[Unicode]  Technical Reports

Unicode Standard Annex #9

The Bidirectional Algorithm

Version Unicode 5.0.0
Authors Mark Davis (mark.davis@google.com)
Date 2006-09-15
This Version http://www.unicode.org/reports/tr9/tr9-17.html
Previous Version http://www.unicode.org/reports/tr9/tr9-15.html
Latest Version http://www.unicode.org/reports/tr9/
Revision 17


This annex describes specifications for the positioning of characters flowing from right to left, such as Arabic or Hebrew.


This document has been reviewed by Unicode members and other interested parties, and has been approved for publication by the Unicode Consortium. This is a stable document and may be used as reference material or cited as a normative reference by other specifications.

A Unicode Standard Annex (UAX) forms an integral part of the Unicode Standard, but is published online as a separate document. The Unicode Standard may require conformance to normative content in a Unicode Standard Annex, if so specified in the Conformance chapter of that version of the Unicode Standard. The version number of a UAX document corresponds to the version of the Unicode Standard of which it forms a part.

Please submit corrigenda and other comments with the online reporting form [Feedback]. Related information that is useful in understanding this annex is found in Unicode Standard Annex #41, “Common References for Unicode Standard Annexes.” For the latest version of the Unicode Standard see [Unicode]. For a list of current Unicode Technical Reports see [Reports]. For more information about versions of the Unicode Standard, see [Versions].


1 Introduction

The Unicode Standard prescribes a memory representation order known as logical order. When text is presented in horizontal lines, most scripts display characters from left to right. However, there are several scripts (such as Arabic or Hebrew) where the natural ordering of horizontal text in display is from right to left. If all of the text has the same horizontal direction, then the ordering of the display text is unambiguous. However, when bidirectional text (a mixture of left-to-right and right-to-left horizontal text) is present, some ambiguities can arise in determining the ordering of the displayed characters.

This annex describes the algorithm used to determine the directionality for bidirectional Unicode text. The algorithm extends the implicit model currently employed by a number of existing implementations and adds explicit format codes for special circumstances. In most cases, there is no need to include additional information with the text to obtain correct display ordering.

However, in the case of bidirectional text, there are circumstances where an implicit bidirectional ordering is not sufficient to produce comprehensible text. To deal with these cases, a minimal set of directional formatting codes is defined to control the ordering of characters when rendered. This allows exact control of the display ordering for legible interchange and ensures that plain text used for simple items like filenames or labels can always be correctly ordered for display.

The directional formatting codes are used only to influence the display ordering of text. In all other respects they should be ignored—they have no effect on the comparison of text or on word breaks, parsing, or numeric analysis.

When working with bidirectional text, the characters are still interpreted in logical order—only the display is affected. The display ordering of bidirectional text depends on the directional properties of the characters in the text.

2 Directional Formatting Codes

Two types of explicit codes are used to modify the standard implicit Unicode Bidirectional Algorithm. In addition, there are implicit ordering codes, the right-to-left and left-to-right marks. All of these codes are limited to the current paragraph; thus their effects are terminated by a paragraph separator. The directional types left-to-right and right-to-left are called strong types, and characters of those types are called strong directional characters. The directional types associated with numbers are called weak types, and characters of those types are called weak directional characters.

Although the term embedding is used for some explicit codes, the text within the scope of the codes is not independent of the surrounding text. Characters within an embedding can affect the ordering of characters outside, and vice versa. The algorithm is designed so that the use of explicit codes can be equivalently represented by out-of-line information, such as stylesheet information. However, any alternative representation will be defined by reference to the behavior of the explicit codes in this algorithm.

2.1 Explicit Directional Embedding

The following codes signal that a piece of text is to be treated as embedded. For example, an English quotation in the middle of an Arabic sentence could be marked as being embedded left-to-right text. If there were a Hebrew phrase in the middle of the English quotation, the that phrase could be marked as being embedded right-to-left. These codes allow for nested embeddings.

RLE Right-to-Left Embedding   Treat the following text as embedded right-to-left.
LRE Left-to-Right Embedding   Treat the following text as embedded left-to-right.

The precise meaning of these codes will be made clear in the discussion of the algorithm. The effect of right-left line direction, for example, can be accomplished by simply embedding the text with RLE...PDF.

2.2 Explicit Directional Overrides

The following codes allow the bidirectional character types to be overridden when required for special cases, such as for part numbers. These codes allow for nested directional overrides.

RLO Right-to-Left Override   Force following characters to be treated as strong right-to-left characters.
LRO Left-to-Right Override   Force following characters to be treated as strong left-to-right characters.

The precise meaning of these codes will be made clear in the discussion of the algorithm. The right-to-left override, for example, can be used to force a part number made of mixed English, digits and Hebrew letters to be written from right to left.

2.3 Terminating Explicit Directional Code

The following code terminates the effects of the last explicit code (either embedding or override) and restores the bidirectional state to what it was before that code was encountered.

PDF Pop Directional Format   Restore the bidirectional state to what it was before the last LRE, RLE, RLO, or LRO.

