wikitextprocessor

Parser and expander for Wikipedia, Wiktionary etc. dump files, with Lua execution support


Keywords
dictionary, wiktionary, wikipedia, data, extraction, wikitext, scribunto, lua, mediawiki, scribuntu
License
Unicode-DFS-2016
Install
pip install wikitextprocessor==0.4.96

Documentation

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

The best way to install this package is from pypi:

pip3 install wikitextprocessor

Alternatively, you can install the master branch from github:

git clone https://github.com/tatuylonen/wikitextprocessor
cd wikitextprocessor
pip3 install -e .

Running tests

This package includes tests written using the unittest framework. They can be run using, for example, nose, which can be installed using pip3 install nose.

To run the tests, use the following command in the top-level directory:

nosetests

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 wikitextprocessor import Wtp, WikiNode, NodeKind
   ctx = Wtp()

   def page_handler(model, title, text):
       if model != "wikitext" or title.startswith("Template:"):
           return None
       tree = ctx.parse(text, pre_expand=True)
       ... process parse tree
         ... value = ctx.node_to_wikitext(node)

   ctx.process("enwiktionary-20201201-pages-articles.xml.bz2", page_handler)

The basic operation of Wtp.process() is as follows:

  • Extract templates, modules, and other pages from the dump file and save them in a temporary 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 function acts as an iterator). This is called the second phase.
  • Optionally, the Wtp.reprocess() function may be used for processing the same data several times (it basically repeats 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. Additionally, ALL_LANGUAGES is exported and is a list that describes all languages (language codes, names, and other data) used in Wiktionary.

class Wtp(object)

def __init__(self, num_threads=None, cache_file=None, quiet=False)

The initializer can usually be called without arguments, but recognizes the following arguments:

  • num_threads - if set to an integer, use that many parallel processes for processing the dump. The default is to use as many processors as there are available cores/hyperthreads. You may need to limit the number of parallel processes if you are limited by available memory; we have found that processing Wiktionary (including templates and Lua macros) requires 3-4GB of memory per process. This MUST be set to 1 on Windows.
  • cache_file can normally be None, in which case a temporary file will be created under /tmp, or a path (str) for the cache file(s). There are two reasons why you might want to set this: 1) you don't have enough space on /tmp (the whole uncompressed dump must fit there, which can easily be 10-20GB), or 2) for testing. If you specify the cache file, if an existing cache file exists, that will be loaded and 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 call Wtp.process() but instead use Wtp.reprocess() or just call Wtp.expand() or Wtp.parse() on wikitext that you have obtained otherwise (e.g., from some file). If the cache file doesn't exist, you will need to call Wtp.process() to parse a dump file, which will initialize the cache file during the first phase. If you wish to re-create cache file, you should remove the old cache file first. The cache file path is actually a prefix for multiple individual files.
  • quiet - if set to True, suppress progress messages during processing

Windows and MacOS note: Setting num_threads to a value other than 1 probably doesn't currently work on Windows and MacOS. It now defaults to 1 on these platforms. This is because these platforms don't use fork() in the Python multiprocessing package, and the current parallelization implementation depends on this.

def process(self, path, page_handler, phase1_only=False)

This function processes a WikiMedia dump, uncompressing and extracing pages (including templates and Lua modules), and calling Wtp.add_page() for each page (phase 1). This then calls Wtp.reprocess() to execute the second phase.

This takes the following arguments:

  • path (str) - path to the WikiMedia dump file to be processed (e.g., "enwiktionary-20201201-pages-articles.xml.bz2"). Note that the compressed file can be used. Dump files can be downloaded here.
  • page_handler (function) - this function will be called for each page in phase 2 (unless phase1_only is set to True). The call takes the form page_handler(model, title, data), where model is the model value for the page in the dump (wikitext for normal wikitext pages and templates, Scribunto for Lua modules; other values are also possible), title is page title (e.g., sample or Template:foobar or Module:mystic), and data is the contents of the page (usually wikitext).
  • phase1_only (boolean) - if set to True, prevents phase 2 processing and the page_handler function will not be called. The Wtp.reprocess() function can be used to run the second phase separately, or Wtp.expand(), Wtp.parse() and other functions can be used.

This function returns an iterator over the values returned by the page_handler function (if page_handler returns None or no value, the iterator does not return those values). Note that page_handler will usually be run in a separate process, and cannot pass any values back in global variables. It can, however, access global variables assigned before calling Wtp.process() (in Linux only).

def reprocess(self, page_handler, autoload=True)

Iterates over all pages in the cache file and calls page_handler for each page. This basically implements phase 2 of processing a dump file (see Wtp.process()). This can be called more than once if desired.

The arguments are:

  • page_handler (function) - this function will be called for every page. The call takes the form page_handler(model, title, data), where model is the model value for the page in the dump (wikitext for normal wikitext pages and templates, Scribunto for Lua modules; other values are also possible), title is page title (e.g., sample or Template:foobar or Module:mystic), and data is the contents of the page (usually wikitext). If autoload is set to False, then data will be None.
  • autoload (boolean) - Normally this function loads the page contents automatically before calling the page handler. If autoload is set to False, then this will not automatically load the contents. This will make the iteration over the pages much faster, and is useful for scanning all pages when only a small fraction of them are likely to be of interest. The Wtp.read_by_title() function can then be used to load the page contents.

