khyang-lab:~$

Introduction to Computational Social Science · an explorable

Humiliation, payback, reveal: measuring the moral formula of micro-dramas

We start with one 86-minute drama and end with 77,000 of them. On the way you will see how computational social scientists turn a hunch about stories into measurements, check whether the measurements can be trusted, and use them to compare formats, languages and years.

A research case study from an ongoing project (Kunhao Yang & Mengyuan Fu). Numbers may change before publication. The example drama contains depictions of violence and humiliation.

Micro-dramas are serialized stories told in one- to two-minute episodes, made for phones. Many people describe them as "all the same story". That is a hunch. Computational social science asks whether the hunch can be measured, at what scale, and with what confidence.

  1. 1One drama, scene by sceneWatch, code it yourself, then compare with the machine.
  2. 2From a hunch to a questionTheory tells us what to measure, and why.
  3. 3Building the dataA sampling frame, transcripts, benchmarks, rules.
  4. 4Four ways to measure a storyAn LLM, word embeddings, a dictionary model, a path.
  5. 5Can we trust the measures?Agreement between methods that fail differently.
  6. 6What we foundFormats, languages, time, genres, attention.

1One drama, scene by scene

Our example is a Chinese micro-drama dubbed into Japanese and uploaded to a Japanese-language YouTube channel on 21 July 2026. It runs 86 minutes and was cut together from many short episodes. It is not part of the study's sample; we use it only to show each step.

The example

Open the drama on YouTube (new tab). The title, as shown on YouTube: 【日本語吹き替え】美しき義母が愛人扱いされ暴行！婚約式で息子が母を庇い、有名人嫁の偽りの顔を暴く！

In English, the title promises: a beautiful stepmother is treated as a mistress and assaulted; at the engagement ceremony her son protects her and exposes the celebrity bride's false face.

Your turn, before you scroll

Watch 0:00 to 8:40. For every moral event you notice, write down four things:

  1. Who wrongs whom? Is the wrongdoer on the hero's side or the villain's side?
  2. Is anyone looked down on because of their status (wealth, class, age, looks, job)?
  3. Is anyone punished, exposed or humiliated for an earlier wrong?
  4. Is a hidden identity, fortune or power revealed?

What the pipeline saw

The drama was transcribed, translated into English, cut into 25 scenes of about 500 words each, and every scene was read by a large language model (LLM) with a written codebook. Click a scene to see what was coded there; each scene links to that moment of the video.

The example drama, scene by scene

  • W a wrong by the villain's side
  • C status contempt
  • R retribution
  • V reveal

Select a scene above.

Each cell is one ~500-word scene, from 0:00 to 1:26:17; the number on top is the minute where it starts. Letters repeat once per coded event. Names are as the model wrote them from the translation, so the same character may appear under different spellings (e.g. Junsok / Jun-seok); a few labels stayed Japanese (婚約者 fiancée, 母 mother, 息子 son).

Where does this drama sit?

The same four numbers, next to the averages of 1,090 English micro-dramas and of soap operas, prime-time TV and films:

Teal = the example drama; blue = English micro-drama; greys = soap, TV, film. Rates are events per 1,000 words of dialogue.

Code it yourself: four events from this drama

For each event, pick the side of the wrongdoer and the event's role in the story. Then check what the model said.

One drama is not data
  • The measure depends on the text. Coding the Japanese text directly found no status contempt at all; coding the English translation found four events. Both versions agree on almost everything else (retribution 1.59 vs 1.51 per 1,000 words). For one drama, small differences in wording matter; across hundreds they average out.
  • The rule decides what counts. We say a story "runs the formula" if contempt, retribution and a reveal all occur in its first ~4,000 words. This drama has the first two early, but its only "reveal" comes at 57 minutes. By our rule it does not count, even though it plainly uses the formula. Every operational rule draws a line somewhere; the point is to draw it before looking at the results and to apply it the same way to every story.