2.4 Implicit Directional Marks

These characters are very light-weight codes. They act exactly like right-to-left or left-to-right characters, except that they do not display or have any other semantic effect. Their use is more convenient than using explicit embeddings or overrides because their scope is much more local.

RLM Right-to-Left Mark   Right-to-left zero-width character
LRM Left-to-Right Mark   Left-to-right zero-width character

There is no special mention of the implicit directional marks in the following algorithm. That is because their effect on bidirectional ordering is exactly the same as a corresponding strong directional character; the only difference is that they do not appear in the display.

3 Basic Display Algorithm

The Bidirectional Algorithm takes a stream of text as input and proceeds in three main phases:

The algorithm reorders text only within a paragraph; characters in one paragraph have no effect on characters in a different paragraph. Paragraphs are divided by the Paragraph Separator or appropriate Newline Function (for guidelines on the handling of CR, LF, and CRLF, see Section 4.4, Directionality, and Section 5.8, Newline Guidelines of [Unicode]). Paragraphs may also be determined by higher-level protocols: for example, the text in two different cells of a table will be in different paragraphs.

Combining characters always attach to the preceding base character in the memory representation. Even after reordering for display and performing character shaping, the glyph representing a combining character will attach to the glyph representing its base character in memory. Depending on the line orientation and the placement direction of base letterform glyphs, it may, for example, attach to the glyph on the left, or on the right, or above.

This annex uses the numbering conventions for normative definitions and rules in Table 1.

Table 1. Normative Definitions and Rules

Numbering Section
BDn Definitions
Pn Paragraph levels
Xn Explicit levels and directions
Wn Weak types
Nn Neutral types
In Implicit levels
Ln Resolved levels

3.1 Definitions

BD1. The bidirectional characters types are values assigned to each Unicode character, including unassigned characters.

BD2. Embedding levels are numbers that indicate how deeply the text is nested, and the default direction of text on that level. The minimum embedding level of text is zero, and the maximum explicit depth is level 61.

Embedding levels are explicitly set by both override format codes and by embedding format codes; higher numbers mean the text is more deeply nested. The reason for having a limitation is to provide a precise stack limit for implementations to guarantee the same results. Sixty-one levels is far more than sufficient for ordering, even with mechanically generated formatting; the display becomes rather muddied with more than a small number of embeddings.

BD3. The default direction of the current embedding level (for the character in question) is called the embedding direction. It is L if the embedding level is even, and R if the embedding level is odd.

For example, in a particular piece of text, Level 0 is plain English text. Level 1 is plain Arabic text, possibly embedded within English level 0 text. Level 2 is English text, possibly embedded within Arabic level 1 text, and so on. Unless their direction is overridden, English text and numbers will always be an even level; Arabic text (excluding numbers) will always be an odd level. The exact meaning of the embedding level will become clear when the reordering algorithm is discussed, but the following provides an example of how the algorithm works.

BD4. The paragraph embedding level is the embedding level that determines the default bidirectional orientation of the text in that paragraph.

BD5. The direction of the paragraph embedding level is called the paragraph direction.

BD6. The directional override status determines whether the bidirectional type of characters is to be reset with explicit directional controls. This status has three states, as shown in Table 2.

Table 2. Directional Override Status

Status Interpretation
Neutral No override is currently active
Right-to-left Characters are to be reset to R
Left-to-right Characters are to be reset to L

BD7. A level run is a maximal substring of characters that have the same embedding level. It is maximal in that no character immediately before or after the substring has the same level (a level run is also known as a directional run).


In this and the following examples, case is used to indicate different implicit character types for those unfamiliar with right-to-left letters. Uppercase letters stand for right-to-left characters (such as Arabic or Hebrew), and lowercase letters stand for left-to-right characters (such as English or Russian).

Memory:            car is THE CAR in arabic

Character types:   LLL-LL-RRR-RRR-LL-LLLLLL

Resolved levels:   000000011111110000000000

Notice that the neutral character (space) between THE and CAR gets the level of the surrounding characters. The level of the neutral characters can also be changed by inserting appropriate directional marks around neutral characters. These marks have no other effects.

Table 3 lists additional abbreviations used in the examples and internal character types used in the algorithm.

Table 3. Abbreviations for Examples and Internal Types

Symbol Description
N Neutral or Separator (B, S, WS, ON)
e The text ordering type (L or R) that matches the embedding level direction (even or odd)
sor The text ordering type (L or R) assigned to the position before a level run.
eor The text ordering type (L or R) assigned to the position after a level run.

3.2 Bidirectional Character Types

The normative bidirectional character types for each character are specified in the Unicode Character Database [UCD] and are summarized in Table 4. This is a summary only: there are exceptions to the general scope. For example, certain characters such as U+0CBF KANNADA VOWEL SIGN I are given Type L (instead of NSM) to preserve canonical equivalence.

