Narrative Graphs: Concepts, Algorithms, Applications

• Narrative graphs are computational representations that encode entities, events, relationships, tropes, and structural features, offering a clear framework for analyzing narratives.
• They are constructed through diverse methodologies such as AMR extraction, clustering, graph grammar evolution, and embedding-based relation discovery to capture temporal and causal dynamics.
• Applications span literary analysis, political discourse, game design, and visual storytelling, with evaluation metrics like clustering scores, centrality measures, and coherence assessments validating their effectiveness.

Narrative graphs are a family of computational representations that encode entities, events, relationships, tropes, and structural features of stories in graph form. They are widely utilized in literary analysis, script event prediction, political discourse analysis, game design, multimodal visual storytelling, news narrative extraction, and event-based knowledge graph enrichment. This article surveys formal definitions, modeling approaches, algorithmic construction, and practical applications of narrative graphs in current research, with particular focus on the diversity of graph structures (event/eventuality graphs, actor–event–perspective graphs, trope-structure graphs, multimodal scene graphs, and coherence graphs) and the algorithms and metrics used for their construction and exploitation.

1. Formal Definitions and Classes of Narrative Graphs

Narrative graphs are instantiated in multiple formal structures, each tailored to specific narrative domains and analytical goals.

• Actor–Event–Perspectivization Graphs: In political discourse and general narrative analysis, a narrative graph $G = (V, E)$ is a directed, labeled graph where $V$ consists of disjoint sets of actor nodes (entities filling semantic roles), event nodes (predicate frames), and perspectivization nodes (events representing attitudes such as "want," "hope," or "need"). Edges $E$ are drawn from events to actors by semantic roles (agent, patient), and between events via thematic or attitude dependencies (Pournaki et al., 1 Nov 2024).
• Event Evolutionary Graphs (Script/Eventuality Graphs): For script event prediction and narrative reasoning, events are represented as nodes (e.g., predicate–GR pairs, or semantic roles with stripped arguments), and edges encode observed temporal and causal transitions in large corpora (e.g., newswire, narrative chains). These graphs are highly sparse and large, capturing complex temporal dynamics (Li et al., 2018, Jiayang et al., 30 Mar 2024).
• Trope-based Graphs (Abstract Narrative Structure): In narrative structure generation for games or literary analysis, narrative graphs encode tropes (e.g., hero, villain, conflict, plot device) as node labels, with edges capturing causal, derivative, conflict, or entailment relations among tropes. Graph connectivity and specialized edge constraints enforce narrative coherence and diversity (Alvarez et al., 2022, Alvarez et al., 2022).
• Multimodal Hierarchical Graphs: In visual storytelling (e.g., comics, multimodal narratives), a hierarchy of graphs is constructed: macro-level (story arcs), mid-level (event segments), and panel-level multimodal graphs, linking visual objects, actions, captions, and dialogues. Edges encode semantic, spatial, and temporal relationships and integrate across levels (Chen, 14 Apr 2025, Ghorbani, 29 Jul 2025).
• Coherence and Causal Graphs (Structural): For comprehension and retrieval tasks, coherence graphs connect context snippets with semantic links (e.g., retrospective "why" or "how" questions), while causal graphs constructed from narrative text nodes are connected by directed edges encoding explicit situation–task–action–consequence (STAC) relations or by causality extracted with hybrid linguistic and LLM-based methods (Xu et al., 21 Feb 2024, Li et al., 10 Apr 2025).
• Visualization-Oriented and Narrative-Enriched Event Graphs: Narrative maps employ weighted DAGs, with nodes as events and edge weights coding coherence (based on lexical/semantic similarity and topic structure). Event-centric knowledge graphs are enriched with subjective narrative attributes via lightweight indexing for viewpoint-specific retrieval (Keith et al., 2020, Plötzky et al., 2022).

This formal diversity reflects the requirements of distinct narrative phenomena (e.g., temporality, causal reasoning, multi-character interaction, narrative tropes, and multimodality).

2. Algorithmic Construction Methodologies

Algorithmic pipelines for building narrative graphs are adapted to source modality (text, multimodal, corpus type) and analytical focus.

