Character Model for the World Wide Web

World Wide Web Consortium Working Draft 29-November-1999

This version:
(ZIP archive)
Latest version:
Previous version:
Martin J. Dürst (W3C) <duerst@w3.org>
François Yergeau (Alis Technologies, Inc.) <yergeau@alis.com>


This model provides authors of specifications, software developers, and content developers a common reference for interoperable text manipulations on the World Wide Web. Topics addressed include encoding identification, early uniform normalization, string identity matching, string indexing, and URI conventions, building on the Universal Character Set (UCS) (refer to [ISO10646] and [Unicode]). Some introductory material on characters and character encodings is also provided.

Status of this document

This is a W3C Working Draft for review by W3C members and other interested parties. It is a draft document and may be updated, replaced, or obsoleted by other documents at any time.

While this document addresses W3C Working Groups explicitly, other groups outside of W3C are strongly encouraged to incorporate the relevant parts of this model into their Web specifications and software.

This document is published as part of the W3C Internationalization Activity by the Internationalization Working Group (I18N WG), with the help of the Internationalization Interest Group (I18N IG). Various parts of this document are in different states of development. The I18N WG will not allow early implementation to constrain its ability to make changes to this specification prior to final release. It is inappropriate to use W3C Working Drafts as reference material or to cite them as other than "work in progress". A list of current W3C working drafts can be found at http://www.w3.org/TR/.

Comments to this Working Draft are very welcome. Comments intended for public discussion and archival should be sent to www-international@w3.org. Comments to the editors should be sent to i18n-editor@w3.org (archived for access by W3C member organizations).

Table of Contents

  1. Introduction
    1. Why is this document necessary?
    2. Document Conventions
  2. Conformance
  3. Characters
    1. Characters as seen by Humans
    2. Characters and their Digital Representation
    3. Choice and Identification of Character Encodings
    4. The UCS as a Common Reference
      1. Reference Processing Model
      2. References to ISO 10646/Unicode
    5. Character Escaping
  4. Character Data Exchange: Early Uniform Normalization
    1. W3C Text Normalization
    2. Application of Early Uniform Normalization
    3. String Identity Matching
    4. Compatibility Equivalents and Control Characters
  5. String Indexing
  6. Character Encoding in URI References
  7. Language Identification
  8. Appendix: Change Log
  9. Glossary
  10. Acknowledgments
  11. References

1. Introduction

The 'character model' described in this document provides authors of specifications, software developers, and content developers with a common reference for interoperable text manipulations on the World Wide Web. Working together, these three groups can build a more international Web. Topics addressed include encoding identification, early uniform normalization, string identity matching, string indexing, and URI conventions, building on the Universal Character Set (UCS) (refer to [ISO10646] and [Unicode]). Some introductory material on characters and character encodings is also provided. The model will allow Web documents authored in the world's scripts (and on different platforms) to be exchanged, read, and searched by Web users around the world.

Although Web developers in general are encouraged to follow the specifications of this document, the document targets W3C Working Groups specifically and lists requirements to ensure interoperability of W3C specifications (refer to the section on conformance). Some of W3C Working Groups and Activities that should be integrating this model into their work are:

Outside of W3C, some areas of work where this document , and in particular Section 4, Character Data Exchange: Early Uniform Normalization, may apply, include:

1.1 Why is this document necessary?

Starting with [RFC 2070], the Web community has recognized the need for a character model for the World Wide Web. W3C's first step towards building this model was the adoption of the UCS (Universal Character Set )(refer to [ISO10646] and [Unicode])as the document character set for HTML 4.0. This choice was motivated by the fact that the UCS:

The UCS meant that HTML documents were not limited to containing ASCII characters. After HTML 4.0, W3C adopted the UCS for other specifications such as XML [XML 1.0] and CSS 2 [CSS2]. UCS now serves as a common reference for W3C's specifications and applications.

Where data transfer on the Web remained unidirectional (from server to browser) , and where the main purpose was to render documents, the use of the UCS without specifying additional details sufficed. However, the Web has grown:

In short, the Web may be seen as a single, very large application [Nicol], rather than as a collection of independent small applications.

While these developments strengthen the requirement that UCS be the basis of a character model for the Web, they also create the need for additional specifications on the application of UCS to the Web. Some properties of the UCS that require additional specification for the Web include:

It should be noted that such properties also exist in legacy encodings, and in many cases have been inherited by the UCS in one way or another from such legacy encodings.

The remainder of this document presents the additional specifications and requirements to ensure an interoperable character model for the Web.

Section 1.2 explains some of the conventions and notation used in the document. Section 2 defines conformance for different consumers of this document.

Section 3 defines a general character model, e.g., in the sense of the reference processing model in [RFC 2070], and general guidelines, e.g., similar to those in [RFC 2130] and [RFC 2277]. Much of section 3 is introductory material for readers who might not be familiar with the complexity of the topic.

Section 4 discusses Early Uniform Normalization, Section 5 string indexing. Section 6 deals with Character Encoding in URI References, Section 7 with Language Identification.

1.2 Document Conventions

UCS codepoints are denoted as U+hhhh, where hhhh is a sequence of hexadecimal digits.

Where this specification contains procedural descriptions, they are understood to be a way to specify the desired external behavior. As long as observable behavior is not affected, implementations may use some other way of achieving the same results.

2. Conformance

In order to conform to this document, all requirements must be satisfied. Requirements vary for content providers, software developers, and specification writers.

Requirements are expressed using the key words "MUST", "MUST NOT", "REQUIRED", "SHALL", and "SHALL NOT". Recommendations are expressed using the key words "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL". Key words are used in accordance with [RFC2119].

3. Characters

The word character is used in many contexts and with different meanings. In the context of texts and digital representations of texts, it can roughly be defined as a small logical unit of a text. See later sections of this chapter for examples.

A text is then defined as sequences of characters. While such a definition is sufficient to create or capture a common understanding in many cases, it is also sufficiently open to create misunderstanding as soon as details start to matter. It is very important to understand where misunderstanding can occur in order to write appropriate specifications, protocol implementations, and software for end users. The term 'character' is used in a variety of contexts and thus leads to confusion when used outside of these contexts.

Why is so difficult to give one definition to the term "character"? There are a number of dimensions to characters that lend themselves to conflicting interpretations:

Consider, for example, how the different dimensions relate to string indexing. Due to the wide variability of scripts and characters, and because of tradeoffs between user friendliness and implementation efficiency, indexing operations, as well as other operations, may be more efficiently carried out at a particular layer.

