Big data has the potential to change the world. The talent exists and the tools are already there. What’s lacking is access to data. Imagine the questions we could answer and the problems we could solve if talented, creative technologists could freely access more big data.
At Common Crawl, we are passionate about getting big open data into the hands of talented and creative people. Increasing access to data enables everything from business innovation to groundbreaking research.
Common Crawl is proud of what we have accomplished in 2014 thanks to our dedicated team and the support of donors like you.
19 academic publications using Common Crawl data were published
Several Open Educational Resources designed to teach big data tools and methods were created
1.3 petabytes containing 18.5 billion web pages were added to the Common Crawl corpus
Numerous presentations and tutorials were given at international conferences, local meetup groups, and academic workshops in six countries
100% of our funding comes from donors like you — Thank you! We accomplish a great deal with a small, dedicated staff on a limited budget so your investment in us has a big impact.
More resources for Common Crawl means more access to data for everyone. In 2015 we have big plans to scale up crawling to more rapidly increase the Common Crawl corpus and to grow our educational programs and catalogue of tutorials in order to invest in the next generation of data-driven technologists.
Whether it’s $10 or $10,000, Common Crawl needs your donation today.
Oskar Singer is a Software Developer and Computer Science student at University of Massachusetts Amherst. He recently did some very interesting text analytics work during his internship at Lexalytics . The post below describes the work, how Common Crawl data was used, and includes a link to code.
At Lexalytics, I have been working with our head of software engineering, Paul Barba, on improving our accuracy with Twitter data for POS-tagging, entity extraction, parsing and ultimately sentiment analysis through building an interesting model-based approach for handling misspelled words.
Our approach involves a spell checker that automatically corrects the input text internally for the benefit of the engine and outputs the original text for the benefit of the engine user, so this must be a different kind of automated spell-correction.
The First Attempt:
Our first attempt was to take the top scoring word from the list of unranked correction suggestions provided by Hunspell, an open-source spell checking library. We calculated each suggestion’s score as word frequency from Common Crawl data divided by string edit distance with consideration for keyboard distance.
The resulting corrections were scored against hand-corrected tweets by counting the number of tokens that differed. Hunspell scored worse than the original tweets. It corrected usernames and hashtags and gave totally unreasonable suggestions. My favorite Hunspell correction was the mapping from “ur” (as in the short-form for “your” or “you’re”) to “Ur” (as in the ancient Mesopotamian city-state).
Hunspell also missed mistakes like misused homophones, which did not count as a misspelling when considered in isolation. This last issue seemed to be the primary issue with our data, so the problem required a method with the ability to consider context.
The Second (and final) Attempt:
We title the next attempt “the Switchabalizer”, and it can be summarized as a multinomial, sliding-window, Naive-Bayes word classifier. On a high level, we classify each of the target words in a piece of text, based on the preceding and succeeding words, as itself or one of its homophones.
The training process starts with a list of bigrams from the Common Crawl data paired with their occurrence counts. We use this data to calculate P(wi-1 | wi) = #(wi-1wi)/#(wi-1) and P(wi+1 | wi) = #(wiwi+1)/#(wi+1) where wi is the current word, wi-1 is the preceding word and wi+1 is the succeeding word. These probabilities are serialized and archived so they can be deserialized into C++ data structures instead of recalculated for each instantiation of the spell check object. In other words, we’re building a set of probabilities that each switchable “generated” the words preceding and succeeding wi.
The inference process starts with a set S of sets and an inverted index. Each s ∈ S represents a group of commonly confused homophones (e.g. two, too, 2, to), and no word is a member of multiple s ∈ S. The inverted index maps each word w in the union of all s ∈ S to the s in which w holds membership. Each word wi in the ordered sequence of words in a document is checked for an entry in the inverted index. If an entry V is found, the algorithm replaces wi with argmaxv∈V P(v) = P(wi-1 | v) + P(wi+1 | v).
As a matter of efficiency, we assumed that Wikipedia articles have perfect use of the target homophones. I wrote a Python script that took in text, randomly replaced target homophones with members of their switchable set, then output the result.
We ran the Switchabalizer on this data and compared to the original Wikipedia data. Comparing the corrections to the words changed by our test generator, Hunspell, even when forced to ignore usernames, had a 216% error rate (i.e. it made false corrections), and the Switchabalizer had a 20% error rate. Although the test data does not match the target data, the massive and varied data set provided by Common Crawl should ensure good results from the Switchabalizer on many types of data, hopefully even the near-nonsense from the bowels of Twitter.
The Switchabalizer approach is clearly superior to a traditional spell checker for our targeted issues, but still requires significant testing, tuning and improvement. The following section provides some possibilities for improvement and expansion. We hope this approach can be of use to other people with the same problem, and we would like to thank Common Crawl for the fantastic resource that they provide!
Possible future experiments include further testing on different types of data, integration of higher-order n-gram features, implementation of a discriminative model, implementation for other languages, and corrections of common misspellings like “ur”, which cannot be included in sets of switchables without risking the model mapping words to non-words.
