XPath Tutorial

Part 1: What is XPath?

XPath is a language

“XPath is a language for addressing parts of an XML document”

(from the XML Path Language 1.0 specification)

In other words: you write an XPath expression as a string, pass it to an XPath engine together with an XML (or HTML) document, and get back the parts of that document it points to – following the data model described below.

Why learn XPath?

  • with XPath, you can navigate everywhere inside a DOM tree

  • it’s a must-have skill for accurate web data extraction

  • XPath is more powerful than CSS selectors

  • it allows selection and filtering with a fine-grained look at the text content

  • XPath allows complex conditioning with axes

  • XPath is extensible with custom functions (we only touch on the bundled EXSLT extension functions, at the end of this tutorial)

XPath data model

XPath’s data model is a tree of nodes representing a document. Nodes can be either:

  • element nodes (<p>This is a paragraph</p>),

  • or attribute nodes (href="page.html" inside an <a> tag),

  • or text nodes ("I have something to say"),

  • or comment nodes (<!-- a comment -->),

  • (or root nodes, or namespace nodes, or processing instructions nodes but we will not cover them here.)

In XPath’s data model, everything is a node : elements, attributes, comments… (but not all nodes are elements.)

And nodes have an order, the document order: the order in which they appear in the XML/HTML source.

In effect, this data model allows you to represent everything inside an XML or HTML document, in a structured, ordered and hierarchical way.

Throughout this tutorial, we’ll use the following sample HTML page to illustrate how XPath works:

<html>
<head>
  <title>This is a title</title>
  <meta content="text/html; charset=utf-8" http-equiv="content-type">
</head>
<body>
  <div>
    <div>
      <p>This is a paragraph.</p>
      <p>Is this <a href="page2.html">a link</a>?</p>
      <br>
      Apparently.
    </div>
    <div class="second">
      Nothing to add.
      Except maybe this <a href="page3.html">other link</a>.
      <!-- And this comment -->
    </div>
  </div>
</body>
</html>

Here is an ASCII tree representation of our toy HTML document for an XPath engine, according to the data model:

# 0--(ROOT)
 +-- # 1--<html>
     +-- # 2--(TXT): '\n'
     +-- # 3--<head>
     |   +-- # 4--(TXT): '\n  '
     |   +-- # 5--<title>
     |   |   +-- # 6--(TXT): 'This is a title'
     |   +-- # 7--(TXT): '\n  '
     |   +-- # 8--<meta>
     |   |   +-- # 9--(ATTR): content: 'text/html; charset=utf-8'
     |   |   +-- #10--(ATTR): http-equiv: 'content-type'
     |   +-- #11--(TXT): '\n'
     +-- #12--(TXT): '\n'
     +-- #13--<body>
     |   +-- #14--(TXT): '\n  '
     |   +-- #15--<div>
     |   |   +-- #16--(TXT): '\n    '
     |   |   +-- #17--<div>
     |   |   |   +-- #18--(TXT): '\n      '
     |   |   |   +-- #19--<p>
     |   |   |   |   +-- #20--(TXT): 'This is a paragraph.'
     |   |   |   +-- #21--(TXT): '\n      '
     |   |   |   +-- #22--<p>
     |   |   |   |   +-- #23--(TXT): 'Is this '
     |   |   |   |   +-- #24--<a>
     |   |   |   |   |   +-- #25--(ATTR): href: 'page2.html'
     |   |   |   |   |   +-- #26--(TXT): 'a link'
     |   |   |   |   +-- #27--(TXT): '?'
     |   |   |   +-- #28--(TXT): '\n      '
     |   |   |   +-- #29--<br>
     |   |   |   +-- #30--(TXT): '\n      Apparently.\n    '
     |   |   +-- #31--(TXT): '\n    '
     |   |   +-- #32--<div>
     |   |   |   +-- #33--(ATTR): class: 'second'
     |   |   |   +-- #34--(TXT): '\n      Nothing to add.\n      Except maybe this '
     |   |   |   +-- #35--<a>
     |   |   |   |   +-- #36--(ATTR): href: 'page3.html'
     |   |   |   |   +-- #37--(TXT): 'other link'
     |   |   |   +-- #38--(TXT): '.\n      '
     |   |   |   +-- #39--(COMM): ' And this comment '
     |   |   |   +-- #40--(TXT): '\n    '
     |   |   +-- #41--(TXT): '\n  '
     |   +-- #42--(TXT): '\n'
     +-- #43--(TXT): '\n'

You can see various tree branches and leaves:

  • e.g. <div> or <p>: these are element nodes

  • (TXT) represent text nodes

  • (ATTR) represent attribute nodes

  • (COMM) represent comment nodes

The #<number> are the document orders of each node.

Note

You can also notice that text with only whitespace (space and newlines in our example) are proper nodes, they do have their document order and can be selected with XPath.

In-browser widget and using parsel

To illustrate and learn XPath, we will use an in-browser widget allowing you to play around with XPath expressions and see the output live. We will also illustrate some Python pattern for data extraction with XPath using the parsel library which powers Scrapy selectors under the hood. It is a Python module written on top of lxml.

Note

The in-browser playground evaluates your input with the browser’s native engine (document.evaluate() over application/xml), so the HTML input must be well-formed XHTML: close every void element (<br/>, <img/>, <meta/>, <hr/>) and keep tags balanced, or you will get a parse error. parsel’s HTML parser is much more forgiving and happily accepts the sloppy HTML found on real pages (including the sample document above). For the same reason the widget’s output is cosmetically slightly different from parsel’s: void elements come back self-closed (<br/> rather than <br>) and numbers without a trailing .0 (2 rather than '2.0').

Note

lxml itself is built using the C library libxml2, which has a conformant XPath 1.0 engine. You should be able to run the same XPath expressions with any XPath 1.0 engine, and get the same results.

This tutorial only showcases XPath 1.0. lxml’s engine – and therefore parsel – implements XPath 1.0, and that’s what we use throughout. (XPath has reached version 3, and pure-Python XPath 2.0/3.x engines such as elementpath do exist, but you can already do a lot with XPath 1.0.)

