Docutils | Overview | About | Users | Reference | Developers

Problems With StructuredText

Author:

David Goodger

Contact:
docutils-develop@lists.sourceforge.net
Revision:
$Revision$
Date:
$Date$

There are several problems, unresolved issues, and areas of controversy within StructuredText (Classic and Next Generation). In order to resolve all these issues, this analysis brings all of the issues out into the open, enumerates all the alternatives, and proposes solutions to be incorporated into the reStructuredText specification.

Formal Specification

The description in the original StructuredText.py has been criticized for being vague. For practical purposes, "the code is the spec." Tony Ibbs has been working on deducing a detailed description from the documentation and code of StructuredTextNG. Edward Loper's STMinus is another attempt to formalize a spec.

For this kind of a project, the specification should always precede the code. Otherwise, the markup is a moving target which can never be adopted as a standard. Of course, a specification may be revised during lifetime of the code, but without a spec there is no visible control and thus no confidence.

Understanding and Extending the Code

The original StructuredText is a dense mass of sparsely commented code and inscrutable regular expressions. It was not designed to be extended and is very difficult to understand. StructuredTextNG has been designed to allow input (syntax) and output extensions, but its documentation (both internal [comments & docstrings], and external) is inadequate for the complexity of the code itself.

For reStructuredText to become truly useful, perhaps even part of Python's standard library, it must have clear, understandable documentation and implementation code. For the implementation of reStructuredText to be taken seriously, it must be a sterling example of the potential of docstrings; the implementation must practice what the specification preaches.

Section Structure via Indentation

Setext required that body text be indented by 2 spaces. The original StructuredText and StructuredTextNG require that section structure be indicated through indentation, as "inspired by Python". For certain structures with a very limited, local extent (such as lists, block quotes, and literal blocks), indentation naturally indicates structure or hierarchy. For sections (which may have a very large extent), structure via indentation is unnecessary, unnatural and ambiguous. Rather, the syntax of the section title itself should indicate that it is a section title.

The original StructuredText states that "A single-line paragraph whose immediately succeeding paragraphs are lower level is treated as a header." Requiring indentation in this way is:

Python's use of significant whitespace is a wonderful (if not original) innovation, however requiring indentation in ordinary written text is hypergeneralization.

reStructuredText indicates section structure through title adornment style (as exemplified by this document). This is far more natural. In fact, it is already in widespread use in plain text documents, including in Python's standard distribution (such as the toplevel README file).

Character Escaping Mechanism

No matter what characters are chosen for markup, some day someone will want to write documentation about that markup or using markup characters in a non-markup context. Therefore, any complete markup language must have an escaping or encoding mechanism. For a lightweight markup system, encoding mechanisms like SGML/XML's '*' are out. So an escaping mechanism is in. However, with carefully chosen markup, it should be necessary to use the escaping mechanism only infrequently.

reStructuredText needs an escaping mechanism: a way to treat markup-significant characters as the characters themselves. Currently there is no such mechanism (although ZWiki uses '!'). What are the candidates?

  1. ! (http://www.zope.org/DevHome/Members/jim/StructuredTextWiki/NGEscaping)

  2. \

  3. ~

  4. doubling of characters

The best choice for this is the backslash (\). It's "the single most popular escaping character in the world!", therefore familiar and unsurprising. Since characters only need to be escaped under special circumstances, which are typically those explaining technical programming issues, the use of the backslash is natural and understandable. Python docstrings can be raw (prefixed with an 'r', as in 'r""'), which would obviate the need for gratuitous doubling-up of backslashes.

(On 2001-03-29 on the Doc-SIG mailing list, GvR endorsed backslash escapes, saying, "'nuff said. Backslash it is." Although neither legally binding nor irrevocable nor any kind of guarantee of anything, it is a good sign.)

The rule would be: An unescaped backslash followed by any markup character escapes the character. The escaped character represents the character itself, and is prevented from playing a role in any markup interpretation. The backslash is removed from the output. A literal backslash is represented by an "escaped backslash," two backslashes in a row.

A carefully constructed set of recognition rules for inline markup will obviate the need for backslash-escapes in almost all cases; see Delimitation of Inline Markup below.