2From a hunch to a question

The social phenomenon

  • Scale. China's broadcasting regulator reported 430,000 micro-drama titles released online in January–August 2026, and more than 90% of the new titles were AI-generated (NRTA statement, 17 September 2026).
  • Export. Platforms such as ReelShort carry Chinese story templates abroad and localise them (Tang & Wang, 2025). On YouTube, we found 77,214 long English micro-drama uploads in 21 months, and 403,544 long uploads on Chinese-language channels.
  • Standardisation at home. A content analysis of popular Chinese urban-romance micro-dramas finds "highly standardized" themes, characters and plots (Hu & Zhu, 2026).

Long micro-drama uploads per month in our YouTube frames

Left: English-language channels (77,214 uploads, 303 channels). Right: Chinese-language channels (403,544 uploads of 20–240 minutes, 721 channels). Separate scales; both rise steeply.

Why moral structure?

A hunch ("they are all the same") is not yet a research question. Theory tells us which sameness matters:

  • Enjoyment depends on moral judgment. Disposition theory: viewers enjoy drama when they like the characters who are rewarded and dislike those who are punished (Raney, 2004; Raney & Bryant, 2002).
  • Producers follow what they think audiences want. The moral domains a story engages predict its appeal, and producers make content to match the intuitions they attribute to audiences (Tamborini et al., 2013; Hahn & Tamborini, 2020).
  • Cultural industries imitate under uncertainty. When nobody knows what will succeed, organisations copy what seems to work (Hirsch, 1972; DiMaggio & Powell, 1983).

From theory to four measurable elements

1 · Who does wrong

The share of wrongs committed by the villain's side: moral clarity, which makes outcomes satisfying (Lewis et al., 2017).

2 · Status contempt

Looking down on someone for their position. Contempt rejects and excludes (Fischer & Roseman, 2007), and social comparison turns status gaps into scorn (Fiske, 2010).

3 · Retribution

The offender is punished or exposed: deserved punishment drives enjoyment (Raney, 2005; Grizzard et al., 2021).

4 · Hidden-status reveal

Hidden worth is revealed, turning contempt into recognition, like the "overbearing CEO" Cinderella fantasy (Song, 2023).

We add a fifth, structural measure, narrative shape: the path a story takes through meaning space (Toubia, Berger & Eliashberg, 2021).

Research questions
  1. Structure and comparison. How does micro-drama's moral structure differ from soap opera, prime-time TV and film? Does the English-language version differ from Chinese micro-drama, where the format began?
  2. Time. How did the stories and their packaging (titles) change between January 2025 and September 2026?
  3. Genre. Did micro-drama converge on one formula instead of exploring diverse genres?
Quick check: which statement is a measurable construct?

3Building the data

A measurement is only as good as the sample it is applied to. Most of the work in this project went into building data that could support a comparison.

  1. Find channels303English micro-drama channels from month-by-month searches, Jan 2025–Sep 2026
  2. List every upload77,214long uploads, via the official YouTube Data API
  3. Draw a sample1,090dramas at random within each month (at most 48 per month, 3 per channel)
  4. Transcribe5-minsegments read by a multimodal model (Gemini)
  5. Code scenes60,102~500-word scenes across all formats
  6. Compare2,727soap, TV and film units, coded the same way

Why not just search YouTube?

In a pilot, YouTube's search found only 0–55% of the long micro-drama uploads that the same channels actually posted on a given day, and it found fewer for older days. A search result is shaped by the platform's ranking, so it is not a sample of what exists. We therefore used search only to discover channels, then listed every upload of those channels.

Comparison groups

  • Legacy formats from public research datasets, passed through the identical pipeline: 420 US daytime soap-opera episodes, 1,829 prime-time TV episodes (SummScreen; Chen et al., 2022) and 478 films (MovieSum; Saxena & Keller, 2024).
  • Chinese micro-drama, two routes on the same platform: 248 Chinese-audio dramas uploaded to the English-market channels, and 252 dramas from Chinese-language channels.
  • Titles (the packaging): 66,620 English titles and 7,471 Chinese micro-drama titles.

4Four ways to measure a story

We measure each element with methods that work in different ways. If they agree, we are more confident that we measure the story and not a quirk of one method.