When showing Python code snippets using Parsel, we assume that we have a Selector – called doc – created with the HTML content, similarly to the following:

import parsel

htmlsample = """<html>
<head>
  <title>This is a title</title>
  <meta content="text/html; charset=utf-8" http-equiv="content-type" />
</head>
<body>
  <div>
    <div>
      <p>This is a paragraph.</p>
      <p>Is this <a href="page2.html">a link</a>?</p>
      <br />
      Apparently.
    </div>
    <div class="second">
      Nothing to add.
      Except maybe this <a href="page3.html">other link</a>.
      <!-- And this comment -->
    </div>
  </div>
</body>
</html>"""

#
# Below is a small "hack" to change the representation of extracted
#  nodes when using parsel.
# This is to represent return values as serialized HTML element or
# string, and not parsel's wrapper objects.
#
parsel.Selector.__str__ = parsel.Selector.extract
parsel.Selector.__repr__ = parsel.Selector.__str__
parsel.SelectorList.__repr__ = lambda x: "[{}]".format(
    "\n ".join("({}) {!r}".format(i, repr(s)) for i, s in enumerate(x, start=1))
).replace(r"\n", "\n")

doc = parsel.Selector(text=htmlsample)

XPath return types

When applied over a document, an XPath expression can return either:

  • a node-set – this is the most common case, and often it’s a set of element nodes

  • a string

  • a number (floating point)

  • a boolean

Note

When an XPath expression returns a node-set, you do get a set of nodes, even if there’s only one node in the set. With parsel, you get a list of nodes though, not a Python set.

XPath expressions

We will now take a look at some example XPath expressions to get a feeling of how they work. We’ll explain the syntax in more details later on.

XPath expressions are passed to an XPath engine as strings.

Selecting the root node (a special case)

One of the shortest XPath expressions, "/" (a string with only a forward slash), selects the root node of the document – the invisible node whose only child is the document element (the top-level <html> element here).

Result

This is very similar to cd / in a Unix shell (going to the root directory).

Warning

This "/" expression does not work as expected with parsel: you get an empty list instead of the root node. It is an lxml limitation (it works with libxml2 directly). In practice this rarely matters – the root node is virtually never used directly.

Selecting elements

Elements build the structure and hierarchy of the document. An element in HTML (and XML) is what you see in the source code between an opening and corresponding closing tag, and everything in between.

  • <title>This is a title</title> is a title element,

  • <p>Is this <a href="page2.html">a link</a>?</p> is a p (paragraph) element.

Selecting elements is probably the most common use-case for XPath on HTML documents.

Elements can have children – the root node being the ancestor of them all. Their children can also have children and so on. Sometimes, elements only have one child. This hierarchy forms a family tree of nodes.

Note

Text nodes are not elements. (They are still nodes, obviously.) They do not have children nodes, but they are always children of some element.

Therefore, text nodes are always leaves of the document tree.

We said earlier that the document element is a child of the root node. In fact, the document element is the only child of the root node. And for our sample HTML document, it’s the top-level <html>...</html> element. Still, selecting it will return a single-node node-set, the XPath expression being /*:

Result

The asterisk here, *, means “any element”. And /* means “any element under the root node”. HTML documents have only one element like this: the <html> element.

Another example: how to get <title> elements? Use /html/head/title:

Result

Again, if you are familiar with the Unix filesystem, you probably intuitively understand what this does:

  • start from the root (of the document)

    • select the <html> node (with /html)

      • select the <head> node under the <html> node (appending /head)

        • select the <title> node under the <head> node (appending /title)

In other words, the XPath expression represents the path from the root node down to the target node(s). Parts of this path are read from left to right, and represent a top-to-bottom direction in the document tree.

Much like a Unix filepath represents the path from the filesystem’s root to the target file(s) or directory(ies). There’s one major difference with a Unix filesystem though: in an HTML or XML document, an element can have multiple children with the same name. For example, the <div> just under the <body> has 2 <div> children:

Result

Another example is getting the paragraphs inside the first child of that <div> under <body>, there are two of them:

Result

Here we’re introducing a positional predicate, [1]. The div[1] part means “the first <div> child under its parent”.

If you recall, earlier we used a * asterisk to mean any element. There are other elements with those two paragraphs under that very <div>. Let’s try and select all of them, regardless of their name:

Result

Note

Continuing the filesystem analogy, * is similar in effect to what you can do in a Unix shell to find files or directories without explicit full names.

See the <br/> being selected? It’s an empty element (i.e. with no child nodes) but it is there nonetheless.

Selecting text nodes

If we stay around these <p> and <br> elements, you may have noticed that the ASCII tree representation from the beginning also shows some text after the <br/> break: the string "Apparently.". It is a text node.

Selecting text nodes is a bit different than selecting elements: you use the special text() syntax. Let’s try it by replacing the last part of our last XPath expression, forming /html/body/div/div[1]/text():

Result

You may have expected only one text result, the last one, "Apparently.". But we got four! And three of them are blank even. Why is that?

In fact, HTML authors usually indent their tags with whitespace for readability. This does not usually change the layout in your browser. But this whitespace counts as text nodes for XPath’s data model, it is not stripped nor filtered.

Let’s represent that <div> as a Python string as it appears in the HTML source:

#
#   text node #1                       text node #2                                           text node #3
#     <------>                           <------>                                               <------>
'<div>\n      <p>This is a paragraph.</p>\n      <p>Is this <a href="page2.html">a link</a>?</p>\n      <br>\n  Apparently.\n    </div>'

We’ve marked the first three text nodes before the non-whitespace only text node.

Another example is to get the text nodes of <title> elements (remember that <title> is an element, and that it happens it contains a text node, with the string content “This is a title”):

Result

Note

Again, there’s only one <title>, and it contains only one text node, but selecting text nodes in <title> returns a single string-value in a list, not one string.

Selecting nodes without a full, explicit path

What we did until now is tell the XPath engine how to get to nodes, node by node, from parent to child, from root node down to target nodes. This assumes that you know the hierarchy of nodes beforehand. This can be the case, but more often than not, either you do not know or you do not want to indicate all the steps from the root node down to the node(s) you are interested in (this can be very error prone – have you put enough div/div/div...?).

XPath provides a handy shortcut when you do not know at what level you expect your target node to be. Say for example that we want to select all <p> paragraph elements inside the <body>. We don’t a-priori know what their parent node is. (For all we know, they can be anywhere under the <body> element.) The shortcut to use is // (two forward slashes). Let’s try this: //body//p

Result

So we got 2 paragraphs, what we expected.

