An updated IDNA criterion for right-to-left
scripts
Google
Beddingen 10
Trondheim
7014
Norway
harald@alvestrand.no
Swedish Museum of Natural History
Frescativ. 40
Stockholm
10405
Sweden
+46 8 5195 4055
ck@nrm.museum
The use of right-to-left scripts in internationalized domain names
has presented several challenges. This memo discusses some problems with
these scripts, and some shortcomings in the 2003 IDNA BIDI criterion.
Based on this discussion, it proposes a new BIDI criterion for IDNA
labels.
This document's purpose is to establish a test that can be applied
to Internationalized Domain Name (IDN) labels in Unicode form
(U-labels) containing right-to-left characters.
When labels pass the test, and when certain other conditions are
satisified, they can be used with a minimal chance of these labels
being displayed in a confusing way by a bidirectional display
algorithm.
This specification does not place any requirements on domain names
that do not contain right-to-left characters.
The IDNA specification "Stringprep"
makes the following statement in its section 6 on the BIDI
algorithm:
3) If a string contains any RandALCat character, a RandALCat
character MUST be the first character of the string, and a
RandALCat character MUST be the last character of the string.
(A RandALCat character is a character with unambiguously
right-to-left directionality.)
The reasoning behind this prohibition was to ensure that every
component of a displayed domain name has an unambiguously preferred
direction. However, this makes certain words in languages written with
right-to-left scripts invalid as IDN labels, and in at least one case
means that all the words of an entire language are forbidden as IDN
labels.
This will be illustrated below with examples taken from the Dhivehi
and Yiddish languages, as written with the Thaana and Hebrew scripts,
respectively.
In investigating this problem, it was realized that the RFC 3454
specification did not exactly specify what the requirement to be
fulfilled was, and therefore, it was impossible to tell whether a
simple relaxation of the rule would continue to fulfil the
requirement. A further investigation led to the conclusion that for
one reasonable set of requirements, IDNA2003's BIDI restriction did
not fulfil the requirements. This document therefore proposes
replacing the RFC 3454 BIDI requirement in its entirety.
While the document proposes completely new text, most reasonable
labels that were allowed under the old criterion will also be allowed
under the new criterion, so the operational impact of the rule change
is limited.
In this memo, we use "network order" to describe the sequence of
characters as transmitted on the wire or stored in a file; the terms
"first", "next", "previous", "before" and "after" are used to refer to
the relationship of characters and labels in network order.
We use "display order" to talk about the sequence of characters as
imaged on a display medium; the terms "left" and "right" are used to
refer to the relationship of characters and labels in display
order.
Most of the time, the examples use the abbreviations for the
Unicode BIDI classes to denote the directionality of the characters;
in some examples, the convention that uppercase characters are of
class R or AL, and lowercase characters are of class L is used - thus,
the example string ABC.abc would consist of 3 right-to-left characters
and 3 left-to-right characters.
The term "paragraph" is used in the sense of the Unicode BIDI
specification - it means "a block of text
that has an overall direction, either left-to-right or right-to-left",
approximately.
"LTR" and "RTL" are abbreviations for "right to left" and "left to
right", respectively.
The other terminology used to describe IDNA concepts is defined in
Dhivehi, the official language of the Maldives, is written with the
Thaana script. This displays some of the characteristics of Arabic
script, including its directional properties, and the indication of
vowels by the diacritical marking of consonantal base characters. This
marking is obligatory, and both double vowels and syllable-final
consonants are indicated by the marking of special unvoiced
characters. Every Dhivehi word therefore ends with a combining
mark.
The word for "computer", which is romanized as "konpeetaru", is
written with the following sequence of Unicode code points:
U+0786 THAANA LETTER KAAFU (AL)
U+07AE THAANA OBOFILI (NSM)
U+0782 THAANA LETTER NOONU (AL)
U+07B0 THAANA SUKUN (NSM)
U+0795 THAANA LETTER PAVIYANI (AL)
U+07A9 THAANA LETTER EEBEEFILI (AL)
U+0793 THAANA LETTER TAVIYANI (AL)
U+07A6 THAANA ABAFILI (NSM)
U+0783 THAANA LETTER RAA (AL)
U+07AA THAANA UBIUFILI (NSM)
The directionality class of U+07AA in the Unicode database is NSM (non-spacing mark), which is not R or
AL; a conformant implementation of the IDNA2003 algorithm will say
that "this is not in RandALCat", and refuse to encode the string.
