The Unicode Bidi Algorithm is a very useful, general, and standard approach for displaying text that contains right-to-left scripts, such as Arabic and Hebrew. But there are situations in which it is awkward to use and/or is visually confusing. This post considers three such situations: math zones, International Resource Identifiers (IRIs), and HTML spans with a directionality attribute. In each of these cases, tailoring the Unicode Bidi Algorithm improves the result. The post starts with a summary of the original case of tailoring, namely the popular keyboard “higher-level protocol” used in rich-text applications.

Keyboard-Driven Bidi Algorithm

For the first example of tailoring the Unicode bidi algorithm, consider keyboard input into rich-text editors, such as Word and RichEdit rich-text edit controls. For this kind of input, the directionality of neutral characters, such as space and parentheses, is determined by the keyboard language. Specifically if the keyboard language is Arabic, Hebrew, or some other right-to-left (RTL) language, the neutrals are treated as RTL characters. If the keyboard language is left-to-right (LTR), the neutrals are treated as LTR characters. Spans of digits, possibly including embedded commas and periods, are always displayed LTR even when entered with an RTL keyboard, but the directionality given to the number they comprise has the directionality of the keyboard language.

Users generally like this approach to bidi text, since it’s simple and predictable. If runs of text could always be reliably stamped by the input language, the Unicode Bidi Algorithm would only have to handle some esoteric text embeddings and ensuring that numbers are displayed LTR. But plain text is commonplace and hence there has to be a way to display bidi text correctly without language attributes. Unfortunately in order to handle the wide variety of special character combinations, the resulting algorithm is very complicated, involving the set of special override codes LRM, RLM, LRO, RLO, LRE, RLE, and PDF. In a real sense, these override codes are used to encode a rich-text directionality attribute in plain text. The justification for this is the Unicode principle that “plain text must contain enough information to permit the text to be rendered legibly and nothing more” (see Plain Text subsection of Sec. 2.2 in Chapter 2 of the Unicode Standard).

Note that plain text pasted into a rich-text instance is treated using the Unicode Bidi Algorithm; the keyboard language doesn’t play a role in pastes.

Bidirectional Text in Math Zones

Consider next the directionality of math text in math zones. The guiding principle is that neutrals other than the period and comma are given the directionality of the math zone. This includes math operators and spaces. Alphabetic spans of characters having the same directionality are displayed with that directionality relative to one another, but the span as a whole is treated as object with the directionality of the math zone. As with the Unicode Bidi Algorithm, spans of digits, possibly including embedded commas and periods, are displayed LTR, but analogously to keyboard language tailoring, the number they comprise has the directionality of the math zone.

Normal text embedded in a math zone obeys the directionality rules of normal text. Typically this is given by the Unicode Bidi Algorithm, but the keyboard language algorithm is used if the text is entered by keyboard.

Internationalized Resource Identifiers

Internationalized Resource Identifiers (IRIs) are a generalization of Universal Resource Identifiers (URIs) that can contain many nonASCII characters, such as most alphabetic characters and Chinese characters. Complications occur when bidi characters are used in IRIs especially when displayed in an RTL context. For example according to the Unicode Bidi Algorithm, http://شس.يب.ثق displays in a right-to-left paragraph as

//:httpشس.يب.ثق

Or more confusing yet, http://exchange.شس.ثق displays in an RTL paragraph as

http://exchange.شس.ثق

As the IRI reference discusses, use of the Unicode Bidi Algorithm at least is consistent in the way such IRIs are displayed in plain text, but in some sense the RTL versions are nearly unreadable. The IRI reference does recommend that IRIs always be displayed in LTR display order. But that requires being able to recognize an IRI in the first place.

So assuming an IRI can be recognized, something that RichEdit does on the fly, we can enforce a more readable display order. Namely we force the delimiters '#', '.', '/', ':', '?', '@', '[', ']' to follow the paragraph (or embedding) direction. With this condition, http://exchange.شس.ثق displays in an RTL paragraph as

http://exchange.ثق.شس

i.e., the alphanumeric spans in between the special delimiters appear in the reverse order from the way they appear in an LTR paragraph and the Unicode Bidi Algorithm is not used to resolve the neutrality of the slash and period. The alphanumeric spans themselves (in effect the “leaves” of the structure) are displayed in the order determined by the Unicode Bidi algorithm.