This also works for text nodes (there are a lot of them in our sample document!). Try //body//text():

Result

Selecting attributes

Elements can also have attributes. In our sample document, we have two <a> elements, each with a href attribute. There’s also a <meta> element with two attributes: content and http-equiv.

This is how you can select these attributes, with an @ prefix before the attribute name:

Result
Result

The * (asterisk) here after @ means the same thing as in /* except that this is for attributes, and not elements: meaning that you want any attributes, whatever their name.

Get a string representation of an element

The XPath language also comes with a few string functions, that you can wrap around an XPath expression selecting elements:

Result

This example uses string(<xpathexpression>), one of several handy functions in XPath. string() will concatenate all text content from the selected node and all of its children, recursively, effectively stripping HTML tags.

You may wonder what’s the difference between string(/html/head/title) and /html/head/title/text() from earlier? Here, in fact, you get the same result because <title> only has one child text node. (Concatenating this list of one text node is the same as getting it directly with text() at the end.)

But string functions can be very handy when you apply them on nodes that have multiple children and multiple text node children or descendant. What happens when you apply string() on the document <body> for example? You get a text representation of the document, without the tags:

Result

Counting elements

We said earlier that XPath expressions could also return numbers. One example of this is counting the number of paragraphs in the document:

Result

Note

With parsel, you get a floating point number back, and in the form of a string. This is specific to parsel. Another XPath engine might return a native floating point number.

Another example: get the number of attributes in the document (whatever their parent element):

Result

Boolean operations

XPath expressions can also return booleans. This is not that useful by itself, but it becomes handy when used in predicates (that we will cover a bit later).

For example, testing the number of paragraphs:

Result
Result

Part 2: Location Paths: how to move inside the document tree

A Location path is the most common XPath expression. It is used to move in any direction from a starting point (the context node) to any node(s) in the tree:

  • It is a string, with a series of “location steps”: "step1 / step2 / step3 ...";

  • It represents the selection and filtering of nodes, processed step by step, from left to right;

  • Each step is of the form axis :: nodetest [predicate]*

    • an axis (implicit or explicit),

    • a node test,

    • zero or more predicates.

So the examples we saw earlier are (or contain) an XPath location path: /html/head/title, //body//p etc.

Tip

Whitespace does NOT matter in XPath.

(Except for “//” and “..”; “/   /” and “.  .” are syntax errors.)

For example, the following three expressions produce the same result:

>>> doc.xpath("/html/head/title")
[(1) '<title>This is a title</title>']

>>> doc.xpath("/    html   / head   /title")
[(1) '<title>This is a title</title>']

>>> doc.xpath("""
...     /html
...         /head
...             /title""")
...
[(1) '<title>This is a title</title>']

So don’t be afraid to indent your XPath expressions to improve readability.

Relative vs. absolute paths

Location paths can be relative or absolute:

  • "step1/step2/step3" is relative

  • "/step1/step2/step3" is absolute

In other words, an absolute path is a relative path starting with “/” (forward slash). Absolute paths are relative to the root node.

Tip

Use relative paths whenever possible. This prevents unexpected selection of duplicate nodes in loop iterations.

For example, in our sample document, only one <div> contains paragraphs. Looping on each <div> and using the absolute location path //p will produce the same result for each iteration: returning ALL paragraphs in the document everytime.

>>> for div in doc.xpath("//body//div"):
...     print(div.xpath("//p"))
...
[(1) '<p>This is a paragraph.</p>'
 (2) '<p>Is this <a href="page2.html">a link</a>?</p>']
[(1) '<p>This is a paragraph.</p>'
 (2) '<p>Is this <a href="page2.html">a link</a>?</p>']
[(1) '<p>This is a paragraph.</p>'
 (2) '<p>Is this <a href="page2.html">a link</a>?</p>']

Compare this with using the relative 'p' or './p' expression that will only look at children <p> under each <div>, and only one of those <div> will show having paragraphs as shown below:

>>> for div in doc.xpath("//body//div"):
...     print(div.xpath("p"))
...
[]
[(1) '<p>This is a paragraph.</p>'
 (2) '<p>Is this <a href="page2.html">a link</a>?</p>']
[]

>>> for div in doc.xpath("//body//div"):
...     print(div.xpath("./p"))
...
[]
[(1) '<p>This is a paragraph.</p>'
 (2) '<p>Is this <a href="page2.html">a link</a>?</p>']
[]

Abbreviated syntax vs. full syntax

What we’ve seen earlier is in fact the “abbreviated syntax” for XPath expressions. The full syntax is quite verbose (but you sometimes need it):

Abbreviated syntax

Full syntax

/html/head/title

/child::html /child:: head /child:: title

//meta/@content

/descendant-or-self::node() /child::meta / attribute::content

//div/div[@class="second"]

/descendant-or-self::node() /child::div /child::div [attribute::class = "second"]

//div/a/text()

/descendant-or-self::node() /child::div /child::a /child::text()

What are these child::, descendant-or-self:: and attribute::, you may ask? They are axes.

Axes: moving around

Important

Remember that each step of an XPath location path is of the form AXIS :: nodetest [predicate]*.

The “axis” is the first part of each location path step. It can be explicit, or implicit in abbreviated syntax. For example, in /html/head/title, the child:: axis is omitted in each step.

In this section, we’ll use explicit axes as much as we can.

Axes give the direction to go next, one location step at a time.

  • self (where you are)

  • parent, child (direct hop up or down the document tree)

  • ancestor, ancestor-or-self, descendant, descendant-or-self (multi-hop)

  • following, following-sibling, preceding, preceding-sibling (document order)

  • attribute, namespace (non-element)

Stay where you are: self

Let’s assume that we have selected the first <div> element in our sample document, the one just under the <body> element:

>>> first_div = doc.xpath("//body/div")[0]
>>> first_div
<div>
    <div>
      <p>This is a paragraph.</p>
      <p>Is this <a href="page2.html">a link</a>?</p>
      <br>
      Apparently.
    </div>
    <div class="second">
      Nothing to add.
      Except maybe this <a href="page3.html">other link</a>.
      <!-- And this comment -->
    </div>
  </div>

The self axis represents the context node, i.e. where you are currently in the Location Path steps. (This may not sound very useful, but we will see later when this can be handy.)

