Language Analysis

In this example, text is sorted according to the default German rules provided by ICU4J.

Locales are typically defined as a combination of language and country, but you can specify just the language if you want. For example, if you specify "de" as the language, you will get sorting that works well for the German language. If you specify "de" as the language and "CH" as the country, you will get German sorting specifically tailored for Switzerland.

In the example above, we defined the strength as "primary". The strength of the collation determines how strict the sort order will be, but it also depends upon the language. For example, in English, "primary" strength ignores differences in case and accents.

Another example:

The type will be used for the fields where the data contains Polish text. The "secondary" strength will ignore case differences, but, unlike "primary" strength, a letter with diacritic(s) will be sorted differently from the same base letter without diacritics.

An example using the "city_sort" field to sort:

There are two approaches to supporting multiple languages: if there is a small list of languages you wish to support, consider defining collated fields for each language and using copyField. However, adding a large number of sort fields can increase disk and indexing costs. An alternative approach is to use the Unicode default collator.

The Unicode default or ROOT locale has rules that are designed to work well for most languages. To use the default locale, simply define the locale as the empty string. This Unicode default sort is still significantly more advanced than the standard Solr sort.

You can define your own set of sorting rules. It’s easiest to take existing rules that are close to what you want and customize them.

In the example below, we create a custom rule set for German called DIN 5007-2. This rule set treats umlauts in German differently: it treats ö as equivalent to oe, ä as equivalent to ae, and ü as equivalent to ue. For more information, see the ICU RuleBasedCollator javadocs.

This example shows how to create a custom rule set for solr.ICUCollationField and dump it to a file:

This rule set can now be used for custom collation in Solr:

As mentioned above, ICU Unicode Collation is better in several ways than JDK Collation, but if you cannot use ICU4J for some reason, you can use solr.CollationField.

The principles of JDK Collation are the same as those of ICU Collation; you just specify language, country and variant arguments instead of the combined locale argument.

Arguments for solr.CollationField, specified as attributes within the <fieldtype> element:

Using a System collator (see Oracle’s list of locales supported in Java):

Using a Tailored ruleset:

This filter converts alphabetic, numeric, and symbolic Unicode characters which are not in the first 127 ASCII characters (the "Basic Latin" Unicode block) into their ASCII equivalents, if one exists. Only those characters with reasonable ASCII alternatives are converted.

This can increase recall by causing more matches. On the other hand, it can reduce precision because language-specific character differences may be lost.

Factory class: solr.ASCIIFoldingFilterFactory

Arguments: None

Example:

In: "Björn Ångström"

Tokenizer to Filter: "Björn", "Ångström"

Out: "Bjorn", "Angstrom"

This filter converts any character in the Unicode "Decimal Number" general category (Nd) into their equivalent Basic Latin digits (0-9).

This can increase recall by causing more matches. On the other hand, it can reduce precision because language-specific character differences may be lost.

Factory class: solr.DecimalDigitFilterFactory

Arguments: None

Example:

The OpenNLP Tokenizer takes two language-specific binary model files as parameters: a sentence detector model and a tokenizer model. The last token in each sentence is flagged, so that following OpenNLP-based filters can use this information to apply operations to tokens one sentence at a time. See the OpenNLP website for information on downloading pre-trained models.

Factory class: solr.OpenNLPTokenizerFactory

Arguments:

Example:

This filter sets each token’s type attribute to the part of speech (POS) assigned by the configured model. See the OpenNLP website for information on downloading pre-trained models.

Factory class: solr.OpenNLPPOSFilterFactory

Arguments:

Examples:

The OpenNLP tokenizer will tokenize punctuation, which is useful for following token filters, but ordinarily you don’t want to include punctuation in your index, so the TypeTokenFilter (described here) is included in the examples below, with stop.pos.txt containing the following:

Index the POS for each token as a payload:

Index the POS for each token as a synonym, after prefixing the POS with "@" (see the TypeAsSynonymFilter description):

Only index nouns - the keep.pos.txt file contains lines NN, NNS, NNP and NNPS:

This filter sets each token’s type attribute based on the output of an OpenNLP phrase chunking model. The chunk labels replace the POS tags that previously were in each token’s type attribute. See the OpenNLP website for information on downloading pre-trained models.

