org.apache.lucene.index.pruning
Class TermPruningPolicy

java.lang.Object
  org.apache.lucene.index.pruning.PruningPolicy
      org.apache.lucene.index.pruning.TermPruningPolicy

Direct Known Subclasses:

CarmelTopKTermPruningPolicy, CarmelUniformTermPruningPolicy, RIDFTermPruningPolicy, TFTermPruningPolicy

public abstract class TermPruningPolicy
extends PruningPolicy

Policy for producing smaller index out of an input index, by examining its terms and removing from the index some or all of their data as follows:

• all terms of a certain field - see pruneAllFieldPostings(String)
• all data of a certain term - see pruneTermEnum(TermEnum)
• all positions of a certain term in a certain document - see #pruneAllPositions(TermPositions, Term)
• some positions of a certain term in a certain document - see #pruneSomePositions(int, int[], Term)

The pruned, smaller index would, for many types of queries return nearly identical top-N results as compared with the original index, but with increased performance.

Pruning of indexes is handy for producing small first-tier indexes that fit completely in RAM, and store these indexes using IndexWriter.addIndexes(IndexReader...)

Interestingly, if the input index is optimized (i.e. doesn't contain deletions), then the index produced via IndexWriter.addIndexes(IndexReader[]) will preserve internal document id-s so that they are in sync with the original index. This means that all other auxiliary information not necessary for first-tier processing, such as some stored fields, can also be removed, to be quickly retrieved on-demand from the original index using the same internal document id. See StorePruningPolicy for information about removing stored fields.

Please note that while this family of policies method produces good results for term queries it often leads to poor results for phrase queries (because postings are removed without considering whether they belong to an important phrase).

Aggressive pruning policies produce smaller indexes - search performance increases, and recall decreases (i.e. search quality deteriorates).

See the following papers for a discussion of this problem and the proposed solutions to improve the quality of a pruned index (not implemented here):

• Pruned query evaluation using pre-computed impacts, V. Anh et al, ACM SIGIR 2006
• A document-centric approach to static index pruning in text retrieval systems, S. Buettcher et al, ACM SIGIR 2006
• Pruning Policies for Two-Tiered Inverted Index with Correctness Guarantee, A. Ntoulas et al, ACM SIGIR 2007.

Field Summary

protected Map<String,Integer>	fieldFlags Pruning operations to be conducted on fields.
protected IndexReader	in

Fields inherited from class org.apache.lucene.index.pruning.PruningPolicy

DEL_ALL, DEL_PAYLOADS, DEL_POSTINGS, DEL_STORED, DEL_VECTOR

Constructor Summary

protected

TermPruningPolicy(IndexReader in, Map<String,Integer> fieldFlags) Construct a policy.

Method Summary

abstract void	initPositionsTerm(TermPositions in, Term t) Called when moving TermPositions to a new Term.
boolean	pruneAllFieldPostings(String field) Pruning of all postings for a field
abstract boolean	pruneAllPositions(TermPositions termPositions, Term t) Prune all postings per term (invoked once per term per doc)
boolean	prunePayload(TermPositions in, Term curTerm) Called when checking for the presence of payload for the current term at a current position
abstract int	pruneSomePositions(int docNum, int[] positions, Term curTerm) Prune some postings per term (invoked once per term per doc).
abstract boolean	pruneTermEnum(TermEnum te) Pruning of all postings for a term (invoked once per term).
abstract int	pruneTermVectorTerms(int docNumber, String field, String[] terms, int[] freqs, TermFreqVector v) Pruning of individual terms in term vectors.
boolean	pruneWholeTermVector(int docNumber, String field) Term vector pruning.

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Field Detail

fieldFlags

protected Map<String,Integer> fieldFlags

Pruning operations to be conducted on fields.

in

protected IndexReader in

Constructor Detail

TermPruningPolicy

protected TermPruningPolicy(IndexReader in,
                            Map<String,Integer> fieldFlags)

Construct a policy.

Parameters:

in - input reader

fieldFlags - a map, where keys are field names and values are bitwise-OR flags of operations to be performed (see PruningPolicy for more details).

Method Detail

pruneWholeTermVector

public boolean pruneWholeTermVector(int docNumber,
                                    String field)
                             throws IOException

Term vector pruning.

Parameters:

docNumber - document number

field - field name

Returns:

true if the complete term vector for this field should be removed (as specified by PruningPolicy.DEL_VECTOR flag).

Throws:

IOException

pruneAllFieldPostings

public boolean pruneAllFieldPostings(String field)
                              throws IOException

Pruning of all postings for a field

Parameters:

field - field name

Returns:

true if all postings for all terms in this field should be removed (as specified by PruningPolicy.DEL_POSTINGS).

Throws:

IOException

initPositionsTerm

public abstract void initPositionsTerm(TermPositions in,
                                       Term t)
                                throws IOException

Called when moving TermPositions to a new Term.

Parameters:

in - input term positions

t - current term

Throws:

IOException

prunePayload

public boolean prunePayload(TermPositions in,
                            Term curTerm)

Called when checking for the presence of payload for the current term at a current position

Parameters:

in - positioned term positions

curTerm - current term associated with these positions

Returns:

true if the payload should be removed, false otherwise.

pruneTermVectorTerms

public abstract int pruneTermVectorTerms(int docNumber,
                                         String field,
                                         String[] terms,
                                         int[] freqs,
                                         TermFreqVector v)
                                  throws IOException

Pruning of individual terms in term vectors.

Parameters:

docNumber - document number

field - field name

terms - array of terms

freqs - array of term frequencies

v - the original term frequency vector

Returns:

0 if no terms are to be removed, positive number to indicate how many terms need to be removed. The same number of entries in the terms array must be set to null to indicate which terms to remove.

Throws:

IOException

pruneTermEnum

public abstract boolean pruneTermEnum(TermEnum te)
                               throws IOException

Pruning of all postings for a term (invoked once per term).

Parameters:

te - positioned term enum.

Returns:

true if all postings for this term should be removed, false otherwise.

Throws:

IOException

pruneAllPositions

public abstract boolean pruneAllPositions(TermPositions termPositions,
                                          Term t)
                                   throws IOException

Prune all postings per term (invoked once per term per doc)

Parameters:

termPositions - positioned term positions. Implementations MUST NOT advance this by calling TermPositions methods that advance either the position pointer (next, skipTo) or term pointer (seek).

t - current term

Returns:

true if the current posting should be removed, false otherwise.

Throws:

IOException

pruneSomePositions

public abstract int pruneSomePositions(int docNum,
                                       int[] positions,
                                       Term curTerm)

Prune some postings per term (invoked once per term per doc).

Parameters:

docNum - current document number

positions - original term positions in the document (and indirectly term frequency)

curTerm - current term

Returns:

0 if no postings are to be removed, or positive number to indicate how many postings need to be removed. The same number of entries in the positions array must be set to -1 to indicate which positions to remove.