This document builds upon on Character Model for the World Wide Web 1.0: Fundamentals [CHARMOD] to provide authors of specifications, software developers, and content developers a common reference on string identity matching on the World Wide Web and thereby increase interoperability.
This section describes the status of this document at the time of its publication. Other documents may supersede this document. A list of current W3C publications and the latest revision of this technical report can be found in the W3C technical reports index at https://www.w3.org/TR/.
This version of the document represents a significant change from the earlier editions. Much of the content is changed and the recommendations are significantly altered. This fact is reflected in a change to the name of the document from "Character Model: Normalization".
Sending comments on this document
If you wish to make comments regarding this document, please raise them as github issues. When reviewing the document, please refer to the latest editor's copy. Only send comments by email if you are unable to raise issues on github (see links below). All comments are welcome.
To make it easier to track comments, please raise separate issues or emails for each comment, and point to the section you are commenting on using a URL.
This document was published by the Internationalization Working Group as a Working Draft. Comments regarding this document are welcome. Please send them to firstname.lastname@example.org (subscribe, archives).
Publication as a Working Draft does not imply endorsement by the W3C Membership. This is a draft document and may be updated, replaced or obsoleted by other documents at any time. It is inappropriate to cite this document as other than work in progress.
This document was produced by a group operating under the W3C Patent Policy. The group does not expect this document to become a W3C Recommendation. W3C maintains a public list of any patent disclosures made in connection with the deliverables of the group; that page also includes instructions for disclosing a patent. An individual who has actual knowledge of a patent which the individual believes contains Essential Claim(s) must disclose the information in accordance with section 6 of the W3C Patent Policy.
This document is governed by the 1 February 2018 W3C Process Document.
The goal of the Character Model for the World Wide Web is to facilitate use of the Web by all people, regardless of their language, script, writing system, or cultural conventions, in accordance with the W3C goal of universal access. One basic prerequisite to achieve this goal is to be able to transmit and process the characters used around the world in a well-defined and well-understood way.
This document builds on Character Model for the World Wide Web: Fundamentals [CHARMOD]. Understanding the concepts in that document are important to being able to understand nd apply this document successfully.
This part of the Character Model for the World Wide Web covers string matching—the process by which a specification or implementation defines whether two string values are the same or different from one another. It describes the ways in which texts that are semantically equivalent can be encoded differently and the impact this has on matching operations important to formal languages (such as those used in the formats and protocols that make up the Web).
The main target audience of this specification is W3C specification developers. This specification and parts of it can be referenced from other W3C specifications and it defines conformance criteria for W3C specifications, as well as other specifications.
Other audiences of this specification include software developers, content developers, and authors of specifications outside the W3C. Software developers and content developers implement and use W3C specifications. This specification defines some conformance criteria for implementations (software) and content that implement and use W3C specifications. It also helps software developers and content developers to understand the character-related provisions in W3C specifications.
The character model described in this specification provides authors of specifications, software developers, and content developers with a common reference for consistent, interoperable text manipulation on the World Wide Web. Working together, these three groups can build a globally accessible Web.
This document defines one of the basic building blocks for the Web related to this problem by defining rules and processes for String Identity Matching in document formats. These rules are designed for the identifiers and structural markup (syntactic content) used in document formats to ensure consistent processing of each and are targeted to Specification writers. This section is targeted to implementers.
This document is divided into two main sections.
The first section lays out the problems involved in string matching; the effects of Unicode and case folding on these problems; and outlines the various issues and normalization mechanisms that might be used to address these issues.
The second section provides requirements and recommendations for string identity matching for use in formal languages, such as many of the document formats defined in W3C Specifications. This primarily is concerned with making the Web functional and providing document authors with consistent results.
This section provides some historical background on the topics addressed in this specification.
At the core of the character model is the Universal Character Set (UCS), defined jointly by the Unicode Standard [Unicode] and ISO/IEC 10646 [ISO10646]. In this document, Unicode is used as a synonym for the Universal Character Set. A successful character model allows Web documents authored in the world's writing systems, scripts, and languages (and on different platforms) to be exchanged, read, and searched by the Web's users around the world.
The first few chapters of the Unicode Standard [Unicode] provide useful background reading.
For information about the requirements that informed the development of important parts of this specification, see Requirements for String Identity Matching and String Indexing [CHARREQ].
This section contains terminology and notation specific to this document.
The Web is built on text-based formats and protocols. In order to describe string matching or searching effectively, it is necessary to establish terminology that allows us to talk about the different kinds of text within a given format or protocol, as the requirements and details vary significantly.
A Unicode code point (or "code point") refers to the numeric value assigned to each Unicode character. Unicode code points range from 0 to 0x10FFFF16. (See Section 4.1 in [CHARMOD] for a deeper discussion of character encoding terminology.)
Unicode code points are denoted as
hhhh is a sequence of at least four, and at most six hexadecimal digits. For example, the character € [U+20AC EURO SIGN] has the code point U+20AC.
Some characters used in this document's examples might not appear as intended on your specific device or display. Usually this is due to lack of a script-specific font installed locally or due to other limitations of your specific rendering system. This document uses a Webfont to provide fallback glyphs for many of the non-Latin characters, but your device might not support displaying the font. To the degree possible, the editors have tried to ensure that the examples nevertheless remain understandable.
A legacy character encoding is a character encoding not based on the Unicode character set.
A transcoder is a process that converts text between two character encodings. Most commonly in this document it refers to a process that converts from a legacy character encoding to a Unicode encoding form, such as UTF-8.
Syntactic content is any text in a document format or protocol that belongs to the structure of the format or protocol. This definition includes values that are typically thought of as "markup" but can also include other values, such as the name of a field in an HTTP header. Syntactic content consists of all of the characters that form the structure of a format or protocol. For example, < and > (as well as the element name and various attributes they surround) are part of the syntactic content in an HTML document.
Syntactic content usually is defined by a specification or specifications and includes both the defined, reserved keywords for the given protocol or format as well as string tokens and identifiers that are defined by document authors to form the structure of the document (rather than the "content" of the document).