Prerequisite: the OpenNLP Tokenizer and the OpenNLP Part-Of-Speech Filter must precede this filter.

Factory class: solr.OpenNLPChunkerFilter

Arguments:

Examples:

Index the phrase chunk label for each token as a payload:

Index the phrase chunk label for each token as a synonym, after prefixing it with "#" (see the TypeAsSynonymFilter description):

This filter replaces the text of each token with its lemma. Both a dictionary-based lemmatizer and a model-based lemmatizer are supported. If both are configured, the dictionary-based lemmatizer is tried first, and then the model-based lemmatizer is consulted for out-of-vocabulary tokens. See the OpenNLP website for information on downloading pre-trained models.

Factory class: solr.OpenNLPLemmatizerFilter

Arguments:

Either dictionary or lemmatizerModel must be provided, and both may be provided - see the examples below:

Examples:

Perform dictionary-based lemmatization, and fall back to model-based lemmatization for out-of-vocabulary tokens (see the OpenNLP Part-Of-Speech Filter section above for information about using TypeTokenFilter to avoid indexing punctuation):

Perform dictionary-based lemmatization only:

Perform model-based lemmatization only, preserving the original token and emitting the lemma as a synonym (see the KeywordRepeatFilterFactory description)):

Solr provides support for the Light-10 (PDF) stemming algorithm, and Lucene includes an example stopword list.

This algorithm defines both character normalization and stemming, so these are split into two filters to provide more flexibility.

Factory classes: solr.ArabicStemFilterFactory, solr.ArabicNormalizationFilterFactory

Arguments: None

Example:

This is a Java filter written specifically for stemming the Brazilian dialect of the Portuguese language. It uses the Lucene class org.apache.lucene.analysis.br.BrazilianStemmer. Although that stemmer can be configured to use a list of protected words (which should not be stemmed), this factory does not accept any arguments to specify such a list.

Factory class: solr.BrazilianStemFilterFactory

Arguments: None

Example:

In: "praia praias"

Tokenizer to Filter: "praia", "praias"

Out: "pra", "pra"

Solr includes a light stemmer for Bulgarian, following this algorithm (PDF), and Lucene includes an example stopword list.

Factory class: solr.BulgarianStemFilterFactory

Arguments: None

Example:

Solr can stem Catalan using the Snowball Porter Stemmer with an argument of language="Catalan". Solr includes a set of contractions for Catalan, which can be stripped using solr.ElisionFilterFactory.

Factory class: solr.SnowballPorterFilterFactory

Arguments:

Example:

In: "llengües llengua"

Tokenizer to Filter: "llengües"(1) "llengua"(2),

Out: "llengu"(1), "llengu"(2)

The default configuration of the ICU Tokenizer is suitable for Traditional Chinese text. It follows the Word Break rules from the Unicode Text Segmentation algorithm for non-Chinese text, and uses a dictionary to segment Chinese words. To use this tokenizer, you must add additional .jars to Solr’s classpath (as described in the section Lib Directives in SolrConfig). See the solr/contrib/analysis-extras/README.txt for information on which jars you need to add to your SOLR_HOME/lib.

Standard Tokenizer can also be used to tokenize Traditional Chinese text. Following the Word Break rules from the Unicode Text Segmentation algorithm, it produces one token per Chinese character. When combined with CJK Bigram Filter, overlapping bigrams of Chinese characters are formed.

CJK Width Filter folds fullwidth ASCII variants into the equivalent Basic Latin forms.

Examples:

Forms bigrams (overlapping 2-character sequences) of CJK characters that are generated from Standard Tokenizer or ICU Tokenizer.

By default, all CJK characters produce bigrams, but finer grained control is available by specifying orthographic type arguments han, hiragana, katakana, and hangul. When set to false, characters of the corresponding type will be passed through as unigrams, and will not be included in any bigrams.

