Schedule (Tentative)

Date Phase
06.29
 1.7.1 Released
07.01
 1.8 Development starts
10.16 1.8 Data submission starts
12.02 1.8 Data submission ends - vetting starts
01.13 1.8 Data resolution, data/structure verification and correction starts
02.10 1.8 Final Candidate available
03.10 1.8 Target release date
For details, see Release Schedule.
See also Notification Services.

Collation Guidelines

Contents

  1. 1 Determining the Order
    1. 1.1 Determining Level Differences
      1. 1.1.1 Primary Test
      2. 1.1.2 Secondary Test1
      3. 1.1.3 Secondary Test2
      4. 1.1.4 Tertiary Test
    2. 1.2 Contractions
  2. 2 Filing a Request
    1. 2.1 Rules:
    2. 2.2 Test Data:
    3. 2.3 Testing Your Request
  3. 3 Minimal Rules
  4. 4 Pitfalls
The locale-based collation rules in Unicode CLDR specify modifications of the standard UCA data, also known as DUCET. Requests to change the collation order for a given locale, or to supply additional variants, need to follow the guidelines in this document.

Determining the Order

The exact collation sequence for a given language may be difficult to determine. The base ordering of characters can be fairly straightforward, but there are quite a few other complications involved.

Most standards that specify collation, such as DIN or CS, are not targeted at algorithmic sorting, and are not complete algorithmic specifications. For example, CSN 97 6030 requires transliteration of foreign scripts, but there are many choices as to how to transliterate, and the exact mechanism is not specified. It also specifies that geometric shapes are sorted by the number of vertices and edges, which is, at a minimum, difficult to determine; and are subject to variation in glyphs.

The CLDR goals are to match the sorting of exemplar letters and common punctuation and leave everything else to the standard UCA ordering. For more information, see UTS #10: Unicode Collation Algorithm (UCA).

Determining Level Differences

It is often tricky to determine the exact relationship between characters. In the UCA, case and similar variant differences are at a third (tertiary) level, while accent and similar differences are at a second (secondary) level, and base letter differences are at the first (primary) level. That results in an order like the following:

  1. cina
  2. Cina
  3. çina
  4. Çina
  5. dina

That is, the difference between c and C is weaker than the difference between c and ç, which in turn is weaker than the difference between c and d. For any two characters α and β, it may be very clear that α < β, but not be clear what the right level difference is. To establish this, see if you can find examples of two words that of the following form.

Primary Test

  1. ...α...Z
  2. ...β...A
That is, the words are identical except for α, β, A, and Z, and you know that A and Z have a clear primary difference. If we get the above ordering in dictionaries and other sources, you know that the difference between α and β is a primary difference. If we get the opposite ordering than 1,2 above, then you only know that the difference between α and β is not a primary difference: it may be secondary or tertiary.

You now need to distinguish which of the non-primary level differences you could have. So try again, this time seeing if you can find examples of two words that of the following form, where you know that A and Á have a clear secondary difference in the script.

Secondary Test1

  1. ...α...Á
  2. ...β...A

Now the ordering of these two strings tells you whether the difference between α and β is a secondary difference, or not. Alternatively, you can look for words of the form:

Secondary Test2

  1. ...B...α
  2. ...b...β

where b < B at a tertiary level. If you get the above ordering for the secondary test2, you also know that the difference between α and β is at a secondary level. The Test2 form is often easier to find examples for.

If you have established that the characters have neither a primary nor secondary difference, the following can be used in a similar fashion to test whether the difference is at a tertiary level or not.

Tertiary Test

  1. ...α...B
  2. ...β...b
If there is no tertiary difference, then the difference is not signficant enough for CLDR to take it into account, so they will be treated as equals (unless someone sorts with a final, codepoint level).

Contractions

Characters may behave differently in different contexts. For example, "ch" in Slovak sorts after H. A sequence of characters that behaves that way is called a contraction. Another common case of contractions is in the case of syllabaries, where a sequence of characters forming a syllable collates as a unit.

