The Anatomy of a Web Scraping Tool

I recently spoke with a resource-limited organization that is investigating government corruption and wants to access various public datasets to monitor politicians and law firms. They don’t have developers in-house, but feel pretty comfortable analyzing datasets in CSV form. While many public datasources are available in structured form, some sources are hidden in what us data folks call the deep web. Amazon is a nice example of a deep website, where you have to enter text into a search box, click on a few buttons to narrow down your results, and finally access relatively structured data (prices, model numbers, etc.) embedded in HTML. Amazon has a structured database of their products somewhere, but all you get to see is a bunch of webpages trapped behind some forms.

A developer usually isn’t hindered by the deep web. If we want the data on a webpage, we can automate form submissions and key presses, and we can parse some ugly HTML before emitting reasonably structured CSVs or JSON. But what can one accomplish without writing code?

This turns out to be a hard problem. Lots of companies have tried, to varying degrees of success, to build a programmer-free interface for structured web data extraction. I had the pleasure of working on one such project, called Needlebase at ITA before Google acquired it and closed things down. David Huynh, my wonderful colleague from grad school, prototyped a tool called Sifter that did most of what one would need, but like all good research from 2006, the lasting impact is his paper rather than his software artifact.

Below, I’ve compiled a list of some available tools. The list comes from memory, the advice of some friends that have done this before, and, most productively, a question on Twitter that Hilary Mason was nice enough to retweet.

The bad news is that none of the tools I tested would work out of the box for the specific use case I was testing. To understand why, I’ll break down the steps required for a working web scraper, and then use those steps to explain where various solutions broke down.

The anatomy of a web scraper

There are three steps to a structured extraction pipeline:

{C}1.      Authenticate yourself. This might require logging in to a website or filling out a CAPTCHA to prove you’re not…a web scraper. Because the source I wanted to scrape required filling out a CAPTCHA, all of the automated tools I’ll review below failed step 1. It suggests that as a low bar, good scrapers should facilitate a human in the loop: automate the things machines are good at automating, and fall back to a human to perform authentication tasks the machines can’t do on their own.

{C}2.      Navigate to the pages with the data. This might require entering some text into a search box (e.g., searching for a product on Amazon), or it might require clicking “next” through all of the pages that results are split over (often called pagination). Some of the tools I looked at allowed entering text into search boxes, but none of them correctly handled pagination across multiple pages of results.

{C}3.      Extract the data. On any page you’d like to extract content from, the scraper has to help you identify the data you’d like to extract. The cleanest example of this that I’ve seen is captured in a video for one of the tools below: the interface lets you click on some text you want to pluck out of a website, asks you to label it, and then allows you to correct mistakes it learns how to extract the other examples on the page.

So that’s it? Nothing works?

Not quite. None of these tools solved the problem I had on a very challenging website: the site clearly didn’t want to be crawled given the CAPTCHA, and the javascript-submitted POST requests threw most of the tools that expected navigation through links for a loop. Still, most of the tools I reviewed have snazzy demos, and I was able to use some of them for extracting content from sites that were less challenging than the one I initially intended to scrape.

All hope is not lost, however. Where pure automation fails, a human can step in. Several proposals suggested paying people on oDesk, Mechanical Turk, or CrowdFlower to extract the content with a human touch. This would certainly get us past the CAPTCHA and hard-to-automate navigation. It might get pretty expensive to have humans copy/paste the data for extraction, however. Given that the tools above are good at extracting content from any single page, I suspect there’s room for a human-in-the-loop scraping tool to steal the show: humans can navigate and train the extraction step, and the machine can perform the extraction. I suspect that’s what is up to, and I’m hopeful they keep the tool available to folks like the ones I initially tried to help.

While we’re on the topic of human-powered solutions, it might make sense to hire a developer on oDesk to just implement the scraper for the site this organization was looking at. While a lot of the developer-free tools I mentioned above look promising, there are clearly cases where paying someone for a few hours of script-building just makes sense.

ARTICLE SOURCE: This factual content has not been modified from the source. This content is syndicated news that can be used for your research, and we hope that it can help your productivity. This content is strictly for educational purposes and is not made for any kind of commercial purposes of this blog.