Daejeon, July 27–31
Narrative plot in long-form fiction emerges through causal, motivational, and thematic relations distributed across events, characters, and scenes. Recent systems for event causality extraction and knowledge graph construction, including CauseNet (Heindorf et al. 2020), ASER (Zhang et al. 2020), and WikiCausal (Hassanzadeh 2024), have improved large-scale event relation modeling. However, these approaches are primarily designed for commonsense reasoning and extraction accuracy, with limited emphasis on interpretable literary narrative organization. Narratological research has similarly emphasized that plot develops through structured relations of causality, motivation, and transformation rather than simple chronology (Bremond 1973; Herman 2004), yet existing computational approaches often struggle to represent long-range thematic progression and subplot organization across full-length literary texts.
To address this gap, we propose a Causal Event Knowledge Graph (CEKG) that models narrative progression through interconnected causal, thematic, and character-centered relations. The CEKG combines Large Language Model (LLM)-driven event extraction with narrative graph construction organized around events to analyze long-range plot structure across literary texts. Applied primarily to Charles Dickens’s Great Expectations, the framework enables thematic filtering, causal navigation, and subplot reconstruction through a browser-based visualization interface designed for exploratory literary analysis.
This study makes three contributions: (1) The CEKG, a framework that models literary plot as a network of causal and thematic event relations; (2) a schema-based methodology that extracts narrative knowledge graphs from literary texts by combining event extraction and ontology-guided annotation; and (3) an interactive interface, demonstrated through Great Expectations, that enables critical inquiry of distributed narrative trajectories.
Both causal graph research and computational narratology address relations extending beyond local event adjacency. Early work on temporal reasoning formalized interval-based representations of event relations (Allen 1983), while later NLP research explored extraction of explicit and implicit cause–effect relations from unstructured text (Oh et al. 2016). More recent event-centric knowledge graph systems and LLM-driven causality frameworks have improved large-scale event relation modeling and long-range inference through neural, ontology-guided, and graph-augmented methods (Heindorf et al. 2020; Zhang et al. 2020; Liu et al. 2023; Hassanzadeh 2024; Su et al. 2025; Wang et al. 2025). Recent studies have additionally explored causal graph generation and temporal-causal consistency in narrative texts, highlighting the importance of structured event relations for long-range reasoning and interpretability (Li et al. 2025; Koupaee et al. 2025; Zhang et al. 2025). However, these approaches remain primarily oriented toward extraction and inference performance, providing limited support for interpretable modeling of long-range literary narrative structure.
Narratological research similarly argues that plot develops through structured relations of causality, motivation, and transformation rather than simple chronology (Bremond 1973; Todorov 1977; Ryan 1991; Herman 2004). Systems such as DramaBank (Elson 2012) operationalized these ideas through structured representations of agency and narrative action, but many existing approaches remain annotation-intensive, corpus-limited, or difficult to scale across full-length literary texts. Existing systems also provide limited support for exploration of thematic progression, scene-level organization, and character-centered narrative structure.
These considerations are particularly significant for literary narratives, where plot development depends on long-range interactions among events, themes, and character trajectories rather than isolated causal pairs. Graph-oriented representations may therefore provide a useful framework for modeling distributed narrative trajectories across multiple interpretive scales.
Causal Event Knowledge Graph (CEKG) is a framework for representing and analyzing literary narratives through event-oriented graph modeling. The system operates in four stages: event extraction, relation classification, thematic and scene organization, and interactive narrative exploration. Literary texts are transformed into interconnected event, character, scene, and thematic relations through schema-guided LLM processing and ontology-guided relation classification. The resulting CEKG is deployed through a browser-based visualization interface supporting thematic filtering, causal navigation, and literary analysis. Figure 1 presents the overall CEKG pipeline from literary text input to navigable graph visualization.
The CEKG organizes literary narratives through linked event, agent, and scene nodes. The pipeline first extracts plot-relevant events from literary texts through structured LLM inference. Characters are represented as agent nodes linked to events through participation edges, while scene nodes group events occurring within the same localized narrative context. Alias resolution merges dispersed character mentions into shared agent identities.
