Compressed collections for simulated crawling

Collections are a fundamental tool for reproducible evaluation of information retrieval techniques. We describe a new method for distributing the document lengths and term counts (a.k.a. within-document frequencies) of a web snapshot in a highly compressed and nonetheless quickly accessible form. Our main application is reproducibility of the behaviour of focused crawlers: by coupling our collection with the corresponding web graph compressed with WebGraph we make it possible to apply text-based machine learning tools to the collection, while keeping the data set footprint small. Finally, we describe a collection based on a crawl of 100Mpages of the .uk domain, publicly available in bundle with a Java open-source implementation of our techniques.