Pith. sign in

REVIEW 3 major objections 4 minor 23 references

Narrative Information Theory

T0 review · 3 major / 4 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The paper proposes that storytelling devices—complexity, pivots, suspense, plot twists—can be measured by entropy and divergence computed over narrative states, and demonstrates the first two on TV shows.

desk verdict A clean but thin proposal: textbook entropy and JSD applied to TV emotion states, with an uncontrolled genre comparison; worth a serious referee but needs statistical work. read the letter →

arxiv 2411.12907 v1 pith:AQS333DV submitted 2024-11-19 cs.MM cs.ITmath.IT

Lion Schulz , Miguel Patrício , Daan Odijk This is my paper · ORCID
classification cs.MMcs.ITmath.IT
keywords narrativeinformationtheorystateentropyascomplexityJensen-ShannondivergencesuspenseandplottwistmetricsTVgenreemotionanalysisfacedistributionsAI-generatedstorybenchmarking
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to make narrative analysis quantitative by translating storytelling devices into information-theoretic quantities. It claims that the entropy of a story state measures complexity, the Jensen-Shannon divergence between consecutive states measures pivots or story beats, and prediction-based quantities formalize predictability, suspense, and plot twists. The empirical demonstration on a corpus of TV shows finds genre-consistent differences: reality and dating formats show broader emotion mixtures and more frequent large shifts, while dramas and thrillers are more tonally focused and change more gradually. If these measures hold up, they give creatives and AI researchers a common ruler for comparing human- and machine-made stories.

What carries the argument

The load-bearing object is the narrative state $s_t$: a probability distribution over whatever features describe the story at time $t$ (in the demonstration, the emotions of visible characters). On this object the paper places five information-theoretic measures: complexity $H(s_t)$, pivot $\mathrm{JSD}(s_t, s_{t-1})$, predictability $I(s_{t+1}; S_t)$, suspense $H(P(s_{t+1} \mid S_t))$, and plot twist $\mathrm{JSD}(P(s_{t+1}), s_{t+1})$. The machinery works because entropy and divergence output dimensionless, comparable numbers for any state definition, so the same formulas transfer across modalities—video, text, audio—without any change to the math.

What would settle it

Recalculate the entropy and pivot values on the same episodes after replacing the face-emotion state with a text-based emotion distribution from the scripts; if the genre ordering shifts or disappears, the reported results depend on the face pipeline rather than on narrative structure itself.

Watch

Extended reading notes

Core claim

The paper's central claim is that a narrative can be turned into a probability distribution over time—one state per moment—and that standard information-theoretic quantities then capture what audiences experience as story structure. It defines complexity as the entropy $H(s_t)$ of the current state, a pivot as the Jensen-Shannon divergence $\mathrm{JSD}(s_t, s_{t-1})$ between consecutive states, and, for prediction-based devices, predictability as $I(s_{t+1}; S_t)$, suspense as the entropy $H(P(s_{t+1} \mid S_t))$ of the predicted next state, and plot twist as $\mathrm{JSD}(P(s_{t+1}), s_{t+1})$ between prediction and realization. Empirically, with states given by emotion distributions read from actors' faces in over 3000 minutes of TV, the paper finds that reality and dating shows have higher entropy and larger average pivots, while dramas and thrillers are lower on both. On these grounds the paper argues the framework quantifies narrative structures and emotional dynamics across genres and can therefore compare human-created and AI-generated stories.

Load-bearing premise

The empirical argument assumes that the emotion mix detected on actors' faces and averaged over rolling windows is a faithful and sufficiently rich proxy for the narrative state at each moment, so all genre differences inherit whatever this mapping gets wrong.