When an expression (requiring backslashes and other characters used for markup) becomes too complicated and therefore unreadable, a literal block may be used instead. Inside literal blocks, no markup is recognized, therefore backslashes (for the purpose of escaping markup) become unnecessary.

We could allow backslashes preceding non-markup characters to remain in the output. This would make describing regular expressions and other uses of backslashes easier. However, this would complicate the markup rules and would be confusing.

Blank Lines in Lists

Oft-requested in Doc-SIG (the earliest reference is dated 1996-08-13) is the ability to write lists without requiring blank lines between items. In docstrings, space is at a premium. Authors want to convey their API or usage information in as compact a form as possible. StructuredText requires blank lines between all body elements, including list items, even when boundaries are obvious from the markup itself.

In reStructuredText, blank lines are optional between list items. However, in order to eliminate ambiguity, a blank line is required before the first list item and after the last. Nested lists also require blank lines before the list start and after the list end.

Bullet List Markup

StructuredText includes 'o' as a bullet character. This is dangerous and counter to the language-independent nature of the markup. There are many languages in which 'o' is a word. For example, in Spanish:

Llamame a la casa
o al trabajo.

(Call me at home or at work.)

And in Japanese (when romanized):

Senshuu no doyoubi ni tegami
o kakimashita.

([I] wrote a letter on Saturday last week.)

If a paragraph containing an 'o' word wraps such that the 'o' is the first text on a line, or if a paragraph begins with such a word, it could be misinterpreted as a bullet list.

In reStructuredText, 'o' is not used as a bullet character. '-', '*', and '+' are the possible bullet characters.

Enumerated List Markup

StructuredText enumerated lists are allowed to begin with numbers and letters followed by a period or right-parenthesis, then whitespace. This has surprising consequences for writing styles. For example, this is recognized as an enumerated list item by StructuredText:

Mr. Creosote.

People will write enumerated lists in all different ways. It is folly to try to come up with the "perfect" format for an enumerated list, and limit the docstring parser's recognition to that one format only.

Rather, the parser should recognize a variety of enumerator styles. It is also recommended that the enumerator of the first list item be ordinal-1 ('1', 'A', 'a', 'I', or 'i'), as output formats may not be able to begin a list at an arbitrary enumeration.

An initial idea was to require two or more consistent enumerated list items in a row. This idea proved impractical and was dropped. In practice, the presence of a proper enumerator is enough to reliably recognize an enumerated list item; any ambiguities are reported by the parser. Here's the original idea for posterity:

The parser should recognize a variety of enumerator styles, mark each block as a potential enumerated list item (PELI), and interpret the enumerators of adjacent PELIs to decide whether they make up a consistent enumerated list.

If a PELI is labeled with a "1.", and is immediately followed by a PELI labeled with a "2.", we've got an enumerated list. Or "(A)" followed by "(B)". Or "i)" followed by "ii)", etc. The chances of accidentally recognizing two adjacent and consistently labeled PELIs, are acceptably small.

For an enumerated list to be recognized, the following must be true:

  • the list must consist of multiple adjacent list items (2 or more)

  • the enumerators must all have the same format

  • the enumerators must be sequential

Definition List Markup

StructuredText uses ' -- ' (whitespace, two hyphens, whitespace) on the first line of a paragraph to indicate a definition list item. The ' -- ' serves to separate the term (on the left) from the definition (on the right).

Many people use ' -- ' as an em-dash in their text, conflicting with the StructuredText usage. Although the Chicago Manual of Style says that spaces should not be used around an em-dash, Peter Funk pointed out that this is standard usage in German (according to the Duden, the official German reference), and possibly in other languages as well. The widespread use of ' -- ' precludes its use for definition lists; it would violate the "unsurprising" criterion.

A simpler, and at least equally visually distinctive construct (proposed by Guido van Rossum, who incidentally is a frequent user of ' -- ') would do just as well:

term 1
    Definition.

term 2
    Definition 2, paragraph 1.

    Definition 2, paragraph 2.

A reStructuredText definition list item consists of a term and a definition. A term is a simple one-line paragraph. A definition is a block indented relative to the term, and may contain multiple paragraphs and other body elements. No blank line precedes a definition (this distinguishes definition lists from block quotes).

