Actor–Event–Perspectivization Graphs

• AEP-Graphs are comprehensive frameworks that unify event-centric data with narrative perspectives, enabling nuanced analysis of public discourse.
• They employ a rigorous tuple structure to encode events, actors, and perspective tokens, effectively linking objective data with subjective attributions.
• Advanced indexing techniques, including Bloom filters, enhance query efficiency by reducing latency by approximately 3x while preserving full recall.

The Actor–Event–Perspectivization Graph (AEP-Graph) is a formalism and data structure designed to capture not only event-centric knowledge but also the subjective narrative and perspective-dependent aspects inherent in public discourse, political narratives, and complex event reporting. These graphs integrate the representation of events, actors, roles, and viewpoint-dependent attributions within a single graph-theoretic framework, supporting both objective properties and narrative (subjective) signals. AEP-Graphs are foundational for advanced narrative analysis, multi-perspective information retrieval, and the empirical study of narrative signals in digital textual corpora (Plötzky et al., 2022, Pournaki et al., 2024).

1. Formal Structure of Actor–Event–Perspectivization Graphs

AEP-Graphs are defined by a rigorous tuple structure, supporting both event-centric and narrative-centric modeling.

Let $G$ denote an AEP-Graph:

$G = (V,\, E,\, \tau,\, \ell,\, \alpha)$

where:

• $V$: the set of vertices, partitioned (depending on use) as follows:

| Component | Node Types (Event-Centric) | Node Types (Narrative-Centric) | |-----------|------------------------------------------------------|----------------------------------| | $Evt$ | Event nodes (e.g., conflicts, negotiations) | $V_{\rm event}$ (AMR predicates) | | $Act$ | Actor/entity nodes (countries, persons, orgs) | $V_{\rm actor}$ (AMR ARG fillers)| | $Typ$ | Event-type/class nodes | | | $View$ | Viewpoint/perspective tokens (e.g., "US", "RU") | $V_{\rm perspec}$ (stance events)| | $G = (V,\, E,\, \tau,\, \ell,\, \alpha)$0 | Collection nodes (event–document sets) | | | $G = (V,\, E,\, \tau,\, \ell,\, \alpha)$1 | Document nodes (textual witnesses) | |

• $G = (V,\, E,\, \tau,\, \ell,\, \alpha)$2: edges are typed, encoding role-based relations (e.g., participation, instantiation, perspectivization).
• $G = (V,\, E,\, \tau,\, \ell,\, \alpha)$3 node-type function, typing nodes as event, actor, viewpoint, etc.
• $G = (V,\, E,\, \tau,\, \ell,\, \alpha)$4: assigns typed labels (e.g., instOf, participatesAs$G = (V,\, E,\, \tau,\, \ell,\, \alpha)$5, attr$G = (V,\, E,\, \tau,\, \ell,\, \alpha)$6).
• $G = (V,\, E,\, \tau,\, \ell,\, \alpha)$7: metadata attribution, distinguishing between objective attributions (unqualified) and subjective (viewpoint-dependent) attributions.

In narrative-centric formulations, each sentence is parsed into an Abstract Meaning Representation (AMR) graph $G = (V,\, E,\, \tau,\, \ell,\, \alpha)$8, from which a document-level AEP-Graph $G = (V,\, E,\, \tau,\, \ell,\, \alpha)$9 is extracted. The graph edges include participation (via PropBank roles, e.g., ARG0, ARG1), perspectivization (stance-taking predicate events), and potential event–event or actor–actor relations (Pournaki et al., 2024).

2. Event, Actor, and Perspective Representation and Linking

In event-centric models, event nodes possess temporal ($V$0) and spatial ($V$1) literals; actor nodes carry entity-specific attributes (e.g., type, temporal facts); and viewpoint nodes are tokens indexing the source or perspective (media outlet, nation-state, etc.).

Within the narrative-centric AMR approach, extraction focuses on three core sets:

• $V$2: Each AMR predicate with outgoing PropBank role edges is an event node.
• $V$3: Each concept node filling a PropBank ARG$V$4 edge is an actor.
• $V$5: Nodes whose frame is one of a manually specified "perspectivization" set (e.g., want‐01, believe‐01).

Edges encode:

• Actor-event participation: $V$6 where $V$7 is actor, $V$8 is event, and $V$9 is PropBank role.
• Perspectivization: $Evt$0 with $Evt$1 and $Evt$2 targeted event; $Evt$3 links actor to perspective.

Subjective attributes (e.g., is_aggressor) are encoded as $Evt$4—attr$Evt$5true, conditioned on viewpoint $Evt$6, with document $Evt$7 acting as a provenance witness if $Evt$8.

