← Dutch Angle Nederlands

Measurement handbook

How Dutch Angle measures the room a film gives to women and men: the sources, the 20 questions in 10 dimensions, the formulas, the angle, the checks and what the measurement cannot do. For coders, researchers, teachers and journalists.

Version 0.5, last updated 31 October 2026. For coders, researchers, teachers, journalists and partners.

Dutch Angle (by The Missing Measure) measures how a film divides room between women and men: who speaks, whose story it is, who decides, and how the story treats them. This handbook defines every question precisely enough that two people, or a person and a language model, working independently on the same film arrive at the same score. It describes question set v0.5 with LICHAAM: 20 questions in 10 dimensions that count equally, plus one signal on the film card (K2). E1 (POWER) and the dimension LICHAAM (BODY) with K1 were added after a test on 31 October 2026 (32-proef-nieuwe-vragen/verslag.md); the dimension LIEFDE (LOVE) with R2 and N1 on 16 November 2026 (41-liefde-archief/verslag.md). The tests behind every choice are documented in the project's question review, available on request.

Status. The story judgements (S2, S3, M1, W1) are made twice, blind, for 1,323 films: over 4,972 pairs of answers, 80% differ by at most a quarter point (Krippendorff's α across the archive: S2 0.76, S3 0.88, M1 0.91, W1 0.82; on the first 60 films 0.91 to 0.96). Every question passed a face-validity test on 20 clearly female-led and male-led films. Three outside checks agree with the measurement: bechdeltest.com (Part D), the screenplay dialogue counts of The Pudding (309 films, correlation 0.97 with our share of women's words, on average 4 percentage points apart) and subtitles with speaker names (260 films, 22,675 lines: the character agrees in 71% of the lines, the gender in 90%). Not yet done: a check by human coders, a Lauzen-style hand coding, and the Vrouwen in Beeld coding for Dutch films. BODY and E1 (31 October 2026): E1 was judged twice, blind, for all films (1,362 films with an answer, α 0.93; in 76 films E1 does not count because a lead is not named in the plot). K1 was judged once across the archive (answers in 685 of 1,464 films, from 3 judgements); a second, blind reading of a spread sample of 327 films gave α 0.98 (143 pairs). The K2 signal was read once (the minor category separately, in 232 films with a search hit); on the same sample the second reading agrees on case or no case in 313 of 327 films (α 0.87; on the number of cases 0.92). LOVE (16 November 2026): R2 and N1 were read twice, blind, for 2,337 films; R2 gives an answer in 1,246 films (α 0.91), N1 in 209 films (α 0.99). There is no check by people yet. Where a rule has not been tested, this handbook says provisional.


Part A. Ground rules

A1. Sources

Source What we take from it Used for
Subtitles of the complete film (OPUS OpenSubtitles v2024, one file per film, chosen by choose.py) Every spoken line with its timestamps P2, P3, St2, L1, I2, I3, J2, A5, F1, K1, K2, R2, N1 (candidate lines)
Wikidata Cast list with character names, the actors' gender and date of birth, billing order, release date, director and writers, country of origin, animation Gender, S1, G1, G3, context
Wikipedia plot summary (English; Dutch when there is no English one) The story as editors summarise it S1, S2, S3, M1, W1, E1, K2, R2, N1