When a CJK character has no adjacent characters to form a bigram, it is output in unigram form. If you want to always output both unigrams and bigrams, set the outputUnigrams argument to true.

In all cases, all non-CJK input is passed through unmodified.

Arguments:

See the example under Traditional Chinese.

For Simplified Chinese, Solr provides support for Chinese sentence and word segmentation with the HMM Chinese Tokenizer. This component includes a large dictionary and segments Chinese text into words with the Hidden Markov Model. To use this tokenizer, you must add additional .jars to Solr’s classpath (as described in the section Lib Directives in SolrConfig). See the solr/contrib/analysis-extras/README.txt for information on which jars you need to add to your SOLR_HOME/lib.

The default configuration of the ICU Tokenizer is also suitable for Simplified Chinese text. It follows the Word Break rules from the Unicode Text Segmentation algorithm for non-Chinese text, and uses a dictionary to segment Chinese words. To use this tokenizer, you must add additional .jars to Solr’s classpath (as described in the section Lib Directives in SolrConfig). See the solr/contrib/analysis-extras/README.txt for information on which jars you need to add to your SOLR_HOME/lib.

Also useful for Chinese analysis:

CJK Width Filter folds fullwidth ASCII variants into the equivalent Basic Latin forms, and folds halfwidth Katakana variants into their equivalent fullwidth forms.

Examples:

For Simplified Chinese, Solr provides support for Chinese sentence and word segmentation with the solr.HMMChineseTokenizerFactory in the analysis-extras contrib module. This component includes a large dictionary and segments Chinese text into words with the Hidden Markov Model. To use this tokenizer, see solr/contrib/analysis-extras/README.txt for instructions on which jars you need to add to your solr_home/lib.

Factory class: solr.HMMChineseTokenizerFactory

Arguments: None

Examples:

To use the default setup with fallback to English Porter stemmer for English words, use:

<analyzer class="org.apache.lucene.analysis.cn.smart.SmartChineseAnalyzer"/>

Or to configure your own analysis setup, use the solr.HMMChineseTokenizerFactory along with your custom filter setup. See an example of this in the Simplified Chinese section.

Solr includes a light stemmer for Czech, following this algorithm, and Lucene includes an example stopword list.

Factory class: solr.CzechStemFilterFactory

Arguments: None

Example:

In: "prezidenští, prezidenta, prezidentského"

Tokenizer to Filter: "prezidenští", "prezidenta", "prezidentského"

Out: "preziden", "preziden", "preziden"

Solr can stem Danish using the Snowball Porter Stemmer with an argument of language="Danish".

Also relevant are the Scandinavian normalization filters.

Factory class: solr.SnowballPorterFilterFactory

Arguments:

Example:

In: "undersøg undersøgelse"

Tokenizer to Filter: "undersøg"(1) "undersøgelse"(2),

Out: "undersøg"(1), "undersøg"(2)

Solr can stem Dutch using the Snowball Porter Stemmer with an argument of language="Dutch".

Factory class: solr.SnowballPorterFilterFactory

Arguments:

Example:

In: "kanaal kanalen"

Tokenizer to Filter: "kanaal", "kanalen"

Out: "kanal", "kanal"

Solr includes support for stemming Finnish, and Lucene includes an example stopword list.

Factory class: solr.FinnishLightStemFilterFactory

Arguments: None

Example:

In: "kala kalat"

Tokenizer to Filter: "kala", "kalat"

Out: "kala", "kala"

Solr includes a stemmer for Galician following this algorithm, and Lucene includes an example stopword list.

Factory class: solr.GalicianStemFilterFactory

Arguments: None

Example:

In: "felizmente Luzes"

Tokenizer to Filter: "felizmente", "luzes"

Out: "feliz", "luz"

Solr includes four stemmers for German: one in the solr.SnowballPorterFilterFactory language="German", a stemmer called solr.GermanStemFilterFactory, a lighter stemmer called solr.GermanLightStemFilterFactory, and an even less aggressive stemmer called solr.GermanMinimalStemFilterFactory. Lucene includes an example stopword list.

