Exercises in Literary Style

Literary scholars are experts in studying style. What do computer scientists understand when they say ‘style’? Style might refer to higher level discourse features like narrative tropes, character agency, and representation of time. In the image generation space, using “style” as a directive is commonplace. Similar control is not available when doing text generation with LLMs. We theorize this is due to a lack of metadata and training data labeled at the literary style level in the corpora LLMs are currently trained on. We propose to survey and categorize existing literature on "style" and develop new systems that operationalize these constructs. What would a benchmark dataset for literary style look like?