Communications of the ACM
Technical Perspective: Building Knowledge Bases from Messy Data
Imagine the task of creating a database of all the high-quality specialty cafés around the world so you never have to settle for an imperfect brew. Relying on reviews from sites such as Yelp will not do the job because there is no restriction on who can post reviews there. You, on the other hand, are interested only in cafés that are reviewed by the coffee intelligentsia. There are several online sources with content relevant to your envisioned database. Cafés may be featured in well-respected coffee publications such as sprudge.com or baristamagazine.com, and data of more fleeting nature may pop up on your social media stream from coffee-savvy friends.
The task of creating such a database is surprisingly difficult. You would begin by deciding which attributes of cafés the database should model. Attributes such as address and opening hours would be obvious even to a novice, but you will need to consult a coffee expert who will suggest more refined attributes such as roast profile and brewing methods. The next step is to write programs that will extract structured data from these heterogeneous sources, distinguish the good extractions from the bad ones, and combine extractions from different sources to create tuples in your database. As part of the data cleaning process, you might want to employ crowd workers to confirm details, such as opening hours that were extracted from text or whether two mentions of cafés in text refer to the same café in the real world. In the extreme case, you might even want to send someone out to a café to check on some of the details in person. The process of creating the database is iterative because your extraction techniques will be refined and because the café scene changes frequently.
No entries found