Table 4. Bidirectional Character Types

Category Type Description General Scope
Strong L Left-to-Right LRM, most alphabetic, syllabic, Han ideographs, non-European or non-Arabic digits, ...
LRE Left-to-Right Embedding LRE
LRO Left-to-Right Override LRO
R Right-to-Left RLM, Hebrew alphabet, and related punctuation
AL Right-to-Left Arabic Arabic, Thaana, and Syriac alphabets, most punctuation specific to those scripts, ...
RLE Right-to-Left Embedding RLE
RLO Right-to-Left Override RLO
Weak PDF Pop Directional Format PDF
EN European Number European digits, Eastern Arabic-Indic digits, ...
ES European Number Separator Plus sign, minus sign
ET European Number Terminator Degree sign, currency symbols, ...
AN Arabic Number Arabic-Indic digits, Arabic decimal and thousands separators, ...
CS Common Number Separator Colon, comma, full stop (period), No-break space, ...
NSM Nonspacing Mark Characters marked Mn (Nonspacing_Mark) and Me (Enclosing_Mark) in the Unicode Character Database
BN Boundary Neutral Most formatting and control characters, other than those explicitly given types above
Neutral B Paragraph Separator Paragraph separator, appropriate Newline Functions, higher-level protocol paragraph determination
S Segment Separator Tab
WS Whitespace Space, figure space, line separator, form feed, General Punctuation spaces, ...
ON Other Neutrals All other characters, including OBJECT REPLACEMENT CHARACTER


3.3 Resolving Embedding Levels

The body of the Bidirectional Algorithm uses character types and explicit codes to produce a list of resolved levels. This resolution process consists of five steps: (1) determining the paragraph level; (2) determining explicit embedding levels and directions; (3) resolving weak types; (4) resolving neutral types; and (5) resolving implicit embedding levels.

3.3.1 The Paragraph Level

P1. Split the text into separate paragraphs. A paragraph separator is kept with the previous paragraph. Within each paragraph, apply all the other rules of this algorithm.

P2. In each paragraph, find the first character of type L, AL, or R.

Because paragraph separators delimit text in this algorithm, this will generally be the first strong character after a paragraph separator or at the very beginning of the text. Note that the characters of type LRE, LRO, RLE, or RLO are ignored in this rule. This is because typically they are used to indicate that the embedded text is the opposite direction than the paragraph level.

P3. If a character is found in P2 and it is of type AL or R, then set the paragraph embedding level to one; otherwise, set it to zero.

Whenever a higher-level protocol specifies the paragraph level, rules P2 and P3 do not apply.

3.3.2 Explicit Levels and Directions

All explicit embedding levels are determined from the embedding and override codes, by applying the explicit level rules X1 through X9. These rules are applied as part of the same logical pass over the input.

Explicit Embeddings

X1. Begin by setting the current embedding level to the paragraph embedding level. Set the directional override status to neutral. Process each character iteratively, applying rules X2 through X9. Only embedding levels from 0 to 61 are valid in this phase.

In the resolution of levels in rules I1 and I2, the maximum embedding level of 62 can be reached.

X2. With each RLE, compute the least greater odd embedding level.

a. If this new level would be valid, then this embedding code is valid. Remember (push) the current embedding level and override status. Reset the current level to this new level, and reset the override status to neutral.

b. If the new level would not be valid, then this code is invalid. Do not change the current level or override status.

For example, level 0 → 1; levels 1, 2 → 3; levels 3, 4 → 5; ...59, 60 → 61; above 60, no change (do not change levels with RLE if the new level would be invalid).

X3. With each LRE, compute the least greater even embedding level.

a. If this new level would be valid, then this embedding code is valid. Remember (push) the current embedding level and override status. Reset the current level to this new level, and reset the override status to neutral.

b. If the new level would not be valid, then this code is invalid. Do not change the current level or override status.

For example, levels 0, 1 → 2; levels 2, 3 → 4; levels 4, 5 → 6; ...58, 59 → 60; above 59, no change (do not change levels with LRE if the new level would be invalid).

Explicit Overrides

An explicit directional override sets the embedding level in the same way the explicit embedding codes do, but also changes the directional character type of affected characters to the override direction.

X4. With each RLO, compute the least greater odd embedding level.

a. If this new level would be valid, then this embedding code is valid. Remember (push) the current embedding level and override status. Reset the current level to this new level, and reset the override status to right-to-left.

b. If the new level would not be valid, then this code is invalid. Do not change the current level or override status.

X5. With each LRO, compute the least greater even embedding level.

a. If this new level would be valid, then this embedding code is valid. Remember (push) the current embedding level and override status. Reset the current level to this new level, and reset the override status to left-to-right.

b. If the new level would not be valid, then this code is invalid. Do not change the current level or override status.

X6. For all types besides RLE, LRE, RLO, LRO, and PDF:

a. Set the level of the current character to the current embedding level.

b. Whenever the directional override status is not neutral, reset the current character type to the directional override status.

If the directional override status is neutral, then characters retain their normal types: Arabic characters stay AL, Latin characters stay L, neutrals stay N, and so on. If the directional override status is R, then characters become R. If the directional override status is L, then characters become L.