Natural language content refers to the language-bearing
content in a document and not to any of the surrounding or embedded
syntactic content that form part of the document structure. You can think
of it as the actual "content" of the document or the "message" in a
given protocol. Note that syntactic content can contain natural
language content, such as when an [HTML]
alt attribute containing a description of
A resource, in the context of this document, is a given
document, file, or protocol "message" which includes both the natural
language content as well as the syntactic content
such as identifiers surrounding or containing it. For example, in an
HTML document that also has some CSS and a few
a file, is a resource. This term is intentionally similar to the term
'resource' as used in [RFC3986], although here the term is applied
A user-supplied value is unreserved syntactic content in a vocabulary that is assigned by users, as distinct from reserved keywords in a given format or protocol. For example, CSS class names are part of the syntax of a CSS style sheet. They are not reserved keywords, predefined by any CSS specification. They are subject to the syntactic rules of CSS. And they may (or may not) consist of natural language tokens.
A vocabulary provides the list of reserved names as well as the set of rules and specifications controlling how user-supplied values (such as identifiers) can be assigned in a format or protocol. This can include restrictions on range, order, or type of characters that can appear in different places. For example, HTML defines the names of its elements and attributes, as well as enumerated attribute values, which defines the "vocabulary" of HTML syntactic content. Another example would be ECMAScript, which restricts the range of characters that can appear at the start or in the body of an identifier or variable name. It applies different rules for other cases, such as to the values of string literals.
A grapheme is a sequence of
one or more characters in a visual representation of some text
that a typical user would perceive as being a single unit (
Graphemes are important for a number of text operations such as
sorting or text selection, so it is necessary to be able to compute
the boundaries between each user-perceived character. Unicode defines
the default mechanism for computing graphemes in Unicode
Standard Annex #29: Text Segmentation [UAX29] and calls
this approximation a grapheme cluster. There are two types
of default grapheme cluster defined. Unless otherwise noted, grapheme
cluster in this document refers to an extended default grapheme
cluster. (A discussion of grapheme clusters is also given in Section 2
of the Unicode Standard, [Unicode]. Cf. near the end of
in version 8.0 of The Unicode Standard)
Because different natural languages have different needs, grapheme clusters can also sometimes require tailoring. For example, a Slovak user might wish to treat the default pair of grapheme clusters "ch" as a single grapheme cluster. Note that the interaction between the language of string content and the end-user's preferences might be complex.
This section illustrates some of the terminology defined above. For illustration purposes we'll use the following small HTML file as an example (line numbers added for reference):
1 <html lang="en" dir="ltr">
3 <meta charset="UTF-8">
7 <img src="shakespeare.jpg" alt="William Shakespeare" id="shakespeare_image">
8 <p>What’s in a name? That which we call a rose by any other name would smell as sweet.</p>
’embedded in the sentence on line 8 is part of the syntactic content.)
William Shakespeare) contains natural language text.
All of the text above (all text in a text file) makes up the resource. It's possible that a given resource will contain no natural language content at all (consider an HTML document consisting of four empty
div elements styled to be orange rectangles). It's also possible that a resource will contain no syntactic content and consist solely of natural language content: for example, a plain text file with a soliloquy from Hamlet in it. Notice too that the HTML entity
’ appears in the natural language content and belongs to both the natural language content and the syntactic content in this resource.
As well as sections marked as non-normative, all authoring guidelines, diagrams, examples, and notes in this specification are non-normative. Everything else in this specification is normative.
The key words MAY, MUST, MUST NOT, RECOMMENDED, SHOULD, and SHOULD NOT are to be interpreted as described in [RFC2119].
This document describes best practices for the authors of other specifications, as well as recommendations for implementations and content authors. These best practices can also be found in the Internationalization Working Group's document [INTERNATIONAL-SPECS], which is intended to serve as a general reference for all Internationalization best practices in W3C specifications.
When a best practice or recommendation appears in this document, it has been styled to like this paragraph. Recommendations for specifications and spec authors are preceded by [S]. Recommendations for implementations and software developers are preceeded by [I]. Recommendations for content and content authors are preceeded by [C].
Best practices in this document use [RFC2119] keywords in order to clarify the Internationalization Working Group's intentions regarding a specific recommendation. Following the recommendations in this document can help avoid issues during the W3C's "wide review" process, during implementation, or in the content that authors produce. This document is not, itself, normative and can be revised from time to time.
Specifications can claim conformance to this document if they:
Requirements placed on specifications might indirectly cause requirements to be placed on implementations or content that claim to conform to those specifications.
Where this specification contains a procedural description, it is to be understood as a way to specify the desired external behavior. Implementations MAY use other means of achieving the same results, as long as observable behavior is not affected.
The Web is primarily made up of document formats and protocols based on character data. These formats or protocols can be viewed as a set of resources consisting mainly of text files that include some form of structural markup or syntactic content. Processing such syntactic content or document data requires string-based operations such as matching (including regular expressions), indexing, searching, sorting, and so forth.
Users, particularly implementers, sometimes have naïve expectations regarding the matching or non-matching of similar strings or of the efficacy of different transformations they might apply to text, particularly to syntactic content, but including many types of text processing on the Web.
Because fundamentally the Web is sensitive to the different ways in which text might be represented in a document, failing to consider the different ways in which the same text can be represented can confuse users or cause unexpected or frustrating results. In the sections below, this document examines the different types of text variation that affect both user perception of text on the Web and the string processing on which the Web relies.
Some scripts and writing systems make a distinction between UPPER, lower, and Title case characters. Most scripts, such as the Brahmic scripts of India, the Arabic script, and the scripts used to write Chinese, Japanese, or Korean do not have a case distinction, but some important ones do. Examples of such scripts include the Latin script used in the majority of this document, as well as scripts such as Greek, Armenian, and Cyrillic.
Case mapping is the process of transforming characters to a specific case, such as upper, lower, or titlecase. For those scripts that have a case distinction, Unicode defines a default UPPER, lower, and Title case character mapping for each Unicode code point. Case mapping, at first, appears simple. However there are variations that need to be considered when treating the full range of Unicode in diverse languages.
Case folding is the process of making two texts which differ only in case identical for comparison purposes, that is, it is meant for the purpose of string matching. This is distinct from case mapping, which is primarily meant for display purposes. As with the default case mappings, Unicode defines default case fold mappings for each Unicode code point. Unicode defines two forms of case fold mapping, which we'll examine below.
