A functional stylistics for text genre and register classification

Text genre classification has shown benefit for many NLP tasks, including domain adaptation of statistical parsers or machine translation. While there is wide acknowledgement on the utility of document characterization by genre, it is quite difficult to determine a definitive set of features or even a comprehensive list of genres. In this talk I will present results from my Master thesis, where I developed a theoretically grounded typology for genre and register analysis and compiled two corpora, one for German register and genre out of DeReKo [1], and another for English literature genre out of Gutenberg. By implementing an extensive (style) feature extraction pipeline I was able to evaluate the efficacy of a wide array of features across different corpus settings and domains, to arrive at high accuracy machine learning algorithms for supervised text genre classification. Finally, the thesis also provides a description of prototypical register classes through the agglomeration of style feature loadings and identifies pervasive functional dimensions such as interpersonal and narrative dimensions.