Joachim Denzler & Sophie Marshall

The goal of this project is the development of a novel distant-reading tool for the quantitative analysis of the use of rhetorical devices. The developed methods will be applied for a comparative stylometric analysis of three medieval genres: the Middle High German adaptations of the trois matières. From the viewpoint of medieval studies, we aim for identifying genre-specific stylistic differences between romances of antiquity, Arthurian romances, and chanson de geste adaptations of the 12 th and 13 th century. Up to now, quantifiable similarities and differences of these genres regarding their use of rhetorical devices have not been analysed. Such an analysis would be valuable for the controversial discussion about the medieval concepts of ‘genre’ and the genre-awareness of medieval authors as well as for assessing the relevance of rhetorical devices taught in medieval schools for the practice of vernacular authors. To this end, we intend to analyse whether selection, use, and frequency of rhetorical devices depend on the scene type and whether characteristic differences can be attributed to genres or just different authors or dates of origin. Such a broad analysis requires computational methods for distant reading, which do not yet exist for the detection of stylistic devices to a satisfactory degree, not to mention for Middle High German texts. For this purpose, we seek to develop novel methods for detecting rhetorical devices based on methods from the fields of anomaly detection as well as active and life-long machine learning. In this context, we consider rhetorical devices as ‘anomalies’, as deviations from the quantitative norm established by the greater part of the corpus, which can be detected using statistical methods (e.g., a chiasmus is an anomaly amidst non-chiastic constructions). To reduce time-consuming manual annotation tasks to a minimum, we propose to employ active learning techniques in a semi-supervised scenario with the human in the loop: Based on an initially unsupervised process of detecting stylistic anomalies, the system refines itself in an interactive process by actively asking the user for annotations for a few informative examples. Furthermore, we intend to integrate prior theoretical knowledge provided by experts into the anomaly detection procedure for filtering out obvious false positives and steering it towards certain rhetorical devices. In this project, we will focus on two exemplary stylistic devices: the chiasmus as a figure of repetition and the metaphor as a trope. The latter can be seen as an anomaly as well due to the unusual interactions between words from different domains, reflected in their word embeddings. Since such embeddings need to be learned from huge corpora, which are not available for low-resource languages such as Middle High German, we also strive for developing a technique to adapt such embeddings learned on New High German corpora to Middle High German.