Since most scripts do not have a case distinction, as with case mappings, most Unicode code points do not require a case fold mapping. For those characters that
have a case fold mapping, the majority have a simple, straight-forward mapping to a single matching (generally lowercase) code point. Unicode
calls these the
common case fold mappings, as they are shared by Unicode's case fold mappings.
A few characters have a case fold mapping that map one Unicode code point to two or more code points during case folding. These are called the
full case fold mappings. The
common case fold mappings are used together to provide the default case fold mapping for all of Unicode. We refer to this form of case folding as full casefolding or Unicode full in this document.
Because some applications cannot allocate additional storage when performing a case fold operation, Unicode provides a
simple case fold mapping that maps characters that would normally map to more or
fewer code points to use a single code point for comparison purposes instead. Unlike the full mapping, this mapping invariably alters the content (and potentially the meaning) of the text. As with full casefolding, the simple casefolding or Unicode simple casefold, is a combination of
common mappings so as to cover the full range of Unicode. Unicode simple is not appropriate for use on the Web.
Note that case folding removes information from a string which cannot be recovered later. For example, two s letters in German do not necessarily represent ß in unfolded text.
Another aspect of case mapping and case folding is that it can be language sensitive. Unicode defines default case mappings and case fold mappings for each encoded character, but these are only defaults and are not appropriate in all cases. Some languages need case-folding to be tailored to meet specific linguistic needs. One common example of this are Turkic languages written in the Latin script:
While the example above (and this document in general) focuses on case folding for the purpose of matching, note that case mapping is also language-specific. The name of the second largest city in Turkey is "
contains both the dotted and dotless letters i.
Some document formats or protocols seek to aid interoperability or provide an aid to content authors by ignoring case variations in the vocabulary they define or in user-supplied values permitted by the format or protocol.
Sometimes case can vary in a way that is not semantically meaningful or is not fully under the user's control. This is particularly true when searching a document, but may sometimes also apply when defining rules for matching user- or content-generated values, such as identifiers. In these situations, case-insensitive matching might be desirable instead.
When defining a vocabulary, one important consideration is whether the values are restricted to the ASCII [ASCII] subset of Unicode or if the vocabulary permits the use of characters (such as accents on Latin letters or a broad range of Unicode including non-Latin scripts) that potentially have more complex case folding requirements. To address these different requirements, there are four types of casefold matching defined by this document for the purposes of string identity matching in document formats or protocols:
Case sensitive matching: code points are compared directly with no case folding.
ASCII case-insensitive matching compares a sequence of code points as if all ASCII code points in the range 0x41 to 0x5A (A to Z) were mapped to the corresponding code points in the range 0x61 to 0x7A (a to z). When a vocabulary is itself constrained to ASCII, ASCII case-insensitive matching can be required.
Unicode case-insensitive matching compares a sequence of code points as if the Unicode full casefolding (see above) had been applied to both input sequences.
Language-sensitive case-sensitive matching is useful in the rare case where a document format or protocol contains information about the language of the syntactic content and where language-sensitive case folding might sensibly be applied. These case-fold mappings are defined in the Common Locale Data Repository [UAX35] project of the Unicode Consortium.
For advice on how to handle case folding see 3.1.4 Choice of Case Folding.
A different kind of variation can occur in Unicode text: sometimes several different Unicode code point sequences can be used to represent the same abstract character. When searching or matching text by comparing code points, these variations in encoding cause text values not to match that users expect to be the same.
Because applications need to find the semantic equivalence in texts that use different code point sequences, Unicode defines a means of making two semantically equivalent texts identical: the Unicode Normalization Forms [UAX15].
Resources are often susceptible to the effects of these variations because their specifications and implementations on the Web do not require Unicode Normalization of the text, nor do they take into consideration the string matching algorithms used when processing the syntactic content and natural language content later. For this reason, content developers need to ensure that they have provided a consistent representation in order to avoid problems later.
However, it can be difficult for users to assure that a given resource or set of resources uses a consistent textual representation because the differences are usually not visible when viewed as text. Tools and implementations thus need to consider the difficulties experienced by users when visually or logically equivalent strings that "ought to" match (in the user's mind) are considered to be distinct values. Providing a means for users to see these differences and/or normalize them as appropriate makes it possible for end users to avoid failures that spring from invisible differences in their source documents. For example, the W3C Validator warns when an HTML document is not fully in Unicode Normalization Form C.
Unicode defines two types of equivalence between characters: canonical equivalence and compatibility equivalence.
Canonical equivalence is a fundamental equivalency between Unicode characters or sequences of Unicode characters that represent the same abstract character. When correctly displayed, these should always have the same visual appearance and behavior. Generally speaking, two canonically equivalent Unicode texts should be considered to be identical as text. Unicode defines a process called canonical decomposition that removes these primary distinctions between two texts.
Examples of canonical equivalence defined by Unicode include:
Compatibility equivalence is a weaker equivalence between Unicode characters or sequences of Unicode characters that represent the same abstract character, but may have a different visual appearance or behavior. Generally the process called compatibility decomposition removes formatting variations, such as superscript, subscript, rotated, circled, and so forth, but other variations also occur. In many cases, characters with compatibility decompositions represent a distinction of a semantic nature; replacing the use of distinct characters with their compatibility decomposition can therefore change the meaning of the text. Texts that are equivalent after compatibility decomposition often were not perceived as being identical beforehand and SHOULD NOT be treated as equivalent by a formal language.
In the above table, it is important to note that the characters illustrated are actual Unicode codepoints, not just presentational variations due to context or style. Each character was encoded into Unicode for compatibility with various legacy character encodings. They should not be confused with the normal kinds of presentational processing used on their non-compatibility counterparts.
For example, most Arabic-script text uses the characters in the Arabic script block of Unicode (starting at U+0600). The actual glyphs used to display the text are selected using fonts and text processing logic based on the position inside a word (initial, medial, final, or isolated), in a process called "shaping". In the table above, the four presentation forms of the Arabic letter ن [U+0646 ARABIC LETTER NOON] are shown. The characters shown are compatibility characters in the U+FE00 block, each of which represents a specific "positional" shape and each of the four code points shown have a compatibility decomposition to the regular Arabic letter ن [U+0646 ARABIC LETTER NOON].
Similarly, the variations in half-width and full-width forms and rotated characters (for use in vertical text) are encoded as separate code points, mainly for compatibility with legacy character encodings. In many cases these variations are associated with the Unicode properties described in East Asian Width [UAX11]. See also Unicode Vertical Text Layout [UTR50] for a discussion of vertical text presentation forms.
