Temporal Narrative Atom (TNA)

• Temporal Narrative Atom (TNA) is a theoretical unit that captures a fully ordered sequence of narrative events in both text and video.
• It employs rule-based and machine learning methods to extract and quantify event sequences using metrics like causal potential and narrative-unit coverage.
• TNA advances narrative event extraction and generative model evaluation by providing a measurable framework for assessing narrative progression.

A Temporal Narrative Atom (TNA) is a theoretical construct deployed to capture the basic unit of temporal progression in both textual and visual narrative mediums. Developed independently in NLP and computer vision benchmarking, the TNA formalizes a fully-ordered, temporally coherent sequence of narrative events or states. In text, this reflects cross-sentence temporal knowledge; in video, it encodes minimal visual segments corresponding to distinct narrative states. The TNA paradigm is foundational for robust event knowledge extraction and for the quantitative evaluation of narrative-rich content generation (Yao et al., 2018, Feng et al., 15 Jul 2025).

1. Formal Definitions in Text and Video

In textual narratives, a TNA is defined as a tuple of events $(e_1, e_2, \ldots, e_n)$, where each $e_i$ is the main verb lemma extracted from a sentence. The tuple is fully ordered: for all $i < j$, the textual order $e_i \rightarrow e_j$ directly implies the temporal "before/after" relation. The double temporality principle—where textual order aligns with real-world event order—underpins this mapping (Yao et al., 2018).

In long video generation assessment, a TNA is the smallest segment within which the narrative state (scene attribute, object attribute, or object action) remains homogeneous or transitions only once. Formally, the video $v(t)$ over $t \in [0, T]$ is segmented into intervals $[T_{i-1}, T_i)$, each corresponding to the $i$th TNA. The TNA count $n$ serves as a direct metric of narrative richness (Feng et al., 15 Jul 2025).

2. Theoretical Motivation and Narratologic Principles

The development of TNA draws heavily on narratological theories. In text, the double temporality characteristic of narratives—where storytelling order mirrors event chronology—allows for atomic extraction of before/after event knowledge without requiring extratextual temporal annotation (Yao et al., 2018). In film theory, TNAs are conceptually proximate to McKee’s "Beat": the minimal unit of narrative progression. By identifying TNAs, researchers operationalize "story richness" as a countable entity.

Furthermore, in the video domain, TNA segmentation is motivated by the need to decompose complex narratives into quantifiable, evaluable slices, reflecting discrete progressions in scene, object, or action (Feng et al., 15 Jul 2025). This metric replaces informal or subjective notions of narrative density with explicit, reproducible units.

3. Methodologies for Extraction and Enumeration

Textual Narratives

The TNA extraction workflow proceeds as follows:

• Narrative paragraph identification is performed via a two-stage process:
  • Rule-based seeding selects paragraphs with $\geq 4$ sentences, $e_i$0 actantial pattern matches, minimal dialog/interrogatives, and a dominant protagonist chain.
  • A bootstrapped Maximum Entropy classifier (LIBLINEAR) generalizes to larger corpora, utilizing features such as grammar rule frequencies, verb-sequence LLM perplexity, protagonist chain statistics, LIWC categories, and POS tag frequencies.
• Event sequence extraction chains main verb lemmas in text order.
• Candidate pair and chain scoring uses the Causal Potential metric:

$e_i$1

where $e_i$2 is pointwise mutual information and $e_i$3 is the empirical probability of temporal succession. Composite chain scores aggregate $e_i$4 across windows up to three event distances.

Visual Narratives

In video evaluation, TNAs are enumerated via controlled prompts:

• Evaluation prompts are constructed to describe $e_i$5 explicit narrative steps.
• An LLM parses prompts to extract the list of TNAs and generate binary existence and transition questions.
• The number of TNAs is the prompt-defined narrative steps, which then structure model evaluation.
• No direct pixel-level segmentation is undertaken; instead, multimodal LLMs interpret video segments relative to prompt-defined TNAs (Feng et al., 15 Jul 2025).

4. Quantitative Metrics and Evaluation Protocols

In NLP, top-ranked TNAs (pairs and chains) are determined by causal potential and evaluated by intrinsic precision (human annotation: pairs $e_i$6, chains $e_i$7), pseudo-recall against gold scripts, and performance improvements in downstream event-temporal-relation tasks (e.g., TimeBank classification accuracy increased by 1.1%, before-relation recall by 4.9%) (Yao et al., 2018).