The second crawl of 2013 is now available! In late November, we published the data from the first crawl of 2013 (see previous blog post for more detail on that dataset). The new dataset was collected at the end of 2013, contains approximately 2.3 billion webpages and is 148TB in size. The new data is located in the commoncrawl bucket at /crawl-data/CC-MAIN-2013-48/
In 2013, we made changes to our crawling and post-processing systems. As detailed in the previous blog post, we switched file formats to the international standard WARC and WAT files. We also began using Apache Nutch to crawl – stay tuned for an upcoming blog post on our use of Nutch. The new crawling method relies heavily on the generous data donations from blekko and we are extremely grateful for blekko’s ongoing support!
In 2014 we plan to crawl much more frequently and publish fresh datasets at least once a month.
I am extremely happy to announce that Professor Jim Hendler has joined the Common Crawl Advisory Board. Professor Hendler is the Head of the Computer Science Department at Rensselaer Polytechnic Institute (RPI) and also serves as the Professor of Computer and Cognitive Science at RPI’s Tetherless World Constellation.
Jim Hendler is a highly respected leader and an early innovator of the Semantic Web. In fact, he has been writing about it for over a decade – since before most of us had even heard the term. The 2001 article in Scientific American that he coauthored with Tim Berners Lee and Ora Lassila has been cited over 15,000 times and to this day is one of the very best explanations of the potential of the Semantic Web. He is one of the editors of Synthesis Lectures on the Semantic Web where he recently published Aaron Swartz’s A Programmable Web: An Unfinished Work. Aaron Swartz’s book is available as a free download. I strongly encourage everyone to read it and to spread the word about it so it reaches as many people as possible.
Professor Hendler is also a strong advocate for open government data and has pushed that movement forward through his work with the data.gov project and his Linking Open Government Data project. His Twitter feed is an excellent source of information about open government data and about all of the important and exciting work he does.
Having Professor Hendler’s insight and guidance will be a tremendous benefit to Common Crawl and everyone on the team is very excited that he has joined us!
URL Search is a web application that allows you to search for any URL, URL prefix, subdomain or top-level domain. The results of your search show the number of files in the Common Crawl corpus that came from that URL and provide a downloadable JSON metadata file with the address and offset of the data for each URL. Once you download the JSON file, you can drop it into your code so that you only run your job against the subset of the corpus you specified. URL Search makes it much easier to find the files you are interested in and significantly reduces the time and money it take to run your jobs since you can now run them across only on the files of interest instead of the entire corpus.
We are excited to see examples of URL Search in action. Are you working with Common Crawl data? Would you like to win $100 in AWS credit for sharing how URL Search makes your life easier? The first five people who share open source code on GitHub that incorporates a JSON file from URL Search will each get $100 in AWS Credit!
Email a link to the GitHub repo to [email protected] for consideration. The code must be accompanied by a ReadMe file that explains. If you would like to write a guest blog post about your work we would be happy to publish it on the Common Crawl blog.
We are very excited to announce that the winners of the Norvig Web Data Science Award Lesley Wevers, Oliver Jundt, and Wanno Drijfhout from the University of Twente!
The Norvig Web Data Science Award was created by Common Crawl and SURFsara to encourage research in web data science and named in honor of distinguished computer scientist Peter Norvig.
There were many excellent submissions that demonstrated how you can extract valuable insight and knowledge from web crawl data. Be sure to check out the work of the winning team, Traitor – Associating Concepts Using The World Wide Web, and the other finalists on the award website. You will find descriptions of the projects as well as links to the code that was used. We hope that these projects will serve as an inspiration for what kind of work can be done with the Common Crawl corpus. All code is open source and we are looking forward to seeing it reused and adapted for other projects.
Jason is the Associate Head of Digital Library Initiatives at North Carolina State University Libraries. He used the Common Crawl Index to look at NCSU Library URLs in the Common Crawl Index. You can see his description of his work and results below and on his blog. Be sure to follow Jason on Twitter and on his blog to keep up to date with other interesting work he does!
Common Crawl URL Index
The Common Crawl now has a URL index available. While the Common Crawl has been making a large corpus of crawl data available for over a year now, if you wanted to access the data you’d have to parse through it all yourself. While setting up a parallel Hadoop job running in AWS EC2 is cheaper than crawling the Web, it still is rather expensive for most. Now with the URL index it is possible to query for domains you are interested in to discover whether they are in the Common Crawl corpus. Then you can grab just those pages out of the crawl segments.
Scott Robertson, who was responsible for putting the index together, writes in the github README about the file format used for the index and the algorithm for querying it. If you’re interested you can read the details there.
Note that because of how the index is constructed you’ll be querying for domains in reverse order. This allows you scope your queries to match everything from a TLD down to a specific subdomain. This will return every URL matching under http://lib.ncsu.edu as well as any subdomains like http://d.lib.ncsu.edu.
As I write this, the index is only partial, while folks provide feedback on the index, so your current results may not reflect everything that is currently in the Common Crawl corpus.
The result is a line delimited file with information about one URL on each line. A space separates the URL from some JSON-like data. (You’ll need to convert the single quotes to double quotes for it to parse as JSON, or just eval the data with Python if you are filled with trust or like to live dangerously.) Again, the URL hostname is in reverse order followed by the path in normal order and finally the protocol. The data is a pointer to the location for the file within a segment of the common crawl dataset. This information can be used toretrieve the page from AWS S3.