Literal Blocks

The StructuredText specification has literal blocks indicated by 'example', 'examples', or '::' ending the preceding paragraph. STNG only recognizes '::'; 'example'/'examples' are not implemented. This is good; it fixes an unnecessary language dependency. The problem is what to do with the sometimes- unwanted '::'.

In reStructuredText '::' at the end of a paragraph indicates that subsequent indented blocks are treated as literal text. No further markup interpretation is done within literal blocks (not even backslash-escapes). If the '::' is preceded by whitespace, '::' is omitted from the output; if '::' was the sole content of a paragraph, the entire paragraph is removed (no 'empty' paragraph remains). If '::' is preceded by a non-whitespace character, '::' is replaced by ':' (i.e., the extra colon is removed).

Thus, a section could begin with a literal block as follows:

Section Title
-------------

::

    print "this is example literal"

Tables

The table markup scheme in classic StructuredText was horrible. Its omission from StructuredTextNG is welcome, and its markup will not be repeated here. However, tables themselves are useful in documentation. Alternatives:

  1. This format is the most natural and obvious. It was independently invented (no great feat of creation!), and later found to be the format supported by the Emacs table mode:

    +------------+------------+------------+--------------+
    |  Header 1  |  Header 2  |  Header 3  |  Header 4    |
    +============+============+============+==============+
    |  Column 1  |  Column 2  | Column 3 & 4 span (Row 1) |
    +------------+------------+------------+--------------+
    |    Column 1 & 2 span    |  Column 3  | - Column 4   |
    +------------+------------+------------+ - Row 2 & 3  |
    |      1     |      2     |      3     | - span       |
    +------------+------------+------------+--------------+

    Tables are described with a visual outline made up of the characters '-', '=', '|', and '+':

    • The hyphen ('-') is used for horizontal lines (row separators).

    • The equals sign ('=') is optionally used as a header separator (as of version 1.5.24, this is not supported by the Emacs table mode).

    • The vertical bar ('|') is used for for vertical lines (column separators).

    • The plus sign ('+') is used for intersections of horizontal and vertical lines.

    Row and column spans are possible simply by omitting the column or row separators, respectively. The header row separator must be complete; in other words, a header cell may not span into the table body. Each cell contains body elements, and may have multiple paragraphs, lists, etc. Initial spaces for a left margin are allowed; the first line of text in a cell determines its left margin.

  2. Below is a simpler table structure. It may be better suited to manual input than alternative #1, but there is no Emacs editing mode available. One disadvantage is that it resembles section titles; a one-column table would look exactly like section & subsection titles.

    ============ ============ ============ ==============
      Header 1     Header 2     Header 3     Header 4
    ============ ============ ============ ==============
      Column 1     Column 2    Column 3 & 4 span (Row 1)
    ------------ ------------ ---------------------------
        Column 1 & 2 span       Column 3    - Column 4
    ------------------------- ------------  - Row 2 & 3
          1            2            3       - span
    ============ ============ ============ ==============

    The table begins with a top border of equals signs with a space at each column boundary (regardless of spans). Each row is underlined. Internal row separators are underlines of '-', with spaces at column boundaries. The last of the optional head rows is underlined with '=', again with spaces at column boundaries. Column spans have no spaces in their underline. Row spans simply lack an underline at the row boundary. The bottom boundary of the table consists of '=' underlines. A blank line is required following a table.

  3. A minimalist alternative is as follows:

    ====  =====  ========  ========  =======  ====  =====  =====
    Old State    Input     Action             New State    Notes
    -----------  --------  -----------------  -----------
    ids   types  new type  sys.msg.  dupname  ids   types
    ====  =====  ========  ========  =======  ====  =====  =====
    --    --     explicit  --        --       new   True
    --    --     implicit  --        --       new   False
    None  False  explicit  --        --       new   True
    old   False  explicit  implicit  old      new   True
    None  True   explicit  explicit  new      None  True
    old   True   explicit  explicit  new,old  None  True   [1]
    None  False  implicit  implicit  new      None  False
    old   False  implicit  implicit  new,old  None  False
    None  True   implicit  implicit  new      None  True
    old   True   implicit  implicit  new      old   True
    ====  =====  ========  ========  =======  ====  =====  =====

    The table begins with a top border of equals signs with one or more spaces at each column boundary (regardless of spans). There must be at least two columns in the table (to differentiate it from section headers). Each line starts a new row. The rightmost column is unbounded; text may continue past the edge of the table. Each row/line must contain spaces at column boundaries, except for explicit column spans. Underlines of '-' can be used to indicate column spans, but should be used sparingly if at all. Lines containing column span underlines may not contain any other text. The last of the optional head rows is underlined with '=', again with spaces at column boundaries. The bottom boundary of the table consists of '=' underlines. A blank line is required following a table.

