Another highlighter implementation based on term vectors.


  • fast for large docs
  • support N-gram fields
  • support phrase-unit highlighting with slops
  • support multi-term (includes wildcard, range, regexp, etc) queries
  • highlight fields need to be stored with Positions and Offsets
  • take into account query boost and/or IDF-weight to score fragments
  • support colored highlight tags
  • pluggable FragListBuilder / FieldFragList
  • pluggable FragmentsBuilder


To explain the algorithm, let's use the following sample text (to be highlighted) and user query:

sample document and query
Sample Text Lucene is a search engine library.
User Query Lucene^2 OR "search library"~1

The user query is a BooleanQuery that consists of TermQuery("Lucene") with boost of 2 and PhraseQuery("search library") with slop of 1.

For your convenience, here is the offsets and positions info of the sample text.

 |        |          1111111111222222222233333|
 |  offset|01234567890123456789012345678901234|
 |document|Lucene is a search engine library. |
 |position|0      1  2 3      4      5        |

Step 1.

In Step 1, Fast Vector Highlighter generates FieldQuery.QueryPhraseMap from the user query. QueryPhraseMap consists of the following members:

 public class QueryPhraseMap {
   boolean terminal;
   int slop;   // valid if terminal == true and phraseHighlight == true
   float boost;  // valid if terminal == true
   Map<String, QueryPhraseMap> subMap;

QueryPhraseMap has subMap. The key of the subMap is a term text in the user query and the value is a subsequent QueryPhraseMap. If the query is a term (not phrase), then the subsequent QueryPhraseMap is marked as terminal. If the query is a phrase, then the subsequent QueryPhraseMap is not a terminal and it has the next term text in the phrase.

From the sample user query, the following QueryPhraseMap will be generated:

 +--------+-+  +-------+-+
 |"Lucene"|o+->|boost=2|*|  * : terminal
 +--------+-+  +-------+-+

 +--------+-+  +---------+-+  +-------+------+-+
 +--------+-+  +---------+-+  +-------+------+-+

Step 2.

In Step 2, Fast Vector Highlighter generates FieldTermStack. Fast Vector Highlighter uses term vector data (must be stored FieldType.setStoreTermVectorOffsets(boolean) and FieldType.setStoreTermVectorPositions(boolean)) to generate it. FieldTermStack keeps the terms in the user query. Therefore, in this sample case, Fast Vector Highlighter generates the following FieldTermStack:

 |"Lucene"(0,6,0)   |
 |"search"(12,18,3) |
 where : "termText"(startOffset,endOffset,position)

Step 3.

In Step 3, Fast Vector Highlighter generates FieldPhraseList by reference to QueryPhraseMap and FieldTermStack.

 |"Lucene"        |[(0,6)]          |w=2|
 |"search library"|[(12,18),(26,33)]|w=1|

The type of each entry is WeightedPhraseInfo that consists of an array of terms offsets and weight.

Step 4.

In Step 4, Fast Vector Highlighter creates FieldFragList by reference to FieldPhraseList. In this sample case, the following FieldFragList will be generated:

 |"Lucene"[(0,6)]                  |
 |"search library"[(12,18),(26,33)]|
 |totalBoost=3                     |

The calculation for each FieldFragList.WeightedFragInfo.totalBoost (weight) depends on the implementation of FieldFragList.add( ... ):

   public void add( int startOffset, int endOffset, List<WeightedPhraseInfo> phraseInfoList ) {
     float totalBoost = 0;
     List<SubInfo> subInfos = new ArrayList<SubInfo>();
     for( WeightedPhraseInfo phraseInfo : phraseInfoList ){
       subInfos.add( new SubInfo( phraseInfo.getText(), phraseInfo.getTermsOffsets(), phraseInfo.getSeqnum() ) );
       totalBoost += phraseInfo.getBoost();
     getFragInfos().add( new WeightedFragInfo( startOffset, endOffset, subInfos, totalBoost ) );

The used implementation of FieldFragList is noted in BaseFragListBuilder.createFieldFragList( ... ):
   public FieldFragList createFieldFragList( FieldPhraseList fieldPhraseList, int fragCharSize ){
     return createFieldFragList( fieldPhraseList, new SimpleFieldFragList( fragCharSize ), fragCharSize );

Currently there are basically to approaches available:

  • SimpleFragListBuilder using SimpleFieldFragList: sum-of-boosts-approach. The totalBoost is calculated by summarizing the query-boosts per term. Per default a term is boosted by 1.0
  • WeightedFragListBuilder using WeightedFieldFragList: sum-of-distinct-weights-approach. The totalBoost is calculated by summarizing the IDF-weights of distinct terms.

Comparison of the two approaches:

query = das alte testament (The Old Testament)
Terms in fragmentsum-of-distinct-weightssum-of-boosts
das alte testament5.3396213.0
das alte testament5.3396213.0
das testament alte5.3396213.0
das alte testament5.3396213.0
das testament2.94556882.0
das alte2.47595952.0
das das das das1.50153574.0
das das das1.30036813.0
das das1.0617462.0

Step 5.

In Step 5, by using FieldFragList and the field stored data, Fast Vector Highlighter creates highlighted snippets!