Awesome
wikitextprocessor
This is a Python package for processing WikiMedia dump files for Wiktionary, Wikipedia, etc., for data extraction, error checking, offline conversion into HTML or other formats, and other uses. Key features include:
- Parsing dump files, including built-in support for processing pages in parallel
- Wikitext syntax parser that converts the whole page into a parse tree
- Extracting template definitions and Scribunto Lua module definitions from dump files
- Expanding selected templates or all templates, and heuristically identifying templates that need to be expanded before parsing is reasonably possible (e.g., templates that emit table start and end tags)
- Processing and expanding wikitext parser functions
- Processing, executing, and expanding Scribunto Lua modules (they are very widely used in, e.g., Wiktionary, for example for generating IPA strings for many languages)
- Controlled expansion of parts of pages for applications that parse overall page structure before parsing but then expand templates on certain sections of the page
- Capturing information from template arguments while expanding them, as template arguments often contain useful information not available in the expanded content.
This module is primarily intended as a building block for other packages that process Wikitionary or Wikipedia data, particularly for data extraction. You will need to write code to use this.
For pre-existing extraction modules that use this package, please see:
- Wiktextract for extracting rich machine-readable dictionaries from Wiktionary. You can also find pre-extracted machine-readable Wiktionary data in JSON format at kaikki.org.
Getting started
Installing
Install from source:
git clone --recurse-submodules --shallow-submodules https://github.com/tatuylonen/wikitextprocessor.git
cd wikitextprocessor
python -m venv .venv
source .venv/bin/activate
python -m pip install -U pip
python -m pip install -e .
Running tests
This package includes tests written using the unittest
framework.
The test dependencies can be installed with command
python -m pip install -e .[dev]
.
To run the tests, use the following command in the top-level directory:
make test
To run a specific test, use the following syntax:
python -m unittest tests.test_[module].[Module]Tests.test_[name]
Python's unittest framework help and options can be accessed through:
python -m unittest -h
Obtaining WikiMedia dump files
This package is primarily intended for processing Wiktionary and Wikipedia dump files (though you can also use it for processing individual pages or other files that are in wikitext format). To download WikiMedia dump files, go to the dump download page. We recommend using the <name>-<date>-pages-articles.xml.bz2 files.
API documentation
Usage example:
from functools import partial
from typing import Any
from wikitextprocessor import Wtp, WikiNode, NodeKind, Page
from wikitextprocessor.dumpparser import process_dump
def page_handler(wtp: Wtp, page: Page) -> Any:
wtp.start_page(page.title)
# process parse tree
tree = wtp.parse(page.body)
# or get expanded plain text
text = wtp.expand(page.body)
wtp = Wtp(
db_path="en_20230801.db", lang_code="en", project="wiktionary"
)
# extract dump file then save pages to SQLite file
process_dump(
wtp,
"enwiktionary-20230801-pages-articles.xml.bz2",
{0, 10, 110, 828}, # namespace id, can be found at the start of dump file
)
for _ in map(
partial(page_handler, wtp), wtp.get_all_pages([0])
):
pass
The basic operation is as follows:
- Extract templates, modules, and other pages from the dump file and save them in a SQLite file
- Heuristically analyze which templates need to be pre-expanded before parsing to make sense of the page structure (this cannot detect templates that call Lua code that outputs wikitext that affects parsed structure). These first steps together are called the "first phase".
- Process the pages again, calling a page handler function for each page. The page handler can extract, parse, and otherwise process the page, and has full access to templates and Lua macros defined in the dump. This may call the page handler in multiple processes in parallel. Return values from the page handler calls are returned to the caller. This is called the second phase.
Most of the functionality is hidden behind the Wtp
object.