Yiddish is one of several languages written with the Hebrew script
(others include Hebrew and Ladino). This is basically a consonantal
alphabet (also termed an "abjad") but Yiddish is written using an
extended form that is fully vocalic. The vowels are indicated in
several ways, of which one is by repurposing letters that are
consonants in Hebrew. Other letters are used both as vowels and
consonants, with combining marks, called "points", used to
differentiate between them. Finally, some base characters can indicate
several different vowels, which are also disambiguated by combining
marks. Pointed characters can appear in word-final position and may
therefore also be needed at the end of labels. This is not an
invariable attribute of a Yiddish string and there is thus greater
latitude here than there is with Dhivehi.
The organization now known as the "YIVO Institute for Jewish
Research" developed orthographic rules for modern Standard Yiddish
during the 1930s on the basis of work conducted in several venues
since earlier in that century. These are given in, "The Standardized
Yiddish Orthography: Rules of Yiddish Spelling" , and are taken as normatively descriptive of
modern Standard Yiddish in any context where that notion is deemed
relevant. They have been applied exclusively in all Yiddish
dictionaries published since their establishment, and are similarly
dominant in academic and bibliographic regards.
It therefore appears appropriate for this repertoire also to be
supported fully by IDNA. This presents no difficulty with characters
in initial and medial positions, but pointed characters are regularly
used in final position as well. All of the characters in the SYO
repertoire appear in both marked and unmarked form with one exception:
the HEBREW LETTER PE (U+05E4). The SYO only permits this with a HEBREW
POINT DAGESH (U+05BC), providing the Yiddish equivalent to the Latin
letter "p", or a HEBREW POINT RAFE (U+05BF), equivalent to the Latin
letter "f". There is, however, a separate unpointed allograph, the
HEBREW LETTER FINAL PE (U+05E3), for the latter character when it
appears in final position. The constraint on the use of the SYO
repertoire resulting from the proscription of combining marks at the
end of RTL strings thus reduces to nothing more, or less, than the
equivalent of saying that a string of Latin characters cannot end with
the letter "p". It must also be noted that the HEBREW LETTER PE with
HEBREW POINT DAGESH is characteristic of almost all traditional
Yiddish orthographies that predate (or remain in use in parallel to)
the SYO, being the first pointed character to appear in any of
them.
A more general instantiation of the basic problem can be seen in
the representation of the YIVO acronym. This is written with the
Hebrew letters YOD YOD HIRIQ VAV VAV ALEF QAMATS, where HIRIQ and
QAMATS are combining points:
U+05D9 HEBREW LETTER YOD (R)
U+05B4 HEBREW POINT HIRIQ (NSM)
U+05D5 HEBREW LETTER VAV (R)
U+05D0 HEBREW LETTER ALEF (R)
U+05B8 HEBREW POINT QAMATS (NSM)
The directionality class of U+05B8 HEBREW POINT QAMATS in
the Unicode database is NSM, which again causes the IDNA2003 algorithm
to reject the string.
It may also be noted that all of the combined characters mentioned
above exist in precomposed form at separate positions in the Unicode
chart. However, by invoking Stringprep, the IDNA2003 algorithm also
rejects those codepoints, for reasons not discussed here.
RFC 3454, in its insistence that the first or last character of a
string be category R or AL, prohibited strings that contained
right-to-left characters and numbers at the end.
Consider the strings ALEF 5 (HEBREW LETTER ALEF + DIGIT FIVE) and 5
ALEF. Displayed in a LTR context, the first one will be displayed from
left to right as 5 ALEF (with the 5 being considered right-to-left
because of the leading ALEF), while 5 ALEF will be displayed in
exactly the same order (5 taking the direction from context). Clearly,
only one of those should be permitted as a registered label,
but barring them both seems to require justification.
