The basic premise underlying authorship attribution studies is that while the form of expression in language is in some respects strictly bound by linguistic rule systems and in others somewhat constrained by topic and genre, it is in some other respects freely available for configuration or preferential choice by author or speaker. This individual variation can be observed, detected, and predicted to some extent, using traditional stylostatistic measures. For instance, word length varies from

more »

... uthor to author [Mendenhall, 1887, e.g.]; sentence length likewise; and some forms of lexical expression are characteristic of speakers, either on an individual level or on a community level [Book of Judges]. Common to most computation of individual difference in authorship is that the features used to characterise and distinguish authors are based on the repeated measurement of some, often clause-internal, property at independent positions in the text and then aggregating these pointwise measures by averaging or normalising the result. In this position paper we claim that by measuring local clause-or even wordinternal properties, and by aggregating in such a way that the relation between individual observations is destroyed, we obtain features that are most likely to have been subject to pressure from conventionalisation and grammaticalisation processes in language. Instead, we want to examine features that capture differences between authors on a level of textual structure where the space for individual choice is wide: the organisation of informational flow and narrative frame. Such features can be obtained by studying configurations and progressions of observable properties above the clause level. We will call this family of aggregated features configurational in contrast to the typical pointwise measurements.