Note that contractions are typically rather expensive in implementations: they take more storage, and are much slower to compare. So they should be avoided where possible. For example, suppose that we have the following sequence in a dictionary (where the uppercase characters represent characters in the target script):

KB
... // combinations of K with consonants
KZ
KA
KE
KI
KO
KU
LB
...

There are two ways to produce this ordering. One is to have KA, KE, KI, etc be contractions. The other is to order all the vowels after all the consonants. Where the latter is sufficient, it is strongly preferred.

Filing a Request

Requests to change the collation order for a given locale, or to supply additional variants should be filed as bugs at CLDR Locale Bugs.

For readability, the rules are presented here in ICU rule format, rather than XML; for the same reason, we prefer the bug reports to also use that format, even though the end result will be in XML. For more information, see ICU Collation Customization. The rules must also be Minimal Rules as described below.

Rules:

& c < cs
& cs <<< ccs / cs

Please supply some short test cases that illustrate the correct sorting behavior as a list of lines in sorted order. Try to include cases that show the boundary behavior by including high suffixes, such as the following:

Test Data:

c
cy
cs
cscs
ccs
cscsy
ccsy
csy
d

Testing Your Request

Please test out any suggested rules before filing a bug, using Locale Explorer:

  1. Go to the ICU Locale Explorer
  2. Pick the appropriate locale
  3. Follow the instructions at the bottom to use your suggested rules on your suggested test data.
  4. Verify that the proper order results.
The core syntax is specified in the table "Specifying Collation Ordering" in http://www.unicode.org/reports/tr35/.

Minimal Rules

The goal is always specify the minimal differences from the DUCET. For example, take the case of Slovak, where everything sorts as in DUCET except for certain characters. The following rules place the characters ä, č, đ, and the sequence "ch" (and their case variants) at the appropriate positions in the sorting sequence, and with the appropriate strengths:

Minimal Rules

ICU Format
 XML Format
& A
< ä
<<< Ä		 <reset>A</reset>
<p>ä</p>
<t>Ä</t>
& C
< č
<<< Č		 <reset>C</reset>
<p>č</p>
<t>Č</t> 
& D
< đ
<<< Đ		 <reset>D</reset>
<p>đ</p>
<t>Đ</t>
& H
< ch
<<< cH
<<< Ch
<<< CH		 <reset>H</reset>
<p>ch</p>
<t>cH</t>
<t>Ch</t>
<t>CH</t>
...
 ...

It would be possible instead to have rules that list every letter used by Slovak [a á ä b c č d ď e é f-h {ch} i í j-l ĺ ľ m n ň o ó ô p-r ŕ s š t ť u ú v-y ý z ž], looking something like the following.

Maximal Rules

ICU Format
 		XML Format
& A
<< á
<<< Á
< ä
<<< Ä
< b
<<< B
< c
<<< C
< č
<<< Č
< d
...
 <reset>A</reset>
<s>á</s>
<t>Á</t>
<p>ä</p>
<t>Ä</t>
<p>b</p>
<t>B</t>
<p>c</t>
<t>C</t>
<p>č</p>
<t>Č</t>
<p>d</p>
...

The Maximal Rules format is not accepted in CLDR. The reasons are:
  1. Every time a character is tailored, the data for that character takes up more room in typical implemenations. That means that the data for collation is larger, downloads of collation libraries with that data are slower, sort keys are longer, and performance is slower; sometimes very much so.
  2. Related characters in the same script are in a peculiar order. For example, if the Slovak tailoring omits ƀ, then it would show up as after z.
Similarly, normally CLDR does not accept submissions that reorder digits, punctuation, or other symbols, following instead the UCA ordering for those characters.