One issue with RFC 3454 was that it did not give an explicit
justification for the BIDI rule, thus it was hard to tell if a modified
rule would continue to fulfil the purpose for which the RFC 3454 rule
was written.
This document proposes an explicit justification, by stating a set of
requirements for which it is possible to test whether or not the
modified rule fulfils the requirement.
All the text in this document assumes that text containing the labels
under consideration will be displayed using the Unicode bidirectional
algorithm .
The requirements proposed are these:
Label Uniqueness: No two labels, when presented in display order
in the same paragraph, should have the same sequence of characters
without also having the same sequence of characters in network
order, both when the paragraph has LTR direction and when the
paragraph has RTL direction. (This is the criterion that is explicit
in RFC 3454). (Note that a label displayed in an RTL paragraph may
display the same as a different label displayed in a LTR paragraph,
and still satisfy this criterion.)
Character Grouping: When displaying a string of labels, using the
Unicode BIDI algorithm to reorder the characters for display, the
characters of each label should remain grouped between the
characters delimiting the labels, both when the string is embedded
in a paragraph with LTR direction and when it's embedded in a
paragraph with RTL direction. [[NOTE IN DRAFT: The need for this
requirement has been challenged. Should be decided by WG.]]
Several stronger statements were considered and rejected, because
they seem to be impossible to fulfil within the constraints of the
Unicode bidirectional algorithm. These include:
The appearance of a label should be unaffected by its embedding
context. This proved impossible even for ASCII labels; the label
"123-456" will have a different display order in an RTL context than
in a LTR context.
The sequence of labels should be consistent with network order.
This proved impossible - a domain name consisting of the labels (in
network order) L1.R1.R2.L2 will be displayed as L1.R2.R1.L2 in an
LTR context. (In a RTL context, it will be displayed as
L2.R2.R1.L1).
The Label Uniqueness property should hold true between LTR
paragraphs and RTL paragraphs. This was shown to be unsound.
No two domain names should be displayed the same, even under
differing directionality. This was shown to be unsound, since the
domain name (network) ABC.abc will have display order CBA.abc in an
LTR context and abc.CBA in an RTL context, while the domain name
(network) abc.ABC will have display order abc.CBA in an LTR context
and CBA.abc in an RTL context.
One specific requirement was thought to be problematic, but turned
out to be satisfied by a string that obeys the proposed rules:
The Character Grouping requirement should be satisfied when
directional controls (LRE, RLE, RLO, LRO, PDF) are used in the same
paragraph (outside of the labels). Because these controls affect
presentation order in non-obvious ways, by affecting the "sor" and
"eor" properties of the Unicode BIDI algorithm, the conditions above
would be very hard to satisfy for an useful set of strings if this
was true. As long as these controls have no influence over the
display of the domain name, no problem will be caused, but the exact
criterion for "will not influence" is hard to codify.
For reference, here are the values that the Unicode BIDI property can
have:
L - Left-to-right - most letters in LTR scripts
R - Right-to-left - most letters in non-Arabic RTL scripts
AL - Arabic letters - most letters in the Arabic script
EN - European Number (0-9, and Extended Arabic-Indic numbers)
ES - European Number Separator (+ and -)
ET - European Number Terminator (currency symbols, the hash sign,
the percent sign and so on)
AN - Arabic Number
CS - Common Number Separator (. , / : et al)
NSM - Non spacing Mark - most combining accents
BN - Boundary Neutral - control characters
B - Paragraph Separator
S - Segment Separator
WS - Whitespace, including the SPACE character
ON - Other Neutrals, including @, &, parentheses, MIDDLE
DOT
LRE, LRO, RLE, RLO, PDF - these are "directional control
characters", and are not used in IDNA labels.
In the following descriptions, first-level bullets are used to
indicate rules or normative statements; second-level bullets are
commentary.
