Fabula-NET: A Deep Neural Network for Automated Multidimensional Assessment of Literacy Fiction and Narratives

Fabula-NET is an interdisciplinary collaboration between literary research, linguistics and informatics with the aim of expanding the scope of machine learning with domain knowledge from the humanities. Literature’s multidimensional and complex texts constitute a particularly challenging material with a potential to develop new models for automated text classification.

The project combines fractal analysis, sentiment classification and advanced language models with deep neural networks to describe the internal coherence of texts based on the hypothesis that a successful literary work exhibits a particular variation between predictability and unpredictability. This variation is particularly reflected in the dynamic properties of the narrative. The overall model can be used to classify texts as high/low quality and successful/unsuccessful, which is supported by preliminary studies of, for example, H. C. Andersen’s fairy tales and J. K. Rowling’s novels.

The project is multilingual and works with a very large corpus in English, Danish and Chinese in collaboration with a number of experts who help to validate the automated analysis results. The application possibilities for this technology are wide-ranging and it will be relevant for both libraries and publishers for searching through and evaluating texts, and in research to understand and compare large collections of texts from world literature at a level higher than compiled single analyzes. The application can also be developed for other types of texts and contribute to increase the quality of e.g. automated text generation.