4.1 LLM event coding: reading like a trained coder

A large language model reads each ~500-word scene with a codebook based on the six moral foundations of the MFQ-2 (Atari et al., 2023): care, equality, proportionality, loyalty, authority and purity. For every concrete moral act it records:

FieldExample from scene 2 of our drama (8:40)
agent → targetthe new fiancé and associates → the grandmother
act (the model's paraphrase)physically assaulted and attempted to strip an older woman
foundation / polaritycare / violation
side of the actorantagonist (villain's) side
narrative functionescalation (other options: initial wrong, retribution, reward, status reveal, reconciliation)
status contempt? basisyes; age

Counting these codes gives rates per 1,000 words: contempt, retribution, reveals, all moral events, and the villain's share of wrongs. Strength: it understands who does what to whom. Weakness: it can misread, and it can be confidently wrong (you saw two such codes in Part 1).

4.2 Seed-based relevance: how much is a scene about contempt?

This method, from Yang & Fu (2025), uses no language model, only word statistics:

  1. All 60,102 scenes form one corpus. A scene × word table is weighted by pointwise mutual information and compressed to 100 dimensions (a kind of word embedding learned from our own data).
  2. A short list of seed words is fixed before any analysis, for example poor, peasant, beggar, trash, servant, commoner for contempt, and projected into the same space.
  3. Each scene gets a similarity score (cosine) to the seeds. Most scenes form a bell-shaped background; scenes really about the topic form a long right tail. A two-part model (a Gaussian for the background, a heavy-tailed Cauchy for the relevant scenes) is fitted to all scores, and each scene's probability of belonging to the tail is its relevance.

Similarity of 60,102 scenes to the seed words

Bars: all scenes of all formats. Dashed grey: fitted background; blue: fitted relevant tail. Teal ticks: the 25 scenes of the example drama. The seeds for the selected construct are listed below.

4.3 Moral sentiment: closer to moral words or to neutral words?

The model of Xie et al. (2019) represents a scene by the average of its word vectors (pretrained fastText). It asks whether that average lies closer to the centre of the Moral Foundations Dictionary vocabulary or to the centre of emotionally neutral words. The result is a log-odds of moral relevance. It cannot tell who did what to whom, but it needs no training and is easy to reproduce.

4.4 Narrative shape: a story's path through meaning space

Toubia, Berger & Eliashberg (2021) treat a story as a path. We cut each drama into ~250-word windows and embed each window as a point in a high-dimensional "meaning space". Then:

  • Speed is how far the story moves between consecutive windows,
  • volume is how much ground it covers overall,
  • circuitousness is how much it returns to earlier themes instead of moving on.

Four story paths, projected onto two dimensions

Each dot is one window of dialogue; the hollow dot is the start. All four panels use the same scale. Hover a dot to see its position in the story. Projections lose information: the statistics are computed in the full space.

5Can we trust the measures?

There is no answer key for "how much contempt is in this drama". Instead, we look for agreement between methods that fail in different ways, and we report where they fail.

CheckWhat it showed
LLM vs seed-based relevance (same dramas)Spearman .52 (contempt), .37 (retribution), .33 (reveal)
LLM moral events vs Xie moral relevance.49
A second model of another company (Qwen3-235B, 1,000 scenes)Same ordering of formats; levels differ (Qwen codes 2–5× more events)
Quotes required for every coded act99.5–99.9% of the quotes are found in the transcripts
Chinese dramas coded in Chinese vs in English translationSame averages (ratios 0.98–1.21)
Xie's second tier, moral negativityFails: −.14 with LLM-coded wrongs. Dropped.
Human codersA double-coded student validation is prepared; it will set absolute levels

Two different LLMs, same conclusion: how many times more than soap, TV, film?

Ratio of micro-drama to each legacy format on the same 1,000 scenes. Blue = Gemini (the study's coder), orange = Qwen. Every ratio is above 1 under both models, although the two models disagree on many individual scenes.
The logic of validation

Levels depend on the instrument; orderings and gaps are what we can defend. That is why the findings below are reported as differences and ratios between groups measured the same way.