The Character Grouping requirement can be more formally stated
as:
Let "Delimiterchars" be a set of characters with the Unicode BIDI
properties CS, WS, ON. (These are commonly used to delimit labels -
both the FULL STOP and the space are included.)
ET, though it commonly occurs next to domain names in
practice, is problematic: the context R CS L EN ET (for instance
A.a1%) makes the label L EN not satisfy the character grouping
requirement.
ES commonly occurs in labels as HYPHEN-MINUS, but could also
be used as a delimiter (for instance, the plus sign). It is left
out here.
Let "unproblematic label" be a label that either satisifies the
requirements, or does not contain any character with the bidi
properties R, AL or AN, and does not begin with a character with the
bidi property EN. (Informally, "it does not start with a
number".)
A label L satisfies the Character Grouping requirement when, for any
Delimiter Character D1 and D2, and for any label S1 and S2 that is
either a label satisfying the requirements or an unproblematic
label.
If the string formed by concatenating S1, D1, L, D2 and S2 is
reordered according to the BIDI algorithm, then all the characters of L
in the reordered string are between D1 and D2, and no other characters
are between D1 and D2, both if the overall paragraph direction is LTR
and if the overall paragraph direction is RTL.
Note that the definition is self-referential, since S1 and S2 are
constrained to be "legal" by this definition; this makes testing changes
to proposed rules a little complex, but does not create problems for
testing whether or not a given proposed rule satisfies the
criterion.
(The "zero-length" case represents the case where a domain name is
next to something that isn't a domain name, separated by a delimiter
character).
The Label Uniqueness requirement can be formally stated as:
If two labels L and L', embedded as for the test above, displayed in
paragraphs with the same directionality, are reordered by the BIDI
algorithm into the same sequence of codepoints, at most one of the
labels L and L' is a legal label.
Based on the above discussion, and on testing, the following rule has
been developed for BIDI domain names:
A label containing a character of type R, AL or AN MUST satisfy all 7
of the rules below.
The main bullets give the rule, subordinate bullets (if any) give
justifications or examples of things that break if this rule is not
present.
Exhaustive testing has verified that all strings that satisfy this
criterion satisfy both the requirements above at least for all strings
up to 6 characters. (TO BE VERIFIED WITH NEW CONDITIONS)
Only characters with the BIDI properties L, R, AL, AN, EN, ES,
BN, ON and NSM are allowed.
B, S and WS are excluded because they are separators or
spaces.
LRE, LRO, RLE, RLO, PDF are excluded because they are BIDI
controls.
ET is excluded because the string L ET does not satisfy the
Character Grouping requirement.
CS is excluded because the string L CS does not satisfy the
Character Grouping requirement.
ES and ON are not allowed in the first position
ES R and ON R do not satisfy the Character Grouping
requirement.
ES and ON, followed by zero or more NSM, is not allowed in the
last position
L ON and L ES both fail the Character Grouping
requirement.
If an R, AL or AN is present, no L may be present.
If an EN is present, no AN may be present, and vice versa.
The first character may not be an NSM
The first character may not be an EN (European Number) or an AN
(Arabic Number).
If the character on both sides of a CS is an EN or an AN, the
labels fail the Character Grouping requirement.
Some domain names where some of the labels use leading EN and
AN may be problem-free, but there's no way of verifying this
while looking at a single label in isolation..
We could achieve stability by barring numbers at the end of
labels, but this may be more disruptive in practice.
There are situations in which labels that satisfy the rule above will
be displayed in a surprising fashion; the most important of these is the
case where a label ending in a character with BIDI property AL, AN or R
occurs before a label beginning with a character of BIDI property EN. In
that case, the number will appear to move into the label containing the
right-to-left character, violating the Character Grouping
requirement.
If the label that occurs after the right-to-left label itself
satisfies the bidi criterion, the requirements will be satisfied in all
cases (this is the reason why the criterion talks about strings
containing L in some cases). However, the WG concluded that this could
not be required for several reasons:
There is a large current deployment of ASCII domain names
starting with digits. These cannot possibly be invalidated.
