Year: 2026
Type: other
Abstract: Automatically extracted narrative networks – graphs with entities as nodes and their relationships as edges – can reveal structures of a narrative through salient entities and how they relate (Labatut and Bost 2019; Tangherlini et al. 2020; Holur et al. 2021). In such an entity-centric approach (as opposed to event-centric), the structure of the network reflects the narrative in terms of associations and events involving the extracted entities. Such networks naturally foreground central narrative structures — salient entities and their strong relations. But a narrative is more than the central structures that everything else revolves around, builds upon and adds to. This work is concerned with the everything else – brief sub-plots, small clusters of descriptions, or associations between minor characters that make a fictional world come alive, but which may go under the radar in computational studies of text. We refer to these as micro-narratives – delimited peripheral narrative structures, localized configurations of entities that together constitute recognizable subplots or thematic clusters, identified as communities of rare entities in the narrative network. We present an approach to unearth such micro-narratives. The micro-narratives involve rare entities with limited textual presence. In network-based approaches to narrative, entity salience is typically determined by frequency of occurrence or network centrality measures, which naturally foregrounds dominant entities and the overarching narrative structures they constitute. But rarer entities are not only overshadowed by dominant ones — they are also lost among each other in the long tail of many but rare entities (Baayen 2001). Identifying meaningful peripheral structures in this abundance is the core challenge this work addresses. After all, as molecules of a larger narrative structure, they help give it depth and weight. We create a narrative graph from a text collection by performing Named Entity Recognition extended by Noun Phrase Chunking with SpaCy (Honnibal et al. 2020) to extract everything from salient to inconspicuous entities in a broad sense — characters, locations, objects, and concepts — and taking these as nodes in the graph. Since we are interested in rare entities, we employ little filtering and are hesitant to resolve different mentions of potentially identical entities too aggressively, so as to not distort the weak signals. Edges between the entity nodes are created by counting co-occurrences between extracted entities within some defined window of text. Since we here favor the rarities, we weight the edges using pointwise mutual information (Church and Hanks 1990) – a measure of association, the primary weakness of which is its inflation for rare events. For this purpose, its weakness becomes its strength. The result is a weighted, undirected graph with heterogeneous entity types as nodes, and PMI-weighted co-occurrence as edges. Given a graph that favors weak relational signals, we employ community detection techniques (both with and without overlapping communities; Palla et al. 2005; Blondel et al. 2008) to identify tightly connected sections of the graph as structural traces of underlying micro-narratives. Still, irrelevant or erroneously extracted entities can dominate in the extracted communities, as we search for meaningful structures among noise in the long tail. To separate the wheat from the chaff, a range of structural measures aid in their identification, including their textual confinement and their connectivity to salient nodes in the broader narrative. The method is demonstrated on The Lord of the Rings as a literary work with rich narrative complexity. We further apply it to a corpus of Reddit comments about ChatGPT in the months after its launch to explore the extent to which the method generalizes beyond fictional narrative to emergent public discourse. We evaluate the method qualitatively by examining extracted micro-narratives and their characteristics across the two registers and with two different community detection algorithms. The two cases highlight structural similarities and differences as well as strengths and weaknesses of the method. The proposed method offers a conceptually simple yet effective way of targeting peripheral narrative structures, using well-known techniques in a novel creative way. In doing so, it contributes to bridging distant and close reading by surfacin
Authors: Fyhn, Kasper, Baglini, Rebekah Brita
Language: English
Subfield: Literature and Literary Theory
Field: Arts and Humanities
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7220564617