This approach appears to be ideal. The only problem is that it’s not trivial to identify IRIs using heuristics. Both Word and RichEdit can identify IRIs, but ambiguities can occur when spaces appear at the ends. Word allows the user to overrule the IRI recognition, but RichEdit 7.0 and earlier versions do not. The IRI and URI specifications require that spaces be “escaped”, that is, replaced by their ASCII hexadecimal code %20.

Additional Requirements for Bidi in HTML

The paper Additional Requirements for Bidi in HTML describes a number of situations in HTML when application of the Unicode Bidi Algorithm yields bizarre, or at least unwanted, results. These situations don’t cause problems in Word or RichEdit rich-text controls, since every format run (similar to <span> in HTML) is assigned directionality. But in HTML, a <span> with a particular directionality influences the way the text next to it is displayed unless that text also has an assigned directionality.

Even simple things can display in surprising ways. For example,

10 main st.

displays in an RTL context as

.main st 10

This happens because the number at the start of an RTL paragraph is classified as an RTL object in an RTL context (even though the digits are displayed LTR with respect to one another), and hence the number comes first (to the right). It’s followed by the LTR text “main st”, but the period is classified as RTL and appears last, namely to the left of “main st”.

Let’s look at a couple of examples in that paper that can be fixed by adding a new <span> attribute called bdi for “bidirectional isolate”. These examples use the notation that upper-case ASCII letters are strong RTL letters and lower-case ASCII letters are strong LTR. The first example consists of the string “PURPLE PIZZA – 3 reviews”, which is displayed LTR in logical order (RTL characters are displayed LTR). In an LTR context, one would like this string to be displayed as

AZZIP ELPRUP – 3 reviews

But because a number is given the directionality of the (in this case RTL) run preceding it, it displays as

3 – AZZIP ELPRUP reviews

The second example in logical order is

USE css (<span dir="ltr">position:relative</span>).

One wants this to display in an RTL paragraph as

.(position:relative) css ESU

But because the LTR span continues the LTR css, it gets displayed as

.(css (position:relative ESU

The proposed fix is to put the embedded text into a <span> with a bdi attribute, which is then “displayed as if it were surrounded with strong-directional characters of the last explicit embedding level within which it appears” (taken from Sec. 2.1 of the paper referenced above). For the “PURPLE PIZZA” case, the “ - 3” would then be treated as if it were preceded by a strong LTR character. For the “USE css” case, the parentheses would resolve to RTL and get mirrored, giving the desired result above. If you’re interested in these problems, please read the full article.

Implementation for Math Zones and Hyperlinks

RichEdit tailors the Unicode Bidi Algorithm in two main steps: 1) get a string of massaged plain text from a range, and 2) break that string into a set of substrings based on currently assigned directionalities. It’s in step 1) that the math, IRI and some other overrides are made. For math, all operators (not including comma and period) are replaced by the letter ‘a’ for LTR math zones and by alef (U+0627) for RTL math zones. Since only Arabic locales use RTL math zones, alef is a good choice to use as an RTL character. For automatically detected IRIs, the delimiters '#', '.', '/', ':', '?', '@', '[', ']' are similarly replaced by ‘a’ or alef according to the paragraph direction. In Hebrew IRIs, it’s probably a good idea to use the Hebrew alef (U+05D0) instead of the Arabic alef, but I haven’t tested to see if it matters. The IRI overrides are in an experimental stage.

In step 2) Uniscribe’s ScriptItemize() function analyses each substring, returning information that may cause the directionality of some characters to be overruled. When this process is completed, the range is passed to a finite state machine to assign embedding levels for use by the LineServices reverse object handler. This handler is responsible for supplying the relevant glyphs in the appropriate display order for the reverse object nesting level. It’s a little tricky. Actually it’s fair to say that it’s very tricky!