Mind the gap: word-embedding and multi-layered literary networks

Abstract In recent years, computational literary studies (CLS) have sought to analyse human interactions using social network analysis in prose literature. However, common methods based on co-occurrences often fail to capture the complex, multidimensional, and inter-textual relationships between characters, as described, for instance, by Mikhail Bakhtin. Computational-assisted literary networks might represent the entire character system in a given work, an entirety that human readers usually can’t otherwise see, but in a somewhat flat manner. This paper presents an innovative discrepancy-based approach to the uncovering of hidden connections between characters, offering a richer understanding of literary texts. By computing different character relation graphs (such as, co-occurrence, word-embedding-similarity, and annotated-based conversational networks), and taking account of their differences, we can generate an integrated visual model that reveals associations between characters beyond the initial relation graphs—to be then presented to a human reader as a ‘teaser’ for further reading. The method was tested in the works of Amos Oz, whose novels are characterized by intricate character interactions. Our findings demonstrate that discrepancy networks can identify subtle connections, providing new perspectives on character dynamics and narrative structure. These results are interpreted through a heuristic framework, emphasizing the importance of integrating computational methods with traditional literary analysis. The approach we outline here, enhances the study of individual texts and offers a method for cross-book analysis, contributing to a broader understanding of literary history and theory.