In video, TNA count $e_i$8 and the associated narrative-unit coverage score $e_i$9 (mean fraction of correctly detected TNAs across multiple MLLM runs) quantify actual narrative expression. The effective number of expressed TNAs is $i < j$0; coherence metrics further assess transition plausibility between adjacent TNAs. These metrics reveal that state-of-the-art long-video generation models reliably express only $i < j$1–$i < j$2 TNAs for $i < j$3 (Feng et al., 15 Jul 2025).

Table: TNA-Associated Metrics in Text and Video

Domain	Metric	Quantification/Protocol
Text	Causal Potential (CP)	PMI + log temporal directionality
Text	Intrinsic Precision	Human-annotated pairs and chains
Video	Narrative-unit coverage	$i < j$4
Video	Effective TNAs expressed	$i < j$5
Video	Coherence	Transition accuracy, existence proportion

5. Illustrative Examples

Textual TNAs

• News Example:
  • Paragraph: “Michael Kennedy graduated… He married… and attended… After receiving… he worked… before… took over… Kennedy expanded… and increased…”
  • Extracted TNA: $i < j$6
• Fiction Example:
  • Paragraph: “Beth paid… She jumped out… headed… reached into… Beth entered… and got undressed… showered… changed… grabbed… left…”
  • Extracted TNA: $i < j$7 (Yao et al., 2018).

Video TNAs

• Action sequence, $i < j$8:
  • Prompt: “Initially, the sea turtle slowly descends toward the corals. Then, the turtle stops and rests on a coral. Finally, the turtle starts swimming upwards.”
  • TNAs: “turtle slowly descends” $i < j$9 “turtle stops and rests” $e_i \rightarrow e_j$0 “turtle swims upwards”
• Attribute change, $e_i \rightarrow e_j$1:
  • Prompt: “A chameleon changes from brown to green.”
  • TNAs: “brown chameleon” $e_i \rightarrow e_j$2 “green chameleon” (Feng et al., 15 Jul 2025).

6. Properties, Limitations, and Extensions

Strengths

• In text, the TNA model exploits double temporality to afford cross-sentence temporal reasoning with minimal supervision, enabling high-precision acquisition of event chains across diverse domains (news, fiction, blogs).
• In video generation evaluation, TNA provides a theoretically grounded, flexible unit for assessing narrative richness, allowing for fine-grained, scalable benchmarking over narrative content.

Limitations

• Textual TNAs are restricted to main-verb events in the past tense, excluding stative and present-tense constructions. The model does not resolve event sense ambiguity or extract argument roles, and is sensitive to coreference quality for protagonist tracking.
• Video TNAs rely on prompt-structured narrative steps rather than automatic segmentation of arbitrary videos; the definition assumes that discrete narrative transitions are detectable and salient via prompt-LMM evaluation.

Potential Extensions

• Text: Integrating event sense disambiguation and frame-semantic role annotation, verb clustering via distributional/embedding methods, scaling TNA extraction to document level to cover narrative discontinuities (e.g., flashbacks), and learning hierarchical TNAs (scene $e_i \rightarrow e_j$3 sequence $e_i \rightarrow e_j$4 event) for multilayered narrative modeling (Yao et al., 2018).
• Video: Automatic detection of TNAs directly from video, extension to handle overlapping or nested narrative structures, and compositional evaluation pipelines for higher-order narrative phenomena.

7. Broader Implications and Current Findings

The TNA construct bridges narratology, NLP, and video understanding by delivering a rigorous atomic unit applicable both for knowledge acquisition and model evaluation. In temporal event extraction, TNA-based methods outperform recent neural models on narrative cloze and temporal relation tasks. In long-form video generation, TNA metrics reveal a significant gap between narrative-element fidelity and the capacity to realize multiple, coherent narrative steps, with most current models plateauing at expressing 1–2 TNAs in complex prompts. Human-LMM alignment in TNA-based metrics far outpaces legacy video benchmarks, underscoring the construct’s empirical validity (Yao et al., 2018, Feng et al., 15 Jul 2025).

A plausible implication is that further refinements in TNA-based modeling and evaluation may catalyze advances in both NLP and generative video systems by enabling more sophisticated representations and measurements of real-world narrative complexity.