WikiNode
objects are used for representing the parse
tree that is returned by the Wtp.parse()
function. NodeKind
is an enumeration type used to encode the type of a WikiNode
.
class Wtp
def __init__(
self,
db_path: Optional[Union[str, Path]] = None,
lang_code="en",
template_override_funcs: Dict[str, Callable[[Sequence[str]], str]] = {},
project: str = "wiktionary",
):
The initializer can usually be called without arguments, but recognizes the following arguments:
db_path
can beNone
, in which case a temporary database file will be created under/tmp
, or a path for the database file which contains page texts and other data of the dump file. There are two reasons why you might want to set this:- you don't have enough space on
/tmp
(3.4G for English dump file), or 2) for testing. If you specify the path and an existing database file exists, that file will be used, eliminating the time needed for Phase 1 (this is very important for testing, allowing processing single pages reasonably fast). In this case, you should not callWtp.process()
but instead useWtp.reprocess()
or just callWtp.expand()
orWtp.parse()
on wikitext that you have obtained otherwise (e.g., from some file). If the file doesn't exist, you will need to callWtp.process()
to parse a dump file, which will initialize the database file during the first phase. If you wish to re-create the database, you should remove the old file first.
- you don't have enough space on
lang_code
- the language code of the dump file.template_override_funcs
- Python functions for overriding expanded template text.project
- "wiktionary" or "wikipedia".
def read_by_title(
self, title: str, namespace_id: Optional[int] = None
) -> Optional[str]:
Reads the contents of the page with the specified title from the cache
file. There is usually no need to call this function explicitly, as
Wtp.process()
and Wtp.reprocess()
normally load the page
automatically. This function does not automatically call Wtp.start_page()
.
Arguments are:
title
- the title of the page to readnamespace_id
- namespace id number, this argument is required iftitle
donesn't have namespace prefix likeTemplate:
.
This returns the page contents as a string, or None
if the page
does not exist.
def parse(
self,
text: str,
pre_expand=False,
expand_all=False,
additional_expand=None,
do_not_pre_expand=None,
template_fn=None,
post_template_fn=None,
) -> WikiNode:
Parses wikitext into a parse tree (WikiNode
), optionally expanding
some or all the templates and Lua macros in the wikitext (using the definitions
for the templates and macros in the cache files, as added by Wtp.process()
or calls to Wtp.add_page()
.
The Wtp.start_page()
function must be called before this function
to set the page title (which may be used by templates, Lua macros, and
error messages). The Wtp.process()
and Wtp.reprocess()
functions will call it automatically.
This accepts the following arguments:
text
(str) - the wikitext to be parsedpre_expand
(boolean) - if set toTrue
, the templates that were heuristically detected as affecting parsing (e.g., expanding to table start or end tags or list items) will be automatically expanded before parsing. Any Lua macros those templates use may also be called.expand_all
- if set toTrue
, expands all templates and Lua macros in the wikitext before parsing.additional_expand
(set orNone
) - if this argument is provided, it should be a set of template names that should be expanded in addition to those specified by the other options (i.e., in addition to to the heuristically detected templates ifpre_expand
isTrue
or just these if it is false; this option is meaningless ifexpand_all
is set toTrue
).
This returns the parse tree. See below for a documentation of the WikiNode
class used for representing the parse tree.
def node_to_wikitext(self, node)
Converts a part of a parse tree back to wikitext.
node
(WikiNode
, str, list/tuple of these) - This is the part of the parse tree that is to be converted back to wikitext. We also allow strings and lists, so thatnode.children
can be used directly as the argument.
def expand(self, text, template_fn=None, post_template_fn=None,
pre_expand=False, templates_to_expand=None,
expand_parserfns=True, expand_invoke=True)
Expands the selected templates, parser functions and Lua macros in the given Wikitext. This can selectively expand some or all templates. This can also capture the arguments and/or the expansion of any template as well as substitute custom expansions instead of the default expansions.
The Wtp.start_page()
function must be called before this function to
set the page title (which may be used by templates and Lua macros). The
Wtp.process()
and Wtp.reprocess()
will call it automatically. The
page title is also used in error messages.