>>> first_div.xpath("self::*")
[(1) '<div>
    <div>
      <p>This is a paragraph.</p>
      <p>Is this <a href="page2.html">a link</a>?</p>
      <br>
      Apparently.
    </div>
    <div class="second">
      Nothing to add.
      Except maybe this <a href="page3.html">other link</a>.
      <!-- And this comment -->
    </div>
  </div>']

self:: is usually seen in its abbreviated form, . (one dot), which means self::node(). Chaining self steps (self::*/self::* or ././.) just keeps you on the same context node, so . is a compact way of saying “right here”:

>>> first_div.xpath(".")
[(1) '<div>
    <div>
      <p>This is a paragraph.</p>
      <p>Is this <a href="page2.html">a link</a>?</p>
      <br>
      Apparently.
    </div>
    <div class="second">
      Nothing to add.
      Except maybe this <a href="page3.html">other link</a>.
      <!-- And this comment -->
    </div>
  </div>']

Move up or down the tree: child, descendant, parent, ancestor

The child axis is for immediate children nodes of the context node. Here, our context node <div> has two <div> children:

>>> first_div.xpath("child::*")
[(1) '<div>
      <p>This is a paragraph.</p>
      <p>Is this <a href="page2.html">a link</a>?</p>
      <br>
      Apparently.
    </div>'
 (2) '<div class="second">
      Nothing to add.
      Except maybe this <a href="page3.html">other link</a>.
      <!-- And this comment -->
    </div>']

child is in fact the default axis, hence it can be omitted (e.g. we saw that /html/head/title is equivalent of /child::html/child::head/child::title.)

The parent axis is the dual of child: you go up one level in the document tree:

>>> first_div.xpath("parent::*")
[(1) '<body>
  <div>
    <div>
      <p>This is a paragraph.</p>
      <p>Is this <a href="page2.html">a link</a>?</p>
      <br>
      Apparently.
    </div>
    <div class="second">
      Nothing to add.
      Except maybe this <a href="page3.html">other link</a>.
      <!-- And this comment -->
    </div>
  </div>
</body>']

There’s an alias for parent::: it’s .. (two dots, much like in a Unix filesystem):

>>> first_div.xpath("..")
[(1) '<body>
  <div>
    <div>
      <p>This is a paragraph.</p>
      <p>Is this <a href="page2.html">a link</a>?</p>
      <br>
      Apparently.
    </div>
    <div class="second">
      Nothing to add.
      Except maybe this <a href="page3.html">other link</a>.
      <!-- And this comment -->
    </div>
  </div>
</body>']

Let’s simplify our ASCII tree representation from earlier to only consider element nodes:

# 0--(ROOT)
 +-- # 1--<html>
     +-- # 3--<head>
     |   +-- # 5--<title>
     |   +-- # 8--<meta>
     +-- #13--<body>
         +-- #15--<div>
             +-- #17--<div>
             |   +-- #19--<p>
             |   +-- #22--<p>
             |   |   +-- #24--<a>
             |   +-- #29--<br>
             +-- #32--<div>
                 +-- #35--<a>

With this simplified tree representation, this is what self, child and parent select:

                # 0--(ROOT)
                 +-- # 1--<html>
                     +-- # 3--<head>
                     |   +-- # 5--<title>
                     |   +-- # 8--<meta>
parent::* ---------> +-- #13--<body>
                         |
self::* ------------->   +-- #15--<div>
                             |
child::*----+----------->    +-- #17--<div>
            |                |   +-- #19--<p>
            |                |   +-- #22--<p>
            |                |   |   +-- #24--<a>
            |                |   +-- #29--<br>
            +----------->    +-- #32--<div>
                                 +-- #35--<a>
Recursively go up or down

The descendant axis is similar to child but also goes deeper down the tree, looking at children of each child, recursively:

>>> first_div.xpath("descendant::*")
[(1) '<div>
      <p>This is a paragraph.</p>
      <p>Is this <a href="page2.html">a link</a>?</p>
      <br>
      Apparently.
    </div>'
 (2) '<p>This is a paragraph.</p>'
 (3) '<p>Is this <a href="page2.html">a link</a>?</p>'
 (4) '<a href="page2.html">a link</a>'
 (5) '<br>'
 (6) '<div class="second">
      Nothing to add.
      Except maybe this <a href="page3.html">other link</a>.
      <!-- And this comment -->
    </div>'
 (7) '<a href="page3.html">other link</a>']

You might guess already what ancestor is for: it is the dual axis of descendant. It goes to the parent of the context node, the parent of this parent, the parent of the parent of this parent, etc.

>>> first_div.xpath("ancestor::*")
[(1) '<html>
<head>
  <title>This is a title</title>
  <meta content="text/html; charset=utf-8" http-equiv="content-type">
</head>
<body>
  <div>
    <div>
      <p>This is a paragraph.</p>
      <p>Is this <a href="page2.html">a link</a>?</p>
      <br>
      Apparently.
    </div>
    <div class="second">
      Nothing to add.
      Except maybe this <a href="page3.html">other link</a>.
      <!-- And this comment -->
    </div>
  </div>
</body>
</html>'
 (2) '<body>
  <div>
    <div>
      <p>This is a paragraph.</p>
      <p>Is this <a href="page2.html">a link</a>?</p>
      <br>
      Apparently.
    </div>
    <div class="second">
      Nothing to add.
      Except maybe this <a href="page3.html">other link</a>.
      <!-- And this comment -->
    </div>
  </div>
</body>']
Special case of descendant-or-/ancestor-or-self axes

The last axes to be aware of when you need to move up or down the document tree are descendant-or-self and ancestor-or-self. They are the same as descendant or ancestor except they also include the context node.

>>> first_div.xpath("./descendant-or-self::node()/text()")
[(1) '
    '
 (2) '
      '
 (3) 'This is a paragraph.'
 (4) '
      '
 (5) 'Is this '
 (6) 'a link'
 (7) '?'
 (8) '
      '
 (9) '
      Apparently.
    '
 (10) '
    '
 (11) '
      Nothing to add.
      Except maybe this '
 (12) 'other link'
 (13) '.
      '
 (14) '
    '
 (15) '
  ']

Move “sideways”: children nodes of the same parent

If nodes can have parents, children, ancestors and descendants, they can also have siblings (to continue the family analogy). Siblings are nodes that have the same parent node.