6What we found

6.1 A moral signature, about one standard deviation apart

Predict first: per 1,000 words, how much more status contempt does micro-drama stage than prime-time TV?

Micro-drama vs soap, TV and film

95% intervals resample channels, shows or films. Micro-drama is weighted to each month's uploads.
  • All methods agree. Micro-drama routes 84% of wrongs to the villain's side (60–62% in soap and TV) and stages several times more contempt, retribution and reveals: about one standard deviation more than every legacy format.
  • Soap talks, micro-drama stages. Seed-based relevance finds soap dialogue close to micro-drama on retribution and reveal words, but the LLM finds soap rarely stages those events.
  • Robust. The gaps hold within genre (romance: contempt .55 vs .10), with each month counted equally, and under the second LLM.
  • Shape. Micro-drama moves more slowly and covers less ground than every legacy format (film is 6% faster), and it is among the most direct: a short, straight corridor packed with moral events.

6.2 A Chinese template, carried intact

If the formula were adapted for foreign audiences, English versions should carry more of it. They do not.

English vs Chinese micro-drama

Route A: Chinese-audio dramas uploaded to the English-market channels. Route B: Chinese-language channels. Hollow diamonds: the same Chinese dramas coded in English translation. Each month counts equally.
  • Chinese dramas carry 80–100% of the gap between English micro-drama and the legacy formats on villain share, contempt and retribution.
  • Within the same month, English minus Chinese is small and mostly not significant (contempt +0.09 SD; retribution about 0).
  • English dubbing does not add intensity: dubbed versions are, if anything, slightly less dense. Western-cast productions match the Chinese originals.

6.3 Stable stories, an inflating package

Over 21 months the stories changed only through which channels grew; within channels they stayed the same. The titles moved.

Stories (quarterly) and titles (monthly)

Titles advertising the formula: two markets

Share of titles naming snobbery, promising revenge or promising a hidden-status reveal. English: 66,620 titles (+12.2 points per year, +3.6 within channel). Chinese-language micro-drama titles: 7,471 sampled titles, already formula-heavy and flat. Bands: 95% intervals over channels.

Do the titles keep their promises? For the same 1,090 English dramas, the share of titles advertising the formula rose by 8.3 points per year, while the share of stories that run it stayed at 16–20%. And when the same drama was uploaded again under a new title, the new title leaned more toward the formula:

Retitling drift: later minus earlier title of the same drama (percentage points)

31,327 pairs of uploads of the same drama under different titles, matched by the opening dialogue. The placebo row checks that this is not just longer titles: neutral setting words (CEO, wife, family) do not drift.

6.4 One engine, many worlds

Did micro-drama collapse into one genre? No. Romance and marriage stay at about half of all titles, and the effective number of genres stays near 4.8. Instead, the formula spread into every genre:

Titles advertising the formula, by genre: 2025 → 2026

Of the 9-point rise, 8 points happened within genres; only 1 came from a shift between genres.

6.5 The formula does not buy attention

Do formula-heavy stories or titles get more views? We compare each drama with the same channel's other uploads that month, and titles with other titles of the same drama:

Titles and views: within channel vs same drama

Percent difference in views per title element. Grey: compared within channel. Blue: compared within the same drama (29,915 pairs of the same drama under different titles). Once the drama is held fixed, no moral element of the title survives correction for multiple tests.
Within channels, titles naming snobbery get about 9% fewer views. Why is that not yet evidence that snobbery in a title costs views?

?Discussion

  1. The formula is a Chinese template that travels intact, while English titles inflate toward it. What does that suggest about how platforms spread cultural formats?
  2. Contempt in these stories is mostly about wealth and class. What might millions of hours of such stories teach viewers about inequality, and how could we study that without assuming an effect?
  3. We used an LLM as a measuring instrument. What would convince you that it measures the same thing in English, Chinese and Japanese?
  4. The formula does not predict views. Why might producers adopt it anyway?

Limitations

  • Human validation is still running; absolute levels depend on the coding model.
  • Legacy texts are human transcripts and screenplays from earlier decades; micro-drama transcripts are machine-made.
  • Our Chinese data are Chinese-language micro-dramas on YouTube, not China's domestic platforms.
  • Views are a single snapshot, and titles are not assigned at random.