Terminating Embeddings and Overrides

There is a single code to terminate the scope of the current explicit code, whether an embedding or a directional override. All codes and pushed states are completely popped at the end of paragraphs.

X7. With each PDF, determine the matching embedding or override code. If there was a valid matching code, restore (pop) the last remembered (pushed) embedding level and directional override.

X8. All explicit directional embeddings and overrides are completely terminated at the end of each paragraph. Paragraph separators are not included in the embedding.

X9. Remove all RLE, LRE, RLO, LRO, PDF, and BN codes.

X10. The remaining rules are applied to each run of characters at the same level. For each run, determine the start-of-level-run (sor) and end-of-level-run (eor) type, either L or R. This depends on the higher of the two levels on either side of the boundary (at the start or end of the paragraph, the level of the “other” run is the base embedding level). If the higher level is odd, the type is R; otherwise, it is L.

For example:

Levels:  0   0   0   1   1   1   2

Runs:   <--- 1 ---> <--- 2 ---> <3>

Run 1 is at level 0, sor is L, eor is R.
Run 2 is at level 1, sor is R, eor is L.
Run 3 is at level 2, sor is L, eor is L.

For two adjacent runs, the eor of the first run is the same as the sor of the second.

3.3.3 Resolving Weak Types

Weak types are now resolved one level run at a time. At level run boundaries where the type of the character on the other side of the boundary is required, the type assigned to sor or eor is used.

Nonspacing marks are now resolved based on the previous characters.

W1. Examine each nonspacing mark (NSM) in the level run, and change the type of the NSM to the type of the previous character. If the NSM is at the start of the level run, it will get the type of sor.


Assume in this example that sor is R:

sor NSM     → sor R

The text is next parsed for numbers. This pass will change the directional types European Number Separator, European Number Terminator, and Common Number Separator to be European Number text, Arabic Number text, or Other Neutral text. The text to be scanned may have already had its type altered by directional overrides. If so, then it will not parse as numeric.

W2. Search backward from each instance of a European number until the first strong type (R, L, AL, or sor) is found. If an AL is found, change the type of the European number to Arabic number.

AL EN    → AL AN


sor N EN → sor N EN

L N EN   → L N EN

R N EN   → R N EN

W3. Change all ALs to R.

W4. A single European separator between two European numbers changes to a European number. A single common separator between two numbers of the same type changes to that type.




W5. A sequence of European terminators adjacent to European numbers changes to all European numbers.




W6. Otherwise, separators and terminators change to Other Neutral.

AN ET    → AN ON



ET AN    → ON AN

W7. Search backward from each instance of a European number until the first strong type (R, L, or sor) is found. If an L is found, then change the type of the European number to L.

L  N EN => L  N  L

R  N EN => R  N  EN


3.3.4 Resolving Neutral Types

Neutral types are now resolved one level run at a time. At level run boundaries where the type of the character on the other side of the boundary is required, the type assigned to sor or eor is used.

The next phase resolves the direction of the neutrals. The results of this phase are that all neutrals become either R or L. Generally, neutrals take on the direction of the surrounding text. In case of a conflict, they take on the embedding direction.

N1. A sequence of neutrals takes the direction of the surrounding strong text if the text on both sides has the same direction. European and Arabic numbers act as if they were R in terms of their influence on neutrals. Start-of-level-run (sor) and end-of-level-run (eor) are used at level run boundaries.

R  N  R  → R  R  R

L  N  L  → L  L  L

R  N  AN → R  R  AN

AN N  R  → AN R  R

R  N  EN → R  R  EN

EN N  R  → EN R  R

Note that any AN or EN remaining after W7 will be in an right-to-left context.

N2. Any remaining neutrals take the embedding direction.

N → e

Assume in this example that eor is L and sor is R. Then an application of N1 and N2 yields the following:

L   N eor → L   L eor

R   N eor → R   e eor

sor N L   → sor e L

sor N R   → sor R R

Examples. A list of numbers separated by neutrals and embedded in a directional run will come out in the run’s order.

  Storage:	he said "THE VALUES ARE 123, 456, 789, OK".

  Display:	he said "KO ,789 ,456 ,123 ERA SEULAV EHT".

In this case, both the comma and the space between the numbers take on the direction of the surrounding text (uppercase = right-to-left), ignoring the numbers. The commas are not considered part of the number because they are not surrounded on both sides (see Section 3.3.3, Resolving Weak Types). However, if there is an adjacent left-to-right sequence, then European numbers will adopt that direction:

Storage:	he said "IT IS A bmw 500, OK."

Display:	he said ".KO ,bmw 500 A SI TI"

3.3.5 Resolving Implicit Levels

In the final phase, the embedding level of text may be increased, based on the resolved character type. Right-to-left text will always end up with an odd level, and left-to-right and numeric text will always end up with an even level. In addition, numeric text will always end up with a higher level than the paragraph level. (Note that it is possible for text to end up at levels higher than 61 as a result of this process.) This results in the following rules:

I1. For all characters with an even (left-to-right) embedding direction, those of type R go up one level and those of type AN or EN go up two levels.

I2. For all characters with an odd (right-to-left) embedding direction, those of type L, EN or AN go up one level.