Some siblings may come before the context node (they appear before in the document, their order is lower), or they can come after the context node. There are two axis for these two directions: preceding-sibling and following-sibling.

Let’s first select this paragraph from our sample document: <p>Is this <a href="page2.html">a link</a>?</p>. It’s the 2nd child of the 1st <div> of the <div> we used above:

paragraph = first_div.xpath("child::div[1]/child::p[2]")[0]

Here we started using 2 new patterns along with the axes:

  • child::div vs. child::*:

    • * means “any element node” (this is a node-test that we’ll cover afterwards),

    • while child::div means “any child that is a <div> element”.

  • [1] and [2]: which mean first and second in the current step’s node-set (this is a kind of predicate that we’ll cover afterwards also)

>>> paragraph.xpath("preceding-sibling::*")
[(1) '<p>This is a paragraph.</p>']

>>> paragraph.xpath("following-sibling::*")
[(1) '<br>']

Again, let’s see which elements were selected in our ASCII tree representation:

                # 0--(ROOT)
                 +-- # 1--<html>
                     +-- # 3--<head>
                     |   +-- # 5--<title>
                     |   +-- # 8--<meta>
                     +-- #13--<body>
                         |
                         +-- #15--<div>
                             |
                             +-- #17--<div>
                             |   |
                             |   |
preceding-sibling::* ----------> +-- #19--<p>
                             |   |
                             |   |
self::* -----------------------> +-- #22--<p>
                             |   |   |
                             |   |   +-- #24--<a>
                             |   |
                             |   |
following-sibling::* ----------> +-- #29--<br>
                             |
                             |
                             +-- #32--<div>
                                 +-- #35--<a>

Earlier we were also able to get text nodes that were siblings of these <p> elements. Why did they not get selected?

The reason is that child::* means “any child element”, not “any node.” (Remember that text nodes are not elements.)

To also get text node siblings, you need to use either child::text() or child::node(). (But we may be getting ahead of ourselves with node tests.)

>>> paragraph.xpath("following-sibling::node()")
[(1) '
      '
 (2) '<br>'
 (3) '
      Apparently.
    ']

>>> paragraph.xpath("following-sibling::text()")
[(1) '
      '
 (2) '
      Apparently.
    ']
Nodes before and after, in document order

preceding and following are two special axes that do not look at the tree hierarchy, but work on the document order of nodes.

Important

Remember, all nodes in XPath data model have an order, called the document order. Node 1 is the first node in the HTML source, node 2 is the node appearing next etc.

  #1    #2    #3   ...
<html><head><title>...

>>> paragraph.xpath("preceding::*")
[(1) '<head>
  <title>This is a title</title>
  <meta content="text/html; charset=utf-8" http-equiv="content-type">
</head>'
 (2) '<title>This is a title</title>'
 (3) '<meta content="text/html; charset=utf-8" http-equiv="content-type">'
 (4) '<p>This is a paragraph.</p>']

>>> paragraph.xpath("following::*")
[(1) '<br>'
 (2) '<div class="second">
      Nothing to add.
      Except maybe this <a href="page3.html">other link</a>.
      <!-- And this comment -->
    </div>'
 (3) '<a href="page3.html">other link</a>']

This is what these axes select in our ASCII-tree representation:

                # 0--(ROOT)
                 +-- # 1--<html>
                     |   |
                +--> +-- # 3--<head>
                |    |   |
                +------> +-- # 5--<title>
                |    |   |
                +------> +-- # 8--<meta>
                |    |
                |    +-- #13--<body>
                |        |
                |        +-- #15--<div>
                |            |
                |            +-- #17--<div>
                |            |   |
                |            |   |
preceding::* ---+--------------> +-- #19--<p>
                             |   |
                             |   |
self::* -----------------------> +-- #22--<p>
                             |   |   |
                             |   |   +-- #24--<a>
                             |   |
                             |   |
following::* -----------+------> +-- #29--<br>
                        |    |
                        |    |
                        +--> +-- #32--<div>
                        |        |
                        +------> +-- #35--<a>

Note

Notice that preceding does not include ancestors and following does not include descendants. This property is mentioned in XPath specs like this:

“The ancestor, descendant, following, preceding and self axes partition a document (ignoring attribute and namespace nodes): they do not overlap and together they contain all the nodes in the document.”

In other words:

document == self ∪ (ancestor ∪ preceding) ∪ (descendant ∪ following)

(∪ denoting the “union” for node-sets.)

Attribute axis

Attributes are nodes too, but they are special: they are not children of their element. This is why none of the axes we have seen so far (child, descendant, following…) ever returned an attribute node. Attributes live on their own axis, attribute.

To reach them, use the attribute axis explicitly:

Result

Because selecting attributes is so common, the attribute:: axis has a short form: the @ sign. So //a/@href means exactly the same thing as //a/attribute::href:

Result

You can select every attribute of an element with @*, and every attribute in the whole document with //@*:

Result

Note

An attribute node has a name and a string value. When you extract an attribute with parsel, you get its value:

>>> doc.xpath("//a/@href")
[(1) 'page2.html'
 (2) 'page3.html']

Attributes are most often used inside predicates, to keep only the elements that carry a given attribute (or a given attribute value). We will use them a lot in the Predicates section below. As a taste, //div[@class] selects the <div> elements that have a class attribute, whatever its value:

Result

Node tests

Important

A “node test” is the second part of each step in a location path.

axis :: NODETEST [predicate]*

Node tests select node types along the step’s axis.

a node test can be:

  • a name test:

    • such as p, title or a for elements: /html/head/title contains 3 steps, each with a name test node-test

    • or href or src for attributes: /a/@href selects “href” attributes of <a> elements

  • a node type test:

    • node(): any node type

    • text(): text nodes

    • comment(): comment nodes

    • * (an asterisk): the meaning depends on the axis:

      • an * step alone selects any element node

      • an @* selects any attribute node

Warning

text() is not a function call that converts a node to it’s text representation, it’s just a test on the node type.

Compare these two expressions:

>>> paragraph.xpath("child::text()")
[(1) 'Is this '
 (2) '?']

