Need help with extraction?
Click the “chat” button below for chat support from the developer who created it, or find similar developers for support.

About the developer

139 Stars 34 Forks MIT License 60 Commits 6 Opened issues


A Python library for extracting titles, images, descriptions and canonical urls from HTML.

Services available


Need anything else?

Contributors list

# 330,035
44 commits
# 480,780
11 commits
# 41,587
1 commit



Extraction is a Python package for extracting titles, descriptions, images and canonical urls from web pages. You might want to use Extraction if you're building a link aggregator where users submit links and you want to display them (like submitting a link to Facebook, Digg or Delicious).

Extraction is not a web crawling or content retrieval mechanism, rather it is a tool to use on data which has always been retrieved or crawled by a different tool.

Updated to work with Python3.

See the last Python 2.x compatible version here 

See on

, or on

Hello World Usage

An extremely simple example of using

>>> import extraction
>>> import requests
>>> url = ""
>>> html = requests.get(url).text
>>> extracted = extraction.Extractor().extract(html, source_url=url)
>>> extracted.title
>>> "Social Hierarchies in Engineering Organizations - Irrational Exuberance"
>>> print extracted.title, extracted.description, extracted.image, extracted.url
>>> print extracted.titles, extracted.descriptions, extracted.images, extracted.urls

Note that

is optional in extract, but is recommended as it makes it possible to rewrite relative urls and image urls into absolute paths.
is not used for fetching data, but can be used for targetting extraction techniques to the correct domain.

More details usage examples, including how to add your own extraction mechanisms, are beneath the installation section.


The simplest way to install Extraction is via PyPi::

pip install extraction

You'll also have to install a parser for

, and while
already pulls down
through it's requirements, I really recommend installing
_ as well, because there are some extremely gnarly issues with
failing to parse XHTML pages (for example, PyPi fails to parse entirely with html5lib::
>>> bs4.BeautifulSoup(text, ["html5lib"]).find_all("title")
>>> bs4.BeautifulSoup(text, ["lxml"]).find_all("title")
[extraction 0.1.3 : Python Package Index]

You should be able to install

_ via pip::
pip install lxml

If you want to develop extraction, then after installing

, you can install from GitHub::
git clone
cd extraction
python3 -m venv env    
. ./env/bin/activate
pip install -r requirements.txt
pip install -e .

Then you can run the tests::

python tests/

All of which should pass in a sane installation.

Using Extraction

This section covers various ways to use extraction, both using the existing extraction techniques as well as add your own.

For more examples, please look in the


Basic Usage

The simplest possible example is the "Hello World" example from above::

>>> import extraction
>>> import requests
>>> url = ""
>>> html = requests.get(url).text
>>> extracted = extraction.Extractor().extract(html, source_url=url)
>>> extracted.title
>>> "Social Hierarchies in Engineering Organizations - Irrational Exuberance"
>>> print extracted.title, extracted.description, extracted.image, extracted.url

You can get the best title, description and such out of an

instance (which are returned by
) by::
>>> print extracted.title
>>> print extracted.description
>>> print extracted.url
>>> print extracted.image
>>> print extracted.feed

You can get the full list of extracted values using the plural versions::

>>> print extracted.titles
>>> print extracted.descriptions
>>> print extracted.urls
>>> print extracted.images
>>> print extracted.feeds

If you're looking for data which is being extracted but doesn't fall into one of those categories (perhaps using a custom technique), then take a look at the

>>> print extracted._unexpected_values

Any type of metadata which isn't anticipated is stored there (look at

Subclassing Extracted to Extract New Types of Data
if this is something you're running into frequently).

Using Custom Techniques and Changing Technique Ordering

The order techniques are run in is significant, and the most accurate techniques should always run first, and more general, lower quality techniques later on.

This is because titles, descriptions, images and urls are stored internally in a list, which is built up as techniques are run, and the

properties simply return the first item from the corresponding list.

Techniques are represented by a string with the full path to the technique, including its class. For example

is a valid representation of a technique.

The default ordering of techniques is within the extraction.Extractor's

class variable, and is::

You can modify the order and inclusion of techniques in three ways. First, you can modify it by passing in a list of techniques to the optional

parameter when initializing an extraction.Extractor::
>>> techniques = ["my_module.MyTechnique", "extraction.techniques.FacebookOpengraphTags"]
>>> extractor = extraction.Extractor(techniques=techniques)

The second approach is to subclass Extractor with a different value of

from extraction import Extractor

class MyExtractor(Extractor): techniques = ["my_module.MyTechnique"]

Finally, the third option is to directly modify the

class variable. This is probably the most unpredictable technique, as it's possible for mutiple pieces of code to perform this modification and to create havoc, if possible use one of the previous two techniques to avoid future debugging::
>>> import extraction
>>> extraction.Extractor.techniques.insert(0, "my_module.MyAwesomeTechnique")
>>> extraction.Extractor.techniques.append("my_module.MyLastReportTechnique")

Again, please try the first two techniques instead if you value sanity.

Writing New Technique

It may be that you're frequently parsing a given website and aren't impressed with how the default extraction techniques are performing. In that case, consider writng your own technique.

Let's take for example a blog entry at 
_, which uses the
tag to represent the overall blogs title, and always uses the first
tag in
for its actual title.

A technique to properly extract this data would look like::

from extraction.techniques import Technique
from bs4 import BeautifulSoup
class LethainComTechnique(Technique):
    def extract(self, html):
        "Extract data from"
        soup = BeautifulSoup(html)
        page_div = soup.find('div', class_='page')
        text_div = soup.find('div', class_='text')
        return { 'titles': [page_div.find('h2').string],
                 'dates': [page_div.find('span', class_='date').string],
                 'descriptions': [" ".join(text_div.find('p').strings)],
                 'tags': [x.find('a').string for x in page_div.find_all('span', class_='tag')],
                 'images': [x.attrs['src'] for x in text_div.find_all('img')],

To integrate your technique, take a look at the

Using Custom Techniques and Changing Technique Ordering
section above.