3.1 Characters as seen by Humans

Perceptions of characters by end users can vary widely based on script, language, function, and context, or just individual differences. In some scripts and languages, in particular in English, such perception differences are almost nonexistent. In other cases the differences are more evident, but even in these cases, the context usually present in human discourse can make it rather difficult to become aware of the fact that even the same person can use the term "character" in slightly or vastly different senses without necessarily been misunderstood. Misunderstanding does not arise due to imperfect technology (i.e., "Unicode just didn't get it right."). Instead, it comes from the high flexibility and creativity of the human mind and the long tradition of writing as an important part of human cultural heritage.

Specification writers using the term "character" MUST specify which meaning(s) they intend. Specification writers SHOULD avoid the use of the term "character" if a more specific term is available. Otherwise, there are many potential sources of misunderstanding, including:

Japanese Hiragana/Katakana are syllabaries, not phonemic alphabets. A character in these scripts is therefore not a phoneme, but a syllable. Korean Hangul combines symbols for phonemes into square syllabic blocks. Depending on the user and the application, both the individual symbols as well as the syllabic clusters can be called characters. Indic scripts use semi-regular or irregular ways to combine consonants and vowels into clusters. Depending on the user and the application, both individual consonants and vowels as well as consonant clusters or consonant-vowel clusters can be seen as characters.
"ö" is considered as a character completely independent from "o" in Swedish, as an extra character quite related to "o" in German, and as an "o" having to be modified in certain contexts in order to make sure it stays an "o" by itself in French or English.
Users interact with characters in a number of ways that raise issues about their definition, including:
Display as a function is discussed separately under Section 3.2.
Note. The current version of this document only gives lists of topics to be addressed for each subsection; the WG plans to address the topics in more detail in the next version of this document.

3.2 Characters and their digital representation

To be of any use in computers, in computer communications and in particular on the World Wide Web, characters must be encoded. In fact, much of the information processed by computers over the last few decades has been encoded text, exceptions being images, audio, video and numeric data. To achieve text encoding, a large variety of encoding schemes have been devised, which can be loosely defined as mappings between the character sequences that users manipulate and the sequences of bits that computers manipulate.

Given the complexity of text encoding and the large variety of schemes for character encoding invented throughout the computer age, however, a more formal description of the encoding process is felt to be useful. Text encoding can be described as follows (see [UTR #17] for a more thorough description):

  1. First, a set of characters to be encoded is identified. The units of encoding, the characters, are pragmatically chosen as appropriate to express text and allow various text processes in one or more target languages. They may not correspond exactly to what users perceive as letters and other characters. The set is called a repertoire.
    Note.Where one way of looking at characters in a given script is predominant, and corresponds for all or most functions, it is rather easy to choose this for the actual digital encoding. For example, nobody has seriously proposed to use syllable-based encoding for the Latin script, and nobody has seriously proposed phoneme-based encoding for Japanese Hiragana and Katakana.
    Where there is more than one way to perceive, identify, and encode the characters of a given script, it is important to realize that:
    • Choosing a single encoding has vast benefits over having to deal with multiple encodings.
    • The choice of encoding has to take into account the various languages and functions without being biased towards one language or function to the extent that the others become overly difficult to realize. In this sense, the choice of encoding is often a technical compromise.
    • In many cases, the choice of encoding is also to a certain extent a political compromise.
  2. Each character of the repertoire is then associated with a (abstract, mathematical) non-negative integer, the character number or code point. The result, a mapping from the repertoire to the set of non-negative integers, is called a coded character set, abbreviated CCS.
  3. To enable use in computers, a suitable base datatype is identified (byte, 16-bit wyde or other) and a character encoding form is devised, which encodes the abstract integers of a CCS into sequences of the base datatype. The encoding form can be extremely simple (for instance, it encodes the integers of the CCS into the natural representation of integers of the chosen datatype of the computing platform) or arbitrarily complex (variable number of base datatype units, value of each unit a non-trivial function of the encoded integer, etc.)
  4. To enable transmission or storage using byte-oriented devices, a serialization scheme or character encoding scheme (CES) is next devised. A CES maps the integers of one or more CCSes to well-defined sequences of bytes, taking into account the necessary specification of byte-order for multi-byte base datatypes and including in some cases switching schemes between multiple CCSes (for instance ISO 2022). A CES, together with the CCSes that it is used with, is what is identified by an IANA charset identifier. Given a sequence of bytes representing text and a charset identifier, one can unambiguously recover the sequence of characters of the text.

In some cases, the whole encoding process can be collapsed to a single step, a trivial one-to-one mapping from characters to bytes; this is the case, for instance, for US-ASCII and ISO/IEC 8859-1. It should be clear, however, that characters and bytes are very different entities that SHOULD NOT be confused: in general, the relationship is many-to-many.

Input and rendering (display, printing) of text are two other areas where complexities occur. In keyboard input, it is not the case that keystrokes and input character correspond one-to-one. Many writing systems have too many characters to allow such a correspondence and must rely on more complex input methods which transform keystroke sequences into character sequences. Thus specification writers and software developers SHOULD NOT assume that a single keystroke results in a single character, nor that a single character can be input with a single keystroke (even with modifiers), nor that keyboards are the same all over the world.

Rendering, or at least the visual forms of rendering, introduces the notion of a glyph, which can be defined as the components used to generate the visible representation of a sequence of characters. This definition points to the unsurprising fact that there is not, again, a one-to-one correspondence between characters and glyphs: a single character can be represented by multiple glyphs (each glyph is then part of the representation of that character); a single glyph may also represent multiple characters (this is the case of ligatures, among others). A set of glyphs makes up a font. Glyphs can be construed as the basic units of organization of the visual rendering of text, just as characters are the basic unit of organization of encoded text.

A few examples will help make sense of all this complexity (which is mostly a reflexion of the complexity of human writing systems). Let us start with a very simple example: a user, equipped with a US-English keyboard, types "Foo", which the computer encodes as 16-bit values (the UTF-16 encoding of the UCS) and displays on the screen.
Example 3.1: Basic Latin
Keystrokes Shift-f o o
Input characters F o o
Encoded characters
(byte values in hex)
0x0046 0x006F 0x006F
Display Foo

The only complexity here is the use of a modifier (Shift) to input the capital F. A slightly more complex example is a user typing "çé" on a French-Canadian keyboard, which the computer encodes in fully decomposed UTF-16 and displays.
Example 3.2: Latin with diacritics
Keystrokes ¸ c é
Input characters ç é
Encoded characters
(byte values in hex)
0x0063 0x0327 0x0065 0x0301
Display çé

A few interesting things are happening here: when the user types the cedilla (¸), nothing happens except for a change of state of the keyboard driver; the cedilla is a dead key. When the driver gets the c, it provides a complete ç character to the system, which encodes it as two characters: a c and a combining cedilla, each represented by a single 16-bit unit. These two characters are then displayed as one ç glyph. The user then presses the dedicated é key, which results in, again, two characters represented by two bytes. Most systems will display this pair as one glyph, but it is also possible to combine two glyphs (the base letter and the accent) to obtain the same rendering.

