elasticsearch/docs/plugins/analysis-kuromoji.asciidoc

[[analysis-kuromoji]]
=== Japanese (kuromoji) Analysis Plugin

The Japanese (kuromoji) Analysis plugin integrates Lucene kuromoji analysis
module into {es}.

:plugin_name: analysis-kuromoji
include::install_remove.asciidoc[]

[[analysis-kuromoji-analyzer]]
==== `kuromoji` analyzer

The `kuromoji` analyzer consists of the following tokenizer and token filters:

* <<analysis-kuromoji-tokenizer,`kuromoji_tokenizer`>>
* <<analysis-kuromoji-baseform,`kuromoji_baseform`>> token filter
* <<analysis-kuromoji-speech,`kuromoji_part_of_speech`>> token filter
* {ref}/analysis-cjk-width-tokenfilter.html[`cjk_width`] token filter
* <<analysis-kuromoji-stop,`ja_stop`>> token filter
* <<analysis-kuromoji-stemmer,`kuromoji_stemmer`>> token filter
* {ref}/analysis-lowercase-tokenfilter.html[`lowercase`] token filter

It supports the `mode` and `user_dictionary` settings from
<<analysis-kuromoji-tokenizer,`kuromoji_tokenizer`>>.

[discrete]
[[kuromoji-analyzer-normalize-full-width-characters]]
==== Normalize full-width characters

The `kuromoji_tokenizer` tokenizer uses characters from the MeCab-IPADIC
dictionary to split text into tokens. The dictionary includes some full-width
characters, such as `ｏ` and `ｆ`. If a text contains full-width characters,
the tokenizer can produce unexpected tokens.

For example, the `kuromoji_tokenizer` tokenizer converts the text
`Ｃｕｌｔｕｒｅ　ｏｆ　Ｊａｐａｎ` to the tokens `[ culture, o, f, japan ]`
instead of `[ culture, of, japan ]`.

To avoid this, add the <<analysis-icu-normalization-charfilter,`icu_normalizer`
character filter>> to a custom analyzer based on the `kuromoji` analyzer. The
`icu_normalizer` character filter converts full-width characters to their normal
equivalents.

First, duplicate the `kuromoji` analyzer to create the basis for a custom
analyzer. Then add the `icu_normalizer` character filter to the custom analyzer.
For example:

[source,console]
----
PUT index-00001
{
  "settings": {
    "index": {
      "analysis": {
        "analyzer": {
          "kuromoji_normalize": {                 <1>
            "char_filter": [
              "icu_normalizer"                    <2>
            ],
            "tokenizer": "kuromoji_tokenizer",
            "filter": [
              "kuromoji_baseform",
              "kuromoji_part_of_speech",
              "cjk_width",
              "ja_stop",
              "kuromoji_stemmer",
              "lowercase"
            ]
          }
        }
      }
    }
  }
}
----
<1> Creates a new custom analyzer, `kuromoji_normalize`, based on the `kuromoji`
analyzer.
<2> Adds the `icu_normalizer` character filter to the analyzer.


[[analysis-kuromoji-charfilter]]
==== `kuromoji_iteration_mark` character filter

The `kuromoji_iteration_mark` normalizes Japanese horizontal iteration marks
(_odoriji_) to their expanded form. It accepts the following settings:

`normalize_kanji`::

    Indicates whether kanji iteration marks should be normalize. Defaults to `true`.

`normalize_kana`::

    Indicates whether kana iteration marks should be normalized. Defaults to `true`


[[analysis-kuromoji-tokenizer]]
==== `kuromoji_tokenizer`

The `kuromoji_tokenizer` accepts the following settings:

`mode`::
+
--

The tokenization mode determines how the tokenizer handles compound and
unknown words.  It can be set to:

`normal`::

    Normal segmentation, no decomposition for compounds. Example output:

    関西国際空港
    アブラカダブラ

`search`::

    Segmentation geared towards search. This includes a decompounding process
    for long nouns, also including the full compound token as a synonym.
    Example output:

    関西, 関西国際空港, 国際, 空港
    アブラカダブラ

`extended`::

    Extended mode outputs unigrams for unknown words. Example output:

    関西, 関西国際空港, 国際, 空港
    ア, ブ, ラ, カ, ダ, ブ, ラ
--

`discard_punctuation`::

    Whether punctuation should be discarded from the output. Defaults to `true`.

