Living in a Unicode World

When Elasticsearch compares one token with another, it does so at the byte level. In other words, for two tokens to be considered the same, they need to consist of exactly the same bytes. Unicode, however, allows you to write the same letter in different ways.

For instance, what’s the difference between é and é? It depends on who you ask. According to Elasticsearch, the first one consists of the two bytes 0xC3 0xA9, and the second one consists of three bytes, 0x65 0xCC 0x81.

According to Unicode, the differences in how they are represented as bytes is irrelevant, and they are the same letter. The first one is the single letter é, while the second is a plain e combined with an acute accent ´.

If you get your data from more than one source, it may happen that you have the same letters encoded in different ways, which may result in one form of déjà not matching another!

Fortunately, a solution is at hand. There are four Unicode normalization forms, all of which convert Unicode characters into a standard format, making all characters comparable at a byte level: nfc, nfd, nfkc, nfkd.

Unicode Normalization Forms

The composed forms—`nfc` and nfkc—represent characters in the fewest bytes possible. So é is represented as the single letter é. The decomposed forms—`nfd` and nfkd—represent characters by their constituent parts, that is e + ´.

The canonical forms—`nfc` and nfd—represent ligatures like ffi or œ as a single character, while the compatibility forms—`nfkc` and nfkd—break down these composed characters into a simpler multiletter equivalent: f + f + i or o + e.

It doesn’t really matter which normalization form you choose, as long as all your text is in the same form. That way, the same tokens consist of the same bytes. That said, the compatibility forms allow you to compare ligatures like ffi with their simpler representation, ffi.

You can use the icu_normalizer token filter to ensure that all of your tokens are in the same form:

PUT /my_index
{
  "settings": {
    "analysis": {
      "filter": {
        "nfkc_normalizer": { (1)
          "type": "icu_normalizer",
          "name": "nfkc"
        }
      },
      "analyzer": {
        "my_normalizer": {
          "tokenizer": "icu_tokenizer",
          "filter":  [ "nfkc_normalizer" ]
        }
      }
    }
  }
}

  1. Normalize all tokens into the nfkc normalization form.

Tip

Besides the icu_normalizer token filter mentioned previously, there is also an icu_normalizer character filter, which does the same job as the token filter, but does so before the text reaches the tokenizer. When using the standard tokenizer or icu_tokenizer, this doesn’t really matter. These tokenizers know how to deal with all forms of Unicode correctly.

However, if you plan on using a different tokenizer, such as the ngram, edge_ngram, or pattern tokenizers, it would make sense to use the icu_normalizer character filter in preference to the token filter.

Usually, though, you will want to not only normalize the byte order of tokens, but also lowercase them. This can be done with icu_normalizer, using the custom normalization form nfkc_cf, which we discuss in the next section.

results matching ""

    No results matching ""