Modeling Prototypicality for Genre Concepts

Perspectival Modeling (Ted Underwood) is among the most powerful methods for investigating the literary change of genres based on machine learning. Concerning the semantic change of loosely ordered genres, further challenges for perspectival modeling arise. This poster provides a new technique for modeling the prototypicality  of genre concepts. Modeling prototypicality in this way is a principled strategy for accessing the vagueness of conceptual boundaries.

The model uses a so-called c@1 score, a variant of the accuracy score accounting for undecidability. Similar to the implementation from the Authorship verification task 2021 (https://github.com/pan-webis-de/pan-code/tree/master/clef20/authorship-verification) and based on grid search, the boundaries of undecidability that lead to optimal c@1-accuracy for prediction are calculated. This strategy is integrated into the extraction of predictive probabilities in logistic regression models for genre classification within perspectival modeling as it has been developed by Underwood (2019b). These results are presented within different types of established and new visualizations. Furthermore, the results can be used for a more principled way of expressing conceptual vagueness in quantitative terms. While it uses two existing methods with perspectival modeling and c@1-score, the model presented in the poster is new to computational literary studies. Its strength lies in its dual function of making conceptual looseness interpretable in the way it has been conceived in literary theory and semantics from the 1960s on.