Skip to content
Find file
Fetching contributors…
Cannot retrieve contributors at this time
429 lines (313 sloc) 15.3 KB


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.

Developed against Python 2.7, but should also work with Python 2.6+ if you install the importlib requirement.

See on Github <>, or on PyPi <>.

Hello World Usage

An extremely simple example of using extraction is:

>>> 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 source_url is optional in extract, but is recommended as it makes it possible to rewrite relative urls and image urls into absolute paths. source_url 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

If you want to develop against extraction, you can install from GitHub:

git clone
cd extraction
virtualenv env
. ./env/bin/activate
pip install -r requirements.txt
python develop

Then you can run the tests:

python tests/

All of which should pass in a sane installation.

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:

    <meta name="description" content="Will Larson&#39;s blog about programming and other things." />
    <link rel="alternate" type="application/rss+xml" title="Page Feed" href="/feeds/" />
    <link rel="canonical" href="">
    <title>Digg v4&#39;s Architecture and Development Processes - Irrational Exuberance</title>

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:

    <meta property="og:title" content="Something"/>
    <meta property="og:url" content=""/>
    <meta property="og:image" content=""/>
    <meta property="og:description" content="Something amazing."/>

as their source of data.


The HTML5 article tag, and also the video 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 HTML5SemanticTags is intentionally much more conservative than SemanticTags, as it provides high quality information in the small number of cases where it hits, and otherwise expects SemanticTags 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 img tags include images, most h1 and h2 tags include titles, and p 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="this_will_be_extracted_as_an_img.png">
    <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 SemanticTags 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.

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 extraction/examples directory.

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 Extracted instance (which are returned by Extractor.extract) 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 Extracted._unexpected_values dictionary:

>>> 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 title, url, image and description 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 "extraction.technique.FacebookOpengraphTags" is a valid representation of a technique.

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

from extraction import Extractor

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

Finally, the third option is to directly modify the techniques 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 H1 tag to represent the overall blogs title, and always uses the first H2 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 extract, which will be available in the Extracted()._unexpected_values 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 extraction/examples/, 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.

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.

Something went wrong with that request. Please try again.