Factory classes: solr.GermanStemFilterFactory, solr.LightGermanStemFilterFactory, solr.MinimalGermanStemFilterFactory

Arguments: None

Examples:

In: "haus häuser"

Tokenizer to Filter: "haus", "häuser"

Out: "haus", "haus"

This filter converts uppercase letters in the Greek character set to the equivalent lowercase character.

Factory class: solr.GreekLowerCaseFilterFactory

Arguments: None

Example:

Solr includes support for stemming Hindi following this algorithm (PDF), support for common spelling differences through the solr.HindiNormalizationFilterFactory, support for encoding differences through the solr.IndicNormalizationFilterFactory following this algorithm, and Lucene includes an example stopword list.

Factory classes: solr.IndicNormalizationFilterFactory, solr.HindiNormalizationFilterFactory, solr.HindiStemFilterFactory

Arguments: None

Example:

Solr includes support for stemming Indonesian (Bahasa Indonesia) following this algorithm (PDF), and Lucene includes an example stopword list.

Factory class: solr.IndonesianStemFilterFactory

Arguments: None

Example:

In: "sebagai sebagainya"

Tokenizer to Filter: "sebagai", "sebagainya"

Out: "bagai", "bagai"

Solr includes two stemmers for Italian: one in the solr.SnowballPorterFilterFactory language="Italian", and a lighter stemmer called solr.ItalianLightStemFilterFactory. Lucene includes an example stopword list.

Factory class: solr.ItalianStemFilterFactory

Arguments: None

Example:

In: "propaga propagare propagamento"

Tokenizer to Filter: "propaga", "propagare", "propagamento"

Out: "propag", "propag", "propag"

Solr can stem Irish using the Snowball Porter Stemmer with an argument of language="Irish". Solr includes solr.IrishLowerCaseFilterFactory, which can handle Irish-specific constructs. Solr also includes a set of contractions for Irish which can be stripped using solr.ElisionFilterFactory.

Factory class: solr.SnowballPorterFilterFactory

Arguments:

Example:

In: "siopadóireacht síceapatacha b’fhearr m’athair"

Tokenizer to Filter: "siopadóireacht", "síceapatacha", "b’fhearr", "m’athair"

Out: "siopadóir", "síceapaite", "fearr", "athair"

Solr includes support for analyzing Japanese, via the Lucene Kuromoji morphological analyzer, which includes several analysis components - more details on each below:

Also useful for Japanese analysis, from lucene-analyzers-common:

Lucene provides support, in addition to UAX#29 word break rules, for Hebrew’s use of the double and single quote characters, and for segmenting Lao, Myanmar, and Khmer into syllables with the solr.ICUTokenizerFactory in the analysis-extras contrib module. To use this tokenizer, see solr/contrib/analysis-extras/README.txt for instructions on which jars you need to add to your solr_home/lib.

See the ICUTokenizer for more information.

Solr includes support for stemming Latvian, and Lucene includes an example stopword list.

Factory class: solr.LatvianStemFilterFactory

Arguments: None

Example:

In: "tirgiem tirgus"

Tokenizer to Filter: "tirgiem", "tirgus"

Out: "tirg", "tirg"

Solr includes two classes for stemming Norwegian, NorwegianLightStemFilterFactory and NorwegianMinimalStemFilterFactory. Lucene includes an example stopword list.

Another option is to use the Snowball Porter Stemmer with an argument of language="Norwegian".

Also relevant are the Scandinavian normalization filters.

Solr provides support for Polish stemming with the solr.StempelPolishStemFilterFactory, and solr.MorphologikFilterFactory for lemmatization, in the contrib/analysis-extras module. The solr.StempelPolishStemFilterFactory component includes an algorithmic stemmer with tables for Polish. To use either of these filters, see solr/contrib/analysis-extras/README.txt for instructions on which jars you need to add to your solr_home/lib.

Factory class: solr.StempelPolishStemFilterFactory and solr.MorfologikFilterFactory

Arguments: None

Example:

In: ""studenta studenci"

Tokenizer to Filter: "studenta", "studenci"

Out: "student", "student"

More information about the Stempel stemmer is available in the Lucene javadocs.