In the case of characters with compatibility decompositions, such as those shown above, the K Unicode Normalization forms convert the text to the "normal" or "expected" Unicode code point. But the existence of these compatibility characters cannot be taken to imply that similar appearance variations produced in the normal course of text layout and presentation are affected by Unicode Normalization. They are not.
These two types of Unicode-defined equivalence are then grouped by another pair of variations: "decomposition" and "composition". In "decomposition", separable logical parts of a visual character are broken out into a sequence of base characters and combining marks and the resulting code points are put into a fixed, canonical order. In "composition", the decomposition is performed and then any combining marks are recombined, if possible, with their base characters. Note that this does not mean that all of the combining marks have been removed from the resulting normalized text.
Roughly speaking, NFC is defined such that each combining character sequence (a base character followed by one or more combining characters) is replaced, as far as possible, by a canonically equivalent precomposed character. Text in a Unicode character encoding form (such as UTF-8 or UTF-16) is said to be in NFC if it doesn't contain any combining sequence that could be replaced with a precomposed character and if any remaining combining sequence is in canonical order.
Users are cautioned that the resulting character sequence can still contain combining marks: not all character sequences have a precomposed equivalent and some scripts depend on combining marks for encoding. There are even cases where a given base character and combining mark is not replaced with a precomposed character because the combination is "blocked" by another combining mark in the sequence.
There are four Unicode Normalization Forms. Each form is named using a letter code:
Unicode Normalization reduces these (and other potential sequences of escapes representing the same character) to just three possible variations. However, Unicode Normalization doesn't remove all textual distinctions and sometimes the application of Unicode Normalization can remove meaning that is distinctive or meaningful in a given context. For example:
NFKC), becomes an ASCII character sequence: 81/2.
Many users are surprised to find that two identical-looking strings—including
those that have had a specific Unicode normalization form applied—might not
in fact use the same underlying Unicode code points. This includes strings that have
had the more-destructive
NFKD compatibility normalization
forms applied to them. Even when strings, tokens, or identifiers appear visually to be the same,
they can be encoded differently.
The Unicode canonical normalization forms are concerned with folding the multiple different code point sequences that can be used for a given abstract character or grapheme cluster to use the same code point sequence. However, logically distinct characters or grapheme clusters can still look the same or very similar. When a pair of graphemes look identical (or very similar), they are called homographs. When a pair of graphemes look similar or are homographs but actually represent logically different characters or character sequences, they are said to be
Examples of identical or identical-seeming appearance can appear even within a single script. This can take the form of similarly shaped characters, such as "0" and "O" or "l" and "1". But other scripts or the use of different compatibility characters can present much less readily distinguished variations. In some cases, Unicode Normalization brings these together, but in many other cases it does not.
Characters that are identical or
confusable in appearance can present spoofing and
other security risks. This can be true within a single script or for similar characters in
separate scripts. For further discussion and examples of homoglyphs and confusability,
one useful reference is [UTS39].
In addition to identical or similar-appearing characters, the opposite problem also exists:
Unicode Normalization, even the
NFKD Compatibility forms,
does not bring together characters that have the same intrinsic meaning or function
but which vary in appearance or usage. For example,
U+3002 (。) both function as sentence ending punctuation,
but the distinction is not removed by normalization because the characters have a distinct identity.
Most document formats or protocols provide an escaping mechanism to permit the inclusion of characters that are otherwise difficult to input, process, or encode. These escaping mechanisms provide an additional equivalent means of representing characters inside a given resource. They also allow for the encoding of Unicode characters not represented in the character encoding scheme used by the document.
For further discussion of character escapes, including guidelines for the definition of escaping mechanisms in specifications, see: Section 4.6 of [CHARMOD].
The expansion of character escapes and includes is dependent on context, that is, on which syntactic content or programming language is considered to apply when the string matching operation is performed. Consider a search for the string suçon in an XML document containing
suçon but not
suçon. If the search is performed in a plain text editor, the context is plain text (no syntactic content or programming language applies), the
ç character escape is not recognized, hence not expanded and the search fails. If the search is performed in an XML browser, the context is
XML, the character escape (defined by XML) is expanded and the search succeeds.
An intermediate case would be an XML editor that purposefully provides a view of an XML document with entity references left unexpanded. In that case, a search over that pseudo-XML view will deliberately not expand entities: in that particular context, entity references are not considered includes and need not be expanded
For example, € U+20AC
EURO SIGN can also be encoded in HTML as the hexadecimal
€ or as the decimal entity
\u20ac or as
while in a CSS stylesheet it can appear as
\20ac. All of
these representations encode the same literal character value: €.
Character escapes are normally interpreted before a document is processed and strings within the format or protocol are matched. Returning to an example we used above:
You would expect that text to display like the following: Hello world!
In order for this to work, the user-agent (browser) had to match two strings representing the class name
héllo, even though the CSS and HTML each used a different escaping mechanism. The above fragment demonstrates one way that text can vary and still be considered "the same" according to a specification: the class name
h\e9llo matched the class name in the HTML mark-up
héllo (and would also match the literal value
héllo using the code point é [U+00E9 LATIN SMALL LETTER E WITH ACUTE]).
Formal languages and document formats often offer facilities for including a piece of text from one resource inside another. An include is a mechanism for inserting content into the body of a resource. Include mechanisms import content into a resource at processing time. This affects the structure of the document and potentially matching against the vocabulary of the document. Examples of includes are entity references in XML, the XInclude [XInclude] specification, and @import rules in CSS.
An include is said to be include normalized if it does not begin with a combining mark (either in the form of a character escape or as a character literal in the included resource).
Unicode provides a number of special-purpose characters that help document authors control the appearance or performance of text. Because many of these characters are invisible or do not have keyboard equivalents, users are not always aware of their presence or absence. As a result, these characters can interfere with string matching when they are part of the encoded character sequence but the expected matching text does not also include them. Some examples of these characters include:
The Unicode control characters U+200D Zero Width Joiner (also known as ZWJ) and U+200C Zero Width Non-Joiner (also known as ZWNJ). While these characters can be used to control ligature formation—either preventing the formation of undesirable ligatures or encouraging the formation of desirable ones—their primary use is to control the joining and shape selection in complex scripts such as the Arabic or various of the Indic scripts. Some Indic scripts use the ZWJ and ZWNJ characters to allow authors to control the shape that certain conjuncts take. See the discussion in Chapter 12 of [Unicode].