Table 5 summarizes the results of the implicit algorithm.

Table 5. Resolving Implicit Levels

Type Embedding Level
Even Odd
AN EL+2 EL+1
EN EL+2 EL+1

3.4 Reordering Resolved Levels

The following rules describe the logical process of finding the correct display order. As opposed to resolution phases, these rules act on a per-line basis and are applied after any line wrapping is applied to the paragraph.

Logically there are the following steps:

L1. On each line, reset the embedding level of the following characters to the paragraph embedding level:

  1. Segment separators,
  2. Paragraph separators,
  3. Any sequence of whitespace characters preceding a segment separator or paragraph separator, and
  4. Any sequence of white space characters at the end of the line.

In combination with the following rule, this means that trailing whitespace will appear at the visual end of the line (in the paragraph direction). Tabulation will always have a consistent direction within a paragraph.

L2. From the highest level found in the text to the lowest odd level on each line, including intermediate levels not actually present in the text, reverse any contiguous sequence of characters that are at that level or higher.

This reverses a progressively larger series of substrings. The following four examples illustrate this. In these examples, the paragraph embedding level for the first and third examples is assumed to be 0 (left-to-right direction), and for the second and fourth examples is assumed to be 1 (right-to-left direction).

Example 1 (embedding level = 0)
Memory:                car means CAR.

Resolved levels:     00000000001110

Reverse level 1:     car means RAC.
Example 2 (embedding level = 1)
Memory:                car MEANS CAR.

Resolved levels:     22211111111111

Reverse level 2:     rac MEANS CAR.

Reverse levels 1-2:  .RAC SNAEM car
Example 3 (embedding level = 0)
Memory:                he said "car MEANS CAR."

Resolved levels:     000000000222111111111100

Reverse level 2:     he said "rac MEANS CAR."

Reverse levels 1-2:  he said "RAC SNAEM car."
Example 4 (embedding level = 1)
Memory:                DID YOU SAY ‘he said "car MEANS CAR"’?

Resolved levels:     11111111111112222222224443333333333211

Reverse level 4:     DID YOU SAY ‘he said "rac MEANS CAR"’?

Reverse levels 3-4:  DID YOU SAY ‘he said "RAC SNAEM car"’?

Reverse levels 2-4:  DID YOU SAY ‘"rac MEANS CAR" dias eh’?

Reverse levels 1-4:  ?‘he said "RAC SNAEM car"’ YAS UOY DID

L3. Combining marks applied to a right-to-left base character will at this point precede their base character. If the rendering engine expects them to follow the base characters in the final display process, then the ordering of the marks and the base character must be reversed.

Many font designers provide default metrics for combining marks that support rendering by simple overhang. Because of the reordering for right-to-left characters, it is common practice to make the glyphs for most combining characters overhang to the left (thus assuming the characters will be applied to left-to-right base characters) and make the glyphs for combining characters in right-to-left scripts overhang to the right (thus assuming that the characters will be applied to right-to-left base characters). With such fonts, the display ordering of the marks and base glyphs may need to be adjusted when combining marks are applied to “unmatching” base characters. See Section 5.13, Rendering Nonspacing Marks of [Unicode], for more information.

L4. A character is depicted by a mirrored glyph if and only if (a) the resolved directionality of that character is R, and (b) the Bidi_Mirrored property value of that character is true.

For example, U+0028 left parenthesis—which is interpreted in the Unicode Standard as an opening parenthesis—appears as “(” when its resolved level is even, and as the mirrored glyph “)” when its resolved level is odd. Note that for backward compatibility the characters U+FD3E (﴾) ORNATE LEFT PARENTHESIS and U+FD3F (﴿) ORNATE RIGHT PARENTHESIS are not mirrored.

3.5 Shaping

Cursively connected scripts, such as Arabic or Syriac, require the selection of positional character shapes that depend on adjacent characters (see Section 8.2, Arabic of [Unicode]). Shaping is logically applied after the Bidirectional Algorithm is used and is limited to characters within the same directional run. Consider the following example string of Arabic characters, which is represented in memory as characters 1, 2, 3, and 4, and where the first two characters are overridden to be LTR. To show both paragraph directions, the next two are embedded, but with the normal RTL direction.













One can use embedding codes to achieve this effect in plain text or use markup in HTML, as in the examples below. (The bold text would be for the right-to-left paragraph direction.)

The resulting shapes will be the following, according to the paragraph direction:

Left-Right Paragraph Right-Left Paragraph

















3.5.1 Shaping and Line Breaking

The process of breaking a paragraph into one or more lines that fit within particular bounds is outside the scope of the Bidirectional Algorithm. Where character shaping is involved, the width calculations must be based on the shaped glyphs.

Note that the soft-hyphen (SHY) works in cursively connected scripts as it does in other scripts. That is, it indicates a point where the line could be broken in the middle of a word. If the rendering system breaks at that point, the display—including shaping—should be what is appropriate for the given language. For more information on this and other line breaking issues, see Unicode Standard Annex #14, “Line Breaking Properties” [UAX14].