>>> paragraph.xpath("string(self::*)")
[(1) 'Is this a link?']

child::text() selects all children nodes that are also text nodes.

The “a” string is part of the <a> inside the paragraph, so it’s not selected. It is not a direct child of the <p> element.

Whereas string(self::*) applies to the paragraph (the context node, selected with self::*) and recursively gets text content of children, children of children and so on.

Abbreviation cheatsheet

Abbreviated step

Meaning

* (asterisk)

all element nodes (i.e. not text nodes, not attribute nodes).

Remember that .//* is not the same as .//node().

Also, there’s no element() node test.

@*

attribute::* (all attribute nodes)

//

/descendant-or-self::node()/ (exactly this, nothing more, nothing less)

so //* is not the same as /descendant-or-self::*

. (a single dot)

self::node(), the context node; useful for making XPaths relative, e.g. .//tr

.. (2 dots)

parent::node()

Why //* is not /descendant-or-self::*

The abbreviation // is a frequent source of confusion. It does not expand to /descendant-or-self::; it expands to /descendant-or-self::node()/ – note the node() node test and the trailing slash, which introduces another step.

So //* is really:

/descendant-or-self::node()/child::*

which reads as “every element that is the child of any node”, i.e. every element in the document.

/descendant-or-self::*, on the other hand, is a single step: “the context node or any of its descendants, provided it is an element”.

From the root node these two happen to select the same set (the root is not an element, so both return every element). The difference becomes visible as soon as you use them relative to an element, because descendant-or-self includes the context node itself, while the child::* step at the end of .//* never can:

>>> paragraph = doc.xpath("//p[2]")[0]
>>> paragraph.xpath(".//*")  # descendants only
[(1) '<a href="page2.html">a link</a>']
>>> paragraph.xpath("descendant-or-self::*")  # the <p> itself, then its descendants
[(1) '<p>Is this <a href="page2.html">a link</a>?</p>'
 (2) '<a href="page2.html">a link</a>']

This is also why .//node() and .//* differ: * is a name test that only matches elements, while node() matches every node type.

Predicates

Important

Predicates are the last part of each step in a location path. Predicates are optional.

axis :: nodetest [PREDICATE]*

They are used to further filter nodes on properties that cannot be expressed with the step’s axis and node test.

Remember that XPath location paths work step by step. Each step produces a node-set for each node from the previous step’s node-set, with possibly more than 1 node in each node set.

You may not be interested in all nodes from a node test. And predicates are used to tell the XPath engine the condition(s) they should meet.

The syntax for predicates is simple: just surround conditions within square brackets. What’s inside the square brackets can be:

  • a number (see positional predicates below)

  • a location path: the predicate will select nodes for which the location path matches at least a node

  • a boolean operation: for example to test a condition on text content or count of children

Positional predicates

The first use-case is selecting nodes based on their position in a node-set. (Node-sets order depends on the axis, but let’s consider that the order of a node in a node-set is the document order.)

Remember the two paragraphs in the <div> we looked at earlier:

>>> doc.xpath("//body/div/div/p")
[(1) '<p>This is a paragraph.</p>'
 (2) '<p>Is this <a href="page2.html">a link</a>?</p>']

Let’s say that we are not interested in the two paragraphs but only the first one. You would use [1] as predicate:

Result

Warning

Positions in XPath start from 1, not 0.

If you want the last node in a node-set, you can use last():

Result

Warning

Because location paths work step by step, from left to right, positional predicates are about the position of a node in a node-set produced by the current step, not about the position of the node in the document tree.

For example, //body//div[1] is NOT the first <div> under the <body> element; it will select all <div> that are the first child of their parent:

Result

This becomes more apparent when you expand the expression to its full syntax:

/descendant-or-self::node()
    /child::body
        /descendant-or-self::node()
                               ^
                               |
            # first child of this parent
            /child::div[1]

You can however select the first <div> (in document order) in a <body> using parentheses to group nodes into a new node-set:

  • first select all <div> elements – //body//div,

  • then group them – ( //body//div ),

  • and finally select the first one – ( //body//div ) [1],

Result
Position ranges

Sometimes you need more than one node in a node-set but not all of them. For that you can use boolean expression in your predicate in conjunction with the position() function that returns the node’s position.

Let’s change our sample HTML document a bit to include a list of five items. Say we need all but the 1st and last one. You can use [position()>1 and position()<last()]:

Result

//body//div//li[position()>1 and position()<last()] correctly selects the 2nd, 3rd and 4th items.

Location paths as predicates

Location paths can also serve as predicates within a parent location path.

It can happen that your HTML markup does not distinguish the elements you are after with any “class” or “id” attributes, but maybe these elements have a structural feature that you can use to identify them.

For example, a <table> element may have rows – table/tr – with or without link anchors in them. Within each <tr> row, td/a selects something on some rows, nothing for others:

Result

Boolean predicates

We saw boolean predicates earlier with positional ranges. But you can craft complex boolean filters based on any features of nodes; structural information on children or parent nodes, text values, position, etc.

A simple example could be selecting a <table> that has a specific number of rows, say, 5. You can simply count the number of rows:

Result
Special case of string value tests

XPath also allows comparing string values of nodes within predicates. If you use an equality operation with a location path and a string, each node of the location path will be converted to its string value and then compared with the string value to match.

This may sound more obscure than it is. Say for example that you have two tables, with different headers. You know the string value of the header in the table you want, “The header I want.”

<div>
  <table>
   <tr><th>The header I want</th></tr>
   <tr><td>second row</td></tr>
   <tr><td>third row</td></tr>
  </table>
</div>
<div>
  <table>
   <tr><th>Another header I do NOT want</th></tr>
   <tr><td>second row</td></tr>
   <tr><td>third row</td></tr>
   <tr><td>fourth row</td></tr>
   <tr><td>fifth row</td></tr>
  </table>
</div>

To select the different headers, you would use //table/tr/th. You want the <table> so you can move the tr/th part inside a predicate and compare it with string “The header I want”.

Result

This kind of predicates also works for attribute values, e.g. testing links to some website:

//body//p [ a/@href="http://www.example.com" ]
Special trick for testing multiple node names

In a predicate you can also test the current node of the node-set. For example if you want to test for several wanted element names. This is when self:: axis can be helpful.

