Explaining Delta, or: How do distance measures for authorship attribution work?

Authorship Attribution is a research area in quantitative text analysis concerned with attributing texts of unknown or disputed authorship to their actual author based on quantitatively measured linguistic evidence (see Juola 2006; Stamatatos 2009; Koppel et al. 2009). Authorship attribution has applications in literary studies, history, forensics and many other fields, e.g. corpus stylistics (Oakes 2009). The fundamental assumption in authorship attribution is that individuals have idiosyncratic habits of language use, leading to a stylistic similarity of texts written by the same person. Many of these stylistic habits can be measured by assessing the relative frequencies of function words or parts of speech, vocabulary richness, and many other linguistic features. Distance metrics between the resulting feature vectors indicate the overall similarity of texts to each other, and can be used for attributing a text of unknown authorship to the most similar of a (usually closed) set of candidate authors. The aim of this paper is to present findings from a larger investigation of authorship attribution methods which centres around the following questions: (a) How and why exactly does authorship attribution based on distance measures work? (b) Why do different distance measures and normalization strategies perform differently? (c) Specifically, why do they perform differently for different languages and language families, and (d) How can such knowledge be used to improve authorship attribution methods? First, we describe current issues in authorship attribution and contextualize our own work. Second, we report some of our earlier research into the question. Then, we present our most recent investigation, which pertains to the effects of normalization methods and distance measures in different languages, describing our aims, data and methods. We conclude with a summary of our results.