At the time I joined Metaweb Technologies, Inc. in 2008, we were building up Freebase in earnest; entity by entity, fact by fact. Now you may know Freebase through its newest incarnation, Google's Knowledge Graph, which powers the "Knowledge panels" on www.google.com.

Building up "the world's database of everything" is a tall order that machines and algorithms alone cannot do, even if raw public domain data exists in abundance. Raw data from multiple sources must be cleaned up, homogenized, and then reconciled with data already in Freebase. Even that first step of cleaning up the data cannot be automated entirely; it takes the common sense of a human reader to know that if both 0.1 and 10,000,000 occur in a column named cost, they are very likely in different units (perhaps millions of dollars and dollars respectively). It also takes a human reader to decide that UCBerkley means the same as University of California in Berkeley, CA, but not the same as Berkeley DB.

If these errors occur often enough, we might as well have given up or just hired enough people to perform manual data entry. But these errors occur often enough to be a problem, and yet not often enough that anyone who has not dealt with such data thinks simple automation is sufficient. But, dear reader, you have dealt with data, and you know how unpredictably messy it can be.

Every dataset that we wanted to load into Freebase became an iterative exercise in programming mixed with manual inspection that led to hard-coding transformation rules, from turning two-digit years into four-digits, to swapping given name and surname if there is a comma in between them. Even for most of us programmers, this exercise got old quickly, and it was painful to start every time.

So, we created Freebase Gridworks, a tool for cleaning up data and making it ready for loading into Freebase. We designed it to be a database-spreadsheet hybrid; it is interactive like spreadsheet software and programmable like databases. It was this combination that made Gridworks the first of its kind.

In the process of creating and then using Gridworks ourselves, we realized that cleaning, transforming, and just playing with data is crucial and generally useful, even if the goal is not to load data into Freebase. So, we redesigned the tool to be more generic, and released its Version 2 under the name "Google Refine" after Google acquired Metaweb.

Since then, Refine has been well received in many different communities; data journalists, open data enthusiasts, librarians, archivists, hacktivists, and even programmers and developers by trade. Its adoption in the early days spread through word of mouth, in hackathons and informal tutorials held by its own users.

Having proven itself through early adopters, Refine now needs better organized efforts to spread and become a mature product with a sustainable community around it. Expert users, open source contributors, and data enthusiast groups are actively teaching how to use Refine on tours and in the classroom. Ruben and Max from the Free Your Metadata team have taken the next logical step in consolidating those tutorials and organizing those recipes into this handy missing manual for Refine.

Stepping back to take in the bigger picture, we may realize that messy data is not anyone's own problem, but it is more akin to ensuring that one's neighborhood is safe and clean. It is not a big problem, but it has implications on big issues such as transparency in government. Messy data discourages analysis and hides real-world problems, and we all have to roll up our sleeves to do the cleaning.

David Huynh

Original creator of OpenRefine