The arguments are as follows:
text
(str) - the wikitext to be expandedtemplate_fn
(function) - if set, this will be called astemplate_fn(name, args)
, wherename
(str) is the name of the template andargs
is a dictionary containing arguments to the template. Positional arguments (and named arguments with numeric names) will have integer keys in the dictionary, whereas other named arguments will have their names as keys. All values corresponding to arguments are strings (after they have been expanded). This function can returnNone
to cause the template to be expanded in the normal way, or a string that will be used instead of the expansion of the template. This can return""
(empty string) to expand the template to nothing. This can also capture the template name and its arguments.post_template_fn
(function) - if set, this will be called aspost_template_fn(name, ht, expansion)
after the template has been expanded in the normal way. This can returnNone
to use the default expansion, or a string to use a that string as the expansion. This can also be used to capture the template, its arguments, and/or its expansion.pre_expand
(boolean) - if set toTrue
, all templates that were heuristically determined as needing to be expanded before parsing will be expanded.templates_to_expand
(None
or set or dictionary) - if this is set, these templates will be expanded in addition to any other templates that have been specified to be expanded. If a dictionary is provided, its keys will be taken as the names of the templates to be expanded. If this has not been set or isNone
, all templates will be expanded.expand_parserfns
(boolean) - Normally, wikitext parser functions will be expanded. This can be set toFalse
to prevent parser function expansion.expand_invoke
(boolean) - Normally, the#invoke
parser function (which calls a Lua module) will be expanded along with other parser functions. This can be set toFalse
to prevent expansion of the#invoke
parser function.
def start_page(self, title)
This function should be called before starting the processing of a new page or file. This saves the page title (which is frequently accessed by templates, parser functions, and Lua macros). The page title is also used in error messages.
The Wtp.process()
and Wtp.reprocess()
functions will automatically
call this before calling the page handler for each page. This needs to be
called manually when processing wikitext obtained from other sources.
The arguments are as follows:
title
(str) - The page title. For normal pages, there is usually no prefix. Templates typically haveTemplate:
prefix and Lua modulesModule:
prefix, and other prefixes are also used (e.g.,Thesaurus:
). This does not care about the form of the name, but some parser functions do.
def start_section(self, title)
Sets the title of the current section on the page. This is
automatically reset to None
by Wtp.start_page()
. The section
title is only used in error, warning, and debug messages.
The arguments are:
title
(str) - the title of the section, orNone
to clear it.
def start_subsection(self, title)
Sets the title of the current subsection of the current section on the
page. This is automatically reset to None
by Wtp.start_page()
and Wtp.start_section()
. The subsection title is only used in error,
warning, and debug messages.
The arguments are:
title
(str) - the title of the subsection, orNone
to clear it.
def add_page(self, title: str, namespace_id: int, body: Optional[str] = None,
redirect_to: Optional[str] = None, need_pre_expand: bool = False,
model: str = "wikitext") -> None:
This function is used to add pages, templates, and modules for
processing. There is usually no need to use this if Wtp.process()
is used; however, this can be used to add templates and pages for
testing or other special processing needs.
The arguments are:
title
- the title of the page to be added (normal pages typically have no prefix in the title, templates begin withTemplate:
, and Lua modules begin withModule:
)namespace_id
- namespace idbody
- the content of the page, template, or moduleredirect_to
- title of redirect pageneed_pre_expand
- set toTrue
if the page is a template that need to be expanded before parsing.model
- the model value for the page (usuallywikitext
for normal pages and templates andScribunto
for Lua modules)
The Wtp.analyze_templates()
function needs to be called after
calling Wtp.add_page()
before pages can be expanded or parsed (it should
preferably only be called once after adding all pages and templates).
def analyze_templates(self)
Analyzes the template definitions in the cache file and determines which
of them should be pre-expanded before parsing because they affect the
document structure significantly. Some templates in, e.g., Wiktionary
expand to table start tags, table end tags, or list items, and parsing
results are generally much better if they are expanded before parsing.
The actual expansion only happens if pre_expand
or some other argument
to Wtp.expand()
or Wtp.parse()
tells them to do so.
