The importance of being earnest (and average)

When the aim of a study is comparing and contrasting texts of the same genre and achieving a good arrangement for a text clustering, we often resort to lexical-based approaches and appropriate measures of similarity/distance (Burrows, 2002; Juola, 2008; Rudman, 1998, Stamatatos, 2009; Labbé and Labbé, 2001; Tuzzi, 2010) between texts, e.g. cosine similarity, Burrows's Delta, Labbé's intertextual distance, etc. Given the properties and the formula of a distance, we obtain a square matrix that includes n×n cells and n(n-1)/2 positive non-zero non-redundant values that can be exploited for an automatic classification of the n available texts. This distance matrix might be read from an alternative perspective, i.e. as a ranking system: for each text we can sort all the other n-1 texts from the closest to the furthest. The distribution of these ranks among texts represents an interesting object of research (Alvo and Yu, 2014) when we consider the whole corpus and also when we observe groups of texts that share some properties (e.g. they belong to the same author). A preliminary experiment involved a large corpus of contemporary Italian novels and showed that we can indentify some novels that systematically top positions in all rankings and prove to be close to most of the available texts; on the contrary, we have novels that do not show strong similarities in any list and systematically lie in the furthest positions. This study compared results achieved through different measures and formulated some hypothesis to understand when in text clustering it is worth either to distinguish "average" and "eccentric" novels or disregard them in in-depth investigations.