`user_dictionary`::
+
--
The Kuromoji tokenizer uses the MeCab-IPADIC dictionary by default. A `user_dictionary`
may be appended to the default dictionary. The dictionary should have the following CSV format:

[source,csv]
-----------------------
<text>,<token 1> ... <token n>,<reading 1> ... <reading n>,<part-of-speech tag>
-----------------------
--

As a demonstration of how the user dictionary can be used, save the following
dictionary to `$ES_HOME/config/userdict_ja.txt`:

[source,csv]
-----------------------
東京スカイツリー,東京 スカイツリー,トウキョウ スカイツリー,カスタム名詞
-----------------------

--

You can also inline the rules directly in the tokenizer definition using
the `user_dictionary_rules` option:

[source,console]
--------------------------------------------------
PUT nori_sample
{
  "settings": {
    "index": {
      "analysis": {
        "tokenizer": {
          "kuromoji_user_dict": {
            "type": "kuromoji_tokenizer",
            "mode": "extended",
            "user_dictionary_rules": ["東京スカイツリー,東京 スカイツリー,トウキョウ スカイツリー,カスタム名詞"]
          }
        },
        "analyzer": {
          "my_analyzer": {
            "type": "custom",
            "tokenizer": "kuromoji_user_dict"
          }
        }
      }
    }
  }
}
--------------------------------------------------
--

`nbest_cost`/`nbest_examples`::
+
--
Additional expert user parameters `nbest_cost` and `nbest_examples` can be used
to include additional tokens that most likely according to the statistical model.
If both parameters are used, the largest number of both is applied.

`nbest_cost`::

    The `nbest_cost` parameter specifies an additional Viterbi cost.
    The KuromojiTokenizer will include all tokens in Viterbi paths that are
    within the nbest_cost value of the best path.

`nbest_examples`::

    The `nbest_examples` can be used to find a `nbest_cost` value based on examples.
    For example, a value of /箱根山-箱根/成田空港-成田/ indicates that in the texts,
    箱根山 (Mt. Hakone) and 成田空港 (Narita Airport) we'd like a cost that gives is us
    箱根 (Hakone) and 成田 (Narita).
--


Then create an analyzer as follows:

[source,console]
--------------------------------------------------
PUT kuromoji_sample
{
  "settings": {
    "index": {
      "analysis": {
        "tokenizer": {
          "kuromoji_user_dict": {
            "type": "kuromoji_tokenizer",
            "mode": "extended",
            "discard_punctuation": "false",
            "user_dictionary": "userdict_ja.txt"
          }
        },
        "analyzer": {
          "my_analyzer": {
            "type": "custom",
            "tokenizer": "kuromoji_user_dict"
          }
        }
      }
    }
  }
}

GET kuromoji_sample/_analyze
{
  "analyzer": "my_analyzer",
  "text": "東京スカイツリー"
}
--------------------------------------------------

The above `analyze` request returns the following:

[source,console-result]
--------------------------------------------------
{
  "tokens" : [ {
    "token" : "東京",
    "start_offset" : 0,
    "end_offset" : 2,
    "type" : "word",
    "position" : 0
  }, {
    "token" : "スカイツリー",
    "start_offset" : 2,
    "end_offset" : 8,
    "type" : "word",
    "position" : 1
  } ]
}
--------------------------------------------------

`discard_compound_token`::
    Whether original compound tokens should be discarded from the output with `search` mode. Defaults to `false`.
    Example output with `search` or `extended` mode and this option `true`:

    関西, 国際, 空港

NOTE: If a text contains full-width characters, the `kuromoji_tokenizer`
tokenizer can produce unexpected tokens. To avoid this, add the
<<analysis-icu-normalization-charfilter,`icu_normalizer` character filter>> to
your analyzer. See <<kuromoji-analyzer-normalize-full-width-characters>>.


