Comparing web-crawled and traditional corpora

Using a multi-dimensional (MD) analysis of register variability, the study compares two corpora of Czech: Koditex, a "traditional" corpus carefully designed using various sources with rich metadata, and Araneum Bohemicum Maximum, a web-crawled corpus with an opportunistic composition representative of the "searchable" web. Both types of corpora are projected onto the space induced by the MD model, with the main objective being to find out whether they overlap in the linguistic variation they cover, or whether one introduces some specific variation which cannot be found in the other. We also document a crucial methodological point which has broader relevance for MD analyses in general, namely that texts have to be of similar lengths in order for their scores on the dimensions to be comparable. Results indicate that some traditional text categories, such as journalism or non-fiction, are characterized by language phenomena which are equally well covered by web-crawled data, though of course traditional corpora keep their edge in terms of the richness of the accompanying metadata. But overall, the range of variation in Koditex is broader as it contains texts which have no adequate substitute (i.e. texts with a comparable set of linguistic characteristics, regardless of their extratextual label) in data acquired through general-purpose web-crawling techniques. These include informal conversations, private correspondence, some types of fiction, but also user-generated content (comments on Facebook, forums etc.).