The analysis is heuristic and is not guaranteed to find every such template. In particular, it cannot detect templates that call Lua modules that output Wikitext control structures (there are several templates in Wiktionary that call Lua code that outputs list items, for example). Such templates may need to be identified manually and specified as additional templates to expand. Luckily, there seem to be relatively few such templates, at least in Wiktionary.
This function is automatically called by Wtp.process()
at the end of
phase 1. An explicit call is only necessary if Wtp.add_page()
has been
used by the application.
Error handling
Various functions in this module, including Wtp.parse()
and
Wtp.expand()
may generate errors and warnings. Those will be displayed
on stdout
as well as collected in Wtp.errors
, Wtp.warnings
, and
Wtp.debugs
. These fields will contain lists of dictionaries, where
each dictionary describes an error/warning/debug message. The dictionary can
have the following keys (not all of them are always present):
msg
(str) - the error messagetrace
(str orNone
) - optional stacktrace where the error occurredtitle
(str) - the page title on which the error occurredsection
(str orNone
) - the section where the error occurredsubsection
(str orNone
) - the subsection where the error occurredpath
(tuple of str) - a path of title, template names, parser function names, or Lua module/function names, giving information about where the error occurred during expansion or parsing.
The fields containing the error messages will be cleared by every call
to Wtp.start_page()
(including the implicit calls during
Wtp.process()
and Wtp.reprocess()
). Thus, the
page_handler
function often returns these lists together with any
information extracted from the page, and they can be collected
together from the values returned by the iterators returned by these
functions. The Wtp.to_return()
function maybe useful for this.
The following functions can be used for reporting errors. These can
also be called by application code from within the page_handler
function as well as template_fn
and post_template_fn
functions
to report errors, warnings, and debug messages in a uniform way.
def error(self, msg, trace=None)
Reports an error message. The error will be added to Wtp.errors
list and
printed to stdout. The arguments are:
- msg (str) - the error message (need not include page title or section)
- trace (str or
None
) - an optional stack trace giving more information about where the error occurred
def warning(self, msg, trace=None)
Reports a warning message. The warning will be added to Wtp.warnings
list
and printed to stdout. The arguments are the same as for Wtp.error()
.
def debug(self, msg, trace=None)
Reports a debug message. The message will be added to Wtp.debugs
list
and printed to stdout. The arguments are the same as for Wtp.error()
.
def to_return(self)
Produces a dictionary containing the error, warning, and debug
messages from Wtp
. This would typically be called at the end of a
page_handler
function and the value returned along with whatever
data was extracted from that page. The error lists are reset by
Wtp.start_page()
(including the implicit calls from
Wtp.process()
and Wtp.reprocess()
), so they should be saved
(e.g., by this call) for each page. (Given the parallelism in
the processing of the pages, they cannot just be accumulated in the
subprocesses.)
The returned dictionary contains the following keys:
errors
- a list of dictionaries describing any error messageswarnings
- a list of dictionaries describing any warning messagesdebugs
- a list of dictionaries describing any debug messages.
class WikiNode
The WikiNode
class represents a parse tree node and is returned by
Wtp.parse()
. This object can be printed or converted to a string
and will display a human-readable format that is suitable for
debugging purposes (at least for small parse trees).
The WikiNode
objects have the following fields:
kind
(NodeKind, see below) - The type of the node. This determines how to interpret the other fields.children
(list) - Contents of the node. This is generally used when the node has arbitrary size content, such as subsections, list items/sublists, other HTML tags, etc.args
(list or str, depending onkind
) - Direct arguments to the node. This is used, for example, for templates, template arguments, parser function arguments, and link arguments, in which case this is a list. For some node types (e.g., list, list item, and HTML tag), this is directly a string.attrs
- A dictionary containing HTML attributes or a definition list definition (under thedef
key).
class NodeKind(enum.Enum)
The NodeKind
type is an enumerated value for parse tree (WikiNode
)
node types. Currently the following values are used (typically these
need to be prefixed by Nodekind.
