Extracting and Visualising Character Associations in Literary Fiction using Association Rule Learning

In many works of fiction, the complexity and evolution of associations between characters is an important aspect of the narrative. Associations between characters are traditionally modeled as undirected networks where vertices are characters in the story and each edge $\{\boldsymbol{a},\ \boldsymbol{b}\}$ represents a pair of associated characters $\boldsymbol{a}$ and $\boldsymbol{b}$, possibly with the strength of the association represented as an edge weight. In this paper, we present a novel application of association rule learning to determine a richer class of character associations in fictional works between (non-empty, non-overlapping) sets of characters $\boldsymbol{A}$ and $\boldsymbol{B}$ in an almost completely automated way. Furthermore, associations are directed (associations $\boldsymbol{A}\Rightarrow \boldsymbol{B}$ and $\boldsymbol{B}\Rightarrow \boldsymbol{A}$ may differ in strength), and we demonstrate that standard metrics (support, confidence and lift) can be used to determine association strength in the context of literary analysis. Association rules can be expressed as Character Association Networks (CANs), and we demonstrate that visualising the evolution of these networks and computing centrality measures for such networks can rapidly provide literary analysts with insights such as identifying protagonists and key clusters of characters.