Parsing XML and HTML with lxml

lxml provides a very simple and powerful API for parsing XML and HTML. It supports one-step parsing as well as step-by-step parsing using an event-driven API (currently only for XML).

Contents

The usual setup procedure:

>>> from lxml import etree
>>> from StringIO import StringIO

Parsers

Parsers are represented by parser objects. There is support for parsing both XML and (broken) HTML. Note that XHTML is best parsed as XML, parsing it with the HTML parser can lead to unexpected results. Here is a simple example for parsing XML from an in-memory string:

>>> xml = '<a xmlns="test"><b xmlns="test"/></a>'

>>> root = etree.fromstring(xml)
>>> print etree.tostring(root)
<a xmlns="test"><b xmlns="test"/></a>

To read from a file or file-like object, you can use the parse() function, which returns an ElementTree object:

>>> tree = etree.parse(StringIO(xml))
>>> print etree.tostring(tree.getroot())
<a xmlns="test"><b xmlns="test"/></a>

Note how the parse() function reads from a file-like object here. If parsing is done from a real file, it is more common (and also somewhat more efficient) to pass a filename:

>>> tree = etree.parse("doc/test.xml")

lxml can parse from a local file, an HTTP URL or an FTP URL. It also auto-detects and reads gzip-compressed XML files (.gz).

If you want to parse from memory and still provide a base URL for the document (e.g. to support relative paths in an XInclude), you can pass the base_url keyword argument:

>>> root = etree.fromstring(xml, base_url="http://where.it/is/from.xml")

Parser options

The parsers accept a number of setup options as keyword arguments. The above example is easily extended to clean up namespaces during parsing:

>>> parser = etree.XMLParser(ns_clean=True)
>>> tree   = etree.parse(StringIO(xml), parser)
>>> print etree.tostring(tree.getroot())
<a xmlns="test"><b/></a>

The keyword arguments in the constructor are mainly based on the libxml2 parser configuration. A DTD will also be loaded if validation or attribute default values are requested.

Available boolean keyword arguments:

  • attribute_defaults - read the DTD (if referenced by the document) and add the default attributes from it
  • dtd_validation - validate while parsing (if a DTD was referenced)
  • load_dtd - load and parse the DTD while parsing (no validation is performed)
  • no_network - prevent network access when looking up external documents
  • ns_clean - try to clean up redundant namespace declarations
  • recover - try hard to parse through broken XML
  • remove_blank_text - discard blank text nodes between tags
  • remove_comments - discard comments
  • compact - use compact storage for short text content (on by default)

Parsing HTML

HTML parsing is similarly simple. The parsers have a recover keyword argument that the HTMLParser sets by default. It lets libxml2 try its best to return something usable without raising an exception. You should use libxml2 version 2.6.21 or newer to take advantage of this feature:

>>> broken_html = "<html><head><title>test<body><h1>page title</h3>"

>>> parser = etree.HTMLParser()
>>> tree   = etree.parse(StringIO(broken_html), parser)

>>> print etree.tostring(tree.getroot(), pretty_print=True)
<html>
  <head>
    <title>test</title>
  </head>
  <body>
    <h1>page title</h1>
  </body>
</html>

Lxml has an HTML function, similar to the XML shortcut known from ElementTree:

>>> html = etree.HTML(broken_html)
>>> print etree.tostring(html, pretty_print=True)
<html>
  <head>
    <title>test</title>
  </head>
  <body>
    <h1>page title</h1>
  </body>
</html>

The support for parsing broken HTML depends entirely on libxml2's recovery algorithm. It is not the fault of lxml if you find documents that are so heavily broken that the parser cannot handle them. There is also no guarantee that the resulting tree will contain all data from the original document. The parser may have to drop seriously broken parts when struggling to keep parsing. Especially misplaced meta tags can suffer from this, which may lead to encoding problems.

Doctype information

The use of the libxml2 parsers makes some additional information available at the API level. Currently, ElementTree objects can access the DOCTYPE information provided by a parsed document, as well as the XML version and the original encoding:

>>> pub_id  = "-//W3C//DTD XHTML 1.0 Transitional//EN"
>>> sys_url = "http://www.w3.org/TR/xhtml1/DTD/xhtml1-transitional.dtd"
>>> doctype_string = '<!DOCTYPE html PUBLIC "%s" "%s">' % (pub_id, sys_url)
>>> xml_header = '<?xml version="1.0" encoding="ascii"?>'
>>> xhtml = xml_header + doctype_string + '<html><body></body></html>'

>>> tree = etree.parse(StringIO(xhtml))
>>> docinfo = tree.docinfo
>>> print docinfo.public_id
-//W3C//DTD XHTML 1.0 Transitional//EN
>>> print docinfo.system_url
http://www.w3.org/TR/xhtml1/DTD/xhtml1-transitional.dtd
>>> docinfo.doctype == doctype_string
True

>>> print docinfo.xml_version
1.0
>>> print docinfo.encoding
ascii

iterparse and iterwalk

As known from ElementTree, the iterparse() utility function returns an iterator that generates parser events for an XML file (or file-like object), while building the tree. The values are tuples (event-type, object). The event types are 'start', 'end', 'start-ns' and 'end-ns'.