[[analysis-kuromoji-baseform]]
==== `kuromoji_baseform` token filter

The `kuromoji_baseform` token filter replaces terms with their
BaseFormAttribute. This acts as a lemmatizer for verbs and adjectives. Example:

[source,console]
--------------------------------------------------
PUT kuromoji_sample
{
  "settings": {
    "index": {
      "analysis": {
        "analyzer": {
          "my_analyzer": {
            "tokenizer": "kuromoji_tokenizer",
            "filter": [
              "kuromoji_baseform"
            ]
          }
        }
      }
    }
  }
}

GET kuromoji_sample/_analyze
{
  "analyzer": "my_analyzer",
  "text": "飲み"
}
--------------------------------------------------

which responds with:

[source,console-result]
--------------------------------------------------
{
  "tokens" : [ {
    "token" : "飲む",
    "start_offset" : 0,
    "end_offset" : 2,
    "type" : "word",
    "position" : 0
  } ]
}
--------------------------------------------------


[[analysis-kuromoji-speech]]
==== `kuromoji_part_of_speech` token filter

The `kuromoji_part_of_speech` token filter removes tokens that match a set of
part-of-speech tags. It accepts the following setting:

`stoptags`::

    An array of part-of-speech tags that should be removed. It defaults to the
    `stoptags.txt` file embedded in the `lucene-analyzer-kuromoji.jar`.

For example:

[source,console]
--------------------------------------------------
PUT kuromoji_sample
{
  "settings": {
    "index": {
      "analysis": {
        "analyzer": {
          "my_analyzer": {
            "tokenizer": "kuromoji_tokenizer",
            "filter": [
              "my_posfilter"
            ]
          }
        },
        "filter": {
          "my_posfilter": {
            "type": "kuromoji_part_of_speech",
            "stoptags": [
              "助詞-格助詞-一般",
              "助詞-終助詞"
            ]
          }
        }
      }
    }
  }
}

GET kuromoji_sample/_analyze
{
  "analyzer": "my_analyzer",
  "text": "寿司がおいしいね"
}
--------------------------------------------------

Which responds with:

[source,console-result]
--------------------------------------------------
{
  "tokens" : [ {
    "token" : "寿司",
    "start_offset" : 0,
    "end_offset" : 2,
    "type" : "word",
    "position" : 0
  }, {
    "token" : "おいしい",
    "start_offset" : 3,
    "end_offset" : 7,
    "type" : "word",
    "position" : 2
  } ]
}
--------------------------------------------------


[[analysis-kuromoji-readingform]]
==== `kuromoji_readingform` token filter

The `kuromoji_readingform` token filter replaces the token with its reading
form in either katakana or romaji. It accepts the following setting:

`use_romaji`::

    Whether romaji reading form should be output instead of katakana.  Defaults to `false`.

When using the pre-defined `kuromoji_readingform` filter, `use_romaji` is set
to `true`. The default when defining a custom `kuromoji_readingform`, however,
is `false`.  The only reason to use the custom form is if you need the
katakana reading form:

[source,console]
--------------------------------------------------
PUT kuromoji_sample
{
  "settings": {
    "index": {
      "analysis": {
        "analyzer": {
          "romaji_analyzer": {
            "tokenizer": "kuromoji_tokenizer",
            "filter": [ "romaji_readingform" ]
          },
          "katakana_analyzer": {
            "tokenizer": "kuromoji_tokenizer",
            "filter": [ "katakana_readingform" ]
          }
        },
        "filter": {
          "romaji_readingform": {
            "type": "kuromoji_readingform",
            "use_romaji": true
          },
          "katakana_readingform": {
            "type": "kuromoji_readingform",
            "use_romaji": false
          }
        }
      }
    }
  }
}

GET kuromoji_sample/_analyze
{
  "analyzer": "katakana_analyzer",
  "text": "寿司" <1>
}

GET kuromoji_sample/_analyze
{
  "analyzer": "romaji_analyzer",
  "text": "寿司" <2>
}
--------------------------------------------------

<1> Returns `スシ`.
<2> Returns `sushi`.