4 Bidirectional Conformance

A process that claims conformance to this specification shall satisfy the the following clauses:


In the absence of a permissible higher-level protocol, a process that renders text shall display all visible representations of characters (excluding format characters) in the order described by Section 3, Basic Display Algorithm, of this annex. In particular, this includes definitions BD1BD7 and steps P1P3, X1X10, W1W7, N1N2, I1I2, and L1L4.


The only permissible higher-level protocols are those listed in Section 4.3, Higher-Level Protocols. They are HL1, HL2, HL3, HL4, HL5, and HL6.

Use of higher-level protocols is discouraged, because it introduces interchange problems and can lead to security problems. For more information, see Unicode Technical Report #36, “Unicode Security Considerations” [UTR36].

4.1 Boundary Neutrals

The goal in marking a format or control character as BN is that it have no effect on the rest of the algorithm. (ZWJ and ZWNJ are exceptions; see X9). Because conformance does not require the precise ordering of format characters with respect to others, implementations can handle them in different ways as long as they preserve the ordering of the other characters.

4.2 Explicit Formatting Codes

As with any Unicode characters, systems do not have to support any particular explicit directional formatting code (although it is not generally useful to include a terminating code without including the initiator). Generally, conforming systems will fall into three classes:

4.3 Higher-Level Protocols

The following clauses are the only permissible ways for systems to apply higher-level protocols to the ordering of bidirectional text. Some of the clauses apply to segments of structured text. This refers to the situation where text is interpreted as being structured, whether with explicit markup such as XML or HTML, or internally structured such as in a word processor or spreadsheet. In such a case, a segment is span of text that is distinguished in some way by the structure. 


Override P3, and set the paragraph embedding level explicitly.

  • A higher-level protocol may set the paragraph level explicitly and ignore P3. This can be done on the basis of the context, such as on a table cell, paragraph, document, or system level.
HL2. Override W2, and set EN or AN explicitly.
  • A higher-level process may reset characters of type EN to AN, or vice versa, and ignore W2. For example, style sheet or markup information can be used within a span of text to override the setting of EN text to be always be AN, or vice versa.
HL3. Emulate directional overrides or embedding codes.
  • A higher-level protocol can impose a directional override or embedding on a segment of structured text. The behavior must always be defined by reference to what would happen if the equivalent explicit codes as defined in the algorithm were inserted into the text. For example, a style sheet or markup can set the embedding level on a span of text.
HL4. Apply the Bidirectional Algorithm to segments.
  • The Bidirectional Algorithm can be applied independently to one or more segments of structured text. For example, when displaying a document consisting of textual data and visible markup in an editor, a higher-level process can handle syntactic elements in the markup separately from the textual data.
HL5. Provide artificial context.
  • Text can be processed by the Bidirectional Algorithm as if it were preceded by a character of a given type and/or followed by a character of a given type. This allows a piece of text that is extracted from a longer sequence of text to behave as it did in the larger context.
HL6. Additional mirroring.
  • Characters with a resolved directionality of R that do not have the Bidi_Mirrored property can also be depicted by a mirrored glyph in specialized contexts. Such contexts include, but are not limited to, historic scripts and associated punctuation, private-use characters, and characters in mathematical expressions. (See Section 6, Mirroring.)

Clauses HL1 and HL3 are not logically necessary; they are covered by applications of clauses HL4 and HL5. However, they are included for clarity because they are more common operations.

As an example of the application of HL4, suppose an XML document contains the following fragment. (Note: This is a simplified example for illustration: element names, attribute names, and attribute values could all be involved.)

ARABICenglishARABIC<e1 type='ab'>ARABICenglish<e2 type='cd'>english

This can be analyzed as being five different segments:

  1. ARABICenglishARABIC
  2. <e1 type='ab'>
  3. ARABICenglish
  4. <e2 type='cd'>
  5. english

To make the XML file readable as source text, the display in an editor could order these elements all in a uniform direction (for example, all left-to-right) and apply the Bidirectional Algorithm to each field separately. It could also choose to order the element names, attribute names, and attribute values uniformly in the same direction (for example, all left-to-right). For final display, the markup could be ignored, allowing all of the text (segments a, c, and e) to be reordered together.

When text using a higher-level protocol is to be converted to Unicode plain text, for consistent appearance formatting codes should be inserted to ensure that the order matches that of the higher-level protocol.

5 Implementation Notes

5.1 Reference Code

There are two versions of BIDI reference code available. Both have been tested to produce identical results. One version is written in Java, and the other is written in C++. The Java version is designed to closely follow the steps of the algorithm as described below. The C++ code is designed to show one of the optimization methods that can be applied to the algorithm, using a state table for one phase.

One of the most effective optimizations is to first test for right-to-left characters and not invoke the Bidirectional Algorithm unless they are present.

There are two directories containing source code for reference implementations: BidiReferenceJava and BidiReferenceCpp [Code9]. Implementers are encouraged to use this resource to test their implementations.