The 'start' and 'end' events represent opening and closing elements and are accompanied by the respective element. By default, only 'end' events are generated:

>>> xml = '''\
... <root>
...   <element key='value'>text</element>
...   <element>text</element>tail
...   <empty-element xmlns="testns" />
... </root>
... '''

>>> context = etree.iterparse(StringIO(xml))
>>> for action, elem in context:
...     print action, elem.tag
end element
end element
end {testns}empty-element
end root

The resulting tree is available through the root property of the iterator:

>>> context.root.tag
'root'

The other event types can be activated with the events keyword argument:

>>> events = ("start", "end")
>>> context = etree.iterparse(StringIO(xml), events=events)
>>> for action, elem in context:
...     print action, elem.tag
start root
start element
end element
start element
end element
start {testns}empty-element
end {testns}empty-element
end root

Selective tag events

As an extension over ElementTree, lxml.etree accepts a tag keyword argument just like element.getiterator(tag). This restricts events to a specific tag or namespace:

>>> context = etree.iterparse(StringIO(xml), tag="element")
>>> for action, elem in context:
...     print action, elem.tag
end element
end element

>>> events = ("start", "end")
>>> context = etree.iterparse(
...             StringIO(xml), events=events, tag="{testns}*")
>>> for action, elem in context:
...     print action, elem.tag
start {testns}empty-element
end {testns}empty-element

Modifying the tree

You can modify the element and its descendants when handling the 'end' event. To save memory, for example, you can remove subtrees that are no longer needed:

>>> context = etree.iterparse(StringIO(xml))
>>> for action, elem in context:
...     print len(elem),
...     elem.clear()
0 0 0 3
>>> context.root.getchildren()
[]

WARNING: During the 'start' event, the descendants and following siblings are not yet available and should not be accessed. During the 'end' event, the element and its descendants can be freely modified, but its following siblings should not be accessed. During either of the two events, you must not modify or move the ancestors (parents) of the current element. You should also avoid moving or discarding the element itself. The golden rule is: do not touch anything that will have to be touched again by the parser later on.

If you have elements with a long list of children in your XML file and want to save more memory during parsing, you can clean up the preceding siblings of the current element:

>>> for event, element in etree.iterparse(StringIO(xml)):
...     # ... do something with the element
...     element.clear()                 # clean up children
...     while element.getprevious() is not None:
...         del element.getparent()[0]  # clean up preceding siblings

The while loop deletes multiple siblings in a row. This is only necessary if you skipped over some of them using the tag keyword argument. Otherwise, a simple if should do. The more selective your tag is, however, the more thought you will have to put into finding the right way to clean up the elements that were skipped. Therefore, it is sometimes easier to traverse all elements and do the tag selection by hand in the event handler code.

The 'start-ns' and 'end-ns' events notify about namespace declarations and generate tuples (prefix, URI):

>>> events = ("start-ns", "end-ns")
>>> context = etree.iterparse(StringIO(xml), events=events)
>>> for action, obj in context:
...     print action, obj
start-ns ('', 'testns')
end-ns None

It is common practice to use a list as namespace stack and pop the last entry on the 'end-ns' event.

iterwalk

A second extension over ElementTree is the iterwalk() function. It behaves exactly like iterparse(), but works on Elements and ElementTrees:

>>> root = etree.XML(xml)
>>> context = etree.iterwalk(
...             root, events=("start", "end"), tag="element")
>>> for action, elem in context:
...     print action, elem.tag
start element
end element
start element
end element

>>> f = StringIO(xml)
>>> context = etree.iterparse(
...             f, events=("start", "end"), tag="element")

>>> for action, elem in context:
...     print action, elem.tag
start element
end element
start element
end element

Python unicode strings

lxml.etree has broader support for Python unicode strings than the ElementTree library. First of all, where ElementTree would raise an exception, the parsers in lxml.etree can handle unicode strings straight away. This is most helpful for XML snippets embedded in source code using the XML() function:

>>> uxml = u'<test> \uf8d1 + \uf8d2 </test>'
>>> uxml
u'<test> \uf8d1 + \uf8d2 </test>'
>>> root = etree.XML(uxml)

This requires, however, that unicode strings do not specify a conflicting encoding themselves and thus lie about their real encoding:

>>> etree.XML(u'<?xml version="1.0" encoding="ASCII"?>\n' + uxml)
Traceback (most recent call last):
  ...
ValueError: Unicode strings with encoding declaration are not supported.

Similarly, you will get errors when you try the same with HTML data in a unicode string that specifies a charset in a meta tag of the header. You should generally avoid converting XML/HTML data to unicode before passing it into the parsers. It is both slower and error prone.

Serialising to Unicode strings

To serialize the result, you would normally use the tostring module function, which serializes to plain ASCII by default or a number of other encodings if asked for:

>>> etree.tostring(root)
'<test> &#63697; + &#63698; </test>'

>>> etree.tostring(root, 'UTF-8', xml_declaration=False)
'<test> \xef\xa3\x91 + \xef\xa3\x92 </test>'

As an extension, lxml.etree has a new tounicode() function that you can call on XML tree objects to retrieve a Python unicode representation:

>>> etree.tounicode(root)
u'<test> \uf8d1 + \uf8d2 </test>'

>>> el = etree.Element("test")
>>> etree.tounicode(el)
u'<test/>'

>>> subel = etree.SubElement(el, "subtest")
>>> etree.tounicode(el)
u'<test><subtest/></test>'

>>> et = etree.ElementTree(el)
>>> etree.tounicode(et)
u'<test><subtest/></test>'

The result of tounicode() can be treated like any other Python unicode string and then passed back into the parsers. However, if you want to save the result to a file or pass it over the network, you should use write() or tostring() with an encoding argument (typically UTF-8) to serialize the XML. The main reason is that unicode strings returned by tounicode() never have an XML declaration and therefore do not specify their encoding. These strings are most likely not parsable by other XML libraries.

In contrast, the tostring() function automatically adds a declaration as needed that reflects the encoding of the returned string. This makes it possible for other parsers to correctly parse the XML byte stream. Note that using tostring() with UTF-8 is also considerably faster in most cases.