[[analysis-kuromoji-stemmer]]
==== `kuromoji_stemmer` token filter

The `kuromoji_stemmer` token filter normalizes common katakana spelling
variations ending in a long sound character by removing this character
(U+30FC). Only full-width katakana characters are supported.

This token filter accepts the following setting:

`minimum_length`::

    Katakana words shorter than the `minimum length` are not stemmed (default
    is `4`).


[source,console]
--------------------------------------------------
PUT kuromoji_sample
{
  "settings": {
    "index": {
      "analysis": {
        "analyzer": {
          "my_analyzer": {
            "tokenizer": "kuromoji_tokenizer",
            "filter": [
              "my_katakana_stemmer"
            ]
          }
        },
        "filter": {
          "my_katakana_stemmer": {
            "type": "kuromoji_stemmer",
            "minimum_length": 4
          }
        }
      }
    }
  }
}

GET kuromoji_sample/_analyze
{
  "analyzer": "my_analyzer",
  "text": "コピー" <1>
}

GET kuromoji_sample/_analyze
{
  "analyzer": "my_analyzer",
  "text": "サーバー" <2>
}
--------------------------------------------------

<1> Returns `コピー`.
<2> Return `サーバ`.


[[analysis-kuromoji-stop]]
==== `ja_stop` token filter

The `ja_stop` token filter filters out Japanese stopwords (`_japanese_`), and
any other custom stopwords specified by the user. This filter only supports
the predefined `_japanese_` stopwords list.  If you want to use a different
predefined list, then use the
{ref}/analysis-stop-tokenfilter.html[`stop` token filter] instead.

[source,console]
--------------------------------------------------
PUT kuromoji_sample
{
  "settings": {
    "index": {
      "analysis": {
        "analyzer": {
          "analyzer_with_ja_stop": {
            "tokenizer": "kuromoji_tokenizer",
            "filter": [
              "ja_stop"
            ]
          }
        },
        "filter": {
          "ja_stop": {
            "type": "ja_stop",
            "stopwords": [
              "_japanese_",
              "ストップ"
            ]
          }
        }
      }
    }
  }
}

GET kuromoji_sample/_analyze
{
  "analyzer": "analyzer_with_ja_stop",
  "text": "ストップは消える"
}
--------------------------------------------------

The above request returns:

[source,console-result]
--------------------------------------------------
{
  "tokens" : [ {
    "token" : "消える",
    "start_offset" : 5,
    "end_offset" : 8,
    "type" : "word",
    "position" : 2
  } ]
}
--------------------------------------------------


[[analysis-kuromoji-number]]
==== `kuromoji_number` token filter

The `kuromoji_number` token filter normalizes Japanese numbers (kansūji)
to regular Arabic decimal numbers in half-width characters. For example:

[source,console]
--------------------------------------------------
PUT kuromoji_sample
{
  "settings": {
    "index": {
      "analysis": {
        "analyzer": {
          "my_analyzer": {
            "tokenizer": "kuromoji_tokenizer",
            "filter": [
              "kuromoji_number"
            ]
          }
        }
      }
    }
  }
}

GET kuromoji_sample/_analyze
{
  "analyzer": "my_analyzer",
  "text": "一〇〇〇"
}
--------------------------------------------------

Which results in:

[source,console-result]
--------------------------------------------------
{
  "tokens" : [ {
    "token" : "1000",
    "start_offset" : 0,
    "end_offset" : 4,
    "type" : "word",
    "position" : 0
  } ]
}
--------------------------------------------------