttribute(TermAttribute.class);
    }
    
    public boolean incrementToken() throws IOException {
      if (!input.incrementToken()) {return false;}
      posAtt.setPartOfSpeech(determinePOS(termAtt.termBuffer(), 0, termAtt.termLength()));
      return true;
    }
    
    // determine the part of speech for the given term
    protected PartOfSpeech determinePOS(char[] term, int offset, int length) {
      // naive implementation that tags every uppercased word as noun
      if (length > 0 && Character.isUpperCase(term[0])) {
        return PartOfSpeech.Noun;
      }
      return PartOfSpeech.Unknown;
    }
  }
Just like the LengthFilter, this new filter accesses the attributes it needs in the constructor and stores references in instance variables. Notice how you only need to pass in the interface of the new Attribute and instantiating the correct class is automatically been taken care of. Now we need to add the filter to the chain:
  public TokenStream tokenStream(String fieldName, Reader reader) {
    TokenStream stream = new WhitespaceTokenizer(reader);
    stream = new LengthFilter(stream, 3, Integer.MAX_VALUE);
    stream = new PartOfSpeechTaggingFilter(stream);
    return stream;
  }
Now let's look at the output:
This
demo
the
new
TokenStream
API
Apparently it hasn't changed, which shows that adding a custom attribute to a TokenStream/Filter chain does not affect any existing consumers, simply because they don't know the new Attribute. Now let's change the consumer to make use of the new PartOfSpeechAttribute and print it out:
  public static void main(String[] args) throws IOException {
    // text to tokenize
    final String text = "This is a demo of the new TokenStream API";
    
    MyAnalyzer analyzer = new MyAnalyzer();
    TokenStream stream = analyzer.tokenStream("field", new StringReader(text));
    
    // get the TermAttribute from the TokenStream
    TermAttribute termAtt = (TermAttribute) stream.addAttribute(TermAttribute.class);
    
    // get the PartOfSpeechAttribute from the TokenStream
    PartOfSpeechAttribute posAtt = (PartOfSpeechAttribute) stream.addAttribute(PartOfSpeechAttribute.class);
    
    stream.reset();

    // print all tokens until stream is exhausted
    while (stream.incrementToken()) {
      System.out.println(termAtt.term() + ": " + posAtt.getPartOfSpeech());
    }
    
    stream.end();
    stream.close();
  }
The change that was made is to get the PartOfSpeechAttribute from the TokenStream and print out its contents in the while loop that consumes the stream. Here is the new output:
This: Noun
demo: Unknown
the: Unknown
new: Unknown
TokenStream: Noun
API: Noun
Each word is now followed by its assigned PartOfSpeech tag. Of course this is a naive part-of-speech tagging. The word 'This' should not even be tagged as noun; it is only spelled capitalized because it is the first word of a sentence. Actually this is a good opportunity for an excerise. To practice the usage of the new API the reader could now write an Attribute and TokenFilter that can specify for each word if it was the first token of a sentence or not. Then the PartOfSpeechTaggingFilter can make use of this knowledge and only tag capitalized words as nouns if not the first word of a sentence (we know, this is still not a correct behavior, but hey, it's a good exercise). As a small hint, this is how the new Attribute class could begin:
  public class FirstTokenOfSentenceAttributeImpl extends Attribute
                   implements FirstTokenOfSentenceAttribute {
    
    private boolean firstToken;
    
    public void setFirstToken(boolean firstToken) {
      this.firstToken = firstToken;
    }
    
    public boolean getFirstToken() {
      return firstToken;
    }

    public void clear() {
      firstToken = false;
    }

  ...