Editorial extensions

If this is right

  • The entropy of a narrative state gives a quantitative complexity score, so scenes dominated by one emotion score low and emotionally mixed scenes score high.
  • The Jensen-Shannon divergence between consecutive states localizes story beats, allowing editors and summarizers to find pivotal moments automatically.
  • Genre profiles from the demonstration—reality shows more emotionally mixed and shift-prone, dramas and thrillers more tonally focused and gradual—can serve as baselines for judging whether an AI-generated story fits a requested genre.
  • Because the framework treats states as arbitrary distributions, the same definitions transfer from video to text or audio without changing the math.
  • Once a generative model of story continuation is supplied, the prediction-based metrics provide formal, computable definitions of predictability, suspense, and plot twist.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Replacing the face-emotion state with a text-based emotion model of the same scripts and checking whether the genre rankings replicate is a testable way to show the results are about narratives rather than about faces.
  • The paper defines but never computes the suspense and plot-twist metrics; an immediate implementation would use an LLM's next-token distribution over scene summaries and check whether cliffhanger endings raise the suspense entropy.
  • If the measures stabilize across studies, they could be inverted into a generation objective—steering a story generator toward a target suspense or twist profile—an application the paper does not claim.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes an information-theoretic framework for measuring narratives, decomposing a story into states s_t and defining entropy-based complexity (Eq. 1), Jensen-Shannon divergence between consecutive states as a pivot metric (Eq. 2), and three prediction-based quantities: predictability (Eq. 3), suspense (Eq. ąd4), and plot twist (Eq. 5). The empirical illustration uses emotion distributions extracted from actor faces via deepface in a corpus of nine TV shows totaling about 3,000 minutes, reporting per-show entropy and per-episode pivot means, and interpreting between-genre differences as evidence that the framework can quantify narrative structure and emotional dynamics. The prediction-focused measures are introduced formally but not computed.

Significance. If the empirical claims were fully supported, this would be a useful contribution: the definitions are simple, principled, and modality-agnostic in principle, and the framework offers concrete quantities for benchmarking AI-generated stories and for human-in-the-loop narrative tools. The paper deserves credit for making the mathematical definitions transparent and for choosing bounded, symmetric divergence measures. However, the central empirical assertion—that the framework 'demonstrate[s] its ability to quantify narrative structures and emotional dynamics across genres'—is not yet established. The reported genre separations rest on a single pooled entropy value per show, lack confidence intervals or significance tests, and depend entirely on an unvalidated deepface emotion state. The framework's theoretical core is sound, but the demonstration needs substantial statistical strengthening before the claims can be accepted.

major comments (3)
  1. [Section 2.1; Appendix A.2] The claim that the framework 'demonstrate[s] its ability to quantify narrative structures and emotional dynamics ... across different genres' is not supported by the reported analysis. Per-show entropy is computed on a single emotion distribution formed by averaging all states per show (Appendix A.2), so each show contributes exactly one scalar with no confidence interval or episode-level replication. With only 9 shows and 1–16 episodes per show, the separation in Fig. 2D may simply reflect show identity or sampling noise rather than a genre-level narrative property. Please report per-episode entropies and their uncertainty, add bootstrap or Bayesian intervals, and, if possible, a mixed-effects model with genre as a fixed effect and show/episode as random effects; alternatively, explicitly reframe this section as a descriptive illustration rather than a demonstrated capability.
  2. [Fig. 2F; Appendix A.2] The pivot metric's per-show means are plotted with standard errors of the mean computed at episode level, but no statistical test is provided for whether genre differences in mean JSD are significant. With repeated frames within episodes and repeated episodes within shows, the effective sample size is much smaller than the number of frames, and naive comparisons across per-show means would have very low power. Please add an explicit repeated-measures or mixed-effects analysis, report effect sizes with confidence intervals, and state the number of independent units underlying the comparison. This is load-bearing because the genre-difference claims in Section 2.1 rest on Fig. 2D and Fig. 2F.
  3. [Appendix A.2; Section 2.1] All empirical results depend on the deepface emotion distribution as the operationalization of narrative state, yet the manuscript provides no validation that this mapping is faithful or sufficiently rich for comparing genres. This is a correctness risk for the genre-difference claims: if the emotion-from-face mapping is noisy or biased across genres (e.g., due to camera framing, makeup, or acting style), the reported differences may be artifacts of the pipeline. A concrete robustness check would be re-running the main comparisons with an alternative state definition, such as text-based sentiment from subtitles or manual emotion annotations, and showing that the qualitative patterns persist. Without such a check, the empirical demonstration remains coupled to an unvalidated choice.
minor comments (4)
  1. [Section 2.1] There are typographical errors: 'scence' should be 'scene', and 'comparitively' should be 'comparatively'.
  2. [Section 2.2] 'suprising' should be 'surprising' in the sentence about the absence of a change being surprising.
  3. [Appendix A.2] The appendix refers to 'Fig. 1A', 'Fig. 1B', 'Fig. 1C', 'Fig. 1D', and 'Fig. 1F', but the corresponding panels appear as Figure 2 in the main text. Please renumber or adjust the cross-references for consistency.
  4. [Section 3] The sentence 'Our metrics may also assists in identifying systemic biases' has a subject-verb agreement error; it should be 'may also assist'.