    This table sums up the features. Using all the features in such a small space is not pretty though:

    ========  ========  ========
              Header 2 & 3 Span
              ------------------
    Header 1  Header 2  Header 3
    ========  ========  ========
    Each      line is   a new row.
    Each row  consists  of one line only.
    Row       spans     are not possible.
    The last  column    may spill over to the right.
    Column spans are possible with an underline joining columns.
    ----------------------------
    The span  is        limited to the row above the underline.
    ========  ========  ========
  4. As a variation of alternative 3, bullet list syntax in the first column could be used to indicate row starts. Multi-line rows are possible, but row spans are not. For example:

    ===== =====
    col 1 col 2
    ===== =====
    - 1   Second column of row 1.
    - 2   Second column of row 2.
          Second line of paragraph.
    - 3   Second column of row 3.
    
          Second paragraph of row 3,
          column 2
    ===== =====

    Column spans would be indicated on the line after the last line of the row. To indicate a real bullet list within a first-column cell, simply nest the bullets.

  5. In a further variation, we could simply assume that whitespace in the first column implies a multi-line row; the text in other columns is continuation text. For example:

    ===== =====
    col 1 col 2
    ===== =====
    1     Second column of row 1.
    2     Second column of row 2.
          Second line of paragraph.
    3     Second column of row 3.
    
          Second paragraph of row 3,
          column 2
    ===== =====

    Limitations of this approach:

    • Cells in the first column are limited to one line of text.

    • Cells in the first column must contain some text; blank cells would lead to a misinterpretation. An empty comment ("..") is sufficient.

  6. Combining alternative 3 and 4, a bullet list in the first column could mean multi-line rows, and no bullet list means single-line rows only.

Alternatives 1 and 5 has been adopted by reStructuredText.

Delimitation of Inline Markup

StructuredText specifies that inline markup must begin with whitespace, precluding such constructs as parenthesized or quoted emphatic text:

"**What?**" she cried.  (*exit stage left*)

The reStructuredText markup specification allows for such constructs and disambiguates inline markup through a set of recognition rules. These recognition rules define the context of markup start-strings and end-strings, allowing markup characters to be used in most non-markup contexts without a problem (or a backslash). So we can say, "Use asterisks (*) around words or phrases to emphasisze them." The '(*)' will not be recognized as markup. This reduces the need for markup escaping to the point where an escape character is almost (but not quite!) unnecessary.

Underlining

StructuredText uses '_text_' to indicate underlining. To quote David Ascher in his 2000-01-21 Doc-SIG mailing list post, "Docstring grammar: a very revised proposal":

The tagging of underlined text with _'s is suboptimal. Underlines shouldn't be used from a typographic perspective (underlines were designed to be used in manuscripts to communicate to the typesetter that the text should be italicized -- no well-typeset book ever uses underlines), and conflict with double-underscored Python variable names (__init__ and the like), which would get truncated and underlined when that effect is not desired. Note that while complete markup would prevent that truncation ('__init__'), I think of docstring markups much like I think of type annotations -- they should be optional and above all do no harm. In this case the underline markup does harm.

Underlining is not part of the reStructuredText specification.

Inline Literals

StructuredText's markup for inline literals (text left as-is, verbatim, usually in a monospaced font; as in HTML <TT>) is single quotes ('literals'). The problem with single quotes is that they are too often used for other purposes:

Alternatives:

'text'    \'text\'    ''text''    "text"    \"text\"    ""text""
#text#     @text@      `text`     ^text^    ``text''    ``text``

The examples below contain inline literals, quoted text, and apostrophes. Each example should evaluate to the following HTML:

Some <TT>code</TT>, with a 'quote', "double", ain't it grand?
Does <TT>a[b] = 'c' + "d" + `2^3`</TT> work?