Note the lower case filter is applied after the Morfologik stemmer; this is because the Polish dictionary contains proper names and then proper term case may be important to resolve disambiguities (or even lookup the correct lemma at all).

The Morfologik dictionary parameter value is a constant specifying which dictionary to choose. The dictionary resource must be named path/to/language.dict and have an associated .info metadata file. See the Morfologik project for details. If the dictionary attribute is not provided, the Polish dictionary is loaded and used by default.

Solr includes four stemmers for Portuguese: one in the solr.SnowballPorterFilterFactory, an alternative stemmer called solr.PortugueseStemFilterFactory, a lighter stemmer called solr.PortugueseLightStemFilterFactory, and an even less aggressive stemmer called solr.PortugueseMinimalStemFilterFactory. Lucene includes an example stopword list.

Factory classes: solr.PortugueseStemFilterFactory, solr.PortugueseLightStemFilterFactory, solr.PortugueseMinimalStemFilterFactory

Arguments: None

Example:

In: "praia praias"

Tokenizer to Filter: "praia", "praias"

Out: "pra", "pra"

Solr can stem Romanian using the Snowball Porter Stemmer with an argument of language="Romanian".

Factory class: solr.SnowballPorterFilterFactory

Arguments:

Example:

Scandinavian is a language group spanning three languages Norwegian, Swedish and Danish which are very similar.

Swedish å, ä, ö are in fact the same letters as Norwegian and Danish å, æ, ø and thus interchangeable when used between these languages. They are however folded differently when people type them on a keyboard lacking these characters.

In that situation almost all Swedish people use a, a, o instead of å, ä, ö. Norwegians and Danes on the other hand usually type aa, ae and oe instead of å, æ and ø. Some do however use a, a, o, oo, ao and sometimes permutations of everything above.

There are two filters for helping with normalization between Scandinavian languages: one is solr.ScandinavianNormalizationFilterFactory trying to preserve the special characters (æäöå) and another solr.ScandinavianFoldingFilterFactory which folds these to the more broad ø/ö->o, etc.

See also each language section for other relevant filters.

Solr includes two stemmers for Spanish: one in the solr.SnowballPorterFilterFactory language="Spanish", and a lighter stemmer called solr.SpanishLightStemFilterFactory. Lucene includes an example stopword list.

Factory class: solr.SpanishStemFilterFactory

Arguments: None

Example:

In: "torear toreara torearlo"

Tokenizer to Filter: "torear", "toreara", "torearlo"

Out: "tor", "tor", "tor"

This filter converts sequences of Thai characters into individual Thai words. Unlike European languages, Thai does not use whitespace to delimit words.

Factory class: solr.ThaiTokenizerFactory

Arguments: None

Example:

Solr includes support for stemming Turkish with the solr.SnowballPorterFilterFactory; support for case-insensitive search with the solr.TurkishLowerCaseFilterFactory; support for stripping apostrophes and following suffixes with solr.ApostropheFilterFactory (see Role of Apostrophes in Turkish Information Retrieval); support for a form of stemming that truncating tokens at a configurable maximum length through the solr.TruncateTokenFilterFactory (see Information Retrieval on Turkish Texts); and Lucene includes an example stopword list.

Factory class: solr.TurkishLowerCaseFilterFactory

Arguments: None

Example:

Another example, illustrating diacritics-insensitive search:

Solr provides support for Ukrainian lemmatization with the solr.MorphologikFilterFactory, in the contrib/analysis-extras module. To use this filter, see solr/contrib/analysis-extras/README.txt for instructions on which jars you need to add to your solr_home/lib.

Lucene also includes an example Ukrainian stopword list, in the lucene-analyzers-morfologik jar.

Factory class: solr.MorfologikFilterFactory

Arguments:

Example:

The Morfologik dictionary parameter value is a constant specifying which dictionary to choose. The dictionary resource must be named path/to/language.dict and have an associated .info metadata file. See the Morfologik project for details. If the dictionary attribute is not provided, the Polish dictionary is loaded and used by default.