Domain names are often constructed piecemeal, for instance by
combining a string with the content of a search list. This may occur
after IDNA processing, and thus in part of the code that is not
IDNA-aware, making detection of the undesirable combination
impossible.
Even if a label is registered under a "safe" label, there may be
a DNAME with an "unsafe" label that
points to the "safe" label, thus creating seemingly-valid names that
would not satisfy the criterion.
Wildcards create the odd situation where a label is "valid" (can
be looked up successfully) without the zone owner's knowing that
this label exists. So an owner of a zone whose name starts with a
digit and contains a wildcard has no way of controlling whether or
not names with RTL labels in them are looked up in his zone.
So rather than trying to create rules that disallow all such
undesirable situations, this document merely warns about the
possibility, and specifies the following behaviour for the BIDI test
defined in [protocol]:
The BIDI test MUST return failure if the BIDI rule is not
satisfied by all labels containing AL, AN or R.
The BIDI test MAY return failure if the BIDI rule is not
satisfied by the label following the label that contains AL, AN or R
in the domain name. For all the reasons given above, it may be
impossible to know the following label, but there seems no or
negative value to allowing the BIDI test to succeed if the following
label is known. [[POSSIBLY CONTROVERSIAL]]
This document concerns itself only with the rules that are needed
when dealing with domain names with characters that have differing BIDI
properties, and considers characters only in terms of their BIDI
properties. All other issues with these scripts have to be considered in
other contexts.
One such issue is the need to keep numbers separate; several scripts,
such as Arabic, have multiple sets of numbers. The algorithm in this
document disallows occurrences of AN-class characters ("Arabic-Indic
digits", U+0660 to U+0669) together with EN-class characters (which
includes "European" digits, U+0030 to U+0039 and "extended Arabic-Indic
digits", U+06F0 to U+06F9), but does not help in preventing the mixing
of, for instance, Bengali digits (U+09E6 to U+09EF) and Gujarati digits
(U+0AE6 to U+0AEF), both of which have BIDI class L. Any registry or
script community that wishes to create rules for the mixing of digits in
a label will have to specify these restrictions at the registry
level.
Another set of issues concerns the proper display of IDNs with a
mixture of LTR and RTL labels, or only RTL labels.
It is unrealistic to expect that applications will display domain
names using embedded formatting codes between their labels (for one
thing, no reliable algorithm for identifying domain names in running
text exist); thus, the display order will be determined by the
bidirectional algorithm. Thus, a sequence (in network order) of
R1.R2.ltr will be displayed in the order 2R.1R.ltr in a LTR context,
which might surprise someone expecting to see labels displayed in
hierarchical order. Again, this memo does not attempt to suggest a
solution to this problem.
As with any change to an existing standard, it is important to
consider what happens with existing implementations when the change is
introduced. The following troublesome cases have been noted:
Old program used to input the newly allowed string. If the old
program checks the input against RFC 3454, the string will not be
allowed, and that domain name will remain inaccessible.
Old program is asked to display the newly allowed string, and
checks it against RFC 3454 before displaying. The program will
perform some kind of fallback, most likely displaying the A-label
of the string.
Old program tries to display the newly allowed string. If the
old program has code for displaying the last character of a string
that is different from the code used to display the characters in
the middle of the string, display may be inconsistent and cause
confusion.
One particular example of the last case is if a program chooses to
examine the last character (in network order) of a string in order to
determine its directionality, rather than its first; if it finds an
NSM character and tries to display the string as if it was a
left-to-right string, the resulting display may be interesting, but
not useful.
The editors believe that these cases will have less harmful impact
in practice than continuing to deny the use of words from the
languages for which these strings are necessary as IDN labels.
This specification does not forbids using leading European numbers
in ASCII-only labels, since this would in conflict with a large
installed base of such labels, and would increase the scope of the
specification from RTL labels to all labels. The harm resulting from
this limitation of scope is described in . Zone managers, both registries and private
zone managers, can check for this particular condition before they
allow registration of any string with right-to-left characters in it;
generally it is best to not allow registration of any right-to-left
strings in a zone where the label at the level above begins with a
digit.
