synonym.md

Using SudachiDict Synonyms

Here we describe a temporary way to use Sudachi Dictionary's synonym resource (Sudachi 同義語辞書) with Elasticsearch. We plan to create a dedicated Sudachi synonym filter for Elasticsearch in the future.

You can convert the synonym file to "Solr synonyms" format, and use it via Elasticsearch's default synonym filters.

Format Converion

You can simply convert the Sudachi synonym file into the Solr synonyms format. The Sudachi format is described in detail here.

You can use our example script (ssyn2es.py) for the conversion;

$ python ssyn2es.py SudachiDict/src/main/text/synonyms.txt > synonym.txt

Expansion Suppresion

You can partially make use of the Sudachi synonym resource's detailed information with the Solr format's => notation, which controls the expansion direction.

# synonym entry
アイスクリーム,ice cream,ice=>アイスクリーム,ice cream,ice,アイス

# expansion example
# `アイスクリーム` => `アイスクリーム`, `ice cream`, `ice`, `アイス`
# `アイス` => `アイス` (**no expansion**)

Punctuation Symbols

You may need to remove certain synonym words such as € and ＆ when you use the analyzer with setting "discard_punctuation": true (Otherwise you will be get an error, e.g., "term: € was completely eliminated by analyzer"). Alternatively, you can set "lenient": true for the synonym filter to ignore the exceptions.

These symbols are defined as punctuations; See SudachiTokenizer.java for the detail.

Synonym Filter

You can use the converted Solr format file with Elasticsearch's default synonym filters, Synonym token filter or Synonym graph filter.

Example: Set up

{
    "settings": {
        "analysis": {
            "filter": {
                "sudachi_synonym": {
                    "type": "synonym_graph",
                    "synonyms_path": "sudachi/synonym.txt"
                }
            },
            "tokenizer": {
                "sudachi_tokenizer": {
                    "type": "sudachi_tokenizer"
                }
            },
            "analyzer": {
                "sudachi_synonym_analyzer": {
                    "type": "custom",
                    "tokenizer": "sudachi_tokenizer",
                    "filter": [
                        "sudachi_synonym"
                    ]
                }
            }
        }
    }
}

Here we assume that the converted synonym file is placed as $ES_PATH_CONF/sudachi/synonym.txt.

If you would like to use sudachi_split filter, set it after the synonym filter (otherwise you will get an error, e.g., term: 不明確 analyzed to a token (不) with position increment != 1 (got: 0)).

Example: Analysis

Case 1.

{
  "analyzer": "sudachi_synonym_analyzer",
  "text": "アイスクリーム"
}

Returns

{
    "tokens": [
        {
            "token": "アイスクリーム",
            "start_offset": 0,
            "end_offset": 7,
            "type": "SYNONYM",
            "position": 0
        },
        {
            "token": "ice cream",
            "start_offset": 0,
            "end_offset": 7,
            "type": "SYNONYM",
            "position": 0
        },
        {
            "token": "ice",
            "start_offset": 0,
            "end_offset": 7,
            "type": "SYNONYM",
            "position": 0
        },
        {
            "token": "アイス",
            "start_offset": 0,
            "end_offset": 7,
            "type": "SYNONYM",
            "position": 0
        }
    ]
}

Case 2.

{
    "analyzer": "sudachi_synonym_analyzer",
    "text": "アイス"
}

Returns

{
    "tokens": [
        {
            "token": "アイス",
            "start_offset": 0,
            "end_offset": 3,
            "type": "word",
            "position": 0
        }
    ]
}