A package and script for extracting article text in html documents.
A python package for extracting article text in html documents. Check out this demo.
At a Glance
pip install eatiht ... easy_install eatiht
Note: On Windows, you may need to install lxml manually using: pip install lxml
Using in Python
import eatiht url = 'http://news.yahoo.com/curiosity-rover-drills-mars-rock-finds-water-122321635.html' print eatiht.extract(url)
NASA's Curiosity rover is continuing to help scientists piece together the mystery of how Mars lost its surface water over the course of billions of years. The rover drilled into a piece of Martian rock called Cumberland and found some ancient water hidden within it. Researchers were then able to test a key ratio in the water with Curiosity's onboard instruments...
Using as a command line tool:
eatiht http://news.yahoo.com/curiosity-rover-drills-mars-rock-finds-water-122321635.html >> out.txt
Note: Window’s users may have to add the C:directory to your “path” so that the command line tool works from any directory, not only the ..directory.
After searching through the deepest crevices of the internet for some tool|library|module that could effectively extract the main content from a website (ignoring text from ads, sidebar links, etc.), I was slightly disheartened by the apparent ambiguity caused by this content-extraction problem.
My survey resulted in some of the following solutions:
- boilerpipe - Boilerplate Removal and Fulltext Extraction from HTML pages. Java library written by Christian Kohlschütter
- “The Easy Way to Extract Useful Text from Arbitrary HTML” - a Python tutorial on implementing a neural network for html content extraction. Written by alexjc
- Pyteaser’s Cleaners module - from what I can tell, it’s a purely heuristic-based process
- “Text Extraction from the Web via Text-to-Tag Ratio” - a thesis on Text-to-Tag-heuristic driven clustering as a solution for the problem at hand. Written by Tim Weninger & William H. Hsu
The number of research papers I found on the subject largely outweighs the number available open-source projects. This is my attempt at balancing out the disparity.
In the process of coming up with a solution, I made two unoriginal observations:
- XPath’s select all (//), parent node (..) queries and functions (‘string-length’) are remarkably powerful when used together
- Unnecessary machine learning is unnecessary
By making an assumption on sentence length, and this is trivial, one can query for text-nodes satisfying said sentence length, then create a frequency distribution (histogram) across the parent-nodes, and the argmax of the resulting distribution is the xpath that is shared amongst likely sentences.
The results were surprisingly good. I personally prefer this approach to the others as it seems to lie somewhere in between the purely rule-based and the drowning-in-ML approaches.