NAME

Elastic::Manual::Analysis - Controlling how your attributes are indexed

VERSION

version 0.52

INTRODUCTION

Analysis plus the inverted index is the magic that makes full text search so powerful.

Inverted index

The inverted index is an index structure which is very efficient for performing fast full text searches. For example, let's say we have three documents (taken from recent headlines):

    1: Apple's patent absurdity exposed at last
    2: Apple patents surfaced last year
    3: Finally a sane judgement on patent trolls

The first stage is to "tokenize" the text in each document. We'll explain how this happens later, but the tokenization process could give us:

    1: [ 'appl',  'patent', 'absurd',    'expos',  'last'  ]
    2: [ 'appl',  'patent', 'surfac',    'last',   'year'  ]
    3: [ 'final', 'sane',   'judgement', 'patent', 'troll' ]

Next, we invert this list, listing each token we find, and the documents that contain that token:

             |  Doc_1  |  Doc_2  |  Doc_3
  -----------+---------+---------+--------
  absurd     |  xxx    |         |
  appl       |  xxx    |  xxx    |
  expos      |  xxx    |         |
  final      |         |         |  xxx
  judgement  |         |         |  xxx
  last       |  xxx    |  xxx    |
  patent     |  xxx    |  xxx    |  xxx
  sane       |         |         |  xxx
  surfac     |         |  xxx    |
  troll      |         |         |  xxx
  year       |         |  xxx    |
  -----------+---------+---------+--------

Now we can search the index. Let's say we want to search for "apple patent troll judgement". First we need to tokenize the search terms in the same way that we did when we created the index. (We can't find what isn't in the index.) This gives us:

    [ 'appl', 'patent', 'troll', 'judgement' ]

If we compare these terms to the index we get:

             |  Doc_1  |  Doc_2  |  Doc_3
  -----------+---------+---------+--------
  appl       |  xxx    |  xxx    |
  judgement  |         |         |  xxx
  patent     |  xxx    |  xxx    |  xxx
  troll      |         |         |  xxx
  -----------+---------+---------+--------
  Relevance  |  2      |  2      | 3

It is easy to see the Doc_3 is the most relevant document, because it has more of the search terms than either of the other two.

This is an example of a simple inverted index. Lucene (the full text search engine used by Elasticsearch) has a much more sophisticated scoring system which can consider, amongst other things:

  • Term frequency

    How often does a term/token appear in a document

  • Inverse document frequency

    How often does the term appear throughout all the documents in the index

  • Coord

    The number of terms in the query that were found in the document

  • Length norm

    Measure of the importance of a term according to the total number of terms in the field

  • Term position

    How close is each term to related terms

Analysis / Tokenization

What is obvious from the above, is that the way we "analyze" (ie convert text into terms) is very important. Also, it is important that we use the same analyzer on the text that we store in the index (ie at index time) and on the search keywords (ie at search time).

You can only find what is stored in the index.

An analyzer is a package containing 0 or more character filters, one tokenizer and 0 or more token filters. Text is fed into the analyzer, and it produces a stream of tokens, with other metadata such as the position of each term relative to the others:

Character filters

Character filters work on a character-by-character basis, effecting some sort of transformation. For instance, you may have a character filter which converts particular characters into others, eg:

    ß  => ss
    œ  => oe
    ff  => ff

or a character filter which strips out HTML and converts HTML entities into their unicode characters.

Tokenizer

A tokenizer then breaks up the stream of characters into individual tokens. Take this sentence, for example:

    To order part #ABC-123 email "joe@foo.org"

The whitespace tokenizer would produce:

    'To', 'order', 'part', '#ABC-123', 'email', '"joe@foo.org"'

The letter tokenizer would produce:

    'To', 'order', 'part', 'ABC', 'email', 'joe', 'foo', 'org'

The keyword tokenizer does NO tokenization - it just passes the input straight through, which isn't terribly useful in this example, but is useful for (eg) a tag cloud, where you don't want 'modern perl' to be broken up into two separate tokens: 'modern' and 'perl':

    'To order part #ABC-123 email "joe@foo.org"'

The standard tokenizer which uses Unicodes word boundary rules would produce:

    'To', 'order', 'part', 'ABC', '123', 'email', 'joe', 'foo.org'

And the UAX URL Email tokenizer, which is like the standard tokenizer, but which recognises emails, URLs etc as single entities would produce:

    'To', 'order', 'part', 'ABC', '123', 'email', 'joe@foo.org'
Token filters

Optionally, the token stream produced above can be passed through one or more token filters, which can remove, change or add tokens. For instance:

The lowercase filter would convert all uppercase letters to lowercase:

    'to', 'order', 'part', 'abc', '123', 'email', 'joe@foo.org'

Then the stopword filter could remove common (and thus generally irrelevant stopwords):

    'order', 'part', 'abc', '123', 'email', 'joe@foo.org'

USING BUILT IN ANALYZERS

There are a number of analyzers that are built in to Elasticsearch, which can be used directly, (see http://www.elasticsearch.org/guide/en/elasticsearch/reference/current/analysis.html for a current list) which can be used directly.

For instance the english analyzer "stems" English words into their root, eg "foxy", "fox" and "foxes" would all be converted to the term "fox".

    has 'title' => (
        is       => 'ro',
        isa      => 'Str',
        analyzer => 'english'
    );

Configurable analyzers

Some analyzers are configurable, for instance, the english analyzer accepts a custom list of stop-words. In your model class, you configure the built-in analyzer under a new name:

    package MyApp;

    use Elastic::Model;

    has_namespace 'myapp'. {
        post => 'MyApp::Post'
    };

    has_analyzer 'custom_english' => (
        type        => 'english',
        stopwords   => ['a','the','for']
    );

Then in your doc class, you specify the analyzer using the new name:

    package MyApp::Post;

    use Elastic::Doc;

    has 'title' => (
        is       => 'ro',
        isa      => 'Str',
        analyzer => 'custom_english'
    );

DEFINING CUSTOM ANALYZERS

You can also combine character filters, a tokenizer and token filters together to form a custom analyzer, For instance, in your model class:

    has_char_filter 'charmap' => (
        type        => 'mapping',
        mapping     => [ 'ß=>ss','œ=>oe','ff=>ff' ]
    );

    has_tokenizer 'ascii_letters', => (
        type        => 'pattern',
        pattern     => '[^a-z]+',
        flags       => 'CASE_INSENSITIVE',
    );

    has_filter 'edge_ngrams' => (
        type        => 'edge_ngrams',
        min_gram    => 2,
        max_gram    => 20
    );

    has_analyzer 'my_custom_analyzer' => (
        type        => 'custom',
        char_filter => ['mapping'],
        tokenizer   => 'ascii_letters',
        filter      => ['lowercase','stop','edge_ngrams']
    );

PUTTING YOUR ANALYZERS LIVE

Analyzers can only be created on a new index. You can't change them on an existing index, because the data stored in the index would be wrong.

    my $myapp = MyApp->new;
    my $index = $myapp->namespace('myapp')->index;

    $index->delete;     # remove old index
    $index->create;     # create new index with new analyzers

It is, however, possible to ADD new analyzers to a closed index:

    $index->close;
    $index->update_analyzers;
    $index->open;

AUTHOR

Clinton Gormley <drtech@cpan.org>

COPYRIGHT AND LICENSE

This software is copyright (c) 2015 by Clinton Gormley.

This is free software; you can redistribute it and/or modify it under the same terms as the Perl 5 programming language system itself.