Package org.apache.lucene.search.similarities
Similarity serves as the base for ranking
functions. For searching, users can employ the models already implemented or create their own by
extending one of the classes in this package.
Table Of Contents
Summary of the Ranking Methods
BM25Similarity is an optimized implementation of
the successful Okapi BM25 model.
ClassicSimilarity is the original Lucene scoring
function. It is based on the Vector
Space Model. For more information, see TFIDFSimilarity.
SimilarityBase provides a basic implementation
of the Similarity contract and exposes a highly simplified interface, which makes it an ideal
starting point for new ranking functions. Lucene ships the following methods built on SimilarityBase:
- Amati and Rijsbergen's DFR framework;
- Clinchant and Gaussier's Information-based models for IR;
- The implementation of two language models from Zhai and Lafferty's paper.
- Divergence from independence models as described in "IRRA at TREC 2012" (Dinçer).
SimilarityBase is not optimized to the same
extent as ClassicSimilarity and BM25Similarity, a difference in performance is to be
expected when using the methods listed above. However, optimizations can always be implemented in
subclasses; see below.
Changing Similarity
Chances are the available Similarities are sufficient for all your searching needs. However,
in some applications it may be necessary to customize your Similarity implementation. For instance, some applications do not need
to distinguish between shorter and longer documents and could set BM25's b parameter to
0.
To switch to a Similarity that encodes the
length normalization differently, one must do so for both indexing and searching, and the changes
must happen before either of these actions take place. Note that all of Lucene's built-in
similarities - and more generally all Similarity
sub-classes that don't override Similarity.computeNorm(org.apache.lucene.index.FieldInvertState)
- encode the length normalization factor the same way, so it is fine to change the similarity at
search-time without recreating the index.
To make this change, implement your own Similarity (likely you'll want to simply subclass SimilarityBase), and then register the new class by
calling IndexWriterConfig.setSimilarity(Similarity) before
indexing and IndexSearcher.setSimilarity(Similarity) before
searching.
Tuning BM25Similarity
BM25Similarity has two parameters that may be
tuned:
k1, which calibrates term frequency saturation and must be positive or null. A value of0makes term frequency completely ignored, making documents scored only based on the value of theIDFof the matched terms. Higher values ofk1increase the impact of term frequency on the final score. Default value is1.2.b, which controls how much document length should normalize term frequency values and must be in[0, 1]. A value of0disables length normalization completely. Default value is0.75.
Extending SimilarityBase
The easiest way to quickly implement a new ranking method is to extend SimilarityBase, which provides basic implementations for
the low level . Subclasses are only required to implement the SimilarityBase.score(BasicStats, double, double) and
SimilarityBase.toString() methods.
Another option is to extend one of the frameworks based on SimilarityBase. These Similarities are implemented
modularly, e.g. DFRSimilarity delegates computation
of the three parts of its formula to the classes BasicModel, AfterEffect and Normalization. Instead of subclassing the Similarity, one
can simply introduce a new basic model and tell DFRSimilarity to use it.
-
ClassDescriptionThis class acts as the base class for the implementations of the first normalization of the informative content in the DFR framework.Model of the information gain based on the ratio of two Bernoulli processes.Model of the information gain based on Laplace's law of succession.Axiomatic approaches for IR.F1EXP is defined as Sum(tf(term_doc_freq)*ln(docLen)*IDF(term)) where IDF(t) = pow((N+1)/df(t), k) N=total num of docs, df=doc freqF1LOG is defined as Sum(tf(term_doc_freq)*ln(docLen)*IDF(term)) where IDF(t) = ln((N+1)/df(t)) N=total num of docs, df=doc freqF2EXP is defined as Sum(tfln(term_doc_freq, docLen)*IDF(term)) where IDF(t) = pow((N+1)/df(t), k) N=total num of docs, df=doc freqF2EXP is defined as Sum(tfln(term_doc_freq, docLen)*IDF(term)) where IDF(t) = ln((N+1)/df(t)) N=total num of docs, df=doc freqF3EXP is defined as Sum(tf(term_doc_freq)*IDF(term)-gamma(docLen, queryLen)) where IDF(t) = pow((N+1)/df(t), k) N=total num of docs, df=doc freq gamma(docLen, queryLen) = (docLen-queryLen)*queryLen*s/avdl NOTE: the gamma function of this similarity creates negative scoresF3EXP is defined as Sum(tf(term_doc_freq)*IDF(term)-gamma(docLen, queryLen)) where IDF(t) = ln((N+1)/df(t)) N=total num of docs, df=doc freq gamma(docLen, queryLen) = (docLen-queryLen)*queryLen*s/avdl NOTE: the gamma function of this similarity creates negative scoresThis class acts as the base class for the specific basic model implementations in the DFR framework.Geometric as limiting form of the Bose-Einstein model.An approximation of the I(ne) model.The basic tf-idf model of randomness.Tf-idf model of randomness, based on a mixture of Poisson and inverse document frequency.Stores all statistics commonly used ranking methods.BM25 Similarity.Simple similarity that gives terms a score that is equal to their query boost.Expert: Historical scoring implementation.Implements the Divergence from Independence (DFI) model based on Chi-square statistics (i.e., standardized Chi-squared distance from independence in term frequency tf).Implements the divergence from randomness (DFR) framework introduced in Gianni Amati and Cornelis Joost Van Rijsbergen.The probabilistic distribution used to model term occurrence in information-based models.Log-logistic distribution.The smoothed power-law (SPL) distribution for the information-based framework that is described in the original paper.Provides a framework for the family of information-based models, as described in Stéphane Clinchant and Eric Gaussier.Computes the measure of divergence from independence for DFI scoring functions.Normalized chi-squared measure of distance from independenceSaturated measure of distance from independenceStandardized measure of distance from independenceBayesian smoothing using Dirichlet priors as implemented in the Indri Search engine (http://www.lemurproject.org/indri.php).Models
p(w|C)as the number of occurrences of the term in the collection, divided by the total number of tokens+ 1.The lambda (λw) parameter in information-based models.Computes lambda asdocFreq+1 / numberOfDocuments+1.Computes lambda astotalTermFreq+1 / numberOfDocuments+1.Bayesian smoothing using Dirichlet priors.Language model based on the Jelinek-Mercer smoothing method.Abstract superclass for language modeling Similarities.A strategy for computing the collection language model.Modelsp(w|C)as the number of occurrences of the term in the collection, divided by the total number of tokens+ 1.Stores the collection distribution of the current term.Implements the CombSUM method for combining evidence from multiple similarity values described in: Joseph A.This class acts as the base class for the implementations of the term frequency normalization methods in the DFR framework.Implementation used when there is no normalization.Normalization model that assumes a uniform distribution of the term frequency.Normalization model in which the term frequency is inversely related to the length.Dirichlet Priors normalizationPareto-Zipf NormalizationProvides the ability to use a differentSimilarityfor different fields.Similarity that returns the raw TF as score.Similarity defines the components of Lucene scoring.Stores the weight for a query across the indexed collection.A subclass ofSimilaritythat provides a simplified API for its descendants.Implementation ofSimilaritywith the Vector Space Model.