Circularity Check

1 steps flagged · score 2.0 of 10

No significant circularity: the information-theoretic measures are explicit definitions computed directly from the emotion-state pipeline; the only mild definitional labeling is interpreting JSD peaks as 'story beats,' which is not load-bearing.

  1. self definitional [Section 2.1, Eq. (2) and Fig. 2E discussion]
    "We describe such pivots as the Jensen-Shannon-divergence (JSD) between two states: Pivot = JSD(st || st−1) (2) ... Peaks indicate moments of significant shifts in emotional state (e.g. in our example at around 25:00) - essentially a story beat."

    The 'pivot' measure is defined in Eq. (2) as the JSD between consecutive states, so the statement that JSD peaks correspond to story beats is true by construction: a peak in the defined quantity is, by definition, a pivot. The paper presents this labeling as an observation rather than as a separate empirical or formal derivation, so it adds interpretative vocabulary but no independently derived content. This step is illustrative and not load-bearing for the mathematical framework or for the genre comparisons.

full rationale

The paper proposes definitions rather than fitting parameters: Complexity = H(st), Pivot = JSD(st||st−1), and the prediction-based measures are introduced theoretically with empirical analysis explicitly deferred. The empirical genre results are direct computations from one emotion distribution per show or episode-level JSD means, so they are not circular in the sense of predicting fitted quantities. The only mildly circular element is the semantic labeling of JSD peaks as 'story beats' or a 'heartbeat,' which follows immediately from Eq. (2); because the underlying quantities remain well-defined and computed directly, this does not undermine the framework. Self-citations (e.g., refs. 15, 18, 20, with author overlap) are used for methodological choices and applications, not as load-bearing justification of the central measures. Statistical robustness of genre separation is a separate correctness concern, not circularity.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The framework's central operation rests on the choice of state representation and the assumption that emotion distributions are a valid proxy for narrative state. No new entities are introduced.

free parameters (2)
  • Narrative state operationalization = Emotion distribution over 7 categories from deepface
    The framework is agnostic about state choice, but all empirical results depend on this choice (Section 2.1, Appendix A.2).
  • Rolling average window = 20 frames
    Used to smooth emotion distributions; affects entropy and JSD values (Appendix A.2).
assumptions (3)
  • domain assumption A story can be decomposed into a sequence of states s_t at each timepoint.
    Assumed in Section 2.1 without justification.
  • domain assumption Emotion distributions from actor faces faithfully represent narrative states.
    The entire empirical analysis relies on this, but no validation is provided.
  • standard math Standard information-theoretic definitions apply to these states.
    Entropy, JSD, and mutual information are used as textbook definitions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Narrative Information Theory." pith.science (2026). https://pith.science/paper/AQS333DV

@misc{pith2026241112907,
  author       = {Pith},
  title        = {Pith review of: Narrative Information Theory},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AQS333DV}},
  note         = {Machine review of arXiv:2411.12907}
}
read the original abstract

We propose an information-theoretic framework to measure narratives, providing a formalism to understand pivotal moments, cliffhangers, and plot twists. This approach offers creatives and AI researchers tools to analyse and benchmark human- and AI-created stories. We illustrate our method in TV shows, showing its ability to quantify narrative complexity and emotional dynamics across genres. We discuss applications in media and in human-in-the-loop generative AI storytelling.

Figures

Figures reproduced from arXiv: 2411.12907 by the authors.