#Reading & methods

Data and code

Data were collected through the official YouTube Data API between September and October 2026 and processed with Gemini (gemini-3.5-flash-lite for transcription and coding, gemini-embedding-001 for narrative shape) and, as a check, Qwen3-235B-A22B-Instruct. Raw transcripts and API data are not shared; this page uses derived codes and aggregates only.

References

  1. Atari, M., et al. (2023). Morality beyond the WEIRD: How the nomological network of morality varies across cultures. Journal of Personality and Social Psychology, 125(5), 1157–1188.
  2. Chen, M., Chu, Z., Wiseman, S., & Gimpel, K. (2022). SummScreen: A dataset for abstractive screenplay summarization. ACL 2022, 8602–8615.
  3. DiMaggio, P. J., & Powell, W. W. (1983). The iron cage revisited. American Sociological Review, 48(2), 147–160.
  4. Fischer, A. H., & Roseman, I. J. (2007). Beat them or ban them: The characteristics and social functions of anger and contempt. Journal of Personality and Social Psychology, 93(1), 103–115.
  5. Fiske, S. T. (2010). Envy up, scorn down: How comparison divides us. American Psychologist, 65(8), 698–706.
  6. Grizzard, M., et al. (2021). Narrative retribution and cognitive processing. Communication Research, 48(4), 527–549.
  7. Hahn, L., & Tamborini, R. (2020). Research on the model of intuitive morality and exemplars. In The international encyclopedia of media psychology. Wiley.
  8. Hirsch, P. M. (1972). Processing fads and fashions. American Journal of Sociology, 77(4), 639–659.
  9. 胡岑岑 & 朱格 [Hu, C., & Zhu, G.] (2026). Narrative conventions of urban romance micro-dramas and their effects on audience emotional experience and engagement. 《全球传媒学刊》, 13(3), 41–61.
  10. Lewis, R. J., Grizzard, M., Mangus, J. M., Rashidian, P., & Weber, R. (2017). Moral clarity in narratives elicits greater cooperation than moral ambiguity. Media Psychology, 20(4), 533–556.
  11. Raney, A. A. (2004). Expanding disposition theory. Communication Theory, 14(4), 348–369.
  12. Raney, A. A. (2005). Punishing media criminals and moral judgment. Media Psychology, 7(2), 145–163.
  13. Raney, A. A., & Bryant, J. (2002). Moral judgment and crime drama. Journal of Communication, 52(2), 402–415.
  14. Saxena, R., & Keller, F. (2024). MovieSum: An abstractive summarization dataset for movie screenplays. Findings of ACL 2024.
  15. Song, G. (2023). The overbearing CEO: Cinderella fantasy and Chinese-style neoliberal femininity. Modern Chinese Literature and Culture, 35(1), 201–226.
  16. Tamborini, R., Eden, A., Bowman, N. D., Grizzard, M., Weber, R., & Lewis, R. J. (2013). Predicting media appeal from instinctive moral values. Mass Communication and Society, 16(3), 325–346.
  17. Tang, W., & Wang, Y. (2025). ReelShort as a new template of international short-drama business. Global Media and China.
  18. Toubia, O., Berger, J., & Eliashberg, J. (2021). How quantifying the shape of stories predicts their success. PNAS, 118(26), e2011695118.
  19. Xie, J. Y., Ferreira Pinto Jr., R., Hirst, G., & Xu, Y. (2019). Text-based inference of moral sentiment change. EMNLP-IJCNLP 2019, 4654–4663.
  20. Yang, K., & Fu, M. (2025). Heat in real time, climate in retrospect: Mapping seasonal gap in Japanese heatwave journalism. Journalism Studies.

Kunhao Yang (Shibaura Institute of Technology) and Mengyuan Fu (Waseda University). Charts are drawn in your browser from the files in data/; no data leave this page. Fonts: IBM Plex and Pixelify Sans (SIL Open Font License). The example drama belongs to its rights holders and is referenced for teaching and commentary only.