5.2 Retaining Format Codes

Some implementations may wish to retain the format codes when running the algorithm. The following provides a summary of how this may be done. Note that this summary is an informative implementation guideline; it should provide the same results as the explicit algorithm above, but in case of any deviation the explicit algorithm is the normative statement for conformance.

Implementations that display visible representations of format characters will want to adjust this process to position the format characters optimally for editing.

5.3 Joiners

As described under X9, the zero width joiner and non-joiner affect the shaping of the adjacent characters—those that are adjacent in the original backing-store order—even though those characters may end up being rearranged to be non-adjacent by the Bidirectional Algorithm. To determine the joining behavior of a particular character after applying the Bidirectional Algorithm, there are two main strategies:

5.4 Vertical Text

In the case of vertical line orientation, the Bidirectional Algorithm is still used to determine the levels of the text. However, these levels are not used to reorder the text, because the characters are usually ordered uniformly from top to bottom. Instead, the levels are used to determine the rotation of the text. Sometimes vertical lines follow a vertical baseline in which each character is oriented as normal (with no rotation), with characters ordered from top to bottom whether they are Hebrew, numbers, or Latin. When setting text using the Arabic script in vertical lines, it is more common to employ a horizontal baseline that is rotated by 90° counterclockwise so that the characters are ordered from top to bottom. Latin text and numbers may be rotated 90° clockwise so that the characters are also ordered from top to bottom.

The Bidirectional Algorithm is used when some characters are ordered from bottom to top. For example, this happens with a mixture of Arabic and Latin glyphs when all the glyphs are rotated uniformly 90° clockwise. The Unicode Standard does not specify whether text is presented horizontally or vertically, or whether text is rotated. That is left up to higher-level protocols.

5.5 Usage

Because of the implicit character types and the heuristics for resolving neutral and numeric directional behavior, the implicit bidirectional ordering will generally produce the correct display without any further work. However, problematic cases may occur when a right-to-left paragraph begins with left-to-right characters, or there are nested segments of different-direction text, or there are weak characters on directional boundaries. In these cases, embeddings or directional marks may be required to get the right display. Part numbers may also require directional overrides.

The most common problematic case is that of neutrals on the boundary of an embedded language. This can be addressed by setting the level of the embedded text correctly. For example, with all the text at level 0 the following occurs:

Memory:  he said "I NEED WATER!", and expired.

Display: he said "RETAW DEEN I!", and expired.

If the exclamation mark is to be part of the Arabic quotation, then the user can select the text I NEED WATER! and explicitly mark it as embedded Arabic, which produces the following result:

Memory:  he said "<RLE>I NEED WATER!<PDF>", and expired.

Display: he said "!RETAW DEEN I", and expired.

A simpler method of doing this is to place a right directional mark (RLM) after the exclamation mark. Because the exclamation mark is now not on a directional boundary, this produces the correct result.

Memory:  he said "I NEED WATER!<RLM>", and expired.

Display: he said "!RETAW DEEN I", and expired.

This latter approach is preferred because it does not make use of the stateful format codes, which can easily get out of sync if not fully supported by editors and other string manipulation. The stateful format codes are generally needed only for more complex (and rare) cases such as double embeddings, as in the following:

Memory:  DID YOU SAY ‘<LRE>he said "I NEED WATER!<RLM>", and expired.<PDF>’?

Display: ?‘he said "!RETAW DEEN I", and expired.’ YAS UOY DID

Migrating from 2.0 to 3.0

In the Unicode Character Database for [Unicode3.0], new bidirectional character types were introduced to make the body of the Bidirectional Algorithm depend only on the types of characters, and not on the character values. The changes from the 2.0 bidirectional types are listed in Table 6.





Table 6. New Bidirectional Types in Unicode 3.0

Characters New Bidirectional Type
All characters with General_Category Me, Mn NSM
All characters of type R in the Arabic ranges (0600..06FF, FB50..FDFF, FE70..FEFE)
(Letters in the Thaana and Syriac ranges also have this value.)
The explicit embedding characters: LRO, RLO, LRE, RLE, PDF LRO, RLO, LRE, RLE, PDF, respectively
Formatting characters and controls (General_Category Cf and Cc) that were of bidirectional type ON BN
Zero Width Space BN

Implementations that use older property tables can adjust to the modifications in the Bidirectional Algorithm by algorithmically remapping the characters in Table 6 to the new types.

6 Mirroring

The mirrored property is important to ensure that the correct character codes are used for the desired semantic. This is of particular importance where the name of a character does not indicate the intended semantic, such as with U+0028 “(” LEFT PARENTHESIS. While the name indicates that it is a left parenthesis, the character really expresses an open parenthesis—the leading character in a parenthetical phrase, not the trailing one.

Some of the characters that do not have the Bidi_Mirrored property may be rendered with mirrored glyphs, according to a higher level protocol that adds mirroring: see Section 4.3, Higher-Level Protocols, especially HL6. Except in such cases, mirroring must be done according to rule L4, to ensure that the correct character code is used to express the intended semantic of the character, and to avoid interoperability and security problems.