3. Construction and Extraction Pipelines

Event-Centric (Knowledge Graph) Pipeline

1. Event/Actor Extraction: Use Open IE or event-extraction modules on raw texts $Evt$9 to extract $V_{\rm event}$0. Insert nodes and edges accordingly.
2. Document Linking: For each event $V_{\rm event}$1, identify relevant documents $V_{\rm event}$2 via keyword/time overlap; create event-specific collection node and connect documents.
3. Subjective Attribution Annotation: For each event and its documents, extract subjective attributions by applying a witness-checking procedure (e.g., extractive QA + entity canonicalization). Add attr$V_{\rm event}$3 edges for validated attributions.
4. Index Construction: Assemble specialized indexes (see Section 4).

Extraction is exemplified by the following pseudocode fragment:

$V_{\rm actor}$5

Narrative Signal Extraction Pipeline (AMR-Based)

1. Sentence Segmentation and AMR Parsing: Segment documents into sentences, parse each with an AMR parser.
2. Graph Substructure Extraction: For each AMR, identify predicates (events), their argument structures (actors), and perspectivization frames.
3. Actor/Event/Perspective Linking: Collapse AMR name structures to entities, assign actors and events, and identify perspectivization events according to a curated frame list.
4. Narrative Trace Table Construction: For each extracted event, record sentence ID, root, role-fillers, and perspectivization status.
5. Global Graph Assembly: Aggregate per-sentence subgraphs and narrative traces into a corpus-level $V_{\rm event}$4 (Pournaki et al., 2024).

4. Indexing and Query Mechanisms for Perspective-Aware Retrieval

To enable efficient perspective- and narrative-aware queries, AEP-Graphs employ specialized index structures:

• For each $V_{\rm event}$5 with $V_{\rm event}$6, $V_{\rm event}$7 (subjective attributes), a Bloom filter $V_{\rm event}$8 indexes actors $V_{\rm event}$9 such that $Act$0—attr$Act$1true exists.
• The index function:

$Act$2

• Queries such as "which actors are aggressors under viewpoint US" use $Act$3 to prune candidates prior to expensive document-level validation.

This yields approximately $Act$4 reduction in query latency with no loss of recall, as Bloom filters have no false negatives (Plötzky et al., 2022).

5. Example Graphs and Query Patterns

A sample subgraph (event-centric prototype "RvU", Crimea Crisis 2014):

$V_{\rm actor}$6

Illustrative query (SPARQL-style, event-centric):

Retrieve all events where "Russia" was an aggressor under the US viewpoint: $V_{\rm actor}$7

AMR-based example (SOTEU 2010): "Emmanuel Barroso wants the European Union to invest more in innovation, technology and the role of science."

• Events: want‐01, invest‐01, innovate‐01
• Actors: Emmanuel Barroso, European Union
• Perspective: want‐01 (stance), links Emmanuel Barroso to European Union’s investment action

Resulting actantial edges include (Emmanuel Barroso, ARG0, want-01), (want-01, perspectivizes, invest-01), (European Union, ARG0, invest-01) (Pournaki et al., 2024).

6. Weighting, Scoring, and Narrative Signal Aggregation

Narrative signal analysis employs actantial networks, where actor–actor directed edges are weighted by their participation in beneficial or adverse event frames (as categorized by VerbAtlas):

• $Act$5: number of beneficial interactions (from $Act$6 and $Act$7 with $Act$8 in “beneficial” frame category)
• $Act$9: number of adverse interactions
• Aggregate, signed edge weights:

$V_{\rm actor}$0

$V_{\rm actor}$1

Values of $V_{\rm actor}$2 near $V_{\rm actor}$3 indicate dominance of beneficial interactions, $V_{\rm actor}$4 dominance of adverse, enabling analysis of alliance/adversarial narrative structures in the source corpus (Pournaki et al., 2024).

7. Limitations and Future Extensions

• Cross-Sentence Phenomena: Base pipelines do not resolve coreference or link events, actors, or causality across sentences. Integrating coreference resolution and discourse relation models is identified as a primary extension.
• Implicit Narrative Relations: Extraction is limited to explicitly encoded AMR structures; implicit temporal, causal, or rhetorical links are not currently captured.
• State Change and Narrative Arc Modeling: No current support for tracking state transitions or high-level narrative arcs across event sequences.
• Coverage Constraints: AMR-based pipelines are restricted by parser limitations (primarily English, PropBank inventory).
• Integration with LLMs: Combining AMR’s deep semantics with LLM-based inference may extend implicit narrative reasoning but risks interpretability challenges.
• Scalability: For event-centric AEP-Graphs, anticipated work includes scaling to larger corpora, incremental indexing, and automated viewpoint detection (e.g., stance detection modules) (Plötzky et al., 2022, Pournaki et al., 2024).

AEP-Graphs thus provide a systematic foundation for integrating perspective, narrative attribution, and empirical event knowledge, supporting both exploratory and formal analyses of narrative phenomena in public, political, and historical text corpora.