A2. Gender

A3. From subtitles to speakers

  1. A reader (a language model, in the session that measures the film) reads the subtitles in order and assigns every line to a character from the cast list, or to unknown. It writes these as line ranges in a source file (17-nieuwe-films/bron/<imdb>.txt, or the annotation files for the first 60 films).
  2. Characters without a name who are clearly a woman or a man (a doctor, a cop, a narrator, an announcer) get a generic role with that gender and naam: false. They count for presence and conversation, not as a name in I2. The gender may come from the text, or from the plot summary when the plot clearly points to that one person ("the old jailer" who narrates, "the teacher"); this holds for one identifiable person only, never for a group or a whole profession (decision 31 October 2026).
  3. Group scenes. When the text gives everyone who speaks in a scene the same gender, those unnamed lines go to a generic group role with that gender (Men, Women, Boys, Girls, Soldiers, Nuns; naam: false). The text has to say so: forms of address ("lads", "gentlemen", "ladies", "girls", "sir", "ma'am"), self-references ("my husband", "I'm a father of three") or a setting that the dialogue names (a convent of nuns, a men's barracks). A mixed or unclear group stays unknown. This rule applies from the first reading in every batch; it is not a later repair (decision 23 October 2026, after batch 10 had first left these scenes unknown). A group role holds per scene, and only where the text in that scene fixes the speakers' gender; a whole profession or family is never given a gender on indirect evidence ("guys", "two men went to check") (decision 29 October 2026). The same goes for a scene in which only known characters of the same gender speak but it is not clear which of them says a line: those lines go to the group role. There is no maximum number of characters, as long as the text in that scene fixes their gender (decision 31 October 2026; before that the limit was two or three, for example the four truck drivers in The Wages of Fear and the four friends in Now and Then). A form of address fixes the gender of the people addressed only; the speaker counts too only when the same sentence uses we, us or our about the group ("We have him boys"). Untranslated forms of address ("signori", "ma chérie", "Caballeros") count like their English equivalents. An SDH label with a gender ("MAN:", "WOMAN 2:") fixes the gender of that one speaker for that line only (decisions 10 November 2026). Exceptions decided per film, after batch 21 (14 November 2026): a crew or group that the text once fixes as men ("brother tankies", "OK, Boys") counts as men in all its scenes (T-34); a meeting the dialogue itself calls "men's business" counts as men for the whole meeting (Phase 7); "my girlfriend" without a name counts as evidence for the partner it clearly refers to (Bad Things). These hold for those films only until Astrid decides about the rest of the archive. Decisions 16 November 2026 (after batch 23): "man" and "bro" as a form of address count like "boys"; "Rabbi" and "best man" are gender words like "Father" for a priest; "the kind of guy" said about oneself is a self-reference; untranslated "bhai" does not count. A sleeping woman counts as present in a scene (unlike an unconscious one).
  4. Songs are left out. Credits and on-screen text are not left out separately: in the whole archive they are about 0.03% of the unknown words (measured 23 October 2026), too little to change any film.
  5. Coverage is the share of lines (and seconds) with a known speaker, that is a speaker who is a woman, a man or identifies otherwise. The aim is at least 0.80 of the words; below it the verdict is provisional. Check each film's coverage right after measuring it, before the next film. Films below it are named in the measurement log (in the current archive: Nomadland 0.75, Past Lives 0.71, whose subtitles contain only the English dialogue).
  6. While reading, the reader also tags lines that are about work (J2) or about the other as a (possible) partner (L1), and records the main couple (G1).
  7. Quotes in the evidence must be literal substrings of the line and at most 25 words; a script rejects any that are not.

A4. Who "she" and "he" are

Term Definition
Women / men For counts: all characters with lines of that gender, named or not.
Named character A character with a personal name in the cast list or the dialogue. "Waitress" or "Man at bar" is unnamed.
Leads For the story questions (S2, S3, W1, E1): at most two women and two men, chosen by the judge from the cast list on the film card, as the characters the plot summary follows, mentions most and who drive events. Gender comes from the card, never from the judge. If no woman (or no man) in the list plays a real part in the plot, none is chosen for that side.
Main couple The female and male character of the main romantic storyline, recorded by the reader (A3). Used by G1 only.

Counts cover all characters with lines. Story judgements cover the leads.

A5. Units

Unit Definition
Line One subtitle segment with a speaker.
Word A word in a line, split on spaces.
Exchange Two consecutive lines by two different known speakers, the second starting within 30 seconds of the first. The conversation network is the count of exchanges between each pair of characters.
Turning point A decision by a named character without which the plot would go differently. Things that happen to someone, and decisions by unnamed groups, are not turning points.

A6. Sign and scale

Every question compares women with men and gives a score from −1 to 1: - 0 = equal. - Positive = the men get more room (more words, more of the story, more decisions, better treatment). - Negative = the women get more room. - ±1 = everything on one side.