Note. If Unicode Normalization Form C (precomposed, see section 4.1) had been used in this example, only two characters would have been encoded. If ISO/IEC 8859-1 encoding had been used, those two characters would have encoded as one byte each; this is certainly simpler, but the representable repertoire is very limited. If UTF-8 encoding (fully decomposed) had been used, the result would have been the same four characters as above, but with the first and third encoded as one byte and the second and fourth encoded as two bytes.

On to a Japanese example: our user employs an input method to type "nihongo in Kanji characters", which the computer encodes in UTF-16 and displays.
Example 3.3: Japanese
Keystrokes n i h o n g o <space><return>
Input characters Kanji character ni Kanji character hon Kanji character go
Encoded characters
(byte values in hex)
0x65E5 0x672C 0x8A9E
Display nihongo in Kanji characters

The interesting aspect here is input, where the user has to type a total of nine keystrokes before the three characters are produced, which are then encoded and displayed rather trivially. An Arabic example will show different phenomena:
Example 3.4: Arabic
Keystrokes Arabic lam Arabic alif Arabic lam-alif Arabic ghayn Arabic ghayn
Input characters Arabic lam Arabic alif Arabic lam Arabic alif Arabic ghayn Arabic ghayn
Encoded characters
(byte values in hex)
0x0644 0x0627 0x0644 0x0627 0x0639 0x0639
Display A few Arabic letters

Here the first two keystrokes each produce an input character and an encoded character, but the pair is displayed as a single glyph (Arabic lam-alif, a lam-alif ligature). The next keystroke is a lam-alif, which some Arabic keyboards have; it produces the same two characters which are displayed similarly, but this second lam-alif is placed to the left of the first one. The last two keystrokes produce two identical characters which are rendered by two different glyphs (a medial form followed to its left by a final form). We thus have 5 keystrokes producing 6 characters and 4 glyphs laid out right-to-left.

A final example in Tamil, typed with an ISCII keyboard, will illustrate some additional phenomena:
Example 3.5: Tamil
Keystrokes Tamil ta Tamil aa Tamil na Tamil virama Tamil ka Tamil o
Input characters Tamil ta Tamil aa Tamil na Tamil virama Tamil ka Tamil o
Encoded characters
(byte values in hex)
0x0B9F 0x0BBE 0x0B99 0x0BCD 0x0B95 0x0BCB
Display Tango in Tamil letters

Here input is straightforward, but note that contrary to the preceding accented Latin example, the diacritic Tamil virama (virama, vowel killer) is entered after the Tamil na to which it applies. Rendering is interesting for the last two characters. The last one (Tamil o) clearly consists of two glyphs which surround the glyph of the next to last character (Tamil ka).

A number of operations routinely performed on text can be impacted by the complexities of the world's writing systems. Let us take selection of on-screen text by the mouse as an example, in a bidirectional (bidi) context. First, let's have some bidi text, in this case Arabic letters (written right-to-left) mixed with Arabic-Hindi digits (left-to-right):
Example 3.6: Bidirectional text
In memory Arabic ayn  Arabic dal  Arabic   dal  <space> Arabic mim  Arabic alif  Arabic   ra  Arabic sin  <space> Arabic-Hindi digit one  Arabic-Hindi digit nine  Arabic-Hindi   digit nine  Arabic-Hindi digit   eight  
On screen Example of Arabic text

In the presence of bidi text, two possible selection modes must be considered. The first is logical selection mode, which selects all the characters logically located between the end-points of the user's mouse gesture. Here the user selects from between the first and second letters of the second word to the middle of the number. Logical selection looks like this:
Example 3.7: Example of logical selection
In memory
Arabic ayn  Arabic dal  Arabic dal  <space> Arabic mim   Arabic alif  Arabic ra  Arabic sin  <space> Arabic-Hindi digit one  Arabic-Hindi digit nine  Arabic-Hindi digit nine  Arabic-Hindi digit eight  
On screen Example of logical selection

It is a consequence of the bidirectionality of the text that a single, continuous logical selection in memory results in a discontinuous selection appearing on the screen. This discontinuity, as well as the somewhat unintuitive behaviour of the cursor, makes many users prefer a visual selection mode, which selects all the characters visually located between the end-points of the user's mouse gesture. With the same mouse gesture as before, we now obtain:
Example 3.8: Example of visual selection
In memory
Arabic ayn  Arabic dal  Arabic dal  <space> Arabic mim   Arabic alif  Arabic ra  Arabic sin  <space>  Arabic-Hindi digit one  Arabic-Hindi digit nine   Arabic-Hindi    
                                                                                                nine  Arabic-Hindi digit eight  
On screen Example of visual selection

In this mode, popular with users, a single visual selection range results in two logical ranges [Issue: can there be more?], which MUST be accommodated by protocols, APIs and implementations.

Note. In the next version of this document, the WG plans to address the following topics in this subsection:

3.3 Choice and Identification of Character Encodings

Section 3.2 discussed the relationship between characters and the underlying representation, called character encoding. Because encoded text cannot be interpreted and processed without knowing the encoding, it is vitally important that the character encoding is known at all times and places where text is exchanged or stored.

The existence of a large number of character encodings often brings to the forefront the question of choosing one (or more) encodings when designing a data format, a protocol, an API, when implementing these or simply when editing a document. One aspect of this question is the following: should multiple encodings be allowed, or only one mandated (by a data format for instance)?

Mandating a unique character encoding has strong virtues of simplicity, efficiency and robustness. After choosing this unique encoding, a protocol or data format does not have to deal with provisions for character encoding tagging, since the encoding is known implicitly from usage of such protocol or data format. If the data is to be transferred other than electronically (e.g. written on an envelope or a billboard, like URIs), then there is no way to carry around encoding tags and a unique encoding is the only solution. With a unique encoding, implementations do not have to deal with recognizing encoding tags, nor with encoding conversions and other complexities inherent in working with multiple encodings. Efficiency can seriously enhanced, especially when small pieces of text are involved such as in short headers where encoding identification tags would occupy space comparable with the data itself.