Figure 1
Figure 1. Overview – information-theoretic measures of narratives. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Results – emotions, as well as the complexity and pivot metrics in an example episode [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Supplementary figure: KL-Divergences for example show and trajectories (left) and [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 23 canonical work pages

  1. [1]

    Language as Reality: A Co-Creative Storytelling Game Experience in 1001 Nights Using Generative AI,

    Y . Sun, Z. Li, K. Fang, C. H. Lee, and A. Asadipour, “Language as Reality: A Co-Creative Storytelling Game Experience in 1001 Nights Using Generative AI,” Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment, vol. 19, pp. 425–434, Oct. 2023. Number: 1

    work page 2023

  2. [2]

    AI in storytelling: Machines as cocreators,

    E. Chu, J. Dunn, D. Roy, G. Sands, and R. Stevens, “AI in storytelling: Machines as cocreators,” McKinsey & Company Media & Entertainment, 2017

    work page 2017

  3. [3]

    Narrative Theory for Computational Narrative Under- standing,

    A. Piper, R. J. So, and D. Bamman, “Narrative Theory for Computational Narrative Under- standing,” in Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, (Online and Punta Cana, Dominican Republic), pp. 298–311, Association for Computational Linguistics, 2021. 4

    work page 2021

  4. [4]

    The Data Science of Hollywood: Using Emotional Arcs of Movies to Drive Business Model Innovation in Entertainment Industries

    M. Del Vecchio, A. Kharlamov, G. Parry, and G. Pogrebna, “The Data Science of Hollywood: Using Emotional Arcs of Movies to Drive Business Model Innovation in Entertainment In- dustries,” Journal of the Operational Research Society , vol. 72, pp. 1110–1137, May 2021. arXiv:1807.02221 [cs]

    work page Pith review arXiv 2021

  5. [5]

    The Emotion Dynamics of Literary Novels

    K. Vishnubhotla, A. Hammond, G. Hirst, and S. M. Mohammad, “The Emotion Dynamics of Literary Novels,” Mar. 2024. arXiv:2403.02474 [cs]

    work page Pith review arXiv 2024

  6. [6]

    Quantifying the narrative flow of imagined versus autobiographical stories,

    M. Sap, A. Jafarpour, Y . Choi, N. A. Smith, J. W. Pennebaker, and E. Horvitz, “Quantifying the narrative flow of imagined versus autobiographical stories,” Proceedings of the National Academy of Sciences, vol. 119, p. e2211715119, Nov. 2022. Publisher: Proceedings of the National Academy of Sciences

    work page 2022

  7. [7]

    Modeling Narrative Revelation,

    A. Piper, H. Xu, and E. D. Kolaczyk, “Modeling Narrative Revelation,” inHR 2023: Computa- tional Humanities Research Conference, (Paris, France), 2023

    work page 2023

  8. [8]

    Exploration and exploitation of Victorian science in Darwin’s reading notebooks,

    J. Murdock, C. Allen, and S. DeDeo, “Exploration and exploitation of Victorian science in Darwin’s reading notebooks,”Cognition, vol. 159, pp. 117–126, Feb. 2017

    work page 2017

Show all 23 references
  1. [9]

    Individuals, institutions, and innovation in the debates of the French Revolution,

    A. T. J. Barron, J. Huang, R. L. Spang, and S. DeDeo, “Individuals, institutions, and innovation in the debates of the French Revolution,” Proceedings of the National Academy of Sciences, vol. 115, pp. 4607–4612, May 2018. Publisher: Proceedings of the National Academy of Sciences

    2018

  2. [10]

    Pizzo, V

    A. Pizzo, V . Lombardo, and R. Damiano,Interactive Storytelling: A Cross-Media Approach to Writing, Producing and Editing with AI. Taylor & Francis, Sept. 2023

    2023

  3. [11]

    Evaluating Emotion Arcs Across Languages: Bridging the Global Divide in Sentiment Analysis,

    D. Teodorescu and S. Mohammad, “Evaluating Emotion Arcs Across Languages: Bridging the Global Divide in Sentiment Analysis,” inFindings of the Association for Computational Linguis- tics: EMNLP 2023, (Singapore), pp. 4124–4137, Association for Computational Linguistics, 2023

    2023

  4. [12]

    Shapes of Emotions: Multimodal Emotion Recognition in Conversations via Emotion Shifts,