On the site the sign is mirrored (since 31 October 2026): women are on the right with a plus sign, men on the left with a minus sign, so a film that gives the men more room shows as, for example, −13.9°. The formulas in this handbook keep the internal convention above; only the display is flipped.

A7. Formula families

w refers to women, m to men.

Family Formula Used by
Share score = 1 − 2·w/(w + m): one minus twice her share P2, P3, S1
Ratio score = (a − b)/(a + b), where a and b are rates, each relative to its own group, and a is the one that favours the men St2, L1, I2, I3, J2, K1
Events score = (n_m − n_w)/n, over n events, each by a man, by a woman, or shared M1, F1
Lead categories Each lead gets a category worth 1, 0.5 or 0; score = mean of the men − mean of the women S2, S3, W1, E1
Judgement One of five anchors: −1, −0.5, 0, 0.5, 1 (Appendix 1) A5
Cast Age difference in years divided by 20, capped at ±1 G1, G3

When one side has no leads. If a film has male leads and no female lead, S2, S3, W1 and E1 score 1 (all room to the men). The reverse gives −1.

A zero denominator is never a score of 0. Below each question's minimum (A8) it is reported as too little data and does not count.

A8. Minimum numbers

Question Scored only when
P2 at least 80% of all words have a speaker with a gender
P3 at least 4 characters in the network's top five have a gender
St2 at least 10 same-gender exchanges, and at least 2 expected for each pair type
S1 billing or plot names are available (either part counts on its own)
S2, S3, W1, E1 at least one lead on one side; for E1 all leads must also be named in the plot
K1 at least 50 lines on each side and at least 3 judgements
R2 there is a love line between a woman and a man in which someone is won over, and both runs give an answer
N1 both runs find an ignored no
M1 at least 3 turning points with a woman, a man or both deciding
I2 at least 5 references (names plus role words) on each side
I3 at least 4 named characters of each gender
J2, L1 at least 50 lines on each side
G1 a main couple with both ages known
G3 ages known for speakers on both sides; not for animation
A5, F1 always scored when the film has dialogue (F1 needs at least one evidence quote, also when the score is 0)

A9. From questions to the angle

  1. Dimension score = the mean of that dimension's scored questions.
  2. Angle = 45 × the mean of the dimension scores, rounded to 0.1°. So, internally, −45° is everything her way, 45° everything his way, 0° in balance (on the site mirrored, see A6).
  3. An angle needs at least 5 of the 10 dimensions. A verdict needs at least 7; with 5 or 6 the film is shown as provisional (voorlopig) and its ball in the archive is see-through. A verdict also needs a speaker with a known gender for at least 80% of all words (the P2 coverage); below that the film is provisional too, with the coverage shown (decision 18 Oct 2026). Before this check, characters without a gender in Wikidata get the gender of the character by hand, recorded as gbron: hand (personage).
  4. Margin of error = 1.96 × the standard deviation of the dimension scores ÷ √(number of dimensions) × 45, rounded to whole degrees. The site says whether the interval stays on one side of zero ("even at the low end this film leans his way") or crosses it ("the direction is not certain yet"). The margin reflects how much the dimensions disagree, not the difference between coders.
  5. On the seesaw every scored question is a ball that falls to the side it leans to, further out the stronger it leans. The colour and shape show the dimension.
  6. The cast is shown separately, as the number of named women and men who speak. It explains a tilt but does not count in it.
  7. Two pillars. On the site the ten dimensions are grouped in two pillars: Room (WHO TALKS, WHOSE STORY, POWER, TOGETHER) and Treatment (NAME, WORK, AGE, SACRIFICE, BODY, LOVE). Each pillar is the mean of its measured dimensions. The angle stays the mean of all dimensions, so with all ten measured the Room pillar weighs four tenths and Treatment six tenths (since LOVE, 16 November 2026); the pillars do not weigh the same, and presence-driven dimensions weigh less than half, on purpose (Astrid's decision, 31 October 2026, confirmed after BODY was measured: no other formula). When the pillars point opposite ways, the film page says so. A film where women speak less than 10% of the words is labelled a world without women (and the reverse a world without men); the label explains the angle and does not change it.