The Zero Width Non-Joiner is used in Persian to prevent certain "normal" Arabic script joining. In these cases, the presence or absence of the character does affect the meaning. For example, the word تنها ("alone") and the word تنها ("bodies" or "corpuses") are encoded as "U+062A U+0646 U+0647 U+0627" and "U+062A U+0646 U+200C U+0647 U+0627" respectively, the only difference being the ZWNJ in the latter word.
The ZWJ character is also used in forming certain emoji sequences, which is discussed in more detail below.
Variation selectors (U+FE00 through U+FE0F) are characters used to select an alternate appearance or glyph (see Character Model: Fundamentals [CHARMOD]). For example, they are used to select between black-and-white and color emoji. These are also used in predefined ideographic variation sequences (IVS). Many examples are given in the "Standardized Variants" portion of the Unicode Character Database (UCD).
A few scripts also provide a way to encode visual variation selection: a prominent example of this are the Mongolian script's free variation selectors (U+180B through U+180D).
The character U+034F Combining Grapheme Joiner, whose name is misleading (as it does not join graphemes or affect line breaking), is used to separate characters that might otherwise be considered a grapheme for the purposes of sorting or to provide a means of maintaing certain textual distinctions when applying Unicode normalization to text.
Whitespace variations can also affect the interpretation and matching of text. For example, the various non-breaking space characters, such as NBSP, NNBSP, etc.
U+200B Zero Width Space is a character used to indicate word boundaries in text where spaces do not otherwise appear. For example, it might be used in a Thai language document to assist with word-breaking.
The U+00AD Soft Hyphen can be used in text to indicate a potential or preferred hyphenation position. It only becomes visible when the text is reflowed to wrap at that position.
The U+2060 WORD JOINER, sometimes called WJ is a zero-width non-breaking space character. Its purpose is to replace the functionality of the character U+FEFF ZERO WIDTH NO-BREAK SPACE because that character also serves as the "Byte Order Mark" character (used as a Unicode signature in plain text files). The Word Joiner is used to separate words in languages that do not use explicit spacing. An example would be the Thai language.
Finally, some scripts, such as Arabic and Hebrew, are written prodominently from right-to-left. Text written in these scripts can also
include character sequences, such as numbers or quotes in another script, that are left-to-right. This intermixing of text direction
is called bidirectional text or
bidi for short. The Unicode Bidirectional Algorithm
[UAX9] describes how such mixed-direction text is processed for display. For most text, the directional handling can be derived
from the text itself. However, there are many cases in which the algorithm needs additional information in order to present text
correctly. For more examples, see [html-bidi].
One of the ways that Unicode defines to address the ambiguity of text direction are a set of invisible control characters to mark the start and end of directional runs. While bidirectional controls can have an affect on the appearance of the text (since they help the Unicode Bidirectional Algorithm with the presentation of text), they might have no effect on the text if the text would naturally have fallen into bidirectional runs without the controls. Because these controls are, like the characters mentioned above, invisible, they can have an unintentional effect on matching.
In almost all of these cases, users may not be aware of or cannot be sure if a given document or text string has included or omitted one of these characters. Because text matching depends on matching the underlying codepoints, variation in the encoding of the text due to these markers can cause matches that ought to succeed to mysteriously fail (from the point of view of the user).
A newer feature of Unicode are the emoji characters. In [UTR51], Unicode describes these as:
Emoji are pictographs (pictorial symbols) that are typically presented in a colorful cartoon form and used inline in text. They represent things such as faces, weather, vehicles and buildings, food and drink, animals and plants, or icons that represent emotions, feelings, or activities.
Emoji can be used with a variety of emoji modifiers, including U+200D ZERO WIDTH JOINER or ZWJ, to form more complex emoji.
For example, the emoji (👪 [U+1F46A FAMILY]) can also be formed by using ZWJ between emoji characters in the sequence U+1F468 U+200D U+1F469 U+200D U+1F466. Altering or adding other emoji characters can alter the composition of the family. For example the sequence 👨👩👧👧 U+1F468 U+200D U+1F469 U+200D U+1F467 U+200D U+1F467 results in a composed emoji character for a "family: man, woman, girl, girl" on systems that support this kind of composition. Many common emoji can only be formed using ZWJ sequences. For more information, see [UTR51].
Emoji characters can be followed by emoji modifier characters. These modifiers allow for the selection of skin tones for emoji that represent people. These characters are normally invisible modifiers that follow the base emoji that they modify. For example: 👨 👨🏻 👨🏼 👨🏽 👨🏾 👨🏿
An emoji character can also be followed by a variation selector to indicate text (black and white, indicated by U+FF0E Variation Selector 15) or color (indicated by U+FF0F Variation Selector 16) presentation of the base emoji.
Still another wrinkle in the use of emoji are flags. National flags can be composed using country codes derived from the [BCP47] registry, such as the sequence 🇿 [U+1F1FF REGIONAL INDICATOR SYMBOL LETTER Z] 🇲 [U+1F1F2 REGIONAL INDICATOR SYMBOL LETTER M], which is the country code (ZM) for the country Zambia: 🇿🇲. Other regional or special purpose flags can be composed using a flag emoji with various symbols or with regional indicator codes terminating in a cancel tag. For example, the flag of Scotland (🏴) can be composed like this:
Each of these mechanisms can be used together, so quite complex sequences of characters can be used to form a single emoji grapheme or image. Even very similar emoji sequences might not use the same exact encoded sequence. Many of the modifiers and combinations mentioned above are generated by the end-user's keyboard (where they are presented as a single emoji "character"), so users may not be aware of the underlying encoding complexity. Emoji sequences are evolving rapidly, so there could be additional developments to either help or hinder matching of emoji in the near future. Currently Unicode normalization does not reorder these sequences or insert or remove any of the modifiers. Users and implementers are therefore cautioned that users who employ emoji characters in namespaces and other matching contexts might encounter unexpected character mismatches.
Resources can use different character encoding schemes, including legacy character encodings, to serialize document formats on the Web. Each character encoding scheme uses different byte values and sequences to represent a given subset of the Universal Character Set.