• Event and Actor Extraction:
  • In AMR-based frameworks, sentences are parsed into Abstract Meaning Representation (AMR) graphs; predicates yield event nodes, filler entities are aggregated into actor nodes, and perspectivization is detected via special roles. Edges inherit AMR argument structure, with additional projection to actor-event and event-event links (Pournaki et al., 1 Nov 2024).
  • For narrative knowledge graphs, custom PERSON taggers, entity de-aliasing, and sentence segmentation precede embedding-centric relation extraction (Mellace et al., 2020).
• Relation Discovery and Label Extraction:
  • Unsupervised approaches use dense sentence embeddings (e.g., SBERT) followed by clustering (k-means, DBSCAN) to group relational sentences expressing similar semantic content. Cluster summarization (via BART) produces interpretable edge labels for the KG. A typical outcome is a triplet $(\text{CHAR}_a, r_j, \text{CHAR}_b)$, with $r_j$ a cluster-derived verb (Mellace et al., 2020).
• Graph Construction from Narrative Chains:
  • Event chains are extracted by coreference-driven aggregation of protagonist-centered event tuples from text, mapping sequences $(T, e_1, e_2, \ldots)$ to nodes and edges with observed bigram transitions. Edge weights reflect empirical frequencies normalized over the entire chain set (Li et al., 2018).
• Partial Matching and Abstraction:
  • To combat sparsity, partial event abstraction is employed: argument-dropping creates hierarchies of normalized events that are then anchored to knowledge graph nodes by embedding similarity. This bootstraps sparse matching into dense subgraph retrieval (Jiayang et al., 30 Mar 2024).
• STAC and Linguistic Feature Augmentation:
  • Hybrid pipelines first extract agent-centered, clause-constrained sentences as vertices, then augment these with a linguistically motivated "Expert Index" of seven binary or categorical features (genericity, eventivity, boundedness, initiativity, temporal aspects, impact). Vertices are labeled using a STAC classifier (Situation, Task, Action, Consequence). A structured five-iteration prompt pipeline produces a fully connected causal graph by iteratively proposing, pruning, and validating edges (Li et al., 10 Apr 2025).
• Inductive Question Graphs for Coherence:
  • Edges are formed by LLM-generated retrospective questions, filtered through answerability criteria; edge weights (number of questions) directly encode snippet coherence (Xu et al., 21 Feb 2024).
• Grammar-based and Mixed-Initiative Graph Evolution:
  • Trope-based narrative graphs are grown via application of graph grammars (production rules defined over trope types and edges) and evolved using MAP-Elites illumination search. Mixed-initiative systems integrate manual designer edits with automated, population-based suggestions (Alvarez et al., 2022, Alvarez et al., 2022).
• Multimodal Graph Assembly:
  • In comics and visual narratives, objects, actions, text, captions, and spatial relationships are detected and encoded as panel-level graphs, which are then linked to event and macro-structure via cross-level edges (Chen, 14 Apr 2025, Ghorbani, 29 Jul 2025).

3. Metrics, Structural Analysis, and Evaluation

Narrative graphs are evaluated and analyzed using multiple metrics and structural properties, reflecting their intended narrative or reasoning function.

• Clustering Metrics: Silhouette score and cluster purity (with hand-labeled gold sets) are used to assess the quality of unsupervised relation clusters; e.g., silhouette score 0.21 and purity 0.78 in literary KGs (Mellace et al., 2020).
• Graph-Theoretic Analysis:
  • Standard measures—degree and betweenness centrality, (weighted) degree, density, clustering, and assortativity—are applied to actantial (actor–actor) networks. Edges often carry polarity scores (beneficial vs. adverse actions), enabling analysis of narrative protagonist/antagonist roles and structural hubs (Pournaki et al., 1 Nov 2024).
• Coherence and Interestingness:
  • In trope-structure graphs, fitness functions combine consistency (quantification of valid motif patterns) with penalization for broken or fake conflicts; coherence is computed as $f_\mathrm{consistency} + [1 - f_\mathrm{cohesion}]$. Interestingness is a weighted sum of normalized counts of active plot devices, plot points, and plot twists (Alvarez et al., 2022). Step distance and coverage metrics inform diversity analysis.
• Edge-Relation Efficacy and Retrieval Performance:
  • NarCo graphs are evaluated by F1@5 on recap identification and nDCG on plot retrieval; precision/recall improvements validate the contribution of explicit coherence links (Xu et al., 21 Feb 2024).
• Automated and Human Judgments:
  • Metrics include macro-F1 for classification (e.g., 0.82 for hybrid RoBERTa+Expert Index STAC classification), paired win rates on “causal structure,” “explicit motivations,” and “granularity,” and MTurk-based comprehension/representational user studies (Li et al., 10 Apr 2025, Keith et al., 2020).