You can see what the UCA currently does with a given script by looking at the charts at Unicode Collation Charts, or at the UCA in ICU-style rules. For example, suppose that U+0D89 SINHALA LETTER IYANNA and U+0D8A SINHALA LETTER IIYANNA needed to come after U+0D96 SINHALA LETTER AUYANNA, in primary order, and that otherwise DUCET was ok. Then you would give the following rules:

& ඖ # U+0D96 SINHALA LETTER AUYANNA
< ඉ # U+0D89 SINHALA LETTER IYANNA
< ඊ # U+0D8A SINHALA LETTER IIYANNA

Pitfalls

There are a number of pitfalls with collation, so be careful. In some cases, such as Hungarian or Japanese, the rules can be fairly complicated (of course, reflecting that the sorting sequence for those languages is complicated).

  1. Only tailor expected data. We focus on the required collation sequence for a given language with normal data. So we don't include full-width characters for a European collation sequence, such as

    • ... CSCS <<< ＣＳＣＳ ...
    • ...  CSCS <<< \uFF23\uFF33\uFF23\uFF33 ... (equivalently)

  2. Tailor trailing contractions. If a sequence of characters is treated as a unit for collation, it should be entered as a contraction.

    & c < ch

    One might think that sequence like "dz" doesn't require that, since it would always come after "d" followed by any other letter; it is a "trailing contraction". But in unusual cases, that wouldn't be true; if "dz" is a unit sorted as if it were a distinct letter after "d", one should get the ordering "dα" < "dz". The correct behavior will only happen if "dz" is a contraction, such as

    & d < dz

  3. Watch out for Expansions. If you have a rule like &cs < d, and "cs" has not occurred in a previous rule as a contraction, then this is automatically considered to be the same as &c < d / s; that is, the d expands as if it were a "cs" (actually, primary greater than a "cs", since we wrote "<"). This expansion takes effect until the next primary difference.

    So suppose that "ccs" is to behave as if it were "cscs", and take case differences into account. You might try to do this with the rules on the left:

    Rules (Wrong) Actual Effect
    & C < cs <<< Cs <<< CS
    & cscs <<< ccs
    <<< Cscs <<< Ccs
    <<< CSCS <<< CCS    		 & C < cs <<< Cs <<< CS
    & cs <<< ccs / cs
    <<< Cscs  / cs <<< Ccs  / cs
    <<< CSCS  / cs <<< CCS / cs

    But since the CSCS has not been made a contraction in previous rules, this produces an automatic expansion, one that continues through the entire sequence of non-primary differences, as shown on the right. This is not what is wanted: each item acts like it expands compared to the previous item. So CCS, for example, will act like it expands to CSCScs!

    What you actually want is the following:

    Rules (Right) Actual Effect
    & C < cs <<< Cs <<< CS
    & cscs <<< ccs
    & Cscs <<< Ccs
    & CSCS <<< CCS    		 & C < cs <<< Cs <<< CS
    & cs <<< ccs / cs
    & Cs <<< Ccs / cs
    & CS <<< CCS / CS

    In short, when you have expansions, it is always safer and clearer to express them with separate resets. There are only a few exceptions to this, notably when CJK characters are interleaved with Hangul Syllables.

  4. Minimal Rules. Example: Maltese was sorting character sequences before a base character using the following style:

    & B
    < ċ
    <<<Ċ
    < c
    <<<C

    The correct rules should be the minimal ones.

    & [before 1] c < ċ <<< Ċ

    This finds the highest primary (that's what the 1 is for) character less than c, and uses that as the reset point. For Maltese, the same technique needs to be used for ġ and ż.

  5. Blocking Contractions. Contractions can be blocked with CGJ, as described in the Unicode Standard and in the Characters and Combining Marks FAQ.

  6. Case Combinations. Normally all combinations of case need to be supplied for contractions. That is, if ch is a contraction, then you would have the rules ... ch < cH < Ch < CH. The reason for this is so that all case variants sort at the same primary level: thus lowercasing a string will not affect its primary order. Cases such as McHugh are handled like other instances where contractions should be blocked.