One example is testing different kind of lists, ordered or unordered:

Result

Note

Here we saw that predicates can also appear in the middle of the location path. Indeed, predicates are an (optional) part of each location step.

Nested predicates

We said that location paths can be used as predicate. And location paths can have predicates. So it’s possible to end up with nested predicates. (And that’s ok.)

>>> #                <------predicate --------->
>>> #                    <-nested predicate->
>>> doc.xpath('//div[p  [a/@href="page2.html"]  ]')
[(1) '<div>
      <p>This is a paragraph.</p>
      <p>Is this <a href="page2.html">a link</a>?</p>
      <br>
      Apparently.
    </div>']

In fact, the above is equivalent to //div[p/a/@href="page2.html"] with no nesting:

Result
Result

Order of predicates is important

You can have multiple predicates in sequence per step, each within its [] brackets, i.e. steps in the form of axis::nodetest[predicate#1][predicate#2][predicate#3]....

Predicates are processed in order, from left to right. And the output of one predicate is fed into the next predicate filter, much like steps produce node-sets for the next step to process.

So the order of predicates is important.

The following 2 location paths produce different results:

  • //div[@class="second"][2]: will output one <div>

  • //div[2][@class="second"]: will select nothing

See for yourself:

Result
Result

The second produces nothing indeed. Why is that?

//div[2][@class="second"] looks at div elements that are the 2nd child of their parent. div means child::div, and [2] will select the 2nd node in the current node-set. In our document this happens only once. The final predicate, [@class="second"], filters nodes that have a “class” attribute with value “second”. This happens to be valid for that 2nd child div.

On the contrary, //div[@class="second"][2] will first produce //div[@class="second"], which only produces single-node node-sets (again, there’s only one div with “class” attribute with value “second”). So the subsequent [2] predicate will never match with single-node node-sets (you cannot select the 2nd element of a 1-element list)

Warning

Beware of position() in chained predicates.

Much like [...][2] is different from [2][...], if you chain positional predicates, remember that the position is relative to the node-set processed by the previous predicate.

For example, we saw that position()>1 would filter out the first nodes in a node-set. Chaining [position()>1] will remove the first node each time it’s used:

Result

String functions

So far we have mostly selected nodes. XPath can also compute values from them, thanks to a small standard library of functions. The ones you will reach for the most when scraping work on strings.

Important

A node test like text() is not a function; it selects text nodes. string(), normalize-space(), contains()… on the other hand are real functions that take arguments and return a string, a number or a boolean.

string(): the string value of a node

string(node) returns the string value of a node: for an element, that is the concatenation of all its descendant text nodes, in document order. This is the easiest way to get “all the text under this element” as a single string, even when it is split by inline tags:

Result

Compare this with //p[2]/text(), which returns the direct text-node children separately and skips the text inside <a>:

>>> doc.xpath("string(//p[2])").get()
'Is this a link?'
>>> doc.xpath("//p[2]/text()").getall()
['Is this ', '?']

Warning

string() (and, as we will see, normalize-space()) applied to a node-set only looks at the first node in document order. Everything else is silently ignored:

>>> doc.xpath("string(//p)").get()  # first <p> only
'This is a paragraph.'

To get the string value of each node, apply the function per node instead, for instance by looping (see Loop on elements (table rows, lists)) or with .getall() on text().

normalize-space(): trim and collapse whitespace

HTML is full of insignificant whitespace: indentation, newlines, runs of spaces. normalize-space() strips leading and trailing whitespace and collapses every inner run of whitespace to a single space. It is invaluable for cleaning up extracted text:

Result

Called with no argument, normalize-space() operates on the string value of the context node, which makes it very handy inside a loop:

>>> for paragraph in doc.xpath("//p"):
...     print(repr(paragraph.xpath("normalize-space()").get()))
...
'This is a paragraph.'
'Is this a link?'

Testing the content: contains() and starts-with()

contains(haystack, needle) and starts-with(string, prefix) return a boolean. They are most useful inside predicates, to filter nodes on their text or on an attribute value.

Keep only the links whose href contains page2:

Result

A very common idiom combines contains() with the string value of an element (.) to find elements “containing some text”:

Result

Tip

contains(@class, "foo") is a classic way to match one class among many, but it also matches foobar. To match a whole class token, test against the class value padded with spaces: contains(concat(" ", normalize-space(@class), " "), " foo "). This is exactly what CSS selectors compile to (see CSS Selectors).

More string helpers

XPath 1.0 ships a few more string functions worth knowing:

Function

Returns

concat(s1, s2, ...)

The concatenation of its arguments.

string-length(s)

The number of characters in s.

substring(s, start[, length])

A substring of s (1-based positions).

substring-before(s, sep) / substring-after(s, sep)

The part of s before / after the first sep.

translate(s, from, to)

s with each character of from replaced by the character at the same position in to (a poor man’s lowercasing tool).

Note

XPath also has number functions (count(), sum(), round(), floor(), ceiling()) and boolean functions (not(), boolean(), true(), false()). We already met count() and last() in the Predicates section.

When a function returns a number, parsel gives it back to you as a string (e.g. count(//a) yields '2.0'); the in-browser widget above shows it as 2. Either way it is the same value.

Part 3: Use-cases for web scraping

Parts 1 and 2 covered XPath the language. This part is a cookbook of the patterns that come up again and again when you extract data from real pages with parsel and Scrapy.

Throughout this part we use parsel’s practical extraction API:

  • .get() returns the first match as a string (or None),

  • .getall() returns every match as a list of strings,

  • .attrib gives a mapping of the attributes of the first matched element.

Text extraction

There are two complementary ways to pull text out of an element:

  • text() selects the direct text-node children, each on its own,

  • string(.) (or normalize-space(.)) returns all the text underneath it, joined into one string.

>>> doc.xpath("//p[2]/text()").getall()  # direct text nodes, split
['Is this ', '?']
>>> doc.xpath("//p[2]//text()").getall()  # every descendant text node
['Is this ', 'a link', '?']
>>> doc.xpath("string(//p[2])").get()  # everything, joined
'Is this a link?'