0. Some code, with a quote, double, ain't it grand?
   Does a[b] = 'c' + "d" + `2^3` work?
1. Some 'code', with a \'quote\', "double", ain\'t it grand?
   Does 'a[b] = \'c\' + "d" + `2^3`' work?
2. Some \'code\', with a 'quote', "double", ain't it grand?
   Does \'a[b] = 'c' + "d" + `2^3`\' work?
3. Some ''code'', with a 'quote', "double", ain't it grand?
   Does ''a[b] = 'c' + "d" + `2^3`'' work?
4. Some "code", with a 'quote', \"double\", ain't it grand?
   Does "a[b] = 'c' + "d" + `2^3`" work?
5. Some \"code\", with a 'quote', "double", ain't it grand?
   Does \"a[b] = 'c' + "d" + `2^3`\" work?
6. Some ""code"", with a 'quote', "double", ain't it grand?
   Does ""a[b] = 'c' + "d" + `2^3`"" work?
7. Some #code#, with a 'quote', "double", ain't it grand?
   Does #a[b] = 'c' + "d" + `2^3`# work?
8. Some @code@, with a 'quote', "double", ain't it grand?
   Does @a[b] = 'c' + "d" + `2^3`@ work?
9. Some `code`, with a 'quote', "double", ain't it grand?
   Does `a[b] = 'c' + "d" + \`2^3\`` work?
10. Some ^code^, with a 'quote', "double", ain't it grand?
    Does ^a[b] = 'c' + "d" + `2\^3`^ work?
11. Some ``code'', with a 'quote', "double", ain't it grand?
    Does ``a[b] = 'c' + "d" + `2^3`'' work?
12. Some ``code``, with a 'quote', "double", ain't it grand?
    Does ``a[b] = 'c' + "d" + `2^3\``` work?

Backquotes (#9 & #12) are the best choice. They are unobtrusive and relatviely rarely used (more rarely than ' or ", anyhow). Backquotes have the connotation of 'quotes', which other options (like carets, #10) don't.

Analogously with *emph* & **strong**, double-backquotes (#12) could be used for inline literals. If single-backquotes are used for 'interpreted text' (context-sensitive domain-specific descriptive markup) such as function name hyperlinks in Python docstrings, then double-backquotes could be used for absolute-literals, wherein no processing whatsoever takes place. An advantage of double-backquotes would be that backslash-escaping would no longer be necessary for embedded single-backquotes; however, embedded double-backquotes (in an end-string context) would be illegal. See Backquotes in Phrase-Links in Record of reStructuredText Syntax Alternatives.

Alternative choices are carets (#10) and TeX-style quotes (#11). For examples of TeX-style quoting, see http://www.zope.org/Members/jim/StructuredTextWiki/CustomizingTheDocumentProcessor.

Some existing uses of backquotes:

  1. As a synonym for repr() in Python.

  2. For command-interpolation in shell scripts.

  3. Used as open-quotes in TeX code (and carried over into plaintext by TeXies).

The inline markup start-string and end-string recognition rules defined by the reStructuredText markup specification would allow all of these cases inside inline literals, with very few exceptions. As a fallback, literal blocks could handle all cases.

Outside of inline literals, the above uses of backquotes would require backslash-escaping. However, these are all prime examples of text that should be marked up with inline literals.

If either backquotes or straight single-quotes are used as markup, TeX-quotes are too troublesome to support, so no special-casing of TeX-quotes should be done (at least at first). If TeX-quotes have to be used outside of literals, a single backslash-escaped would suffice: ``TeX quote''. Ugly, true, but very infrequently used.

Using literal blocks is a fallback option which removes the need for backslash-escaping:

like this::

    Here, we can do ``absolutely'' anything `'`'\|/|\ we like!

No mechanism for inline literals is perfect, just as no escaping mechanism is perfect. No matter what we use, complicated inline expressions involving the inline literal quote and/or the backslash will end up looking ugly. We can only choose the least often ugly option.

reStructuredText will use double backquotes for inline literals, and single backqoutes for interpreted text.