Adding new techniques incorporating microformats is an interesting area for some consideration. Most microformats have very limited usage, but where they are in use they tend to be high quality sources of information.

Subclassing Extracted to Extract New Types of Data

Your techniques can return non-standard keys in the dictionary returned by

, which will be available in the
dictionary. In this way you could fairly easily add support for extracting addresses or whatnot.

For a contrived example, we'll extract my address from 
_, which is in no way a realistic example of extracting an address, and is only meant as an example of how to add a new type of extracted data.

As such, to add support for extracting address should look like (a fuller, commented version of this example is available in

, I've written this as concisely as possible to fit into this doc more cleanly)::
from extraction.techniques import Technique
from extraction import Extractor, Extracted
from bs4 import BeautifulSoup

class AddressExtracted(Extracted): def init(self, addresses=None, args, **kwargs): self.addresses = addresses or [] super(AddressExtracted, self).init(*args, *kwargs)

def address(self):
    return self.addresses[0] if self.addresses else None

class AddressExtractor(Extractor): "Extractor which supports addresses as first-class data." extracted_class = AddressExtracted text_types = ["titles", "descriptions", "addresses"]

class AddressTechnique(Technique): def extract(self, html): "Extract address data from" soup = BeautifulSoup(html) return {'addresses': [" ".join(soup.find('div', id='address').strings)]}

Usage would then look like::

>>> import requests
>>> from extraction.examples.new_return_type import AddressExtractor
>>> extractor = AddressExtractor()
>>> extractor.techniques = ["extraction.examples.new_return_type.AddressTechnique"]
>>> extracted = extractor.extract(requests.get(""))
>>> extracted.address
"Cole Valey San Francisco, CA USA"

There you have it, extracted addresses as first class extracted data.

Passing Parameters to Techniques

There isn't a mechanism for passing parameters to Techniques when they are initialized, but it is possible to customize the behavior of Techniques in a couple of ways.

First, you can simply subclass the Technique with the specific behavior you want, perhaps pulling the data from Django settings or what not::

class MyTechnique(Technique):
    def __init__(self, *args, **kwargs):
        if 'something' in kwargs:
            self.something = kwargs['something']
        del kwargs['something']
            self.something = "something else"
        return super(MyTechnique, self).__init__(*args, **kwargs)

def extract(html, source_url=None):
    print self.something
    return super(MyTechnique, self).extract(html, source_url=source_url)

Second, all techniques are passed in the Extractor being used to process them, so you can bake the customization into an extraction.Extractor subclass::

from extraction import Extractor
from extraction.techniques import Technique

class MyExtractor(Extractor): techniques = ["my_module.MyTechnique"] def init(self, something, args, **kwargs): self.something = something super(MyExtractor, self).init(*args, *kwargs)

class MyTechnique(Technique): class extract(self, html, source_url=None): print self.extractor.something return super(MyTechnique, self).extract(html, source_url=source_url)

Between these two techniques, it should be feasible to get the customization of behavior you need.

Extraction Techniques

This section lists the current techniques used by extraction. To rerank the techniques, remove techniques or add new techniques of your own, look at the

Using Extraction
section below.


Every webpage's head tag contains has a title tag, and many also include additional data like descriptions, RSS feeds and such. This technique parses data that looks like::

    Digg v4's Architecture and Development Processes - Irrational Exuberance

While the head tag is authoritative source of canonical URLs and RSS, it's often very hit or miss for titles, descriptions and such. At worst, it's better than nothing.


For better or for worse, the highest quality source of page data is usually the

Facebook Opengraph meta tags 
_. This technique uses Opengraph tags, which look like this::

as their source of data.


Another, increasingly common set of meta tags is the

Twitter Card tags 
_. This technique parses those tags, which look like::

For sites with cards integration (which many high quality sites have, because it's necessary for rendering with images in the Twitter feed), this will be a very high quality source of data.

One oddity is that Twitter cards don't include a URL tag, so they don't help much with canonicalizing articles.



tag, and also the
tag give us some useful hints for extracting page information for the sites which happen to utilize these tags.

This technique will extract information from pages formed like::

<h1>This is not a title to HTML5SemanticTags</h1>
  <h1>This is a title</h1>
  <p>This is a description.</p>
  <p>This is not a description.</p>
  <source src="this_is_a_video.mp4">

Note that

is intentionally much more conservative than
, as it provides high quality information in the small number of cases where it hits, and otherwise expects
to run sweep behind it for the lower quality, more abundant hits it discovers.


This technique relies on the basic tags themselves--for example, all

tags include images, most
tags include titles, and
tags often include text usable as descriptions::
<h1>This will be extracted as a title.</h1>
<h2>So will this, but after all H1s.</h2>
<img src="">
<p>And this as a description.</p>
<p>This as another possible description.</p>
<p>This as a third possible description.</p>

There is a limit, defined within

of how many tags of a given type will be consumed, and is usually 3-5, with the exception of images, where it is 10 (as this is actually a valid way to detect images, unlike the others).

This is a true last resort technique.

Implementation Details

I've tried to comment the classes and modules themselves in a fairly indepth fashion, and would recommend reading them for the most details, the recommended reading order is::


Hopefully all questions are answered therein.

Contributions, Questions, Concerns

Please open a GitHub pull-request with any improvements, preferably with tests, and I'll be glad to merge it in.

We use cookies. If you continue to browse the site, you agree to the use of cookies. For more information on our use of cookies please see our Privacy Policy.