A10. Verdicts

The verdict follows from the angle, equally far on both sides:

Angle Dutch English
−5° to 5° In balans Level
5° to 15° Scheef naar hem / haar Tilted towards him / her
15° to 25° Flink scheef naar hem / haar Strongly tilted towards him / her
beyond 25° Zwaar scheef naar hem / haar Heavily tilted towards him / her

On the site the verdict is also summed up in three boxes: It's a man's world, It's a shared world and It's a woman's world, with each measured dimension in the box it leans to (shared: at most 5° either way).

A11. Who judges, and how often

Kind Questions Who How it is checked
Counted P2, P3, St2, S1, I2 Code Deterministic; depends on the speaker attribution (A3)
Cast data G1, G3 Code, from Wikidata Dates as on Wikidata on the day of fetching
Plot judgement S2, S3, M1, W1, E1 (and the K2 signal) A judge who reads only the film card: title, year, the Wikipedia plot and the cast list. No scores, no other judgements, no memory of the film. Two runs, the second blind to the first, averaged (1,323 films; E1: all 1,464 films; see Status). K2: one run, plus a second on the sample of 327 films. Where only one run exists, the site says so.
Dialogue judgement A5, I3, F1, and the J2/L1 line tags The reader of the subtitles (A3) Quotes checked literally; no second run yet in v0.5
Candidate-line judgement K1 A reader who sees only the candidate lines a word list found (02-prompts/lichaam/woorden.py), each with two lines of context, and the speaker list One run across the archive; a second, blind run on a spread sample of 327 films (α 0.98)
Love reading R2, N1 A reader who sees only the film card: the cast, the fixed leads, the Wikipedia plot and candidate lines from the subtitles (words about courting, kissing and marrying) Two runs, the second blind to the first; quotes checked literally; scored only if both runs give an answer

A12. Evidence

Every scored question shows what it rests on: for subtitle questions a quote with its timestamp, for plot questions the sentence from the plot summary (prefixed Plot:), for counts the numbers behind the score. A quote that does not literally appear in its source is rejected. A judgement without evidence is not shown.

A13. Status labels on the site

Label Meaning
Telling / Count Counted by code from the subtitles or the cast list
Castgegevens / Cast data From Wikidata
Oordeel uit de plotsamenvatting (Wikipedia) Plot judgement; with twee blinde beoordelaars or één beoordelaar (A11)
Oordeel van het taalmodel / Language-model judgement Judged from the subtitles, with quotes
Niet gemeten / Not measured No score; the reason is given (too little data, not applicable)

Part B. The 20 questions

The dimension names are the ones on the site since 31 October 2026, named after the question they ask (the earlier names in brackets, which the code still uses): WHO TALKS / WIE PRAAT (AANWEZIG), WHOSE STORY / WIENS VERHAAL (VERHAAL), POWER / MACHT, TOGETHER / ONDER ELKAAR (STEM), NAME / NAAM (IDENTITEIT), WORK / WERK, AGE / LEEFTIJD, SACRIFICE / OFFER (LOT), BODY / LICHAAM (since 31 October 2026), LOVE / LIEFDE (since 16 November 2026). The headings below use the earlier names.

AANWEZIG: who is there, and who gets the words?

P2. Who gets the words?

P3. Who is central?

VERHAAL: whose story is it?

S1. Who comes first, in the credits and in the plot?

S2. Who has a goal of their own?

S3. Who is needed?

MACHT: who decides, and who has the money?

M1. Who makes the turning-point decisions?

A5. Who has the money?

E1. Who is the expert?

STEM: who talks with whom, and about what?

St2. Do women talk with each other as often as men do?

L1. Who talks about the other as a partner?

IDENTITEIT: who is someone in their own right?

I2. Called by name or by a role?

I3. Whose relationship status is mentioned?

WERK: whose work matters?

W1. Whose work matters to the story?

J2. Who talks about work?

LEEFTIJD: who may grow older?

G1. Who is older in the main couple?

G3. How old are the women and men who speak the most?

LOT: how does the story treat them?

F1. Whose suffering is used for the other's story?

LICHAAM: who is judged on their body?

K1. Who is judged as an object of desire?