Choosing a Unicode character encoding, such as UTF-8, for all documents, formats, and protocols is a strongly encouraged recommendation, since there is no additional utility to be gained from using a legacy character encoding and the considerations in the rest of this section would be completely avoided.
For example, € [U+20AC EURO SIGN]) is encoded as the byte sequence
UTF-8 character encoding. This same
character is encoded as the byte sequence
0x80 in the
legacy character encoding
(Other legacy character encodings may not provide any byte sequence to
encode the character.)
Specifications mainly address these resulting variations by considering each document to be a sequence of Unicode characters after converting from the document's character encoding (be it a legacy character encoding or a Unicode encoding such as UTF-8) and then unescaping any character escapes before proceeding to process the document.
Even within a single legacy character encoding there can
be variations in implementation. One famous example is the legacy
transcoder implementations faced choices about how to map specific
byte sequences to Unicode. So the byte sequence
0x2141 in the JIS X 0208 character set) was mapped by
some implementations to U+301C
WAVE DASH while others chose U+FF5E
FULL WIDTH TILDE. This means that two reasonable,
self-consistent, transcoders could produce different Unicode character
sequences from the same input. The Encoding [Encoding]
specification exists, in part, to ensure that Web implementations use
interoperable and identical mappings. However, there is no guarantee
that transcoders consistent with the Encoding specification will be
applied to documents found on the Web or used to process data
appearing in a particular document format or protocol.
One additional consideration in converting to Unicode is the existence of legacy character encodings of bidirectional scripts (such as Hebrew and Arabic) that use a visual storage order. That is, unlike Unicode and other modern encodings, the characters are stored in memory in the order that they are printed on the screen from left-to-right (as with a line printer). When converting these encodings to Unicode or when comparing text in these encodings, care must be taken to place both the source and target text into logical order. For more information, see Section 3.3.1 of [CHARMOD]
There are additional kinds of equivalence or processing that are appropriate when performing natural language searching or "find" features. These are described in another part of the Character Model series of documents ([STRING-SEARCH]). Specifications for a vocabulary or which define a matching algorithm for use in a formal syntax SHOULD avoid trying to apply additional custom folding, mapping, or processing such as described in that document, since these interfere with producing consistent, predictable results.
In the Web environment, where strings can be encoded in different encodings, using different character sequences, and with variations such as case, it's important to establish a consistent process for evaluating string identity.
This chapter defines the implementation and requirements for string matching in syntactic content.
This section defines the algorithm for matching strings in a formal language or syntax. Specifications need to specify certain options called out below. Recommendations are provided for best practices in the sub-sections below.
Expand all character escapes and includes.
[C] Content authors SHOULD enter and store resources in a Unicode character encoding (generally UTF-8 on the Web).
[C] Content authors SHOULD choose a normalizing transcoder when converting legacy encoded text or resources to Unicode unless the mapping of specific characters interferes with the meaning.
[S] Specifications MUST allow a Unicode character encoding.
[S] Specifications MUST specify a default character encoding and SHOULD specify UTF-8 as the default encoding.
[S] Specifications SHOULD disallow encodings other than UTF-8.
The first step in comparing text is to ensure that both use the same digital representation. This means that implementations need to convert any text in a legacy character encoding to a sequence of Unicode code points. Normally this is done by applying a transcoder to convert the data to a consistent Unicode encoding form (such as UTF-8 or UTF-16). This allows bitwise comparison of the strings in order to determine string equality.
A normalizing transcoder is a transcoder that performs a conversion from a legacy character encoding to Unicode and ensures that the result is in Unicode Normalization Form C (NFC). For most legacy character encodings, it is possible to construct a normalizing transcoder (by using any transcoder followed by a normalizer); it is not possible to do so if the legacy character encoding's repertoire contains characters not represented in Unicode. While normalizing transcoders only produce character sequences that are in NFC, the converted character sequence might not be include normalized (for example, if it begins with a combining mark).
Because document formats on the Web often interact with or are processed using additional, external resources (for example, a CSS style sheet being applied to an HTML document), the consistent representation of text becomes important when matching values between documents that use different character encodings. Use of a normalizing transcoder helps ensure interoperability by making legacy encoded documents match the normally expected Unicode character sequence for most languages.
Most transcoders used on the Web produce NFC as their output, but several do not. This is usually to allow the transcoder to be round-trip compatible with the source legacy character encoding, to preserve other character distinctions, or to be consistent with other transcoders in use in user-agents. This means that the Encoding specification [Encoding] and various other important transcoding implementations include a number of non-normalizing transcoders. Indeed, most compatibility characters in Unicode exist solely for round-trip conversion from legacy encodings and a number of these have singleton canonical mappings in NFC. You saw an example of this earlier in the document with Å [U+212B ANGSTROM SIGN].
Bear in mind that most transcoders produce NFC output and that even those transcoders that do not produce NFC for all characters produce NFC for the preponderence of characters. In particular, there are no commonly-used transcoders that produce decomposed forms where precomposed forms exist or which produce a different combining character sequence from the normalized sequence (and this is true for all of the transcoders in [Encoding]).
Most document formats and protocols provide a means for encoding characters as an escape sequence or including external data, including text, into a resource. This is discussed in detail in Section 4.6 of [CHARMOD] as well as above.
When performing matching, it is important to know when to interpret character escapes so that a match succeeds (or fails) appropriately. Normally, escapes, references, and includes are processed or expanded before performing matching (or match-sensitive processing), since these syntaxes exist to allow difficult-to-encode sequences to be put into a document conveniently, yet allowing the characters to behave as-if they were directly encoded as a codepoint sequence in the document in question.
One area where this can be complicated is deciding how syntactic content and natural language content interact. For example, consider the following snippet of HTML:
Although technically the combining mark ̀ [U+0300 COMBINING GRAVE ACCENT] combines with the preceding quote mark, HTML does not consider the character (whether or not it is encoded as an entity) to form part of the HTML syntax.
When performing a matching operation on a resource, the general rule is to expand escapes on the same "level" as the user is interacting with. For example, when considering the above example, a tool to view the source of the HTML would show the escape sequence
U+0300 as the value of the attribute
When processing the syntax of a document format, escapes are usually converted to the character sequence they represent before the processing of the syntax, except where explicitly forbidden by the format's processing rules. This allows resources to include characters of all types into the resource's syntactic structures.