This desirable solution, however, is often felt to be unacceptable because of the need for compatibility with existing data, systems, protocols and applications, which use various encodings. Nevertheless, specification writers and implementors are strongly encouraged to consider its adoption: it is often the case that multiple encodings can be dealt with at the boundaries or outside a protocol or API, as was done for the [DOM], resulting in much greater simplicity and uniformity in the API itself.

If more than a single encoding is allowed, the question of granularity arises: in a data format or a protocol, should a single 'entity' (document, message, etc.) be allowed to be encoded in multiple encodings, with encoding transitions within the entity? In an API, should a single argument to a procedure be allowed to be so multi-encoded? Should different arguments of the same procedure be allowed to be in different encodings? In an object model, should a text-containing object be allowed to contain multi-encoded text? Should distinct objects representing a 'document' (such as a DOM tree) be allowed to be in different encodings? In general the answer is no, unless there exists a very strong requirement to do so.

Multi-encoding requires provisions for in-text encoding identifications and breaks the model of text-based protocols and data formats, since in the absence of an underlying encoding, one cannot rely on characters for effecting the encoding transitions and the requisite identification. There exist examples of such protocols, MIME being a prominent one. Since a MIME message may contain any number of attachments in different encodings, it has provisions for multipart messages that take care of proper encoding labeling, using headers outside the parts themselves. But MIME also has a way to use multiple encodings in a single header (see RFC 2047 in [MIME]). The in-text encoding identification mechanism of RFC 2047 is delicate and hard to implement correctly, and experience has shown that it does not work reliably, thereby showing prominently the limits of such multi-encoding schemes. Consequently, specification writers SHOULD restrict the basic entities of their specifications to a single encoding.

If the unique encoding approach is adopted, the chosen encoding MUST be such that it covers the needs of the largest possible audience, including coverage for as many human languages as possible. In practice, this will most likely mean that the choice will be one of the standard encodings of ISO 10646/Unicode. If some measure of compatibility with ASCII is desired, UTF-8 (see [RFC 2279]) is most probably the UCS encoding of choice; on the Internet, the IETF Charset Policy [RFC 2277] specifies that "Protocols MUST be able to use the UTF-8 charset". Another UCS encoding very worthy of consideration, especially for APIs, is UTF-16 (see [UTF-16]).

If the unique encoding approach is not chosen, then it is crucial to provide for proper identification of character encoding at all times. For protocols, the [MIME] Internet specification has established a standard way to proceed. First specified for Internet email, the MIME mechanism has also been adopted by HTTP [RFC 2616] and some aspects of it adapted to other protocols. The MIME mechanism consists in having in the protocol a Content-Type header which indicates the nature or format of the protocol payload; when the content type is textual, a parameter called charset is added to the Content-Type header, with a value indicating the character encoding. The charset parameter is defined such that it provides sufficient information to unambiguously decode the sequence of bytes of the payload into a sequence of characters. The values are drawn from the [IANA] registry.

Note. ISO 10646/Unicode designate certain ranges of code points as the Private Use Area (PUA), a set of positions guaranteed never to be allocated to standard characters and available for use by private arrangement between creator and consumer. Since the characters in the PUA are not, by definition, standardized, the values of the charset parameter registered by IANA for UCS encodings do not take them into account and do not contain enough information to unambiguously decode a data stream containing PUA characters. Consequently, the standard, registered charset parameter values may not be used to label entities containing PUA characters.

The term charset derives from "character set", an expression with a long and tortured history that is best avoided (see [Connolly] for a discussion). Specification writers SHOULD avoid using the expression "character set", as well as the term "charset" except when referring to the MIME charset parameter or its IANA-registered values.

Given the importance of proper character encoding identification, specification writers MUST provide mechanisms such that the encoding of text can always be reliably determined. When choosing which encodings may or must be supported in a specification, designers MUST make sure that the UTF-8 and/or UTF-16 encodings of ISO 10646/Unicode are admissible encodings and SHOULD choose at least one of UTF-8 or UTF-16 as mandated encodings (encodings that MUST be supported by implementations of the specification). Reliance on defaults and, most of all, on heuristics MUST be avoided; an exception is defaulting to UTF-8 or UTF-16. Implementors of software MUST fully support such mechanisms and SHOULD make it easy to use them (for instance in HTTP servers). Content developers MUST make use of the offered facilities by always indicating character encoding (for instance in XML encoding declarations).

The [IANA] registry constitutes the closest thing there is to a standard for character encoding names. Specification writers SHOULD mandate the use of those names, and in particular of the MIME preferred names, to designate character encodings in protocols, data formats and APIs. The use of the "x-" convention for unregistered names SHOULD be discouraged, having led to abuse (use of x- for character encodings that were widely used, even long after there was an official registration) in the past. Content developers and software that tags textual data MUST use one of the names mandated by the appropriate specification and SHOULD use the MIME preferred name of an encoding to tag data in that encoding. An IANA-registered charset name MUST NOT be used to tag textual data in another encoding than the one identified in the IANA registration of that name.

Receiving software (which must determine the encoding from available information) MUST be able to recognize the name(s) of any encoding(s) mandated by the specification it implements and SHOULD be able to recognize as many names (in particular aliases of the MIME preferred name) as practicable; it is best to provide a field-upgradable aliasing mechanism for this purpose. In addition, receiving software SHOULD recognize the names and support as many non-mandated encodings as practicable. [ISSUE: should we list a set of encodings that implementors really, really SHOULD support?] When a charset name is recognized, receiving software MUST interpret the received data according to the encoding associated with the name in the IANA registry.

In the absence of suitable information from the protocol, data format or API, receiving software MAY use heuristics to attempt to determine the encoding, realizing however that such heuristics cannot be 100% reliable and may not be advisable in mission-critical applications, especially in situations where there is not a human user present to verify de visu the correct encoding identification. Such heuristics can be a nice feature in a browser, but may be unadvisable when automatically processing a purchase order.

Note. In the next version of this document, the WG plans to address the following topics in this subsection:

3.4 The UCS as a Common Reference

Many Internet protocols and data formats, most notably the very important Web formats HTML, CSS and XML, are based on text. In those formats, everything is text but the relevant specifications impose a structure on the text, giving meaning to certain constructs so as to obtain functionality in addition to that provided by plain text. HTML and XML are markup languages, defining entities entirely composed of text but with conventions allowing the separation of this text into markup and character data. Citing from the [XML 1.0], section 2.4:

Text consists of intermingled character data and markup. Markup takes the form of start-tags, end-tags, empty-element tags, entity references, character references, comments, CDATA section delimiters, document type declarations, and processing instructions. All text that is not markup constitutes the character data of the document.

For the purposes of this section, the important aspect is that everything is text, that is, a sequence of characters. HTML is similarly constructed; in fact, both XML and HTML inherit this structure from SGML. A CSS style sheet, although not SGML-based, is also defined as a sequence of characters, as are formats such as TEX, troff/nroff and many others. Protocols are also often defined in terms of text, with the important benefit that the protocol is then far easier to debug using simple tools such as telnet.

3.4.1 Reference Processing Model

In the early days of the Web, HTML was defined in terms of ISO Latin-1 (see [ISO 8859]), which severely limited the repertoire of characters usable in Web documents and, consequently, the human languages that could be accommodated. During efforts to internationalize HTML and break free of that restriction, it was realized that advantage could be taken of the SGML concept of document character set to decouple the definition in terms of characters from the encoding of those characters. This led to the development of a Reference Processing Model for HTML, first described in [RFC 2070], in which the document character set is defined to be ISO 10646 while actual entities (documents) are allowed to be encoded in any character encoding compatible with the UCS (i.e. any encoding whose repertoire is a subset of that of the UCS). This model has been embraced by XML and CSS and is applicable to any data format or protocol that is text-based as described above, not only formats derived from SGML.

The essence of the model can be described as follows:

Note. It is noteworthy that for a specification to use the Reference Processing Model does not require that implementations actually use ISO 10646/Unicode. The requirement is only that the implementations behave as if the processing took place as described above.
Note. All specifications that derive from XML automatically inherit this Reference Processing Model. XML is entirely defined in terms of UCS characters and mandates the UTF-8 and UTF-16 encodings while allowing any other encoding for parsed entities.

[To do: make normative statements.]

3.4.2 References to ISO 10646/Unicode

Specifications often need to make references to the ISO 10646 or Unicode standards. Such references must be made with care, especially when normative. One problematic area is the distinction between ISO 10646 and Unicode and the question of whether a specification should reference one, the other or both. Another is the fact that both standards are still evolving, in particular with new characters being added to achieve the goal of a truly Universal Character Set.

ISO 10646 (actually ISO/IEC 10646) is a de jure standard, developed and published jointly by ISO (the International Organisation for Standardisation) and IEC (the International Electrotechnical Commission). Unicode is a de facto standard developed and published by the Unicode Consortium, an organization of major computer corporations, software producers, database vendors, research institutions, international agencies, various user groups, and interested individuals. Unicode is therefore much more than a vendor "standard", is comparable in influence to W3C recommendations but does not have a de jure status.

ISO 10646 and Unicode define exactly the same CCS (same repertoire, same character numbers) and encodings. This synchronism is actively maintained by liaisons and common membership between the relevant technical committees. But there are differences between the two standards; in fact, were it not for those differences, Unicode would be pointless given the status of ISO 10646 as an International Standard. In addition to the jointly defined CCS and encodings, the Unicode Standard adds normative and informative lists of character properties, normative character equivalence and normalization specifications, a normative algorithm for bidirectional text and a large amount of useful implementation information. In short, Unicode adds semantics to the characters that ISO 10646 merely enumerates.

Since specifications in general need both a definition for their characters and the semantics associated with these characters, specification writers SHOULD include normative references to both ISO/IEC 10646 and the Unicode Standard. If a normative reference to Unicode is not considered appropriate, an informative reference should nevertheless be provided so that implementors can still benefit from the wealth of information provided in the standard and on the Unicode Consortium Web site.

The fact that both ISO 10646 and Unicode are evolving (in synchronism) raises the issue of versioning: should a specification refer to a specific version of the standard, or should it make a generic reference, so that the normative reference is to the version current at the time of reading the specification? In general the answer is both. A generic reference MUST be made so that characters allocated after a specification is published are usable with that specification. But a specific reference MAY be included to ensure that functionality depending on a particular version is available and will not change over time (an example would be the set of characters acceptable as Name characters in XML 1.0, which is an enumerated list that parsers must implement to validate names).

One notable aspect of the history of ISO 10646 is Amendment 5. This amendment, published in 1998 and reflected in Unicode 2.0 (1996), created an incompatible change in the standard: a few thousand characters (all Korean Hangul syllables) were moved, i.e. their character numbers were changed, causing incompatibility with any anterior implementations and data involving Hangul. The reason such a change was deemed acceptable was that no serious amount of such data, and no widely available implementations, did exist at the time. Because of this incompatibility all references to ISO 10646 and Unicode MUST explicitly refer to Amendment 5 and all references to Unicode MUST be to version 2.0 or later.

This incompatible change also points to the danger of generic references: a specification could be endangered, should such a change occur in the future. The relevant committees have pledged never to let that happen again, but specification writers SHOULD protect against this risk by qualifying generic references to either ISO 10646 or Unicode with a statement to the effect that the reference excludes incompatible changes. An example is the reference to ISO 10646 from HTML 4.01 [To do: add link when HTML 4.01 becomes Rec]:

"Information Technology -- Universal Multiple-Octet Coded Character Set (UCS) -- Part 1: Architecture and Basic Multilingual Plane", ISO/IEC 10646-1:1993. This reference refers to a set of codepoints that may evolve as new characters are assigned to them. This reference therefore includes future amendments as long as they do not change character assignments up to and including the first five amendments to ISO/IEC 10646-1:1993. Also, this reference assumes that the character sets defined by ISO 10646 and Unicode remain character-by-character equivalent. This reference also includes future publications of other parts of 10646 (i.e., other than Part 1) that define characters in planes 1-16.

Whether such a reference is included in the bibliography section of a specification, or a simpler reference with explanatory text in the body of the specification, is an editorial matter best left to each specification. Examples of the latter, as well as a discussion of the versioning issue with respect to MIME charset parameters for UCS encodings, can be found in [RFC 2279] and [UTF-16].

3.5 Character Escaping

In a document format such as HTML or an XML application and in protocols, some characters act as text data while others are used to indicate protocol or format functions. In order to work correctly, a protocol or format has to make a clear distinction between the text data characters and the functional characters. In some cases, the distinction is based on position, for example in TCP headers, or the first character in each line in a Fortran program.

In other cases, certain characters are designated as having certain specific protocol/format functions in certain contexts (e.g. the "<" and the "&" have very specific functions in HTML and XML). If there are such syntactically relevant characters, then these characters cannot at the same time serve to represent themselves in text data in the same way as all other characters do. Also, often formats are represented in an encoding that does not allow to represent all characters directly. For such cases, a technique called escaping is used.

Escaping a character in a protocol or format means representing it by other characters. This works by creating an additional syntactic construct, defining additional characters, or defining character sequences that have special meaning.

For specification authors: When designing potential future W3C protocols and formats, the following points MUST be followed. When revising existing W3C protocols and formats, the following points SHOULD be followed.

Note. In the next version of this document, the WG plans to address the following topics in this subsection:

Make explicit cases of interaction between escaping, legacy encodings with decomposed notations, code table switching and early normalization,...

4. Character data exchange: Early Uniform Normalization

Character data interchange using W3C protocols and formats is based on the principle of early normalization, which defines the exact form to which text data has to be normalized, and the cases in which normalization must be applied. This document encourages Web components that generate content to normalize text for the following reasons:

4.1 W3C Text Normalization

Text data is in normalized form according to this specification if all of the following apply:

Text data is also considered to be in normalized form for the purpose of this specification if all of the following apply:

Note. It is possible that legacy encodings also exhibit the problem of duplicate encodings. In this case, it would be appropriate if a corresponding normalization were applied. Examples are ISO 6937 with multiple accents, and some uses of ISO 2022.

Note. There are legacy encodings that do not exhibit the problem of duplicate encodings, but that do not allow one-to-one conversion of individual codepoints. An example is ISO 6937, which does not use precomposition and places combining marks before the base character. [Issue: Should back-conversion use escaping, or should it do one-to-many/many-to-many conversion?]

4.2 Application of Early Uniform Normalization

The term "normalized form" in this section refers to the form defined in Section 4.1.

Applications or tools transcoding from a legacy encoding to an encoding based on UCS MUST ensure that their output is in normalized form.

The producer of text data MUST ensure that data is produced or sent out in normalized form. For the purpose of W3C specifications and their implementations, the producer of text data is the sender of the data in the case of protocols. In the case of formats, it is the tool that produces the data.

Implementors of producer software in the above sense are encouraged to delegate normalization to their respective data sources wherever possible. (Examples of data sources would be: operating system, libraries, keyboard drivers.)

If any intermediate recipient of text data applies any operations, it MUST ensure that the results of these operations is again in normalized form, provided the incoming data is in that form. Intermediate recipients may provide additional normalization towards the normalization form, as a side-effect of their operations and/or as an additional service.

If intermediate recipients do not touch the data but just pass it on, they are not required to check normalization or to normalize data. (Example: caching proxies)

The recipients of text data SHOULD assume that data is normalized. Recipients MAY provide normalization as an add-on service and safety measure. Recipients SHOULD [Issue: MUST?] provide a way to switch off normalization.[Issue: change to MUST NOT, to always fail?] If a producer and a recipient work as one unit, normalization MUST be applied in the producer part but MUST be switched off in the recipient part.

Example: Authoring tools frequently are extensions to browsers, and the browser software components are used to display the results of editing. Such tools should normalize all the text that is produced, but must not provide normalization for the operations that are the same as in a browser, in order to catch potential normalization problems, e.g. differences with documents not edited by the tool, early.

Tools or operations that just do string identity matching, and that have both strings to be matched available in the same encoding, SHOULD do so by binary comparison.

Note. Producers, intermediate recipients, and transcoders must support a repertoire of Unicode codepoints that is complete with respect to normalization, i.e. if any arbitrary sequence of codepoints in the repertoire is normalized, the codepoints needed for the normalization must also be part of the repertoire.

4.3 String Identity Matching

One important operation that depends on early normalization is string identity matching [CharReq]. String identity matching (a frequent operation) is a subset of the more general problem of string matching. There are various degrees of specificity for string matching, from approximate matching such as regular expressions or phonetic matching for English, to more specific matches such as accent-insensitive or case-insensitive matching. String identity matching is concerned only with strings that contain no user-identifiable distinctions.

At various places in the Web infrastructure, strings, and in particular identifiers, are compared for identity. If different places use different definitions of string identity matching, or if they rely on different mechanisms to test identity, the results are undesired unpredictability and unnecessary conversions. To solve the problem of string identity matching, the following issues have to be addressed:

  1. Which representations to treat as equivalent (and which not)
  2. Which components in the Web architecture to make responsible for equivalence:
    1. Each individual component that performs a string identity check has to take equivalents into account (late normalization)
    2. Duplicates and ambiguities are removed as close to their source as possible (early normalization)
  3. Which way to normalize (in the case that early normalization is needed)

String identity matching on the World Wide Web is based on the following steps:

Conversion to UCS, and to the same encoding for both strings, assures that text strings and not just bytes are compared. Early normalization gives the responsibility of avoiding duplicate encodings to the data producer; it ensures that a minimum of effort is spent on solving the problem.

4.4 Compatibility Equivalents and Control Characters

This specification does not address compatibility equivalents. Compatibility as listed in the Unicode database covers a wide range of similarities/distinctions. Depending on the situation, some distinctions are needed, and others will be confusing. To specify all these situations in a single place seems premature. In the absence of any further specifications, implementations are advised to generate the non-compatibility equivalent if they do not explicitly need the compatibility character. A compatibility character here is a character that disappears when applying Unicode Compatibility Composition (Normalization Form CC of Unicode [UTR #15]). A non-compatibility equivalent is the character resulting from applying Unicode Compatibility Composition. [Issue: This needs to be worded more carefully, not all compatibility characters have canonical equivalents.] [Issue: Some details are already available in [UXML]. How much do we need to leave here?]

Specifications are advised to exclude compatibility characters in the syntactic elements of the formats they define if this is reasonable (e.g. exclusion of compatibility characters for GIs in XML). In the future, compatibility characters should be replaced by appropriate style or markup information wherever possible.

This specification does not address any further equivalents, such as case equivalents, the equivalence between katakana and hiragana, the equivalence between accented and un-accented characters, the equivalence between full characters and fallbacks (e.g., "ö" vs. "oe" in German), and the equivalence between various spellings and morphological forms (e.g., color vs. colour). Such equivalence is on a higher level; whether and where it is needed depends on the language, the application, and the preferences of the user.

5. String Indexing

On many occasions, in order to access a substring or a character, it is necessary to identify positions (between "characters"or other appropriate units) in a text string/sequence/array. Where such indices are exchanged between components of the Web, there is a need for a uniform definition of string indexing in order to ensure consistent behavior. The requirements for string indexing are discussed in [CharReq, section 4].

Depending on the requirements, indexing can occur on any of the Layers defined at the start of Section 3. In particular, Layer 1 is recommended if the primary concern is efficient internal operation. Layer 2 is recommended if the primary concern is interaction with the formal definition of a format, e.g. XML. Layer 3 will be recommended (once it is well defined) where the primary concern is user interaction while maintaining interoperability.

Note: In many cases, it is highly preferable to use non-numeric ways of identifying substrings. The specification of string indexing for the Web should not be seen as a general recommendation for the use of string indexing for substring identification. As an example, in the case of translation of a document from one language to another, identification of substrings based on document structure can be expected to be much more stable than identification based on string indexing.

Note: The issue of indexing origin, i.e. whether the first character in a string is indexed as character number 0 or as character number 1, is not addressed explicitly here. In general, even individual characters should be understood and processed as substrings, identified by a position before and a position after the substring. Understanding indices as boundary positions between the units indexed makes it easier to relate the indices for the various Layers to each other. In the case of using boundary points, starting with an index of 0 for the position at the start of the string is the best solution.

6. Character Encoding in URI References

According to the current definition [RFC 2396], URI references are restricted to a subset of US-ASCII. There is also an escaping mechanism to encode arbitrary byte values using the %HH convention. However, because [RFC 2396] does not define the mapping from characters to bytes, the %HH convention by itself is of limited use. To avoid future incompatibilities, W3C specifications for new protocol/format elements MUST include the following paragraph by reference:

For all syntactic elements in the format/protocol which are being interpreted as URI references, characters that are syntactically not allowed by the generic URI syntax (all non-ASCII characters, plus the excluded characters in [RFC 2396, Section 2.4.] except "#" and "%") MUST be treated as follows: Each such character is converted to UTF-8 as one or more bytes, each of these bytes is escaped with the URI escaping mechanism (i.e. converted to %HH, where HH is the hexadecimal notation of the byte value), and the original character is replaced by the resulting character sequence.

Example: In the URI <http://www.w3.org/People/Dürst/>, the character "ü" is not allowed. The representation of "ü" in UTF-8 consists of two bytes with the values 0xC3 and 0xBC. The URI is therefore converted to <http://www.w3.org/People/D%C3%BCrst/>.

Note: The intent of this is not to freeze the definitions of URI references to a subset of US-ASCII characters forever, but to assure that W3C technology correctly and predictably interacts with systems that are based on the current definition of URI references while not inhibiting a future extension of the URI reference definition.

Note: This provision does not affect the ability to use URI references with other encodings than UTF-8. However, in such a case, the URI reference has to always be given in its escaped form. As an example, if the http server at www.w3.org would only use ISO-8859-1, the above URI would always have to be given as <http://www.w3.org/People/D%FCrst/>, because "ü" in ISO-8859-1 is 0xFC.

Note: Current W3C specifications already contain provisions in accordance with the above. For [XML 1.0], please see Section 4.2.2, External Entities. For [HTML 4.0], please see Appendix B.2.1: Non-ASCII characters in URI attribute values, which also contains some provisions for backwards compatibility. Further information and links can be found at [I18NURI].

[Issue: What to do about [] for IPV6 addresses?]

7. Language Identification

In the next version of this document, the WG plans to address the following topics in this subsection:

Language tagging benefits Web accessibility. When content developers mark up natural language changes in a document, speech synthesizers and Braille devices can automatically switch to the new language, making the document more accessible to multilingual users. Content developers SHOULD always identify the predominant script of a document's content (e.g., through markup or HTTP headers).

In addition to helping assistive technologies, natural language markup allows search engines to find key words and identify documents in a desired language. Natural language markup also improves readability of the Web for all people, including those with learning disabilities, cognitive disabilities, or people who are deaf.

Appendix: Change Log

Changes since http://www.w3.org/TR/1999/WD-charmod-19990225:

Added François Yergeau as an editor. Changed sequence of lists for comments in Status of this Document, mention how i18n-editor@w3.org is archived. Some rewriting of Abstract and Introduction.

Fleshed out Section 3.2 "Characters and their digital representation" and Section 3.3 "Identification of Character Encodings". A couple of existing paragraphs in 3.3 (about benefits of unique encoding) moved to 3.2. Fleshed out section 3.4.

Changed normalization to not mention Unicode Version 3.0 normatively anymore, because this is part of UTR#15 now. Added annotation characters to list of prohibited characters in. Later moved to [UXML], and removed some material. Added issue about normalization and escaping.

Moved definition of layers to Section 3, changed to 4 layers. Made string indexing refer to the general model in Section 3. Clarified use of boundary points for indexing and use of 0-based indexing (needs more detail work).

In Section 6: Changed "represented" to "converted". Added Note to explain that URIs (always escaped) can still be used with servers not using UTF-8. Changed to apply to URI references so that it includes fragment identifiers. Added # and % as characters allowed in URI references. Changed "W3C specifications MUST" to "W3C specifications, for new protocol/format elements, MUST". Changed "in accordance with" to "very similar to".

Added "Language identification" as Section 2.5, later moved to Section 7.

Minor work in Acknowledgements, Glossary.

Added references to UTR #17, Unicode 3.0, HTTP 1.1, UTF-16 I-D, Connolly's "Character Set Considered Harmful". Expanded reference to MIME to include all five RFCs (needed 2047 for section 3.3). Separated out Normative and Other references.

Added <acronym> elements throughout for accessibility. Various typos fixed throughout.


This glossary does not provide exact definitions of terms but gives some background on how certain words are used in this document.

Used in a loose sense to denote small units of text, where the exact definition of these units is still open.
Early Normalization
See Early Uniform Normalization.
Early Uniform Normalization
Duplicates and ambiguities are removed as close to their source as possible. This is done by normalizing them to a single representation. Because the normalization is not done by the component that carries out the identity check, normalization has to be done uniformly for all the components of the Web.
Encoding based on UCS
An encoding that uses UCS codepoints in a reasonably simple manner. Examples: UTF-8, UTF-7 (deprecated), UTF-16, UCS-2, UCS-4. (Note: For the later three, an encoding definition also needs to include provisions for defining and identifying the serialization of 16-bit or 31-bit values into byte sequences.)
Late Normalization
Each individual component that performs a string identity check has to take equivalence into account. This would be done by normalizing each string to a preferred representation that eliminates duplicates and ambiguities. Because, with late normalization, normalization is done locally and on the fly, there is no need to specify a webwide uniform normalization.
Legacy Encoding
An encoding not based on UCS. Examples: ISO-8859-1, EUC-KR, and many others.
String Identity Matching
Exact matching of strings, except for encoding duplicates indistinguishable to the user. See Section 4.3.
String Indexing
Indexing into a string to address a character or a sequence of characters. See Section 5.
The process of changing text data from one character encoding to another.
Universal Character Set, the Coded Character Set (CCS) defined in parallel by [ISO 10646] and [Unicode].
Uniform Resource Identifier, see [RFC 2396]. According to RFC 2396, an URI does not include a fragment identifier (the part after a potential '#'). In colloquial usage, URI is often used for URI reference.
URI reference
URI with potentially attached fragment identifier (the part after '#'). See [RFC 2396].
World Wide Web, the collection of technologies built up starting with HTML, HTTP, and URIs, the corresponding software (servers, browsers,...), and/or the corresponding content.