LIEFDE (LOVE): who has to be won over?

LOVE was added on 16 November 2026 (Astrid's decision), after three tests on the test films (40-proef-liefde/) and a measurement across the whole archive (41-liefde-archief/verslag.md). The definition has not yet been checked by people; a manual rating pack is ready.

R2. Who is won over?

N1. Whose no is ignored?


Part C. Outside the angle

Shown as context, not counted

Signal on the film card: K2, sexual misconduct

Not measured yet

Question Why not What it needs
M2 power and agency in the language (Sap et al. 2017) Too few verbs in plot summaries; on dialogue it does not separate Script action lines
I1 defined through a partner ("his wife") Too few films with enough cases More text, or scripts
F3 who is harmed, who harms (risk ratio) Plot summaries rarely describe violence between named women and men in detail Scripts, or dialogue judgement
F2 punished for what the other gets away with Does not separate clear cases Rework
W2 care work Answers in 22 of 59 films; mostly reflects that women are present Dialogue or script action lines
BEELD screen time and close-ups Needs a lawful copy of the film and a blind hand check Decision by the project lead

Part D. Checked against bechdeltest.com

Why the Bechdel test is not a ball. Its idea is already in the angle, made continuous and two-sided: St2 (do women talk with each other as often as men do, corrected for the cast) and L1 (who talks about the other as a partner, the test's third step). Adding the test itself would count that idea twice, and the test is a yes/no question about women only, with ratings entered by the site's visitors. So we use it the other way round: as an outside check on the angle.

How. The ratings (0 to 3; 3 = two named women talk to each other about something other than a man) come from bechdeltest.com, read from the site's year lists in October 2026 (its old API answers "410 Gone"). 243 films in the archive are listed; 221 of them have a verdict (not provisional) and are compared below. Ratings marked dubious by the site are kept. Angles in this part use the internal sign: positive is his way.

Result.

Bechdel rating Films Mean angle
3 (passes) 161 2.8°
2 (women talk, but only about a man) 17 12.5°
1 (women do not talk to each other) 33 18.5°
0 (fewer than two named women) 10 16.8°

What this does and does not show. The angle and a well-known, independent test point the same way, and the angle adds what the test cannot: degrees, both directions, and how much room a film gives, not only whether two women talk. It does not show that the angle is right film by film; a Bechdel pass is a low bar (161 of 221 films pass).

Part E. Known limitations

  1. The speaker attribution is done by a language model. It is checked by coverage, by quotes and, where they exist, by subtitles with speaker names (Status), not yet by a human coder on a sample.
  2. The plot summaries are Wikipedia's. They are free and cover almost every film, but they are written by editors with their own focus. Films without a summary get no plot questions (Spijt!).
  3. The plot judge is a language model too. The second run is blind to the first, but both are the same kind of judge; people check samples on an ongoing basis, since 19 November 2026.
  4. The archive is a choice, not a sample. Its average says something about these films, not about film in general.
  5. Gender only, and binary in the comparison. Non-binary characters are left out of the formulas, not reassigned.
  6. Outside checks so far: bechdeltest.com (Part D; also via the FiveThirtyEight data, CC BY 4.0, and the 2021 snapshot via TidyTuesday), The Pudding's screenplay counts, subtitles with speaker names (Status), Hollywood Age Gap (age in the main couple), the Cornell Movie-Dialogs Corpus and MovieSum (CC BY-NC 4.0) for the share of women's words, and the power and agency word list of Sap et al. (2017; riveter-nlp, GPL-3.0). That last one barely relates to the angle (r 0.11). The figures are on the site under How we measure (November 2026). None yet with human coding.
  7. The method changes. Scores can move when a rule is improved. Cite a result with the date you read it.

Appendix 1. Anchors for judgements (A5, F1)

Value Meaning
1 Clear and central: the pattern works against the women and drives the story, with no counterweight
0.5 Present but limited, or contradicted by other scenes
0 Absent, or equal on both sides
−0.5 The reverse pattern is present but limited
−1 The reverse pattern is clear and central

Always one quote for the chosen value and one for the strongest argument the other way.