In some cases, pre-processing escapes creates problems. For example, expanding the sequence
< before parsing an HTML document would produce document errors.
A specific Unicode normalization form is not always appropriate or available to content authors and the text encoding choices of users might not be obvious to downstream consumers of the data. As shown in this document, there are many different ways that content authors or applications could choose to represent the same semantic values when inputting or exchanging text. Normalization can remove distinctions that the users applied intentionally. Therefore:
[S] Specifications SHOULD NOT specify the Unicode normalization in string matching for vocabularies.
[I] Implementations MUST NOT alter the normalization form of syntactic or natural language content being exchanged, read, parsed, or processed except when required to do so as a side-effect of text transformation such as transcoding the content to a Unicode character encoding, case mapping or folding, or other user-initiated change, as consumers or the content itself might depend on the de-normalized representation.
[I] Authoring tools SHOULD provide a means of normalizing resources and warn the user when a given resource is not in Unicode Normalization Form C.
[S] Specifications of text-based formats and protocols that as part of their syntax definition require the text be in a normalized form MUST define string matching in terms of normalized string comparison and MUST define the normalized form to be NFC.
A specification that defines comparison of text in a normalized format needs to address the requirements in 126.96.36.199 Considerations When Requiring Normalization.
Specifications are generally discouraged from requiring formats or protocols to store or exchange data in a normalized form unless there are specific, clear reasons why the additional requirement is necessary. As many document formats on the Web do not require normalization, content authors might occasionally rely on denormalized character sequences. A normalization step could negatively affect such content.
The canonical normalization forms (form NFC or form NFD) are intended to preserve the meaning and presentation of the text to which they are applied. This is not always the case, which is one reason why normalization is not recommended. NFC has the advantage that almost all legacy data (if transcoded trivially, one-to-one, to a Unicode encoding), as well as data created by current software or entered by users on most (but not all) keyboards, is already in this form. NFC also has a slight compactness advantage and is a better match to user expectations in most languages with respect to the relationship between characters and graphemes.
[S] Specifications SHOULD NOT specify compatibility normalization forms (NFKC, NFKD).
[I] Implementations MUST NOT apply compatibility normalization forms (NFKC, NFKD) unless specifically requested by the end user.
The compatibility normalization forms (form NFKC and form NFKD) change the structure and lose the meaning of the text in important ways. Users sometimes use characters with a compatibility mapping in Unicode on purpose or they use characters in a legacy character encoding that have a compatibility mapping when converted to Unicode. This has to be considered intentional on the part of the content author. Although NFKC/NFKD can sometimes be useful in "find" operations or string searching natural language content, erasing compatibility differences is harmful.
Requiring NFC requires additional care on the part of the specification developer, as content on the Web generally is not in a known normalization state. Boundary and error conditions for denormalized content need to be carefully considered and well-specified in these cases.
[S] Specifications MUST document or provide a health-warning if canonically equivalent but disjoint Unicode character sequences represent a security issue.
[C] Content authors SHOULD use Unicode Normalization Form C (NFC) wherever possible for content. Note that NFC is not always appropriate to the content or even available to content authors in some languages.
[C] Content authors SHOULD always encode text using consistent Unicode character sequences to facilitate matching, even if a Unicode normalization form is included in the matching performed by the format or implementation.
In order for their content to be processed consistently, content authors should try to use a consistent sequence of code points to represent the same text. While content can be in any normalization form or might use a de-normalized (but valid) Unicode character sequence, inconsistency of representation will cause implementations to treat the different sequences as different. The best way to ensure consistent selection, access, extraction, processing, or display is to always use NFC.
[C] Content authors SHOULD NOT include combining marks without a preceding base character in a resource.
There can be exceptions to this. For example, when making a list of characters (such as a list of [Unicode] characters), an author might want to use combining marks without a corresponding base character. However, use of a combining mark without a base character can cause unintentional display or, with naive implementations that combine the combining mark with adjacent syntactic content or other natural language content, processing problems. For example, if you were to use a combining mark, such as the character ́ [U+0301 COMBINING ACUTE ACCENT], as the start of a
class attribute value in HTML, the class name might not display properly in your editor and be difficult to edit.
Some recommended base characters include ◌ [U+25CC DOTTED CIRCLE] (when the base character needs to be visible) or [U+00A0 NO-BREAK SPACE] (when the base character should be invisible).
Since content authors do not always following these guidelines:
[S] Specifications of vocabularies MUST define the boundaries between syntactic content and character data as well as entity boundaries (if the language has any include mechanism). These need to include any boundary that may create conflicts when processing or matching content when instances of the language are processed, while allowing for character escapes designed to express arbitrary characters.
When a specification requires Unicode normalization for storage, transmission, or string matching, some additional considerations need to be addressed by the specification authors as well as by implementers of that specification:
[S] Where operations can produce denormalized output from normalized text input, specifications MUST define whether the resulting output is required to be normalized or not. Specifications MAY state that performing normalization is optional for some operations; in this case the default SHOULD be that normalization is performed, and an explicit option SHOULD be used to switch normalization off.
[S] Specifications that require normalization MUST NOT make the implementation of normalization optional. Interoperability of matching cannot be achieved if some implementations normalize while others do not.
An implementation that is required to perform normalization needs to consider these requirements:
[I] Normalization-sensitive operations MUST NOT be performed unless the implementation has first either confirmed through inspection that the text is in normalized form or it has re-normalized the text itself. Private agreements MAY be created within private systems which are not subject to these rules, but any externally observable results MUST be the same as if the rules had been obeyed.
[I] A normalizing text-processing component which modifies text and performs normalization-sensitive operations MUST behave as if normalization took place after each modification, so that any subsequent normalization-sensitive operations always behave as if they were dealing with normalized text.
[I] Authoring tool implementations SHOULD warn users or prevent the input or creation of syntactic content starting with a combining mark that could interfere with processing, display, or interchange.
One important consideration in string identity matching is whether the comparison is case sensitive or case insensitive.
[C] Content authors SHOULD always spell identifiers using consistent upper, lower, and mixed case formatting to facilitate matching, even if case-insensitive matching is supported by the format or implementation.
