Automatic authorship attribution based on character n-grams in Swiss German

Abstract

Automatic authorship attribution aims to train computers to identify the author of a disputed text based on idiolectal language features. When confronted with nonstandard data – in the present study Swiss German instant messages – languagespecific NLP toolkits are often unavailable, limiting the availability of features to classify texts. Thus, the approach I propose for Swiss German is based on character ngrams, which not only avoids the problem of a lack of available NLP tools, but – in addition to being a proven successful feature for authorship attribution – allows the capturing of orthographical idiosyncrasies. It thus allows the exploitation of Swiss German’s lack of standardised spelling rules, turning the challenge that Swiss German presents as non-standard data into an advantage. Different lengths of n-grams as features of a Na¨ıve Bayes classifier combined with varying sizes of training and test corpora were tested, and 6- and 7-grams were found to faultlessly identify authors for all combinations considered. The number of distinctive n-grams in an author’s data set was found to be a determining factor for the classifier’s success, highlighting the benefits of exploiting Swiss German’s non-standard nature for authorship identification.