What I’m interested in is what NCSU Libraries URLs are represented in the index. In total the URL index has 4033 URLs that all look to be from a crawl in early September. Here’s the breakdown for subdomains:
Query for edu.ncsu.lib in Common Crawl URL IndexFull Text
The results here are interesting as I’m always trying to raise the discoverability of NCSU Libraries’ digital collections. At the top of the list is the main web site for NCSU Libraries. The hostnames www.lib.ncsu.edu and lib.ncsu.edu both point to the same resources. Looking closer we find that of the 2427 URLs there, many are for digital collections related pages. 636 are under the Special Collections Research Center, and some of these are pages for some legacy collections. 407 URLs are for pages in our collection guides application, many of them for individual guides or, strangely the EAD XML for the guides. Some of those collection guides do link to online digital collections.
The institutional repository (Dspace instances) is also well represented at the top of this list. The Technical Reports Repository accounts for 159 of those URLs, and the NCSU Institutional Repository accounts for just 3. The digital collections in the repository, mainly special collections, accounts for 626 URLs. 719 of the 801 repository URLs are directly to the PDFs. Evidently the PDFs rank higher than the landing pages.
NCSU Libraries has been providing Geospatial Data Services and paying attention to SEO for those pages for a long time, so it isn’t completely surprising that this directory of files has gotten indexed: http://geodata.lib.ncsu.edu. (Note that this server may not be accessible from off-campus.) Many of the URLs under www.lib.ncsu.edu are also GIS pages, so GIS data services and collections pages are even better represented–and human-friendly–than at first appears.
http://d.lib.ncsu.edu/ This is the root page of a subdomain that includes a growing number of digital collections sites. This index page was just updated to be more than a single unstyled link.
http://d.lib.ncsu.edu/collections/ The home page for NCSU Libraries’ Digital Collections: Rare and Unique Materials which includes over 63,700 resources. It has been a focus of my own work to try to improve the discoverability of this content on the open Web. I implemented embedded semantic markup, Microdata and Schema.org, on this site.
http://d.lib.ncsu.edu/collections/catalog/0228376 This is an image of Mary Travers singing live on stage. Looking in Google Analytics for this page as a landing page for referrals, Google is the top referrer. Since Google is not likely to have been crawled to discover this URL, it is more likely that the next referrer is responsible for this getting in the index. This post on the Peter, Paul & Mary Love Tumblr was reblogged and liked a number of times. That particular post is the only one from this Tumblr which is in the Common Crawl index.
So it appears that the Common Crawl probably hasn’t (at least in this half of the index!), decided to crawl this site to any extent. Instead it appears it is only deciding to crawl pages that have been linked up. Once the rest of the index comes out, I’ll have to take a look, and consider how to improve that number. The key though is obviously getting more links into the site.
Further down in the list there are a bunch of funny looking URLs. I think these are all proxy URLs for user authentication to restricted resources.
While the Common Crawl URL index is useful if you need the whole page, in many cases just extracted embedded semantic markup may be enough. The Web Data Commons is already extracting Microdata and RDFa data, and makes indexes available, though it takes a bit more effort to parse through their indexes. (I’d like a service or script to query for an N-Quad context and get back all the related triples. Anyone know if there is already such a service? Do I have to write one?) They do have a helpful page on how to download the extracted data in whole or in part.
First, how much of your content is in the Common Crawl corpus? I’d be interested in hearing what your results are like.
We need to figure out how to get more cultural heritage content crawled and indexed by the Common Crawl. Without our stuff in the Common Crawl we are missing many opportunities to broaden the reach of our content. It doesn’t appear that Common Crawl accepts sitemaps. It works off of page rank and the link graph of popular sites. While my sites for rare and unique digital collections get most of their traffic from search engines, mainly Google, an increasing amount of traffic is due to referrals. Referrals, links from other sites, seem like the key for getting our stuff into the corpus. Efforts to add links to library special or digital collections to appropriate Wikipedia articles and the like would seem to be a good starting point.
Social sites are in the corpus and may also be a good way to get inbound links to our collections. There are 134,928+ Pinterest URLs in the Common Crawl index, and folks are actively pinning content from d.lib.ncsu.edu. Will the content pinned and repinned on Pinterest begin showing up in the crawl? Where else are crawlers likely to find links from people who make use of our content?
If more cultural heritage content is a part of the index, then there are all sorts of things we can begin to do. For web archiving projects it would be possible to begin with data in the corpus, potentially saving some crawling expense. New targeted search engines (or aggregations) can be created for different slices of content. Implement Microdata (or RDFa Lite) with Schema.org vocabularies and richer metadata can be extracted from your pages by the Web Data Commons and understood by many. This data can then be used in a variety of interfaces to save the time of the user in finding the content they really want.
What are some other ways that libraries, archives, and museums might be able to use the Common Crawl?
You can see the simple Ruby scripts I used for parsing the Common Crawl URL index out and the Web Data Commons N-Quads in this gist.