[S] Case-sensitive matching is RECOMMENDED for matching syntactic content, including user-defined values.
Vocabularies usually put a premium on predictability for content authors and users. Case-sensitive matching is the easiest to implement and introduces the least potential for confusion, since it generally consists of a comparison of the underlying Unicode code point sequence. Because it is not affected by considerations such as language-specific case mappings, it produces the least surprise for document authors that have included words, such as the Turkish examples above, in their syntactic content.
Case insensitivity is usually reserved for processing natural language content, such as running a feature for searching text. However, cases exist in which case-insensitivity is desirable. When case-insensitive matching is necessary, there are several implementation choices that a formal language needs to consider.
[S] Specifications that define case-insensitive matching in vocabularies that include more than the Basic Latin (ASCII) range of Unicode MUST specify Unicode full casefold matching.
[S] Specifications SHOULD allow the full range of Unicode for user-defined values.
Vocabularies generally should allow for a wide range of Unicode characters, particularly for user-supplied values, so as to enable use by the broadest range of languages and cultures without disadvantage. As a result, text operations such as case folding need to address the full range of Unicode and not just selected portions. When case-insensitive matching is desired, this means using Unicode case folding:
The Unicode simple casefolding form is not appropriate for string identity matching on the Web.
[S] Specifications that define case-insensitive matching in vocabularies limited to the Basic Latin (ASCII) subset of Unicode MAY specify ASCII case-insensitive matching.
A formal language whose vocabulary is limited to ASCII and which does not allow user-defined names or identifiers can specify ASCII case-insensitive matching. An example of this is HTML, which defines the use of ASCII case-insensitive comparison for element and attribute names defined by the HTML specification.
A vocabulary is considered to be "ASCII-only" if and only if all tokens and identifiers are defined by the specification directly and these identifiers or tokens use only the Basic Latin subset of Unicode. If user-defined identifiers are permitted, the full range of Unicode characters (limited, as appropriate, for security or interchange concerns, see [UTR36]) should be allowed and Unicode case insensitivity used for identity matching.
An ASCII-only vocabulary can exist inside a document format or protocol that allows a larger range of Unicode in identifiers or values. For example [CSS-SYNTAX-3] defines the format of CSS style sheets in a way that allows the full range of Unicode to be used for identifiers and values. However, CSS specifications always define CSS keywords using a subset of the ASCII range. The vocabulary of CSS is thus ASCII-only, even though many style sheets contain identifiers or data values that are not ASCII.
Locale- or language-specific tailoring is most appropriate when it is part of natural language processing operations (which is beyond the scope of this document). Because language-specific tailoring of case mapping or case folding produces different results from the generic case folding rules, these should be avoided in formal languages, where predictability is at a premium.
[S] Specifications that define case-insensitive matching in vocabularies SHOULD NOT specify language-sensitive case-insensitive matching.
[S] If language-sensitive case-sensitive matching is specified, Unicode case-fold mappings SHOULD be tailored according to language and the source of the language used for each tailoring MUST be specified.
Two strings being matched can be in different languages and might appear in yet a third language context. Which language to use for case folding therefore depends on the application and user expectations.
Language specific tailoring is not recommended for formal languages because the language information can be hard to obtain, verify, or manage and because the resulting operations can produce results that frustrate users or which fail for some users and succeed for others depending on the language configuration that they are using or the configuration of the system where the match is performed.
[S] Operations that are language-specific SHOULD include language-specific case folding where appropriate.
For example, the CSS operation
text-transform is language-sensitive when used to case map strings.
Although Unicode case folding is the preferred case-insensitive matching for document formats and protocols, content authors and users of languages that have mappings different from the default can still be surprised by the results, since their expectations are generally consistent with the languages that they speak.
Language-sensitive string comparison is often referred to as being locale-sensitive, since most programming languages and operating environments access language-specific tailoring using their respective locale-based APIs. For example, see the
java.text.Collator class in the Java programming language or
[S] Specificiations MUST clearly define any additional tailoring done as part of the matching process.
Some specifications might wish to include additional tailoring to assist with matching in a given vocabulary. Examples of this might include removing additional textual differences described in Section 2, mapping together or removing characters that are part of the syntax, or performing a whitespace trim.
Any additional tailoring needs to avoid interfering with the way that different languages are represented in Unicode. For example, a process that attempts to remove accents from letters by decomposing the text and then removing all of the combining characters will break languages that rely on combining marks. An example of this would be as the Devanagari text in Example 2. (Such a process would also fail to remove all of the potential accents and probably do harm to the meaning and representation of the text.)
While matching strings and tokens in a formal language is the primary concern of this document, sometimes a specification needs to consider additional types of matching beyond pure string equality.
[S] Specifications that define a regular expression syntax MUST provide at least Basic Unicode Level 1 support per [UTS18] and SHOULD provide Extended or Tailored (Levels 2 and 3) support.
Regular expression syntaxes are sometimes useful in defining a format or protocol, since they allow users to specify values that are only partially known or which can vary in predictable ways. As seen in the various sections of this document, there is variation in the different ways that characters can be encoded in Unicode and this potentially interferes with how strings are specified or matched in expressions. For example, counting characters might need to depend on grapheme boundaries rather than the number of Unicode code points used; caseless matching might need to consider variations in case folding; or the Unicode normalization of the expression or text being processed might need to be considered.
Unicode Regular Expressions Level 1 support includes the ability to specify Unicode code points in regular expressions, including via the use of escapes, and to access Unicode character properties as well as certain kinds of boundaries common to most regular expression syntaxes.
Level 2 extends this with a number of important capabilities, notably the ability to select text on certain kinds of grapheme cluster boundary and support for case conversion (two topics mentioned extensively above). Level 3 provides for locale [LTLI] based tailoring of regular expressions, which are less useful in formal languages but can be useful in processing natural language content.
This document has been extensively revised and rewritten since the Working Draft of 2014-07-15. Please see the github commit log for more details.
The W3C Internationalization Working Group and Interest Group, as well as others, provided many comments and suggestions. The Working Group would like to thank: Mati Allouche, Ebrahim Byagowi, John Cowan, Martin Dürst, Behdad Esfahbod, Asmus Freitag, John Klensin, Peter Saint-Andre, Amir Sarabadani, and all of the CharMod contributors over the twenty (!!) years of this document's development.
The previous version of this document was edited by: