Skip to content

tc39/proposal-intl-messageformat

Folders and files

NameName
Last commit message
Last commit date

Latest commit

f4cb3db · Jul 25, 2024

History

74 Commits
Dec 16, 2023
Jan 14, 2022
Dec 2, 2023
Jan 14, 2022
Jan 14, 2022
May 6, 2024
Jul 25, 2024
Dec 16, 2023
Jul 25, 2024

Repository files navigation

Intl.MessageFormat

Status

Champions: Eemeli Aro (Mozilla/OpenJS Foundation), Ujjwal Sharma (Igalia)

Former Champions: Daniel Minor (Mozilla)

Stage: 1

Presentations

Motivation

This proposal aims to make it easier to localize the web, increasing the openness and accessibility of the web for speakers of all languages. Currently, localization relies on a collection of mostly proprietary message formatting specifications that are limited in their features and/or challenging for translators to work with. Furthermore, localization often relies on parsing these custom formats during the runtime or rendering of an application.

To help with this, we introduce Intl.MessageFormat as a native parser and formatter for MessageFormat 2.0 (aka “MF2”) messages. MF2 is a specification currently being developed under the Unicode Consortium, with wide industry support. This will allow for using MF2 messages to localize web sites, enabling the localization of the web using industry standard tooling and processes.

In addition to a syntax that is designed to be accessible by both developers and translators, MF2 defines a message data model that may be used to represent messages defined in any existing syntax. This enables Intl.MessageFormat to be used within existing systems and workflows, providing a shared message formatting runtime for all users.

Use cases

The primary use case is the retrieval and resolution of localized text (i.e. a “message”) given a message source text, the locale and other options, and optionally a set of runtime values.

Put together, this allows for any message ranging from the simplest to the most complex to be defined by a developer, translated to any number of locales, and displayed to a user.

For instance, consider a relatively simple message such as

You have 3 new notifications

In practice, this would need to account for any number of notifications, and the plural rules of the current locale. Using MF2 syntax, this could be defined as:

.match {$count :number}
0   {{You have no new notifications}}
one {{You have {$count} new notification}}
*   {{You have {$count} new notifications}}

Some parts of the full message are explicitly repeated for each case, as this makes it significantly easier for translators to work with the message.

In code, with the API proposed below, this would be used like this:

const source = ... // string source of the message as above
const mf = new Intl.MessageFormat('en', source);
const notifications = mf.format({ count: 1 });
// 'You have 1 new notification'

As a majority of messages do not require multiple variants, those are of course also supported by the proposed API:

// A plain message
const mf1 = new Intl.MessageFormat('en', 'Hello!');
mf1.format(); // 'Hello!'

// A parametric message, formatted to parts
const mf2 = new Intl.MessageFormat('en', 'Hello {$place}!');
const greet = mf2.formatToParts({ place: 'world' });
/* [
  { type: 'text', value: 'Hello ' },
  { type: 'string', source: '$place', value: 'world' },
  { type: 'text', value: '!' }
] */

More complex use cases and usage patterns are described within the API description.

API Description

Though the MF2 specification is still being developed by the working group, the API presented here is representative of its current consensus. In particular, the exact shape of MessageData is still being discussed in the working group.

This proposal introduces one new primordial to ECMAScript, Intl.MessageFormat. The other interface descriptions below are intended to represent plain objects.

MessageData

The MessageData interface is defined by the MF2 data model developed by the Unicode MessageFormat working group. It contains a parsed representation of a single message for a particular locale.

type MessageData = PatternMessage | SelectMessage;

interface PatternMessage {
  type: 'message';
  declarations: Declaration[];
  pattern: Pattern;
}

interface SelectMessage {
  type: 'select';
  declarations: Declaration[];
  selectors: Expression[];
  variants: Variant[];
}

The full and exact definition of the message data model is given by its JSON Schema definition.

MessageFormat

The Intl.MessageFormat constructor creates a MessageFormat instance for a source message, one or more locale identifiers, and an optional MessageFormatOptions object. If a string is used as the source argument, it will be parsed as a MF2 syntax representation of a message.

Calling the constructor may throw an error if the source includes an MF2 syntax or data model error.

interface MessageFormat {
  new (
    locales: string | string[] | undefined,
    source: MessageData | string,
    options?: MessageFormatOptions
  ): MessageFormat;

  format(
    values?: Record<string, unknown>,
    onError?: (error: Error) => void
  ): string;

  formatToParts(
    values?: Record<string, unknown>,
    onError?: (error: Error) => void
  ): MessagePart[];

  resolvedOptions(): ResolvedMessageFormatOptions;
}

Constructor options and resolvedOptions()

MessageFormatOptions contains configuration options for the creation of MessageFormat instances. The ResolvedMessageFormatOptions object contains the options resolved during the construction of the MessageFormat instance.

As messages may contain placeholders resolving to strings with different directionality than the message as a whole (as in, left-to-right vs. right-to-left), the bidiIsolation option defines a strategy by which these parts will be isolated from each other in the output to avoid spillover effects. The default 'compatibility' strategy will include Unicode isolate code points at the boundaries of all expressions that are not known to match the message's directionality. The 'none' strategy will not provide any bidirectional isolation.

By default the message's directionality is determined from the script corresponding to the first locale, but this may be overridden by dir. Its "auto" value corresponds to messages with unknown directionality, for which the direction is determined by the first strongly directional character.

Custom user-defined message formatting and selection functions may defined by the functions option. These allow for any data types to be handled by custom functions. Such functions may be referenced within messages, and then called with the resolved values of their arguments and options.

interface MessageFormatOptions {
  bidiIsolation?: 'compatibility' | 'none';
  dir?: 'ltr' | 'rtl' | 'auto';
  functions?: { [key: string]: MessageFunction };
  localeMatcher?: 'best fit' | 'lookup';
}

interface ResolvedMessageFormatOptions {
  bidiIsolation: 'compatibility' | 'none';
  dir: 'ltr' | 'rtl' | 'auto';
  functions: { [key: string]: MessageFunction };
  localeMatcher: 'best fit' | 'lookup';
}

format(values?, onError?)

As with other Intl formatters, format() returns a string. This method has the following optional arguments:

  • values provides variable values for the message's variable references.
  • onError defines an error handler that will be called if message resolution or formatting fails. If onError is not defined, a warning will be issued for each error and a fallback representation used for the corresponding message part.

To determine the value res returned by the format() method, the message is first resolved to a list of MessageValue instances. Starting with an empty string res, for each MessageValue mv:

  1. Let msgDir be the base direction of the message.
  2. Let bidiIsolation be the resolved value of the bidiIsolation option.
  3. Let dir be mv.dir.
  4. Let strval be the result of calling mv.toString().
  5. If the call fails or strval is not a string:
    1. Set strval to be the concatenation of {, mv.source, and }.
    2. Set dir to be "auto".
  6. Let bidi be the { start: string, end: string } result of calling ApplyBidiIsolation(bidiIsolation, msgDir, dir).
  7. Append bidi.start, strval, and bidi.end to the end of res.

The ApplyBidiIsolation abstract operation will take as arguments the current bidi isolation strategy and the message and part directions. From these it will determine start and end as sequences of Unicode code points which will, if necessary, isolate parts from each other. With the default "compatibility" strategy, the result matches this TS type:

type BidiIsolation =
  | { start: ''; end: '' }
  | {
      start: '\u2066' | '\u2067' | '\u2068'; // LRI | RLI | FSI
      end: '\u2069'; // PDI
    };

formatToParts(values?, onError?)

For formatting a message to non-string targets, the formatToParts() method is provided, returning an array of MessagePart objects. This method has the following optional arguments:

  • values provides variable values for the message's variable references.
  • onError defines an error handler that will be called if message resolution or formatting fails. If onError is not defined, a warning will be issued for each error and a fallback representation used for the corresponding message part.

To determine the value res returned by the formatToParts() method, the message is first resolved to a list of MessageValue instances. Starting with an empty array res, for each MessageValue mv:

  1. Let msgDir be the base direction of the message.
  2. Let bidiIsolation be the resolved value of the bidiIsolation option.
  3. Let dir be mv.dir.
  4. Let parts be the result of calling mv.toParts().
  5. If the call fails or parts is not an array:
    1. Set parts to be [{ type: "fallback", source: mv.source }].
    2. Set dir to be "auto".
  6. Let bidi be the { start: string, end: string } result of calling ApplyBidiIsolation(bidiIsolation, msgDir, dir).
  7. If bidi.start is not an empty string:
    1. Append { type: 'bidiIsolation', value: bidi.start } to res.
  8. For each part or parts:
    1. Append part to res.
  9. If bidi.end is not an empty string:
    1. Append { type: 'bidiIsolation', value: bidi.end } to res.

MessageValue

When formatting a message, the selectors and placeholders of a message are each resolved first to an intermediate MessageValue representation. This can be thought of as an immutable object with properties and methods, though its JavaScript representation is only available to custom functions.

interface MessageValue {
  type: string;
  locale: string;
  dir: 'ltr' | 'rtl' | 'auto';
  source: string;
  options?: { [key: string]: unknown };
  selectKeys?: (keys: string[]) => string[];
  toParts?: () => MessagePart[];
  toString?: () => string;
  valueOf?: () => unknown;
}

type MessagePart =
  | { type: 'text'; value: string }
  | {
      type: 'bidiIsolation';
      value: '\u2066' | '\u2067' | '\u2068' | '\u2069'; // LRI | RLI | FSI | PDI
    }
  | ({
      type: string;
      source: string;
      locale?: string;
      dir?: 'ltr' | 'rtl' | 'auto';
    } & (
      | { value?: unknown }
      | { parts: Array<{ type: string; value: unknown; source?: string }> }
    ));

A MessageValue is an object with a string type, a string locale identifier, and an opaque source string identifying its origin. All other fields are optional; they determine how the value may be used in MF2 expressions.

In order to be usable as a formatted placeholder, the object (or its prototype chain) MUST include a toString method returning a string and a toParts method returning an array of MessageParts. All of the built-in implementations of this method return an array with exactly one value, but user-defined functions may return any number of parts, or none.

Except for parts corresponding to literal values, each MessagePart MUST include the type and source from the MessageValue. It MAY also include a string locale identifier, and optionally either an explicit value of any type or its own sequence of parts.

In order to be usable as a variant selector, the MessageValue object MUST include a selectKeys method. When called with an array of string keys, it MUST return an array whose elements are a subset of those keys. The returned keys will be considered to match the selector and be in preferential order.

Within a message, the value of an expression may be assigned to a message-local variable, and such a variable may be used as an input argument to another expression, or as an option value. Some functions (including the default number) accept objects with a valueOf method and an options field in addition to type and locale as input. These MAY also be defined on the returned object.

Literal Text

Text in patterns outside expressions is always literal. While its resolved value is never presented as JS, for the sake of simplicity it may be thought of as having the following resolved value:

interface MessageText {
  type: 'text';
  source: string;
  locale: string;
  dir: 'ltr' | 'rtl' | 'auto';
  toParts(): [MessageTextPart];
  toString(): string;
}

interface MessageTextPart {
  type: 'text';
  value: string;
}

For MessageText, the value returned by toString() and the value field of the object returned by toParts() corresponding to the text source. Its locale is always the same as the message's base locale.

Expressions

Expressions are used as selectors and as pattern placeholders. A local variable declaration may assign the value of an expression to a local variable, allowing for the same expression to be used in multiple places and potentially with different roles.

An expression may have one of three forms:

  • An operand (either a literal value or a variable reference).
  • An operand with an annotation.
  • An annotation with no operand.

The resolution of annotations using the : prefix is customisable using the constructor's functions option, which takes MessageFunction function values that are applied when the annotation's name (without the :) corresponds to the functions key.

Markup

In addition to expressions, placeholders may also be markup; content corresponding to HTML elements or other markup syntax. Unlike expressions, markup does not accept a positional input argument and its resolution is not customizable by the functions option.

Markup placeholders take three different forms:

  • "standalone" markup for non-textual content such as inline images,
  • "open" markup that starts a markup span, and
  • "close" markup that end a markup span.

The syntax used by markup is somewhat similar to that of XML, though with curly braces {} instead of angle brackets <> and with # as a prefix for "standalone" and "open": {#img /}, {#b}, {/b}.

Markup placeholders are not required to be paired or nest cleanly; within the formatter each is only considered by itself, and any higher-level validation is the responsibility of the caller.

A markup placeholder cannot be used as a selector. In format(), all markup is ignored, with each being formatted to an empty string. In formatToParts(), each markup placeholder is formatted to a single part:

interface MessageMarkupPart {
  type: 'markup';
  kind: 'open' | 'standalone' | 'close';
  source: string;
  name: string;
  options?: { [key: string]: unknown };
}

The type of the part is always "markup", and its kind is one of "open", "standalone", or "close". The name matches the name of the markup, without the # or / prefixes or suffixes. The source matches the name of the markup placeholder, prefixed and suffixed with the appropriate # and / characters.

The options correspond to the resolved literal and variable values of the options included in the placeholder. For example, when formatting {#open foo=42 bar=$baz} with formatToParts({ baz: 13 }), the formatted part's options would be { foo: '42', bar: 13 }. For options with variable reference values, if the resolved value is an object with a valueOf() method, the returned value is used. The options are only supported for "open" and "standalone" markup placeholders and are never included for a "close" markup placeholder.

MessageFunction

Fundamentally, messages are formed by concatenating values together. In order to support the handling of user-defined value types with user-defined formatting options, as well as other needs, user-provided message functions may be provided via the constructor's functions option to complement or replace the default ones.

type MessageFunction = (
  msgCtx: MessageFunctionContext,
  options: { [key: string]: unknown },
  input?: unknown
) => MessageValue;

interface MessageFunctionContext {
  locales: string[];
  dir: 'ltr' | 'rtl' | 'auto';
  source: string;
}

The msgCtx value defines the context in which the expression is being resolved, with the locales and dir of the whole message as well as the source fallback string representation of the expression.

The input and options values are constructed as follows:

  • If the value is a literal defined in the message syntax, the value is its string value.
  • If the value is a variable referring to a local variable declaration, the value is the MessageValue that the declaration's expression resolves to.
  • Otherwise, the value is a variable referring to an external value, and its type and value are that of the external value.

As function options are often set by literal values, and as MF2 considers all literals to be strings, the JSON string representation should be supported for numerical and boolean values when used as an input or as an option value. Each function will need to parse these separately from their string representations.

If a function is locale-dependent, it should accept an extra option key called locale which resolves to a string overriding the message's base locale. This option value should always be parsed as an array or (if a single string) a comma-delimited list of BCP 47 locale identifiers.

Default Functions

Two commonly used message functions number and string are provided as a starting point, and as handlers for placeholders without an annotation, such as variable references like {$foo} or literal values like {|the bar|}.

Variable references are resolved by first looking for a local variable declaration matching its name, then by looking in the values argument. Literal values always resolve to strings.

As selector expressions must have an annotation or contain a variable reference that references a variable declaration within the same message with an annotation, un-annotated expressions only need to be considered for formattable placeholders.

If a placeholder expression contains a variable reference without an annotation and the variable resolves to a number or bigint value or a Number instance, it resolves instead to the result of calling the number function with the numerical value as input and no options.

If a placeholder expression would resolve to a string value or a String instance, it resolves instead to the result of calling the string function with the value as input and no options.

Otherwise, un-annotated values resolve to the following shape:

interface MessageUnknownValue {
  type: 'unknown';
  source: string;
  locale: string;
  dir: 'ltr' | 'rtl' | 'auto';
  toParts(): [MessageUnknownPart];
  toString(): string;
  valueOf(): unknown;
}

interface MessageUnknownPart {
  type: 'unknown';
  source: string;
  value: unknown;
}

With MessageUnknownValue, the toString() method uses the equivalent of String() to format the value while valueOf() returns the original value.

If the constructor options of a MessageFormat instance include a functions value that overrides the default number or string functions, those will be called instead of the default ones, as appropriate.

number

Accepts as input any of the following:

  • A number or a bigint. This is used directly as the value.
  • An object with a valueOf() method that returns a number or a bigint, which is then used as the value.
  • The JSON string representation of a number, to accommodate literal values as in {42 :number}. The value is determined by the equivalent of calling JSON.parse() on the string, and asserting that it returns a number or a bigint.

Returns a fallback value if invoked without such an input, or if an error occurs while determining value.

Internally, constructs a locales array of strings as used by the Intl.NumberFormat and Intl.PluralRules constructors. This will include, in order:

  1. Any locales set by the expression's "locale" option.
  2. The locale of the input, if it is an object and has a string or string array property "locale".
  3. The base locale or locale chain of the message.

To determine the formatting and selection options for the number, a new empty options object is created. If the input is an object with a property "options" with an object value, the options object is extended with its values. Then, a primitive value is set for each of the expression's options except "locale":

  • For each option value, if it is an object, coerce it to a string.
  • For useGrouping, convert "true" and "false" to their corresponding boolean values.
  • For roundingIncrement, minimumIntegerDigits, and (minimum|maximum)(Fraction|Significant)Digits, parse a string value as a non-negative integer number.

Returns a value with the following shape:

interface MessageNumber {
  type: 'number';
  source: string;
  locale: string;
  dir: 'ltr' | 'rtl' | 'auto';
  options: Intl.NumberFormatOptions & Intl.PluralRulesOptions;
  selectKeys(keys: string[]): string[];
  toParts(): [MessageNumberPart];
  toString(): string;
  valueOf(): number | bigint;
}

interface MessageNumberPart {
  type: 'number';
  source: string;
  locale?: string;
  dir?: 'ltr' | 'rtl' | 'auto';
  parts: Intl.NumberFormatPart[];
}

When a MessageNumber is used as a selector (calling its selectKeys() method), a key with an exact numeric match to the value will be preferred over a key matching the value's plural category (some of zero, one, two, few, many, and other, depending on the locale).

When a MessageNumber is being formatted, calling its toString() will return a string corresponding to calling

new Intl.NumberFormat(locales, options).format(value);

and calling its toParts() method will return an array with a single object member where the parts will correspond to the results of calling

new Intl.NumberFormat(locales, options).formatToParts(value);

string

Accepts any input, and parses any non-string value using String(). For no input, resolves its value to an empty string. On error, resolves to a fallback value.

Accepts only the locale option as an override for the message's locale.

Returns a value with the following shape:

interface MessageString {
  type: 'string';
  source: string;
  locale: string;
  dir: 'ltr' | 'rtl' | 'auto';
  selectKeys(keys: string[]): [] | [string];
  toParts(): [MessageStringPart];
  toString(): string;
  valueOf(): string;
}

interface MessageStringPart {
  type: 'string';
  source: string;
  locale?: string;
  dir?: 'ltr' | 'rtl' | 'auto';
  value: string;
}

When a MessageString is used as a selector, the returned array may include at most one entry, if one of the keys was an exact string match for the value.

Fallback Values

It's possible for a MessageFunction call to throw an error, or for a format() or formatToParts() call to throw an error. When this happens, the error is caught and a user-provided onError handler is called. If no such handler is defined, the default behaviour is to issue a warning rather than throwing the error. This allows for some formatted representation to always be provided.

In such a case, a fallback representation is used instead for the value:

interface MessageFallback {
  type: 'fallback';
  locale: 'und';
  dir: 'auto';
  source: string;
  toParts(): [MessageFallbackPart];
  toString(): string;
}

interface MessageFallbackPart {
  type: 'fallback';
  source: string;
}

This representation is also used when resolving MF2 expressions that include "reserved" or "private-use" annotations.

The source of the MessageFallback corresponds to the source of the MessageValue. When MessageFallback is formatted to a string, its value is the concatenation of a left curly brace {, the source value, and a right curly brace }.

Comparison

The MF2 specification is being developed based upon lessons learned from existing systems including ICU MessageFormat and Fluent.

The implementation of Fluent within Firefox mostly relies upon a declarative syntax in the DOM, but it does provide an API for retrieving messages directly from Fluent when not being used to localize the DOM.

Implementations

Polyfill/transpiler implementations

The MessageFormat 2.0 specification is under development. A polyfill implementation of this proposal is available under the messageformat project.