, e.g., NodeKind.LEVEL2
):
ROOT
- The root node of the parse tree.LEVEL2
- Level 2 subtitle (==). Theargs
field contains the title andchildren
field contains any contents that are within this sectionLEVEL3
- Level 3 subtitle (===)LEVEL4
- Level 4 subtitle (====)LEVEL5
- Level 5 subtitle (=====)LEVEL6
- Level 6 subtitle (======)ITALIC
- Italic, content is inchildren
BOLD
- Bold, content is inchildren
HLINE
- A horizontal line (no arguments or children)LIST
- Indicates a list. Each list and sublist will start with this kind of node.args
will contain the prefix used to open the list (e.g.,"##"
- note this is stored directly as a string inargs
). List items will be stored inchildren
.LIST_ITEM
- A list item in the children of aLIST
node.args
is the prefix used to open the list item (same as for theLIST
node). The contents of the list item (including any possible sublists) are inchildren
. If the list is a definition list (i.e., the prefix ends in";"
), thenchildren
contains the item label to be defined anddefinition
contains the definition.PREFORMATTED
- Preformatted text where markup is interpreted. Content is inchildren
. This is used for lines starting with a space in wikitext.PRE
- Preformatted text where markup is not interpreted. Content is inchildren
. This is indicated in wikitext by <pre>...</pre>.LINK
- An internal wikimedia link ([[...]] in wikitext). The link arguments are inargs
. This tag is also used for media inclusion. Links with a trailing word end immediately after the link have the trailing part inchildren
.TEMPLATE
- A template call (transclusion). Template name is in the first argument and template arguments in subsequent arguments inargs
. Thechildren
field is not used. In wikitext templates are marked up as {{name|arg1|arg2|...}}.TEMPLATE_ARG
- A template argument. The argument name is in the first item inargs
followed by any subsequet arguments (normally at most two items, but I've seen arguments with more - probably an error in those template definitions). Thechildren
field is not used. In wikitext template arguments are marked up as {{{name|defval}}}.PARSER_FN
- A parser function invocation. This is also used for built-in variables such as {{PAGENAME}}. The parser function name is in the first element ofargs
and parser function arguments in subsequent elements.URL
- An external URL. The first argument is the URL. The second optional argument (inargs
) is the display text. Thechildren
field is not used.TABLE
- A table. Content is inchildren
. In wikitext, a table is encoded as {| ... |}.TABLE_CAPTION
- A table caption. This can only occur underTABLE
. The content is inchildren
. Theattrs
field contains a dictionary of any HTML attributes given to the table.TABLE_ROW
- A table row. This can only occur underTABLE
. The content is inchildren
(normally the content would beTABLE_CELL
orTABLE_HEADER_CELL
nodes). Theattrs
field contains a dictionary of any HTML attributes given to the table row.TABLE_HEADER_CELL
- A table header cell. This can only occur underTABLE_ROW
. Content is in children. Theattrs
field contains a dictionary of any HTML attributes given to the table row.TABLE_CELL
- A table cell. This can only occur underTABLE_ROW
. Content is inchildren
. Theattrs
field contains a dictionary of any HTML attributes given to the table row.MAGIC_WORD
- A MediaWiki magic word. The magic word is assigned directly toargs
as a string (i.e., not in a list).children
is not used. An example of a magic word would be__NOTOC__
.HTML
- A HTML tag (or a matched pair of HTML tags).args
is the name of the HTML tag directly (not in a list and always without a slash).attrs
is set to a dictionary of any HTML attributes from the tag. The contents of the HTML tag is inchildren
.
Expected performance
This can generally process a few Wiktionary pages per second per processor core, including expansion of all templates, Lua macros, parsing the full page, and analyzing the parse. On a multi-core machine, this can generally process a few dozen to a few hundred pages per second, depending on the speed and the number of the cores.
Most of the processing effort goes to expanding Lua macros. You can elect not to expand Lua macros, but they are used extensively in Wiktionary and for important information. Expanding templates and Lua macros allows much more robust and complete data extraction, but does not come cheap.
Contributing and bug reports
Please create an issue on github to report bugs or to contribute!