Use //text() (or .//text() relative to an element) when you want to keep the pieces, and string(.) / normalize-space(.) when you want a single, clean string:

>>> doc.xpath('normalize-space(//div[@class="second"])').get()
'Nothing to add. Except maybe this other link.'

Tip

Prefer normalize-space() over calling Python’s str.strip() on the result: it also collapses the newlines and indentation that HTML sources are riddled with, in a single step.

Attributes extraction

To read an attribute value, select it with @name and extract it:

>>> doc.xpath("//a/@href").getall()
['page2.html', 'page3.html']
>>> doc.xpath("//a/@href").get()
'page2.html'

parsel also exposes attributes as a Python mapping through .attrib, which is often more readable when you already have an element in hand:

>>> link = doc.xpath("//a")[0]
>>> link.attrib["href"]
'page2.html'
>>> link.attrib.get("rel")  # missing attributes don't raise
>>> doc.css("a::attr(href)").getall()  # the CSS equivalent
['page2.html', 'page3.html']

Attribute names extraction

Sometimes you care about the attribute names rather than their values – for instance to discover data-* attributes. Select the attribute nodes with @* and read their name with the name() function:

>>> doc.xpath("//meta/@*").getall()  # attribute values
['text/html; charset=utf-8', 'content-type']
>>> doc.xpath("name(//meta/@*[1])").get()  # name of the first attribute
'content'

Loop on elements (table rows, lists)

Repeated structures – table rows, list items, cards – are the bread and butter of scraping. The pattern is always the same: select the repeating elements, then iterate and extract from each one with a relative XPath.

>>> products = parsel.Selector(text="""
...     <ul class="products">
...       <li><span class="name">Laptop</span><span class="price">1200</span></li>
...       <li><span class="name">Mouse</span><span class="price">25</span></li>
...     </ul>""")
...
>>> for li in products.xpath('//ul[@class="products"]/li'):
...     name = li.xpath('.//span[@class="name"]/text()').get()
...     price = li.xpath('.//span[@class="price"]/text()').get()
...     print(name, price)
...
Laptop 1200
Mouse 25

Warning

Inside the loop, the leading dot in .//span matters a lot. Without it, //span is absolute: it restarts from the document root and returns the same nodes on every iteration, ignoring the current li:

>>> for li in products.xpath("//li"):
...     print(li.xpath('//span[@class="name"]/text()').get())  # BUG: no dot
...
Laptop
Laptop

This is the single most common XPath mistake in scraping. When you loop, make your inner expressions relative with a leading ..

CSS Selectors

parsel lets you use CSS selectors too, via .css(). Under the hood they are translated to XPath by the cssselect library, so CSS is really a friendlier syntax for a subset of what XPath can do. You can even see the translation:

>>> from parsel import css2xpath
>>> css2xpath("div > a")
'descendant-or-self::div/a'

CSS is more concise for the common cases – classes, ids, descendants, child combinators – and parsel adds the non-standard ::text and ::attr(name) pseudo-elements so you can extract without dropping back to XPath:

>>> doc.css("div.second a::attr(href)").getall()
['page3.html']
>>> doc.css("title::text").get()
'This is a title'

You can also mix both languages by chaining, using each where it shines – CSS to reach a container, XPath for the parts CSS cannot express (text predicates, axes, positions):

>>> doc.css("div.second").xpath(".//a/@href").get()
'page3.html'

Note

CSS selectors cannot do everything XPath can: there is no way in CSS to select by text content, to walk up to an ancestor, or to reach preceding siblings. When you hit those walls, switch to XPath.

Element boundaries & XPath buckets (advanced)

Some pages are “flat”: instead of wrapping each group in its own container, they put headings and their content side by side as siblings.

<div class="content">
  <h2>Fruits</h2>
  <p>Apple</p>
  <p>Banana</p>
  <h2>Vegetables</h2>
  <p>Carrot</p>
</div>

There is no element that contains “everything under Fruits”, so a plain descendant query cannot group the paragraphs by heading. The trick is to identify each paragraph’s bucket by counting the headings before it with count(preceding-sibling::h2):

Result

Every <p> that has exactly one <h2> before it belongs to the first heading. Iterating over the headings gives you the groups:

>>> flat_html = """
...     <div class="content">
...       <h2>Fruits</h2>
...       <p>Apple</p>
...       <p>Banana</p>
...       <h2>Vegetables</h2>
...       <p>Carrot</p>
...     </div>"""
...
>>> sel = parsel.Selector(text=flat_html)
>>> for i, h2 in enumerate(sel.xpath("//h2"), start=1):
...     title = h2.xpath("normalize-space()").get()
...     items = sel.xpath(f"//p[count(preceding-sibling::h2)={i}]/text()").getall()
...     print(title, "->", items)
...
Fruits -> ['Apple', 'Banana']
Vegetables -> ['Carrot']

Related tools for the same problem are the sibling axes (following-sibling::, preceding-sibling::) and positional predicates, which we covered earlier.

EXSLT extensions

libxml2 (and therefore parsel, lxml and Scrapy) supports a couple of EXSLT extension namespaces on top of plain XPath 1.0. parsel registers them for you under two prefixes:

The most useful one is re:test(), which filters nodes with a regular expression – far more powerful than contains() or starts-with():

>>> links = parsel.Selector(
...     text="<html><body>"
...     '<a href="page2.html">a</a>'
...     '<a href="page3.html">b</a>'
...     '<a href="contact.html">c</a>'
...     "</body></html>"
... )
>>> links.xpath(r'//a[re:test(@href, "page\d")]/@href').getall()
['page2.html', 'page3.html']
>>> links.xpath(r'//a[re:test(@href, "PAGE\d", "i")]/@href').getall()  # case-insensitive
['page2.html', 'page3.html']

The set namespace helps when you need to subtract one node-set from another, for example to collect the itemprop of an element while excluding those that belong to nested scopes. See the selectors documentation for a worked set:difference example.

Note

EXSLT functions are not part of standard XPath 1.0, so the in-browser widget in this tutorial (which uses the browser’s native engine) does not understand them. They do work in parsel, lxml and Scrapy.

Summary of tips

Tip

  • Use relative XPath expressions whenever possible

  • Know your axes!

  • Don’t forget that XPath has string() and normalize-space() functions

  • text() is a node test, not a function call

  • CSS selectors are very handy, easier to maintain, but also less powerful than XPath