The system then assigns causal edges between events through ontology-informed LLM classification constrained by predefined relation dictionaries. Event pairs are classified according to causal relation types including enabling, revelation, escalation, reversal, emotional triggering, and reinterpretation. Implausible or structurally inconsistent relations are filtered before incorporation into the CEKG.
Beyond causal relations, the CEKG incorporates thematic links connecting events across shared narrative trajectories. Events are annotated through ontology-guided thematic classification across five categories: WEALTH, KNOWLEDGE, KINSHIP, JUSTICE, and POWER, enabling thematic filtering and subplot reconstruction through graph navigation.
The resulting narrative graphs are exported as web visualization data and deployed through a browser-based interface supporting traversal, filtering, thematic overlay, and character-centered analysis without requiring direct interaction with the underlying graph database infrastructure.
The CEKG is deployed through a browser-based visualization interface for interactive analysis of literary relations. Users may switch between causal, subplot, and agent-centered views, search event descriptions, and examine narrative structures through multiple graph layouts and filtering operations. The interface supports theme and character filtering, chapter restriction, and selective display of causal and thematic relations across large-scale literary networks. Figure 2 shows the global CEKG interface linking event nodes through causal and thematic connections in Great Expectations.
Selecting nodes or edges opens an inspection panel displaying event descriptions, thematic annotations, confidence scores, participating agents, and neighboring causal relations. These operations allow users to move between localized event structures and broader plot developments through dynamic examination rather than static visualization alone.
The interface supports thematic filtering across five ontology-guided categories: WEALTH, KNOWLEDGE, KINSHIP, JUSTICE, and POWER. In subplot mode, selecting a thematic category isolates events associated with that theme above a user-defined confidence threshold, enabling examination of distributed thematic patterns across distant sections of the novel.
Users may additionally trace causal chains according to selected edge types and confidence thresholds. Selecting an event node opens an inspection panel on the right side of the interface displaying event descriptions, thematic annotations, confidence scores, participating agents, why-factors, and neighboring causal relations. Selecting an edge reveals the relation type, confidence level, and source–target provenance associated with the selected connection.
The CEKG reconnects disperesed long-range event sequences distributed across Great Expectations through event, agent, scene, causal, and thematic relations. Figure 3 presents a selected WEALTH-related CEKG subgraph from Chapters 39 to 42, generated through thematic filtering, agent-based filtering centered on Philip Pirrip (Pip), and keyword-guided causal exploration using the query term “Magwitch.”
Thematic filtering of WEALTH-related events reveals a coherent pattern of aspiration, inheritance, mistaken belief, revelation, and reinterpretation. In Great Expectations, this sequence connects the protagonist Pip’s shame regarding his lower-class background, his desire to become a gentleman after encountering Estella and Miss Havisham, the announcement of his great expectations, his false assumption that Miss Havisham is the source of his fortune, Magwitch’s later revelation as the true benefactor, and Pip’s reinterpretation of his own identity and upward mobility. Although these events are distributed across distant sections of the novel, thematic filtering and causal analysis reconnect them as a unified structure of character transformation within the CEKG.
The resulting graph further suggests that Pip’s rise in status is shaped through concealment, mistaken inference, and delayed revelation as much as through economic change itself. Wealth in the novel therefore functions simultaneously as a material and epistemic structure, linking aspiration, misrecognition, and identity formation across long narrative distances.
This study presents the Causal Event Knowledge Graph (CEKG), a framework that combines structured LLM-assisted event extraction with ontology-guided modeling of causal, thematic, and character relations to reconstruct literary plot organization across long-form fiction. Applied primarily to Great Expectations, the resulting CEKG and web visualization interface demonstrate how dispersed subplots and long-range event connections can be reconstructed and examined through graph-based literary analysis. These results suggest the potential of event knowledge graphs as interpretive frameworks for both computational and interpretive literary analysis.
CEKG webpage link: https://xynovitch.github.io/Narrative-Causal-Graph/web/