This text is, intentionally, specified strictly in terms of the
Unicode BIDI properties. The determination that the condition is
sufficient to fulfil the criteria depends on the Unicode BIDI
algorithm; it is unlikely that drastic changes will be made to this
algorithm.
However, the determination of validity for any string depends on
the Unicode BIDI property values, which are not declared immutable by
the Unicode Consortium. Furthermore, the behaviour of the algorithm
for any given character is likely to be linguistically and culturally
sensitive, so that it's not unlikely that later versions of the
Unicode standard may change the BIDI properties assigned to certain
Unicode characters.
This memo does not propose a solution for this problem.
This document makes no request of IANA.
Note to RFC Editor: this section may be removed on publication as an
RFC.
This modification will allow some strings to be used in Stringprep
contexts that are not allowed today. It is possible that differences in
the interpretation of the specification between old and new
implementations could pose a security risk, but it is difficult to
envision any specific instantiation of this.
Any rational attempt to compute, for instance, a hash over an
identifier processed by Stringprep would use network order for its
computation, and thus be unaffected by the changes proposed here.
While it is not believed to pose a problem, if display routines had
been written with specific knowledge of the RFC 3454 Stringprep
prohibitions, it is possible that the potential problems noted under
"backwards compatibility" could cause new kinds of confusion.
While the listed editors held the pen, this document represents the
joint work and conclusions of an ad hoc design team. In addition to the
editors this consisted of, in alphabetic order, Tina Dam, Patrik
Faltstrom, and John Klensin. Many further specific contributions and
helpful comments were received from the people listed below, and others
who have contributed to the development and use of the IDNA
protocols.
The team wishes in particular to thank Roozbeh Pournader for calling
its attention to the issue with the Thaana script, Paul Hoffmann for
pointing out the need to be explicit about backwards compatibility
considerations, Ken Whistler for suggesting the basis of the formalized
"character grouping" requirement, Mark Davis for commentary, Erik van
der Poel for careful review, comments and verification of the rulesets,
and Marcos Sanz, Andrew Sullivan and Pete Resnick for reviews.
Unicode Standard Annex #9: The Bidirectional Algorithm,
revision 15
Unicode Consortium
The Unicode Standard - version 5.1
The Standardized Yiddish Orthography: Rules of Yiddish
Spelling, 6th ed., , New York, ISBN 0-914512-25-0",
This appendix is intended to be removed by the RFC Editor when this
document is published as an RFC.
Suggested a possible new algorithm.
Multiple smaller changes.
Date of publication updated.
Change log added.
Intro changed to reflect addressing the deeper issues with the BIDI
algorithm.
Gave formalized criteria for "valid strings", and documented the
new set of requirements for strings that satisfy the criteria.
Removed most of section 5, "Other problems", and noted that this
memo focuses ONLY on issues that can be evaluated by looking at the
BIDI properties of characters.
Added back AN to the list of allowed characters; it had been left
out by accident in -03.
Removed some rules that were redundant.
Added some considerations for backwards compatibility and
interaction with ASCII labels that start with a number.
Mentioned the issue with DNAME pointing to a zone containing RTL
labels in the security considerations section.
Wording updates in multiple places, including some spelling
errors.
Rewrote the introduction section.
Split references into "normative" and "informative".
Changed name of draft.
Added a couple of "note in draft" statements to remind the WG of
open issues.
Noted that BIDI controls in the paragraph are unproblematic with
the given ruleset.
Added text to section 5 describing issues with mixture of numbers
in labels
Addressed some of the issues raised by Mark Davis in March 2008 in
regard to document clarity.
Changed the formulation of the label uniqueness requirement to be
consistent with the text under "Labels with numbers".
Spell-checked document.
Changed the domain of applicability to be only labels containing
RTL characters, described the conditions under which harm may result
from putting RTL labels next to other labels, and how to detect
them.
A number of clarification and formatting changes in response to
reviews.