    H. Agarwal, K. Bansal, A. Joshi, and A. Modi, “Shapes of Emotions: Multimodal Emotion Recognition in Conversations via Emotion Shifts,” Nov. 2022. arXiv:2112.01938 [cs]

    2022 arXiv

  5. [13]

    Frowning Frodo, Wincing Leia, and a Seriously Great Friend- ship: Learning to Classify Emotional Relationships of Fictional Characters,

    E. Kim and R. Klinger, “Frowning Frodo, Wincing Leia, and a Seriously Great Friend- ship: Learning to Classify Emotional Relationships of Fictional Characters,” Apr. 2019. arXiv:1903.12453 [cs]

    2019 arXiv

  6. [14]

    Emotion dynamics in movie dialogues,

    W. E. Hipson and S. M. Mohammad, “Emotion dynamics in movie dialogues,” PLOS ONE, vol. 16, p. e0256153, Sept. 2021. Publisher: Public Library of Science

    2021

  7. [15]

    Multimodal Emotion Recognition for Visualizing Storyline in a TV Series,

    T. Crijns, M. Doyran, O. S. Kayhan, R. Klein, V . Koops, C. Laugs, D. Odijk, A. A. Salah, A. Serebrenik, Y . Tımar, and A. V olk, “Multimodal Emotion Recognition for Visualizing Storyline in a TV Series,” 2020

    2020

  8. [16]

    The Shapes of Stories: Sentiment Analysis for Narrative,

    K. Elkins, “The Shapes of Stories: Sentiment Analysis for Narrative,” Elements in Digital Literary Studies, July 2022. ISBN: 9781009270403 9781009270397 Publisher: Cambridge University Press

    2022

  9. [17]

    Relating objective complexity, subjective complexity and beauty,

    S. S. Nath, F. Brändle, E. Schulz, P. Dayan, and A. A. Brielmann, “Relating objective complexity, subjective complexity and beauty,” 2023. Publisher: PsyArXiv

    2023

  10. [18]

    RADio* – An Introduction to Measuring Normative Diversity in News Recommendations,

    S. Vrijenhoek, G. Bénédict, M. Gutierrez Granada, and D. Odijk, “RADio* – An Introduction to Measuring Normative Diversity in News Recommendations,”ACM Trans. Recomm. Syst., vol. 3, pp. 5:1–5:29, Aug. 2024

    2024

  11. [19]

    Frontiers: In-Consumption Social Listening with Moment-to- Moment Unstructured Data: The Case of Movie Appreciation and Live Comments,

    Q. Zhang, W. Wang, and Y . Chen, “Frontiers: In-Consumption Social Listening with Moment-to- Moment Unstructured Data: The Case of Movie Appreciation and Live Comments,” Marketing Science, vol. 39, pp. 285–295, Mar. 2020. Publisher: INFORMS

    2020

  12. [20]

    Find the Cliffhanger: Multi- modal Trailerness in Soap Operas,

    C. Bretti, P. Mettes, H. V . Koops, D. Odijk, and N. van Noord, “Find the Cliffhanger: Multi- modal Trailerness in Soap Operas,” in MultiMedia Modeling (S. Rudinac, A. Hanjalic, C. Liem, M. Worring, B. P. Jónsson, B. Liu, and Y . Yamakata, eds.), (Cham), pp. 199–212, Springer ...

    2024

  13. [21]

    How quantifying the shape of stories predicts their success,

    O. Toubia, J. Berger, and J. Eliashberg, “How quantifying the shape of stories predicts their success,” Proceedings of the National Academy of Sciences, vol. 118, p. e2011695118, June 2021

    2021

  14. [22]

    Information theory and music,

    J. E. Cohen, “Information theory and music,” Behavioral Science, vol. 7, no. 2, pp. 137–163,

  15. [1962]

    Soap" in our plots). • 12 episodes of a dating format where couples test the strength of their relationships by living with attractive singles on a tropical island (

    _eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1002/bs.3830070202. A Appendix / supplemental material A.1 Dataset For the analysis that is illustrated in Figure 1, we had access to an internal dataset from a large European network. This dataset encompassed over 3000 minut...

    doi

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.