4. Applications Across Domains

Applications of narrative graphs span a wide conceptual and technical spectrum.

• Literary and Script Analysis: Relation clustering and event evolution graphs supply open-domain relation extraction, script event prediction, and automated construction of literary KGs for downstream tasks such as reasoning and question answering (Mellace et al., 2020, Li et al., 2018, Jiayang et al., 30 Mar 2024).
• Political Discourse and Public Narratives: AMR-based narrative graphs enable actor–event–perspective modeling and diachronic or comparative analysis of political texts. Longitudinal snapshots reveal evolving central actors, narrative polarity, and perspectivization frequency (Pournaki et al., 1 Nov 2024).
• Game and Narrative Structure Design: Trope-based narrative graphs formalize the canonical makeup of interactive stories (heroes, conflicts, devices) and enable mixed-initiative, co-creative design workflows in systems such as TropeTwist and Story Designer (Alvarez et al., 2022, Alvarez et al., 2022).
• Visual Storytelling and Multimodal Reasoning: Hierarchical, multimodal narrative graphs support complex symbolic reasoning—action retrieval, dialogue tracing, panel ordering—thus enabling structural content analysis, timeline reconstruction, and character continuity in comics and interactive narratives (Chen, 14 Apr 2025, Ghorbani, 29 Jul 2025).
• Information Visualization and Event-centric Knowledge Graphs: Narrative graphs undergird narrative map visualizations, incorporate subjective viewpoints for event-centric KGs, and provide efficient query indexing (e.g., by Bloom filter structures for subjective attributions) for rapid narrative-relevant retrieval (Keith et al., 2020, Plötzky et al., 2022).
• Textual Coherence, Question Answering, and Retrieval: Graphs encoding explicit coherence dependencies via retrospective questions enable enhanced snippet ranking, context enrichment in QA, and improved retrieval (e.g., local context enrichment for long-document QA and story understanding) (Xu et al., 21 Feb 2024).

5. Limitations, Challenges, and Open Questions

Despite wide adoption, narrative graphs face substantive challenges.

• Annotation and Supervision: Literary domain graphs often lack annotated corpora, rendering supervised relation extraction infeasible and motivating unsupervised clustering approaches (Mellace et al., 2020).
• Coreference and Role Resolution: Most systems perform sentence-level analysis (e.g., dependency or AMR parse), with only limited or no cross-sentence coreference resolution, affecting completeness in multi-sentence narrative chains (Pournaki et al., 1 Nov 2024, Li et al., 10 Apr 2025).
• Event Sparsity and Normalization: Event extraction from free text faces sparsity and anchoring issues; partial abstraction and argument-dropping significantly increase KG coverage but may decrease specificity (Jiayang et al., 30 Mar 2024).
• Scalability and Complexity: Large real-world narrative graphs (order $10^5$ nodes, $10^6$ edges) necessitate subgraph-based inference and efficient indexing for practical computation (Li et al., 2018, Plötzky et al., 2022).
• Subjectivity and Viewpoint Representation: Integrating subjective or multi-viewpoint attributions in event-centric KGs requires layered indexing, pragmatic corpus annotation, and robust stance/argument mining; negation handling and topical drift remain problematic (Plötzky et al., 2022).
• Multimodality and Symbolic–Neural Integration: Multimodal scene graphs and narrative representations require dynamic, temporally consistent construction across modalities, with neural modules (LLMs, image detectors) providing both input and validation, but necessitate strong consistency enforcement protocols (Ghorbani, 29 Jul 2025, Chen, 14 Apr 2025).

6. Future Directions

Open questions and active research topics center on multi-hop narrative reasoning, integration of more robust temporal and causal relation detection, cross-document and cross-media narrative alignment, and scaling to interactive, real-time or user-in-the-loop narrative systems. Notably, applications integrating narrative graphs with LLMs, graph neural networks, or hybrid symbolic–statistical models are under active development for higher-level narrative understanding, creative assistive tools, and multimodal generative tasks (Ghorbani, 29 Jul 2025, Li et al., 10 Apr 2025, Xu et al., 21 Feb 2024, Jiayang et al., 30 Mar 2024).