This function returns an iterator that iterates over the return values of page_handler. If the return value is None or page_handler returns no value, no value is returned by the iterator for such calls.

This calls the page_handler using subprocesses (unless num_threads was set to 1 in the initializer). It may be necessary to set it to 1 on Windows and MacOS due to operating system/python limitations on those platforms.

def read_by_title(self, title):

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. However, this can be useful if calling Wtp.reprocess() with autoload set to False. This function does not automatically call Wtp.start_page().

Arguments are:

  • title (str) - the title of the page to read

This returns the page contents as a string, or None if the page does not exist. If a transient page has been added with that title (see Wtp.add_page()), then this returns the transient page.

def parse(self, text, pre_expand=False, expand_all=False,
          additional_expand=None)

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 parsed
  • pre_expand (boolean) - if set to True, 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 to True, expands all templates and Lua macros in the wikitext before parsing.
  • additional_expand (set or None) - 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 if pre_expand is True or just these if it is false; this option is meaningless if expand_all is set to True).

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 that node.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 expanded
  • template_fn (function) - if set, this will be called as template_fn(name, args), where name (str) is the name of the template and args 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 return None 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 as post_template_fn(name, ht, expansion) after the template has been expanded in the normal way. This can return None 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 to True, 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 is None, all templates will be expanded.
  • expand_parserfns (boolean) - Normally, wikitext parser functions will be expanded. This can be set to False 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 to False 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 have Template: prefix and Lua modules Module: 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, or None to clear it.
def start_subsection(self, title)

Sets the title of the current subsection of the current section on the page. This is autimatically 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, or None to clear it.
def add_page(self, model, title, text, transient=False)

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. This can also be used for adding transient pages that are not stored in the cache file but could be used for debugging or information extraction. An example would be overriding some Lua module from the dump file for debugging the Lua code, or adding a new Lua module that can import other modules and then dump data from them, for example to extract the category hierarchy of Wiktionary pages.

The arguments are:

  • model (str) - the model value for the page (usually wikitext for normal pages and templates and Scribunto for Lua modules)
  • title (str) - the title of the page to be added (normal pages typically have no prefix in the title, templates begin with Template:, and Lua modules begin with Module:)
  • text (str) - the content of the page, template, or module
  • transient (boolean) - normally, the added pages will be stored in the cache file. If this is set to True, the page will not be stored in the cache file and will only affect the current run. Transient pages will override any pages in the cache file, and are useful for debugging and data extraction.

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 message
  • trace (str or None) - optional stacktrace where the error occurred
  • title (str) - the page title on which the error occurred
  • section (str or None) - the section where the error occurred
  • subsection (str or None) - the subsection where the error occurred
  • path (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 messages
  • warnings - a list of dictionaries describing any warning messages
  • debugs - a list of dictionaries describing any debug messages.

class WikiNode(object)

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 on kind) - 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 the def 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 (==). The args field contains the title and children field contains any contents that are within this section
  • LEVEL3 - Level 3 subtitle (===)
  • LEVEL4 - Level 4 subtitle (====)
  • LEVEL5 - Level 5 subtitle (=====)
  • LEVEL6 - Level 6 subtitle (======)
  • ITALIC - Italic, content is in children
  • BOLD - Bold, content is in children
  • 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 in args). List items will be stored in children.
  • LIST_ITEM - A list item in the children of a LIST node. args is the prefix used to open the list item (same as for the LIST node). The contents of the list item (including any possible sublists) are in children. If the list is a definition list (i.e., the prefix ends in ";"), then children contains the item label to be defined and attrs["def"] contains the definition.
  • PREFORMATTED - Preformatted text where markup is interpreted. Content is in children. This is used for lines starting with a space in wikitext.
  • PRE - Preformatted text where markup is not interpreted. Content is in children. This is indicated in wikitext by <pre>...</pre>.
  • LINK - An internal wikimedia link ([[...]] in wikitext). The link arguments are in args. This tag is also used for media inclusion. Links with a trailing word end immediately after the link have the trailing part in children.
  • TEMPLATE - A template call (transclusion). Template name is in the first argument and template arguments in subsequent arguments in args. The children 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 in args followed by any subsequet arguments (normally at most two items, but I've seen arguments with more - probably an error in those template definitions). The children 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 of args and parser function arguments in subsequent elements.
  • URL - An external URL. The first argument is the URL. The second optional argument (in args) is the display text. The children field is not used.
  • TABLE - A table. Content is in children. In wikitext, a table is encoded as {| ... |}.
  • TABLE_CAPTION - A table caption. This can only occur under TABLE. The content is in children. The attrs field contains a dictionary of any HTML attributes given to the table.
  • TABLE_ROW - A table row. This can only occur under TABLE. The content is in children (normally the content would be TABLE_CELL or TABLE_HEADER_CELL nodes). The attrs field contains a dictionary of any HTML attributes given to the table row.
  • TABLE_HEADER_CELL - A table header cell. This can only occur under TABLE_ROW. Content is in children. The attrs field contains a dictionary of any HTML attributes given to the table row.
  • TABLE_CELL - A table cell. This can only occur under TABLE_ROW. Content is in children. The attrs 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 to args 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 in children.

Expected performance

This can generally process a few Wiktionary pages 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!