Appendix 2. Role and endearment words (I2)

The lists in 13-ondertitels/scripts/counts_film.py (ROLE) are the reference. Each word belongs to one gender, for example: - EN, women: sweetheart, honey, baby, darling, girl, lady, miss, ma'am, wife, girlfriend, mother, daughter, sister, princess. Men: buddy, pal, son, boy, guy, sir, husband, boyfriend, father, brother, dude, prince. - NL, women: schatje, lieverd, meisje, meid, mevrouw, vrouw, moeder, dochter, zus, vriendin. Men: jongen, meneer, vent, kerel, vader, zoon, broer, vriend, echtgenoot, maat. - FR, women: chérie, mademoiselle, madame, fille, femme, mère, copine. Men: chéri, monsieur, garçon, homme, père, fils, copain, mec.

Appendix 3. Where the code is

The code is not public yet; it is available on request for research. For reference: | Step | Script | |---|---| | Subtitles chosen and turned into transcripts | 13-ondertitels/scripts/choose.py, build_transcripts.py | | Speakers and tags written per film | 17-nieuwe-films/scripts/lees.py, film.py (new films); annotation files for the first 60 | | Counts and cast questions (I2, G1, coverage) | 13-ondertitels/scripts/counts_film.py, measure_film.py | | P2, P3, St2, S1, G3 | 15-vragen-v05/scripts/auto_v05.py | | Film cards for the plot judges | 15-vragen-v05/scripts/kaarten.py | | Plot judgements checked and saved | 17-nieuwe-films/scripts/plotoordeel.py (prompt: 15-vragen-v05/prompts_v05.md) | | Dimension scores and angle | 15-vragen-v05/scripts/hoek_v05.py, export_v05.py; the site recomputes the angle in dutch-angle-site/src/index.html (angleOf) | | LOVE (R2, N1) | 41-liefde-archief/scripts/prep.py (cards), score_archief.py, export_liefde.py | | Country of origin for the filters | dutch-angle-site/scripts/landen.py |

Appendix 4. Sources and inspiration

The classic tests are credited as inspiration only. Dutch Angle uses its own questions. - Anderson, H. & Daniels, M. (2016). Film Dialogue from 2,000 screenplays, Broken Down by Gender and Age. Polygraph / The Pudding. Summary: https://flowingdata.com/2016/06/27/gender-equality-in-the-movies-a-screenplay-analysis - Kagan, D., Chesney, T. & Fire, M. (2020). Using data science to understand the film industry's gender gap. Palgrave Communications 6. https://arxiv.org/abs/1903.06469v1 - Lauzen, M. (2026). It's a Man's (Celluloid) World: Portrayals of Female Characters in the Top Grossing U.S. Films of 2025. Center for the Study of Women in Television and Film, San Diego State University. https://womenintvfilm.sdsu.edu/wp-content/uploads/2026/02/2025-Its-a-Mans-World-Report.pdf - Lippman, J. R. (2018; online 2015). "I did it because I never stopped loving you": the effects of media portrayals of persistent pursuit on beliefs about stalking. Communication Research. - Sanders, W. & Becker, S. (2025). 'Gewoon' een Nederlandse vrouw. Universiteit Utrecht / Stichting Vrouwen in Beeld. https://www.vrouweninbeeld.nl/content/uploads/2025/10/Onderzoek-REPRESENTATIE-in-Beeld_251015.pdf - Sap, M., Prasettio, M. C., Holtzman, A., Rashkin, H. & Choi, Y. (2017). Connotation Frames of Power and Agency in Modern Films. Proceedings of EMNLP 2017, 2329–2334. https://aclanthology.org/D17-1247/ - bechdeltest.com: crowd-sourced Bechdel ratings, https://bechdeltest.com (read 17 October 2026). - Inspiration: the Bechdel-Wallace test (1985), the Mako Mori test (2013), the Sexy Lamp test (Kelly Sue DeConnick), "women in refrigerators" (Simone, 1999), the Smurfette principle (Pollitt, 1991). - Data: OPUS OpenSubtitles v2024 (opus.nlpl.eu); Wikidata; Wikipedia plot summaries (CC BY-SA).