PMRC Narrative Framework

• PMRC Narrative Framework is a psychologically motivated schema that categorizes narrative comprehension tasks along local and global scopes with distinct narrative elements.
• It employs a two-dimensional typology to diagnose benchmark coverage and identify underexplored areas, such as global-setting inference.
• The framework bridges cognitive theories and narratology to guide robust benchmark design and propel future research in automated narrative analysis.

The PMRC (Psychologically-Motivated Narrative Comprehension) Narrative Framework provides a principled, task-oriented schema for analyzing and designing machine narrative reading comprehension assessments. Rooted in cognitive and narrative theories, it formalizes the taxonomy of narrative-comprehension tasks along two orthogonal axes: representational scope (local vs. global) and target narrative element (event, character, setting, functional). The resulting two-dimensional typology yields a grid with eight cells, each of which corresponds to a qualitatively distinct subspace of narrative inference. This framework directly connects theoretical distinctions from classic psycholinguistics and narratology with practical considerations in benchmark construction and automated narrative analysis (Sang et al., 2022).

1. Foundational Dimensions

1.1 Scope of Representation

Building on Kintsch’s Construction–Integration (CI) model and the situation-model literature, the Framework distinguishes between microstructures—coherent narrative representations accessible within a small window of text—and macrostructures—story-wide integrations that span the full discourse. Formally, with a narrative $N$ segmented into $\langle s_1, s_2, ..., s_n \rangle$ and a task $T$ with minimal answer-deriving sub-sequences $W \subseteq N$:

• Task $T$ is local if $\exists\, \delta \ll n$ such that $\forall\, W \in \text{Context}(T), |W| \le \delta$.
• Task $T$ is global if $$\forall\, \delta \ll n, \exists\, W \in \text{Context}(T)$$ such that $|W| > \delta$.

Local tasks require only restricted context (e.g., coreference within a scene), while global tasks mandate integration over large textual extents (e.g., narrative arc resolution).

1.2 Target Narrative Element

The Framework specifies four core narrative elements, drawing from story-grammar and situation-model accounts:

1. Event: Activities, transformations, and their causal/temporal structures.
2. Character: Agents, their motivations, affective states, personalities.
3. Setting: Spatiotemporal and contextual background, including atmosphere.
4. Functional Structure: Abstract plot phases (e.g., complicating action, resolution).

Any task is defined as targeting an element $$e \in E = \{ e_{event}, e_{character}, e_{setting}, e_{functional} \}$$ when its primary assessment requires extracting and evaluating $e$ (Sang et al., 2022).

2. The PMRC Typology Matrix

The Cartesian product of the two axes yields eight distinct task archetypes:

Scope	Event	Character	Setting	Functional
Local	ESTER, HiEve	LiSCU, LitBank NER, TV-script coref	LitBank (location-NER)	CompRes, Labov clause-typ.
Global	NarrativeQA, BookSum, ScreenSum	TVSG, personality prediction	(No standard benchmark)	TRIPOD, screenplay summarization

• Richly populated cells include local/global event and character tasks.
• Sparse cells include setting-global and local-functional (beyond clause-type tagging).
• Formally: $M: I \times J \to 2^{\{\text{datasets}\}}$ with $$M(\text{Global}, \text{Event}) = \{\text{NarrativeQA}, \text{BookSum}, \text{ScreenSum}\}$$.

This organization enables diagnostic analysis of benchmark coverage and reveals neglected task types such as global-setting inference, where elements like “the spirit of an era” emerge only across extended narrative contexts.

3. Psychological and Theoretical Motivations

The PMRC grid is “psychologically motivated” in two senses:

• It operationalizes Kintsch’s distinction between microstructure (local) and macrostructure (global), which cognitive accounts recognize as foundational for measuring the depth of narrative understanding.
• It adheres to well-established narrative dimensions (event, character, setting, functional structure) specified by story-schemata theory and situation models, ensuring correspondence with human comprehension constructs (Sang et al., 2022).

This alignment grounds the typology in both cognitive science and narratology, making it a robust framework for evaluating and designing comprehension tasks with theoretical fidelity.

4. Implications for Assessment and Dataset Design

By enumerating and instantiating each cell, the Framework reveals:

• Existing resource concentrations: Event-local, event-global, character-local, and character-global are richly benchmarked.
• Resource gaps: Setting-global and deep functional-structure tasks remain underpopulated; the Framework thus exposes opportunities for new dataset construction (e.g., benchmarking extraction of emergent social or historical milieus from extended narratives).
• Joint cell analyses: The structure supports complex multi-cell tasks; for example, one could combine local-event inference with global-functional evaluation (identifying turning points among salient scene events).

Examining a model or benchmark’s cell coverage provides a diagnostic for the inference types required (local/propositional vs. global/explanatory) and guides model development (e.g., need for external knowledge, theory-of-mind) (Sang et al., 2022).

5. Applications and Extensions

The PMRC Narrative Framework has yielded a systematic approach to mapping current machine reading comprehension tasks and identifying future research directions. Applications include:

• Benchmark diagnostics: Determining which narrative comprehension skills are actually assessed by a given dataset.
• Blueprint for new benchmarks: Informing the design of tasks/tests by identifying underexplored cell combinations and supporting multi-dimensional challenge sets that integrate several deeply-motivated narrative elements.
• Evaluation of narrative comprehension models: Providing the matrix as a required coverage map for architectural and methodological innovations.

A plausible implication is that PMRC, by clarifying the granular intersections of scope and element, could standardize the taxonomy of narrative assessment for both experimental and theoretical work in NLP and cognitive modeling (Sang et al., 2022).

6. Relation to Broader Narrative and Framing Frameworks

The PMRC Narrative Framework contrasts with narrative framing taxonomies in political/media discourse (e.g., (Otmakhova et al., 31 May 2025)), which focus on protagonist roles (hero/villain/victim), conflict orientation, and cultural story. PMRC’s emphasis is on internal textual representation (events, characters, settings, functions) and the inferential scope (local/global) required to access them, thus providing orthogonal yet complementary analytic perspectives—the former facilitating interpretive media studies, the latter fundamental computational evaluation of comprehension.

In summary, the PMRC Narrative Framework offers a comprehensive, psychologically grounded matrix for narrative-comprehension task assessment, enabling objective mapping, principled benchmark design, and rigorous analysis of inference types in machine reading (Sang et al., 2022).