Implementing rule L4 calls for mirrored glyphs. These glyphs may not be exact graphical mirror images. For example, clearly an italic parenthesis is not an exact mirror image of another— “(” is not the mirror image of “)”. Instead, mirror glyphs are those acceptable as mirrors within the normal parameters of the font in which they are represented.

In implementation, sometimes pairs of characters are acceptable mirrors for one another—for example, U+0028 “(” LEFT PARENTHESIS and U+0029 “)” RIGHT PARENTHESIS or U+22E0 “⋠” DOES NOT PRECEDE OR EQUAL and U+22E1 “⋡” DOES NOT SUCCEED OR EQUAL. Other characters such as U+2231 “∱” CLOCKWISE INTEGRAL do not have corresponding characters that can be used for acceptable mirrors. The informative Bidi Mirroring data file [Data9], lists the paired characters with acceptable mirror glyphs. A comment in the file indicates where the pairs are “best fit”: they should be acceptable in rendering, although ideally the mirrored glyphs may have somewhat different shapes.



Mark Davis is the author of the initial version and has added to and maintained the text of this annex.

Thanks to the following people for their contributions to the Bidirectional Algorithm or for their feedback on earlier versions of this annex: Alaa Ghoneim (علاء غنيم), Ahmed Talaat (أحمد طلعت), Asmus Freytag, Avery Bishop, Behdad Esfahbod (بهداد اسفهبد), Doug Felt, Eric Mader, Ernest Cline, Gidi Shalom-Bendor (גידי שלום-בן דור), Isai Scheinberg, Israel Gidali (ישראל גידלי), Joe Becker, John McConnell, Jonathan Kew, Jonathan Rosenne (יונתן רוזן), Khaled Sherif (خالد شريف), Kamal Mansour (كمال منصور), Kenneth Whistler, Maha Hassan (مها حسن), Markus Scherer, Martin Dürst, Mati Allouche (מתתיהו  אלוש), Michel Suignard, Mike Ksar (ميشيل قصار), Murray Sargent, Paul Nelson, Rick McGowan, Roozbeh Pournader (روزبه پورنادر), Steve Atkin, and Thomas Milo (تُومَاسْ مِيلُو).


For references for this annex, see Unicode Standard Annex #41, “Common References for Unicode Standard Annexes.”


For details of the change history, see the online copy of this annex at http://www.unicode.org/reports/tr9/.

The following summarizes modifications from previous revisions of this annex.

Revision 17

  • This revision incorporates successive changes. The latest changes were on 2006-02-24.
  • Modified L4 and HL6, in conjunction with proposed property change to Bidi_Mirrored (PRI #80)
  • Used new format for conformance clause numbering.
  • Added caution on use of higher-level protocols, after UAX9-C2.
  • Some wording changes in 6  Mirroring, for consistency with new L4 and HL6.
  • Moved text to Shaping_and_line_breaking, and added note on SHY in 3.4 there
  • Removed two notes indicating that the conformance clauses override clause C13 of Unicode 4.0.
  • Changed some references to Unicode4.0

Revision 16 being a proposed update, only changes between versions 17 and 15 are noted here.

Revision 15:

  • Minor editing
  • Fixed section Number for Mirroring
  • Changed “Tracking Number” to Revision
  • Added note on U+0CBF KANNADA VOWEL SIGN I
  • Added note after N1, and clarified example after N2.
  • Fixed references to sections of the Unicode Standard

Revision 14:

  • Aliased directional run and level run
  • Pointed to DerivedBidiClass.txt for unassigned character assignments.

Revision 13:

  • 4. Bidirectional Conformance: added explicit clauses.
  • 4.3. Higher-Level Protocols:
    • Added clarifying text, and renumbered options.
    • Removed option regarding number shaping (because it was irrelevant to bidirectional ordering).
    • Broadened the ability to override on the basis of context, and clarified number handling.
    • Made clear that bidi could be applied to segments
  • 1. Introduction: added note that the changes in 4. Bidirectional Conformance override clause C13 of Unicode 4.0 [Unicode], and tighten the conformance requirements from what they had been previously.
  • Minor editing for clarification.

Revision 11:

  • Updated for Unicode 4.0.
  • Added note on canonical equivalence
  • Added Joiners section on ZWJ and ZWNJ
  • Clarified L2 and examples following.
  • Added a section on the interaction of shaping and bidirectional reordering.
  • Moved lists for unassigned characters into UCD.html (also now explicit in DerivedBidiClass.txt)
  • Updated references for Newline Guidelines (because the UAX is incorporated into the 4.0 book)
  • The first two sections were rearranged, with Reference Code going into Implementation Notes, and Mirroring in its own section at the end.
    • This is not highlighted in the proposed text.
  • Sections were renumbered and the table of contents is more detailed.
    • This is not highlighted in the proposed text.
  • Misc editing.

Revision 10:

  • Updated for Unicode 3.2.
  • Updated UAX boilerplate in the status section.

Revision 9:

  • Clarified the language of P2
  • Corrected the implementation note on “Retaining Format Codes” in Implementation Notes
  • Minor editing