Special thanks go to Ian Jakobs for serious help with editing. Tim Berners-Lee and James Clark provided important details in the section on URIs. The W3C I18N WG and IG, as well as others, provided many comments and suggestions.


Normative References

(Internet Assigned Numbers Authority) K. Simonsen et al., Official Names for Character Sets, See ftp://ftp.isi.edu/in-notes/iana/assignments/character-sets.
[ISO 10646]
ISO/IEC 10646-1:1993, Information technology -- Universal Multiple-Octet Coded Character Set (UCS) -- Part 1: Architecture and Basic Multilingual Plane.
[RFC 2119]
S. Bradner, Key words for use in RFCs to Indicate Requirement Levels, <http://www.ietf.org/rfc/rfc2119.txt>.
[UTR #15]
Mark Davis, Martin Dürst, Unicode Normalization Forms, Unicode Technical Report #15, September 1999, <http://www.unicode.org/unicode/reports/tr15/tr15-17.html>.
[Unicode 3.0]
The Unicode Consortium, The Unicode Standard -- Version 3.0, ISBN 0-201-61633-5.

Other References

D. Connolly, Character Set Considered Harmful, W3C Note 2-May-1995 <http://www.w3.org/MarkUp/html-spec/charset-harmful>
Bert Bos, Håkon Wium Lie, Chris Lilley, Ian Jacobs, Eds., Cascading Style Sheets, level 2 (CSS2 Specification), W3C Recommendation 12-May-1998, <http://www.w3.org/TR/REC-CSS2>.
Vidur Apparao et al., Document Object Model (DOM) Level 1 Specification, W3C Recommendation 1 October, 1998, <http://www.w3.org/TR/REC-DOM-Level-1/>.
Internationalization: URIs and other identifiers <http://www.w3.org/International/O-URL-and-ident>.
[ISO 6937]
ISO/IEC 6937:1994, Information technology -- Coded graphic character set for text communication -- Latin alphabet.
[ISO 8859]
ISO 8859 (various parts and publication dates), Information technology -- 8-bit single-byte coded graphic character sets.
[HTML 4.0]
Dave Raggett, Arnaud Le Hors, Ian Jacobs, Eds., HTML 4.0 Specification, W3C Recommendation 18-Dec-1997 (revised on 24-Apr-1998), <http://www.w3.org/TR/REC-html40/>.
Multipurpose Internet Mail Extensions (MIME). Part One: Format of Internet Message Bodies, N. Freed, N. Borenstein, RFC 2045, November 1996, <http://www.ietf.org/rfc/rfc2045.txt>. Part Two: Media Types, N. Freed, N. Borenstein, RFC 2046, November 1996. Part Three: Message Header Extensions for Non-ASCII Text, K. Moore, RFC 2047, November 1996. Part Four: Registration Procedures, N. Freed, J. Klensin, J. Postel, RFC 2048, November 1996. Part Five: Conformance Criteria and Examples, N. Freed, N. Borenstein, RFC 2049, November 1996.
Gavin Nicol, The Multilingual World Wide Web, Chapter 2: The WWW As A Multilingual Application, <http://www.mind-to-mind.com/documents/i18n/multilingual-www.html#ID-2A08F773>.
Martin J. Dürst, Requirements for String Identity and Character Indexing Definitions for the WWW, <http://www.w3.org/TR/WD-charreq>.
[RFC 2616]
R. Fielding, J. Gettys, J. Mogul, H. Frystyk, T. Berners-Lee, Hypertext Transfer Protocol -- HTTP/1.1, <http://www.ietf.org/rfc/rfc2616.txt>.
[RFC 2070]
F. Yergeau, G. Nicol, G. Adams, M. Dürst, Internationalization of the Hypertext Markup Language, RFC 2070, January 1997, <http://www.ietf.org/rfc/rfc2070.txt>.
[RFC 2130]
C. Weider, C. Preston, K. Simonsen, H. Alvestrand, R. Atkinson, M. Crispin, P. Svanberg, The Report of the IAB Character Set Workshop held 29 February - 1 March, 1996, RFC 2130, April 1997, <http://www.ietf.org/rfc/rfc2130.txt>.
[RFC 2277]
H. Alvestrand, IETF Policy on Character Sets and Languages, RFC 2277 / BCP 18, January 1998, <http://www.ietf.org/rfc/rfc2277.txt>.
[RFC 2279]
F. Yergeau, UTF-8, a transformation format of ISO 10646, RFC 2279, January 1998, <http://www.ietf.org/rfc/rfc2279.txt>.
[RFC 2396]
T. Berners-Lee, R. Fielding, L. Masinter, Uniform Resource Identifiers (URI): Generic Syntax, August 1998, <http://www.ietf.org/rfc/rfc2396.txt>.
P. Hoffman, F. Yergeau, UTF-16, an encoding of ISO 10646, Work in progress, <draft-hoffman-utf16-05.txt>.
[Unicode 2.0]
The Unicode Consortium, The Unicode Standard, Version 2.0, Addison-Wesley, Reading, MA, 1996.
[Unicode 2.1]
Lisa Moore, Unicode Technical Report # 8, The Unicode Standard, Version 2.1, September 1998, <http://www.unicode.org/unicode/reports/tr8.html>.
[UTR #17]
Ken Whistler, Mark Davis, Character Encoding Model, Proposed Draft Unicode Technical Report #17, October 1998, <http://www.unicode.org/unicode/reports/tr17/>.
Martin Dürst, Mark Davis, Hideki Hiura, and Asmus Freytag, Unicode in XML and other Markup Languages, Proposed DRAFT Unicode Technical Report #20 and W3C Working Draft, <http://www.w3.org/TR/unicode-xml>.
[XML 1.0]
Tim Bray, Jean Paoli, C. M. Sperberg-McQueen, Eds., Extensible Markup Language (XML) 1.0, W3C Recommendation 10-February-1998, <http://www.w3.org/TR/REC-xml>.